ceph/qa/tasks/cephfs/test_forward_scrub.py

   1
   2 """
   3 Test that the forward scrub functionality can traverse metadata and apply
   4 requested tags, on well formed metadata.
   5
   6 This is *not* the real testing for forward scrub, which will need to test
   7 how the functionality responds to damaged metadata.
   8
   9 """
  10 import json
  11
  12 import logging
  13 from collections import namedtuple
  14 from textwrap import dedent
  15
  16 from teuthology.orchestra.run import CommandFailedError
  17 from tasks.cephfs.cephfs_test_case import CephFSTestCase
  18
  19 import struct
  20
  21 log = logging.getLogger(__name__)
  22
  23
  24 ValidationError = namedtuple("ValidationError", ["exception", "backtrace"])
  25
  26
  27 class TestForwardScrub(CephFSTestCase):
  28     MDSS_REQUIRED = 1
  29
  30     def _read_str_xattr(self, pool, obj, attr):
  31         """
  32         Read a ceph-encoded string from a rados xattr
  33         """
  34         output = self.fs.rados(["getxattr", obj, attr], pool=pool)
  35         strlen = struct.unpack('i', output[0:4])[0]
  36         return output[4:(4 + strlen)]
  37
  38     def _get_paths_to_ino(self):
  39         inos = {}
  40         p = self.mount_a.run_shell(["find", "./"])
  41         paths = p.stdout.getvalue().strip().split()
  42         for path in paths:
  43             inos[path] = self.mount_a.path_to_ino(path)
  44
  45         return inos
  46
  47     def test_apply_tag(self):
  48         self.mount_a.run_shell(["mkdir", "parentdir"])
  49         self.mount_a.run_shell(["mkdir", "parentdir/childdir"])
  50         self.mount_a.run_shell(["touch", "rfile"])
  51         self.mount_a.run_shell(["touch", "parentdir/pfile"])
  52         self.mount_a.run_shell(["touch", "parentdir/childdir/cfile"])
  53
  54         # Build a structure mapping path to inode, as we will later want
  55         # to check object by object and objects are named after ino number
  56         inos = self._get_paths_to_ino()
  57
  58         # Flush metadata: this is a friendly test of forward scrub so we're skipping
  59         # the part where it's meant to cope with dirty metadata
  60         self.mount_a.umount_wait()
  61         self.fs.mds_asok(["flush", "journal"])
  62
  63         tag = "mytag"
  64
  65         # Execute tagging forward scrub
  66         self.fs.mds_asok(["tag", "path", "/parentdir", tag])
  67         # Wait for completion
  68         import time
  69         time.sleep(10)
  70         # FIXME watching clog isn't a nice mechanism for this, once we have a ScrubMap we'll
  71         # watch that instead
  72
  73         # Check that dirs were tagged
  74         for dirpath in ["./parentdir", "./parentdir/childdir"]:
  75             self.assertTagged(inos[dirpath], tag, self.fs.get_metadata_pool_name())
  76
  77         # Check that files were tagged
  78         for filepath in ["./parentdir/pfile", "./parentdir/childdir/cfile"]:
  79             self.assertTagged(inos[filepath], tag, self.fs.get_data_pool_name())
  80
  81         # This guy wasn't in the tag path, shouldn't have been tagged
  82         self.assertUntagged(inos["./rfile"])
  83
  84     def assertUntagged(self, ino):
  85         file_obj_name = "{0:x}.00000000".format(ino)
  86         with self.assertRaises(CommandFailedError):
  87             self._read_str_xattr(
  88                 self.fs.get_data_pool_name(),
  89                 file_obj_name,
  90                 "scrub_tag"
  91             )
  92
  93     def assertTagged(self, ino, tag, pool):
  94         file_obj_name = "{0:x}.00000000".format(ino)
  95         wrote = self._read_str_xattr(
  96             pool,
  97             file_obj_name,
  98             "scrub_tag"
  99         )
 100         self.assertEqual(wrote, tag)
 101
 102     def _validate_linkage(self, expected):
 103         inos = self._get_paths_to_ino()
 104         try:
 105             self.assertDictEqual(inos, expected)
 106         except AssertionError:
 107             log.error("Expected: {0}".format(json.dumps(expected, indent=2)))
 108             log.error("Actual: {0}".format(json.dumps(inos, indent=2)))
 109             raise
 110
 111     def test_orphan_scan(self):
 112         # Create some files whose metadata we will flush
 113         self.mount_a.run_python(dedent("""
 114             import os
 115             mount_point = "{mount_point}"
 116             parent = os.path.join(mount_point, "parent")
 117             os.mkdir(parent)
 118             flushed = os.path.join(parent, "flushed")
 119             os.mkdir(flushed)
 120             for f in ["alpha", "bravo", "charlie"]:
 121                 open(os.path.join(flushed, f), 'w').write(f)
 122         """.format(mount_point=self.mount_a.mountpoint)))
 123
 124         inos = self._get_paths_to_ino()
 125
 126         # Flush journal
 127         # Umount before flush to avoid cap releases putting
 128         # things we don't want in the journal later.
 129         self.mount_a.umount_wait()
 130         self.fs.mds_asok(["flush", "journal"])
 131
 132         # Create a new inode that's just in the log, i.e. would
 133         # look orphaned to backward scan if backward scan wisnae
 134         # respectin' tha scrub_tag xattr.
 135         self.mount_a.mount()
 136         self.mount_a.run_shell(["mkdir", "parent/unflushed"])
 137         self.mount_a.run_shell(["dd", "if=/dev/urandom",
 138                                 "of=./parent/unflushed/jfile",
 139                                 "bs=1M", "count=8"])
 140         inos["./parent/unflushed"] = self.mount_a.path_to_ino("./parent/unflushed")
 141         inos["./parent/unflushed/jfile"] = self.mount_a.path_to_ino("./parent/unflushed/jfile")
 142         self.mount_a.umount_wait()
 143
 144         # Orphan an inode by deleting its dentry
 145         # Our victim will be.... bravo.
 146         self.mount_a.umount_wait()
 147         self.fs.mds_stop()
 148         self.fs.mds_fail()
 149         self.fs.set_ceph_conf('mds', 'mds verify scatter', False)
 150         self.fs.set_ceph_conf('mds', 'mds debug scatterstat', False)
 151         frag_obj_id = "{0:x}.00000000".format(inos["./parent/flushed"])
 152         self.fs.rados(["rmomapkey", frag_obj_id, "bravo_head"])
 153
 154         self.fs.mds_restart()
 155         self.fs.wait_for_daemons()
 156
 157         # See that the orphaned file is indeed missing from a client's POV
 158         self.mount_a.mount()
 159         damaged_state = self._get_paths_to_ino()
 160         self.assertNotIn("./parent/flushed/bravo", damaged_state)
 161         self.mount_a.umount_wait()
 162
 163         # Run a tagging forward scrub
 164         tag = "mytag123"
 165         self.fs.mds_asok(["tag", "path", "/parent", tag])
 166
 167         # See that the orphan wisnae tagged
 168         self.assertUntagged(inos['./parent/flushed/bravo'])
 169
 170         # See that the flushed-metadata-and-still-present files are tagged
 171         self.assertTagged(inos['./parent/flushed/alpha'], tag, self.fs.get_data_pool_name())
 172         self.assertTagged(inos['./parent/flushed/charlie'], tag, self.fs.get_data_pool_name())
 173
 174         # See that journalled-but-not-flushed file *was* tagged
 175         self.assertTagged(inos['./parent/unflushed/jfile'], tag, self.fs.get_data_pool_name())
 176
 177         # Run cephfs-data-scan targeting only orphans
 178         self.fs.mds_stop()
 179         self.fs.mds_fail()
 180         self.fs.data_scan(["scan_extents", self.fs.get_data_pool_name()])
 181         self.fs.data_scan([
 182             "scan_inodes",
 183             "--filter-tag", tag,
 184             self.fs.get_data_pool_name()
 185         ])
 186
 187         # After in-place injection stats should be kosher again
 188         self.fs.set_ceph_conf('mds', 'mds verify scatter', True)
 189         self.fs.set_ceph_conf('mds', 'mds debug scatterstat', True)
 190
 191         # And we should have all the same linkage we started with,
 192         # and no lost+found, and no extra inodes!
 193         self.fs.mds_restart()
 194         self.fs.wait_for_daemons()
 195         self.mount_a.mount()
 196         self._validate_linkage(inos)
 197
 198     def _stash_inotable(self):
 199         # Get all active ranks
 200         ranks = self.fs.get_all_mds_rank()
 201
 202         inotable_dict = {}
 203         for rank in ranks:
 204             inotable_oid = "mds{rank:d}_".format(rank=rank) + "inotable"
 205             print "Trying to fetch inotable object: " + inotable_oid
 206
 207             #self.fs.get_metadata_object("InoTable", "mds0_inotable")
 208             inotable_raw = self.fs.get_metadata_object_raw(inotable_oid)
 209             inotable_dict[inotable_oid] = inotable_raw
 210         return inotable_dict
 211
 212     def test_inotable_sync(self):
 213         self.mount_a.write_n_mb("file1_sixmegs", 6)
 214
 215         # Flush journal
 216         self.mount_a.umount_wait()
 217         self.fs.mds_asok(["flush", "journal"])
 218
 219         inotable_copy = self._stash_inotable()
 220
 221         self.mount_a.mount()
 222
 223         self.mount_a.write_n_mb("file2_sixmegs", 6)
 224         self.mount_a.write_n_mb("file3_sixmegs", 6)
 225
 226         inos = self._get_paths_to_ino()
 227
 228         # Flush journal
 229         self.mount_a.umount_wait()
 230         self.fs.mds_asok(["flush", "journal"])
 231
 232         self.mount_a.umount_wait()
 233
 234         with self.assert_cluster_log("inode table repaired", invert_match=True):
 235             self.fs.mds_asok(["scrub_path", "/", "repair", "recursive"])
 236
 237         self.mds_cluster.mds_stop()
 238         self.mds_cluster.mds_fail()
 239
 240         # Truncate the journal (to ensure the inotable on disk
 241         # is all that will be in the InoTable in memory)
 242
 243         self.fs.journal_tool(["event", "splice",
 244             "--inode={0}".format(inos["./file2_sixmegs"]), "summary"])
 245
 246         self.fs.journal_tool(["event", "splice",
 247             "--inode={0}".format(inos["./file3_sixmegs"]), "summary"])
 248
 249         # Revert to old inotable.
 250         for key, value in inotable_copy.iteritems():
 251            self.fs.put_metadata_object_raw(key, value)
 252
 253         self.mds_cluster.mds_restart()
 254         self.fs.wait_for_daemons()
 255
 256         with self.assert_cluster_log("inode table repaired"):
 257             self.fs.mds_asok(["scrub_path", "/", "repair", "recursive"])
 258
 259         self.mds_cluster.mds_stop()
 260         table_text = self.fs.table_tool(["0", "show", "inode"])
 261         table = json.loads(table_text)
 262         self.assertGreater(
 263                 table['0']['data']['inotable']['free'][0]['start'],
 264                 inos['./file3_sixmegs'])
 265
 266     def test_backtrace_repair(self):
 267         """
 268         That the MDS can repair an inodes backtrace in the data pool
 269         if it is found to be damaged.
 270         """
 271         # Create a file for subsequent checks
 272         self.mount_a.run_shell(["mkdir", "parent_a"])
 273         self.mount_a.run_shell(["touch", "parent_a/alpha"])
 274         file_ino = self.mount_a.path_to_ino("parent_a/alpha")
 275
 276         # That backtrace and layout are written after initial flush
 277         self.fs.mds_asok(["flush", "journal"])
 278         backtrace = self.fs.read_backtrace(file_ino)
 279         self.assertEqual(['alpha', 'parent_a'],
 280                          [a['dname'] for a in backtrace['ancestors']])
 281
 282         # Go corrupt the backtrace
 283         self.fs._write_data_xattr(file_ino, "parent",
 284                                   "oh i'm sorry did i overwrite your xattr?")
 285
 286         with self.assert_cluster_log("bad backtrace on inode"):
 287             self.fs.mds_asok(["scrub_path", "/", "repair", "recursive"])
 288         self.fs.mds_asok(["flush", "journal"])
 289         backtrace = self.fs.read_backtrace(file_ino)
 290         self.assertEqual(['alpha', 'parent_a'],
 291                          [a['dname'] for a in backtrace['ancestors']])