root / daemons / ganeti-watcher @ 7b195d9b
History | View | Annotate | Download (12 kB)
1 |
#!/usr/bin/python |
---|---|
2 |
# |
3 |
|
4 |
# Copyright (C) 2006, 2007 Google Inc. |
5 |
# |
6 |
# This program is free software; you can redistribute it and/or modify |
7 |
# it under the terms of the GNU General Public License as published by |
8 |
# the Free Software Foundation; either version 2 of the License, or |
9 |
# (at your option) any later version. |
10 |
# |
11 |
# This program is distributed in the hope that it will be useful, but |
12 |
# WITHOUT ANY WARRANTY; without even the implied warranty of |
13 |
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 |
# General Public License for more details. |
15 |
# |
16 |
# You should have received a copy of the GNU General Public License |
17 |
# along with this program; if not, write to the Free Software |
18 |
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA |
19 |
# 02110-1301, USA. |
20 |
|
21 |
|
22 |
"""Tool to restart erronously downed virtual machines. |
23 |
|
24 |
This program and set of classes implement a watchdog to restart |
25 |
virtual machines in a Ganeti cluster that have crashed or been killed |
26 |
by a node reboot. Run from cron or similar. |
27 |
|
28 |
""" |
29 |
|
30 |
import os |
31 |
import sys |
32 |
import re |
33 |
import time |
34 |
import fcntl |
35 |
import errno |
36 |
import simplejson |
37 |
from optparse import OptionParser |
38 |
|
39 |
from ganeti import utils |
40 |
from ganeti import constants |
41 |
from ganeti import ssconf |
42 |
from ganeti import errors |
43 |
|
44 |
|
45 |
MAXTRIES = 5 |
46 |
BAD_STATES = ['stopped'] |
47 |
HELPLESS_STATES = ['(node down)'] |
48 |
NOTICE = 'NOTICE' |
49 |
ERROR = 'ERROR' |
50 |
KEY_RESTART_COUNT = "restart_count" |
51 |
KEY_RESTART_WHEN = "restart_when" |
52 |
KEY_BOOT_ID = "bootid" |
53 |
|
54 |
|
55 |
class Error(Exception): |
56 |
"""Generic custom error class.""" |
57 |
|
58 |
|
59 |
class NotMasterError(Error): |
60 |
"""Exception raised when this host is not the master.""" |
61 |
|
62 |
|
63 |
def Indent(s, prefix='| '): |
64 |
"""Indent a piece of text with a given prefix before each line. |
65 |
|
66 |
Args: |
67 |
s: The string to indent |
68 |
prefix: The string to prepend each line. |
69 |
|
70 |
""" |
71 |
return "%s%s\n" % (prefix, ('\n' + prefix).join(s.splitlines())) |
72 |
|
73 |
|
74 |
def DoCmd(cmd): |
75 |
"""Run a shell command. |
76 |
|
77 |
Args: |
78 |
cmd: the command to run. |
79 |
|
80 |
Raises CommandError with verbose commentary on error. |
81 |
|
82 |
""" |
83 |
res = utils.RunCmd(cmd) |
84 |
|
85 |
if res.failed: |
86 |
raise Error("Command %s failed:\n%s\nstdout:\n%sstderr:\n%s" % |
87 |
(repr(cmd), |
88 |
Indent(res.fail_reason), |
89 |
Indent(res.stdout), |
90 |
Indent(res.stderr))) |
91 |
|
92 |
return res |
93 |
|
94 |
|
95 |
class WatcherState(object): |
96 |
"""Interface to a state file recording restart attempts. |
97 |
|
98 |
""" |
99 |
def __init__(self): |
100 |
"""Open, lock, read and parse the file. |
101 |
|
102 |
Raises StandardError on lock contention. |
103 |
|
104 |
""" |
105 |
# The two-step dance below is necessary to allow both opening existing |
106 |
# file read/write and creating if not existing. Vanilla open will truncate |
107 |
# an existing file -or- allow creating if not existing. |
108 |
f = os.open(constants.WATCHER_STATEFILE, os.O_RDWR | os.O_CREAT) |
109 |
f = os.fdopen(f, 'w+') |
110 |
|
111 |
try: |
112 |
fcntl.flock(f.fileno(), fcntl.LOCK_EX|fcntl.LOCK_NB) |
113 |
except IOError, x: |
114 |
if x.errno == errno.EAGAIN: |
115 |
raise StandardError("State file already locked") |
116 |
raise |
117 |
|
118 |
self.statefile = f |
119 |
|
120 |
try: |
121 |
self.data = simplejson.load(self.statefile) |
122 |
except Exception, msg: |
123 |
# Ignore errors while loading the file and treat it as empty |
124 |
self.data = {} |
125 |
sys.stderr.write("Empty or invalid state file. " |
126 |
"Using defaults. Error message: %s\n" % msg) |
127 |
|
128 |
if "instance" not in self.data: |
129 |
self.data["instance"] = {} |
130 |
if "node" not in self.data: |
131 |
self.data["node"] = {} |
132 |
|
133 |
def __del__(self): |
134 |
"""Called on destruction. |
135 |
|
136 |
""" |
137 |
if self.statefile: |
138 |
self._Close() |
139 |
|
140 |
def _Close(self): |
141 |
"""Unlock configuration file and close it. |
142 |
|
143 |
""" |
144 |
assert self.statefile |
145 |
|
146 |
fcntl.flock(self.statefile.fileno(), fcntl.LOCK_UN) |
147 |
|
148 |
self.statefile.close() |
149 |
self.statefile = None |
150 |
|
151 |
def GetNodeBootID(self, name): |
152 |
"""Returns the last boot ID of a node or None. |
153 |
|
154 |
""" |
155 |
ndata = self.data["node"] |
156 |
|
157 |
if name in ndata and KEY_BOOT_ID in ndata[name]: |
158 |
return ndata[name][KEY_BOOT_ID] |
159 |
return None |
160 |
|
161 |
def SetNodeBootID(self, name, bootid): |
162 |
"""Sets the boot ID of a node. |
163 |
|
164 |
""" |
165 |
assert bootid |
166 |
|
167 |
ndata = self.data["node"] |
168 |
|
169 |
if name not in ndata: |
170 |
ndata[name] = {} |
171 |
|
172 |
ndata[name][KEY_BOOT_ID] = bootid |
173 |
|
174 |
def NumberOfRestartAttempts(self, instance): |
175 |
"""Returns number of previous restart attempts. |
176 |
|
177 |
Args: |
178 |
instance - the instance to look up. |
179 |
|
180 |
""" |
181 |
idata = self.data["instance"] |
182 |
|
183 |
if instance.name in idata: |
184 |
return idata[instance.name][KEY_RESTART_COUNT] |
185 |
|
186 |
return 0 |
187 |
|
188 |
def RecordRestartAttempt(self, instance): |
189 |
"""Record a restart attempt. |
190 |
|
191 |
Args: |
192 |
instance - the instance being restarted |
193 |
|
194 |
""" |
195 |
idata = self.data["instance"] |
196 |
|
197 |
if instance.name not in idata: |
198 |
inst = idata[instance.name] = {} |
199 |
else: |
200 |
inst = idata[instance.name] |
201 |
|
202 |
inst[KEY_RESTART_WHEN] = time.time() |
203 |
inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
204 |
|
205 |
def RemoveInstance(self, instance): |
206 |
"""Update state to reflect that a machine is running, i.e. remove record. |
207 |
|
208 |
Args: |
209 |
instance - the instance to remove from books |
210 |
|
211 |
This method removes the record for a named instance. |
212 |
|
213 |
""" |
214 |
idata = self.data["instance"] |
215 |
|
216 |
if instance.name in idata: |
217 |
del idata[instance.name] |
218 |
|
219 |
def Save(self): |
220 |
"""Save state to file, then unlock and close it. |
221 |
|
222 |
""" |
223 |
assert self.statefile |
224 |
|
225 |
self.statefile.seek(0) |
226 |
self.statefile.truncate() |
227 |
|
228 |
simplejson.dump(self.data, self.statefile) |
229 |
|
230 |
self._Close() |
231 |
|
232 |
|
233 |
class Instance(object): |
234 |
"""Abstraction for a Virtual Machine instance. |
235 |
|
236 |
Methods: |
237 |
Restart(): issue a command to restart the represented machine. |
238 |
|
239 |
""" |
240 |
def __init__(self, name, state, autostart): |
241 |
self.name = name |
242 |
self.state = state |
243 |
self.autostart = autostart |
244 |
|
245 |
def Restart(self): |
246 |
"""Encapsulates the start of an instance. |
247 |
|
248 |
""" |
249 |
DoCmd(['gnt-instance', 'startup', '--lock-retries=15', self.name]) |
250 |
|
251 |
def ActivateDisks(self): |
252 |
"""Encapsulates the activation of all disks of an instance. |
253 |
|
254 |
""" |
255 |
DoCmd(['gnt-instance', 'activate-disks', '--lock-retries=15', self.name]) |
256 |
|
257 |
|
258 |
def _RunListCmd(cmd): |
259 |
"""Runs a command and parses its output into lists. |
260 |
|
261 |
""" |
262 |
for line in DoCmd(cmd).stdout.splitlines(): |
263 |
yield line.split(':') |
264 |
|
265 |
|
266 |
def GetInstanceList(with_secondaries=None): |
267 |
"""Get a list of instances on this cluster. |
268 |
|
269 |
""" |
270 |
cmd = ['gnt-instance', 'list', '--lock-retries=15', '--no-headers', |
271 |
'--separator=:'] |
272 |
|
273 |
fields = 'name,oper_state,admin_state' |
274 |
|
275 |
if with_secondaries is not None: |
276 |
fields += ',snodes' |
277 |
|
278 |
cmd.append('-o') |
279 |
cmd.append(fields) |
280 |
|
281 |
instances = [] |
282 |
for fields in _RunListCmd(cmd): |
283 |
if with_secondaries is not None: |
284 |
(name, status, autostart, snodes) = fields |
285 |
|
286 |
if snodes == "-": |
287 |
continue |
288 |
|
289 |
for node in with_secondaries: |
290 |
if node in snodes.split(','): |
291 |
break |
292 |
else: |
293 |
continue |
294 |
|
295 |
else: |
296 |
(name, status, autostart) = fields |
297 |
|
298 |
instances.append(Instance(name, status, autostart != "no")) |
299 |
|
300 |
return instances |
301 |
|
302 |
|
303 |
def GetNodeBootIDs(): |
304 |
"""Get a dict mapping nodes to boot IDs. |
305 |
|
306 |
""" |
307 |
cmd = ['gnt-node', 'list', '--lock-retries=15', '--no-headers', |
308 |
'--separator=:', '-o', 'name,bootid'] |
309 |
|
310 |
ids = {} |
311 |
for fields in _RunListCmd(cmd): |
312 |
(name, bootid) = fields |
313 |
ids[name] = bootid |
314 |
|
315 |
return ids |
316 |
|
317 |
|
318 |
class Message(object): |
319 |
"""Encapsulation of a notice or error message. |
320 |
|
321 |
""" |
322 |
def __init__(self, level, msg): |
323 |
self.level = level |
324 |
self.msg = msg |
325 |
self.when = time.time() |
326 |
|
327 |
def __str__(self): |
328 |
return self.level + ' ' + time.ctime(self.when) + '\n' + Indent(self.msg) |
329 |
|
330 |
|
331 |
class Watcher(object): |
332 |
"""Encapsulate the logic for restarting erronously halted virtual machines. |
333 |
|
334 |
The calling program should periodically instantiate me and call Run(). |
335 |
This will traverse the list of instances, and make up to MAXTRIES attempts |
336 |
to restart machines that are down. |
337 |
|
338 |
""" |
339 |
def __init__(self): |
340 |
sstore = ssconf.SimpleStore() |
341 |
master = sstore.GetMasterNode() |
342 |
if master != utils.HostInfo().name: |
343 |
raise NotMasterError("This is not the master node") |
344 |
self.instances = GetInstanceList() |
345 |
self.bootids = GetNodeBootIDs() |
346 |
self.messages = [] |
347 |
|
348 |
def Run(self): |
349 |
notepad = WatcherState() |
350 |
self.CheckInstances(notepad) |
351 |
self.CheckDisks(notepad) |
352 |
notepad.Save() |
353 |
|
354 |
def CheckDisks(self, notepad): |
355 |
"""Check all nodes for restarted ones. |
356 |
|
357 |
""" |
358 |
check_nodes = [] |
359 |
for name, id in self.bootids.iteritems(): |
360 |
old = notepad.GetNodeBootID(name) |
361 |
if old != id: |
362 |
# Node's boot ID has changed, proably through a reboot. |
363 |
check_nodes.append(name) |
364 |
|
365 |
if check_nodes: |
366 |
# Activate disks for all instances with any of the checked nodes as a |
367 |
# secondary node. |
368 |
for instance in GetInstanceList(with_secondaries=check_nodes): |
369 |
try: |
370 |
self.messages.append(Message(NOTICE, |
371 |
"Activating disks for %s." % |
372 |
instance.name)) |
373 |
instance.ActivateDisks() |
374 |
except Error, x: |
375 |
self.messages.append(Message(ERROR, str(x))) |
376 |
|
377 |
# Keep changed boot IDs |
378 |
for name in check_nodes: |
379 |
notepad.SetNodeBootID(name, self.bootids[name]) |
380 |
|
381 |
def CheckInstances(self, notepad): |
382 |
"""Make a pass over the list of instances, restarting downed ones. |
383 |
|
384 |
""" |
385 |
for instance in self.instances: |
386 |
# Don't care about manually stopped instances |
387 |
if not instance.autostart: |
388 |
continue |
389 |
|
390 |
if instance.state in BAD_STATES: |
391 |
n = notepad.NumberOfRestartAttempts(instance) |
392 |
|
393 |
if n > MAXTRIES: |
394 |
# stay quiet. |
395 |
continue |
396 |
elif n < MAXTRIES: |
397 |
last = " (Attempt #%d)" % (n + 1) |
398 |
else: |
399 |
notepad.RecordRestartAttempt(instance) |
400 |
self.messages.append(Message(ERROR, "Could not restart %s for %d" |
401 |
" times, giving up..." % |
402 |
(instance.name, MAXTRIES))) |
403 |
continue |
404 |
try: |
405 |
self.messages.append(Message(NOTICE, |
406 |
"Restarting %s%s." % |
407 |
(instance.name, last))) |
408 |
instance.Restart() |
409 |
except Error, x: |
410 |
self.messages.append(Message(ERROR, str(x))) |
411 |
|
412 |
notepad.RecordRestartAttempt(instance) |
413 |
elif instance.state in HELPLESS_STATES: |
414 |
if notepad.NumberOfRestartAttempts(instance): |
415 |
notepad.RemoveInstance(instance) |
416 |
else: |
417 |
if notepad.NumberOfRestartAttempts(instance): |
418 |
notepad.RemoveInstance(instance) |
419 |
msg = Message(NOTICE, |
420 |
"Restart of %s succeeded." % instance.name) |
421 |
self.messages.append(msg) |
422 |
|
423 |
def WriteReport(self, logfile): |
424 |
"""Log all messages to file. |
425 |
|
426 |
Args: |
427 |
logfile: file object open for writing (the log file) |
428 |
|
429 |
""" |
430 |
for msg in self.messages: |
431 |
print >> logfile, str(msg) |
432 |
|
433 |
|
434 |
def ParseOptions(): |
435 |
"""Parse the command line options. |
436 |
|
437 |
Returns: |
438 |
(options, args) as from OptionParser.parse_args() |
439 |
|
440 |
""" |
441 |
parser = OptionParser(description="Ganeti cluster watcher", |
442 |
usage="%prog [-d]", |
443 |
version="%%prog (ganeti) %s" % |
444 |
constants.RELEASE_VERSION) |
445 |
|
446 |
parser.add_option("-d", "--debug", dest="debug", |
447 |
help="Don't redirect messages to the log file", |
448 |
default=False, action="store_true") |
449 |
options, args = parser.parse_args() |
450 |
return options, args |
451 |
|
452 |
|
453 |
def main(): |
454 |
"""Main function. |
455 |
|
456 |
""" |
457 |
options, args = ParseOptions() |
458 |
|
459 |
if not options.debug: |
460 |
sys.stderr = sys.stdout = open(constants.LOG_WATCHER, 'a') |
461 |
|
462 |
try: |
463 |
try: |
464 |
watcher = Watcher() |
465 |
except errors.ConfigurationError: |
466 |
# Just exit if there's no configuration |
467 |
sys.exit(constants.EXIT_SUCCESS) |
468 |
watcher.Run() |
469 |
watcher.WriteReport(sys.stdout) |
470 |
except NotMasterError: |
471 |
if options.debug: |
472 |
sys.stderr.write("Not master, exiting.\n") |
473 |
sys.exit(constants.EXIT_NOTMASTER) |
474 |
except errors.ResolverError, err: |
475 |
sys.stderr.write("Cannot resolve hostname '%s', exiting.\n" % err.args[0]) |
476 |
sys.exit(constants.EXIT_NODESETUP_ERROR) |
477 |
except Error, err: |
478 |
print err |
479 |
|
480 |
|
481 |
if __name__ == '__main__': |
482 |
main() |