root / daemons / ganeti-watcher @ 8a20c732
History | View | Annotate | Download (15.6 kB)
1 | a8083063 | Iustin Pop | #!/usr/bin/python |
---|---|---|---|
2 | a8083063 | Iustin Pop | # |
3 | a8083063 | Iustin Pop | |
4 | fc428e32 | Michael Hanselmann | # Copyright (C) 2006, 2007, 2008 Google Inc. |
5 | a8083063 | Iustin Pop | # |
6 | a8083063 | Iustin Pop | # This program is free software; you can redistribute it and/or modify |
7 | a8083063 | Iustin Pop | # it under the terms of the GNU General Public License as published by |
8 | a8083063 | Iustin Pop | # the Free Software Foundation; either version 2 of the License, or |
9 | a8083063 | Iustin Pop | # (at your option) any later version. |
10 | a8083063 | Iustin Pop | # |
11 | a8083063 | Iustin Pop | # This program is distributed in the hope that it will be useful, but |
12 | a8083063 | Iustin Pop | # WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | a8083063 | Iustin Pop | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 | a8083063 | Iustin Pop | # General Public License for more details. |
15 | a8083063 | Iustin Pop | # |
16 | a8083063 | Iustin Pop | # You should have received a copy of the GNU General Public License |
17 | a8083063 | Iustin Pop | # along with this program; if not, write to the Free Software |
18 | a8083063 | Iustin Pop | # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA |
19 | a8083063 | Iustin Pop | # 02110-1301, USA. |
20 | a8083063 | Iustin Pop | |
21 | a8083063 | Iustin Pop | |
22 | a8083063 | Iustin Pop | """Tool to restart erronously downed virtual machines. |
23 | a8083063 | Iustin Pop | |
24 | a8083063 | Iustin Pop | This program and set of classes implement a watchdog to restart |
25 | a8083063 | Iustin Pop | virtual machines in a Ganeti cluster that have crashed or been killed |
26 | a8083063 | Iustin Pop | by a node reboot. Run from cron or similar. |
27 | a8083063 | Iustin Pop | |
28 | 5a3103e9 | Michael Hanselmann | """ |
29 | a8083063 | Iustin Pop | |
30 | a8083063 | Iustin Pop | import os |
31 | a8083063 | Iustin Pop | import sys |
32 | a8083063 | Iustin Pop | import time |
33 | 438b45d4 | Michael Hanselmann | import logging |
34 | a8083063 | Iustin Pop | from optparse import OptionParser |
35 | a8083063 | Iustin Pop | |
36 | a8083063 | Iustin Pop | from ganeti import utils |
37 | a8083063 | Iustin Pop | from ganeti import constants |
38 | 67fe61c4 | Michael Hanselmann | from ganeti import serializer |
39 | 89e1fc26 | Iustin Pop | from ganeti import errors |
40 | e125c67c | Michael Hanselmann | from ganeti import opcodes |
41 | e125c67c | Michael Hanselmann | from ganeti import cli |
42 | 7dfb83c2 | Iustin Pop | from ganeti import luxi |
43 | a8083063 | Iustin Pop | |
44 | a8083063 | Iustin Pop | |
45 | 5a3103e9 | Michael Hanselmann | MAXTRIES = 5 |
46 | b7309a0d | Iustin Pop | BAD_STATES = ['ERROR_down'] |
47 | cbfc4681 | Iustin Pop | HELPLESS_STATES = ['ERROR_nodedown', 'ERROR_nodeoffline'] |
48 | 5a3103e9 | Michael Hanselmann | NOTICE = 'NOTICE' |
49 | 5a3103e9 | Michael Hanselmann | ERROR = 'ERROR' |
50 | 7b195d9b | Michael Hanselmann | KEY_RESTART_COUNT = "restart_count" |
51 | 7b195d9b | Michael Hanselmann | KEY_RESTART_WHEN = "restart_when" |
52 | 7b195d9b | Michael Hanselmann | KEY_BOOT_ID = "bootid" |
53 | 5a3103e9 | Michael Hanselmann | |
54 | 5a3103e9 | Michael Hanselmann | |
55 | e125c67c | Michael Hanselmann | # Global client object |
56 | e125c67c | Michael Hanselmann | client = None |
57 | e125c67c | Michael Hanselmann | |
58 | e125c67c | Michael Hanselmann | |
59 | 7bca53e4 | Michael Hanselmann | class NotMasterError(errors.GenericError): |
60 | 38242904 | Iustin Pop | """Exception raised when this host is not the master.""" |
61 | a8083063 | Iustin Pop | |
62 | a8083063 | Iustin Pop | |
63 | a8083063 | Iustin Pop | def Indent(s, prefix='| '): |
64 | a8083063 | Iustin Pop | """Indent a piece of text with a given prefix before each line. |
65 | a8083063 | Iustin Pop | |
66 | c41eea6e | Iustin Pop | @param s: the string to indent |
67 | c41eea6e | Iustin Pop | @param prefix: the string to prepend each line |
68 | 38242904 | Iustin Pop | |
69 | a8083063 | Iustin Pop | """ |
70 | a8083063 | Iustin Pop | return "%s%s\n" % (prefix, ('\n' + prefix).join(s.splitlines())) |
71 | a8083063 | Iustin Pop | |
72 | a8083063 | Iustin Pop | |
73 | 7dfb83c2 | Iustin Pop | def StartMaster(): |
74 | 7dfb83c2 | Iustin Pop | """Try to start the master daemon. |
75 | 7dfb83c2 | Iustin Pop | |
76 | 7dfb83c2 | Iustin Pop | """ |
77 | 7dfb83c2 | Iustin Pop | result = utils.RunCmd(['ganeti-masterd']) |
78 | 7dfb83c2 | Iustin Pop | if result.failed: |
79 | 7dfb83c2 | Iustin Pop | logging.error("Can't start the master daemon: output '%s'", result.output) |
80 | 7dfb83c2 | Iustin Pop | return not result.failed |
81 | 7dfb83c2 | Iustin Pop | |
82 | 7dfb83c2 | Iustin Pop | |
83 | c4f0219c | Iustin Pop | def EnsureDaemon(daemon): |
84 | c4f0219c | Iustin Pop | """Check for and start daemon if not alive. |
85 | c4f0219c | Iustin Pop | |
86 | c4f0219c | Iustin Pop | """ |
87 | c4f0219c | Iustin Pop | pidfile = utils.DaemonPidFileName(daemon) |
88 | c4f0219c | Iustin Pop | pid = utils.ReadPidFile(pidfile) |
89 | c4f0219c | Iustin Pop | if pid == 0 or not utils.IsProcessAlive(pid): # no file or dead pid |
90 | c4f0219c | Iustin Pop | logging.debug("Daemon '%s' not alive, trying to restart", daemon) |
91 | c4f0219c | Iustin Pop | result = utils.RunCmd([daemon]) |
92 | c4f0219c | Iustin Pop | if not result: |
93 | c4f0219c | Iustin Pop | logging.error("Can't start daemon '%s', failure %s, output: %s", |
94 | c4f0219c | Iustin Pop | daemon, result.fail_reason, result.output) |
95 | c4f0219c | Iustin Pop | |
96 | c4f0219c | Iustin Pop | |
97 | 5a3103e9 | Michael Hanselmann | class WatcherState(object): |
98 | a8083063 | Iustin Pop | """Interface to a state file recording restart attempts. |
99 | a8083063 | Iustin Pop | |
100 | a8083063 | Iustin Pop | """ |
101 | a8083063 | Iustin Pop | def __init__(self): |
102 | 5a3103e9 | Michael Hanselmann | """Open, lock, read and parse the file. |
103 | 5a3103e9 | Michael Hanselmann | |
104 | 7bca53e4 | Michael Hanselmann | Raises exception on lock contention. |
105 | 5a3103e9 | Michael Hanselmann | |
106 | 5a3103e9 | Michael Hanselmann | """ |
107 | a8083063 | Iustin Pop | # The two-step dance below is necessary to allow both opening existing |
108 | a8083063 | Iustin Pop | # file read/write and creating if not existing. Vanilla open will truncate |
109 | a8083063 | Iustin Pop | # an existing file -or- allow creating if not existing. |
110 | fc428e32 | Michael Hanselmann | fd = os.open(constants.WATCHER_STATEFILE, os.O_RDWR | os.O_CREAT) |
111 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
112 | a8083063 | Iustin Pop | |
113 | eb0f0ce0 | Michael Hanselmann | utils.LockFile(self.statefile.fileno()) |
114 | a8083063 | Iustin Pop | |
115 | 5a3103e9 | Michael Hanselmann | try: |
116 | 2c404217 | Iustin Pop | state_data = self.statefile.read() |
117 | 2c404217 | Iustin Pop | if not state_data: |
118 | 2c404217 | Iustin Pop | self._data = {} |
119 | 2c404217 | Iustin Pop | else: |
120 | 2c404217 | Iustin Pop | self._data = serializer.Load(state_data) |
121 | 5a3103e9 | Michael Hanselmann | except Exception, msg: |
122 | 5a3103e9 | Michael Hanselmann | # Ignore errors while loading the file and treat it as empty |
123 | b76f660d | Michael Hanselmann | self._data = {} |
124 | 2c404217 | Iustin Pop | logging.warning(("Invalid state file. Using defaults." |
125 | 438b45d4 | Michael Hanselmann | " Error message: %s"), msg) |
126 | 5a3103e9 | Michael Hanselmann | |
127 | b76f660d | Michael Hanselmann | if "instance" not in self._data: |
128 | b76f660d | Michael Hanselmann | self._data["instance"] = {} |
129 | b76f660d | Michael Hanselmann | if "node" not in self._data: |
130 | b76f660d | Michael Hanselmann | self._data["node"] = {} |
131 | 5a3103e9 | Michael Hanselmann | |
132 | 26517d45 | Iustin Pop | self._orig_data = serializer.Dump(self._data) |
133 | 2fb96d39 | Michael Hanselmann | |
134 | fc428e32 | Michael Hanselmann | def Save(self): |
135 | fc428e32 | Michael Hanselmann | """Save state to file, then unlock and close it. |
136 | 5a3103e9 | Michael Hanselmann | |
137 | 5a3103e9 | Michael Hanselmann | """ |
138 | fc428e32 | Michael Hanselmann | assert self.statefile |
139 | fc428e32 | Michael Hanselmann | |
140 | 26517d45 | Iustin Pop | serialized_form = serializer.Dump(self._data) |
141 | 26517d45 | Iustin Pop | if self._orig_data == serialized_form: |
142 | 2fb96d39 | Michael Hanselmann | logging.debug("Data didn't change, just touching status file") |
143 | 2fb96d39 | Michael Hanselmann | os.utime(constants.WATCHER_STATEFILE, None) |
144 | 2fb96d39 | Michael Hanselmann | return |
145 | 2fb96d39 | Michael Hanselmann | |
146 | fc428e32 | Michael Hanselmann | # We need to make sure the file is locked before renaming it, otherwise |
147 | fc428e32 | Michael Hanselmann | # starting ganeti-watcher again at the same time will create a conflict. |
148 | fc428e32 | Michael Hanselmann | fd = utils.WriteFile(constants.WATCHER_STATEFILE, |
149 | 26517d45 | Iustin Pop | data=serialized_form, |
150 | eb0f0ce0 | Michael Hanselmann | prewrite=utils.LockFile, close=False) |
151 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
152 | 5a3103e9 | Michael Hanselmann | |
153 | fc428e32 | Michael Hanselmann | def Close(self): |
154 | 5a3103e9 | Michael Hanselmann | """Unlock configuration file and close it. |
155 | 5a3103e9 | Michael Hanselmann | |
156 | 5a3103e9 | Michael Hanselmann | """ |
157 | 5a3103e9 | Michael Hanselmann | assert self.statefile |
158 | 5a3103e9 | Michael Hanselmann | |
159 | fc428e32 | Michael Hanselmann | # Files are automatically unlocked when closing them |
160 | 5a3103e9 | Michael Hanselmann | self.statefile.close() |
161 | 5a3103e9 | Michael Hanselmann | self.statefile = None |
162 | 5a3103e9 | Michael Hanselmann | |
163 | 5a3103e9 | Michael Hanselmann | def GetNodeBootID(self, name): |
164 | 5a3103e9 | Michael Hanselmann | """Returns the last boot ID of a node or None. |
165 | a8083063 | Iustin Pop | |
166 | 5a3103e9 | Michael Hanselmann | """ |
167 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
168 | 5a3103e9 | Michael Hanselmann | |
169 | 7b195d9b | Michael Hanselmann | if name in ndata and KEY_BOOT_ID in ndata[name]: |
170 | 7b195d9b | Michael Hanselmann | return ndata[name][KEY_BOOT_ID] |
171 | 5a3103e9 | Michael Hanselmann | return None |
172 | 5a3103e9 | Michael Hanselmann | |
173 | 5a3103e9 | Michael Hanselmann | def SetNodeBootID(self, name, bootid): |
174 | 5a3103e9 | Michael Hanselmann | """Sets the boot ID of a node. |
175 | 5a3103e9 | Michael Hanselmann | |
176 | 5a3103e9 | Michael Hanselmann | """ |
177 | 5a3103e9 | Michael Hanselmann | assert bootid |
178 | a8083063 | Iustin Pop | |
179 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
180 | a8083063 | Iustin Pop | |
181 | 5a3103e9 | Michael Hanselmann | if name not in ndata: |
182 | 5a3103e9 | Michael Hanselmann | ndata[name] = {} |
183 | 5a3103e9 | Michael Hanselmann | |
184 | 7b195d9b | Michael Hanselmann | ndata[name][KEY_BOOT_ID] = bootid |
185 | 5a3103e9 | Michael Hanselmann | |
186 | 5a3103e9 | Michael Hanselmann | def NumberOfRestartAttempts(self, instance): |
187 | a8083063 | Iustin Pop | """Returns number of previous restart attempts. |
188 | a8083063 | Iustin Pop | |
189 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
190 | c41eea6e | Iustin Pop | @param instance: the instance to look up |
191 | 38242904 | Iustin Pop | |
192 | a8083063 | Iustin Pop | """ |
193 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
194 | a8083063 | Iustin Pop | |
195 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
196 | 7b195d9b | Michael Hanselmann | return idata[instance.name][KEY_RESTART_COUNT] |
197 | a8083063 | Iustin Pop | |
198 | a8083063 | Iustin Pop | return 0 |
199 | a8083063 | Iustin Pop | |
200 | 5a3103e9 | Michael Hanselmann | def RecordRestartAttempt(self, instance): |
201 | a8083063 | Iustin Pop | """Record a restart attempt. |
202 | a8083063 | Iustin Pop | |
203 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
204 | c41eea6e | Iustin Pop | @param instance: the instance being restarted |
205 | 38242904 | Iustin Pop | |
206 | a8083063 | Iustin Pop | """ |
207 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
208 | a8083063 | Iustin Pop | |
209 | 5a3103e9 | Michael Hanselmann | if instance.name not in idata: |
210 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] = {} |
211 | 5a3103e9 | Michael Hanselmann | else: |
212 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] |
213 | a8083063 | Iustin Pop | |
214 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_WHEN] = time.time() |
215 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
216 | a8083063 | Iustin Pop | |
217 | 5a3103e9 | Michael Hanselmann | def RemoveInstance(self, instance): |
218 | c41eea6e | Iustin Pop | """Update state to reflect that a machine is running. |
219 | a8083063 | Iustin Pop | |
220 | c41eea6e | Iustin Pop | This method removes the record for a named instance (as we only |
221 | c41eea6e | Iustin Pop | track down instances). |
222 | a8083063 | Iustin Pop | |
223 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
224 | c41eea6e | Iustin Pop | @param instance: the instance to remove from books |
225 | 38242904 | Iustin Pop | |
226 | a8083063 | Iustin Pop | """ |
227 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
228 | a8083063 | Iustin Pop | |
229 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
230 | 5a3103e9 | Michael Hanselmann | del idata[instance.name] |
231 | a8083063 | Iustin Pop | |
232 | a8083063 | Iustin Pop | |
233 | a8083063 | Iustin Pop | class Instance(object): |
234 | a8083063 | Iustin Pop | """Abstraction for a Virtual Machine instance. |
235 | a8083063 | Iustin Pop | |
236 | a8083063 | Iustin Pop | """ |
237 | 5a3103e9 | Michael Hanselmann | def __init__(self, name, state, autostart): |
238 | a8083063 | Iustin Pop | self.name = name |
239 | a8083063 | Iustin Pop | self.state = state |
240 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart |
241 | a8083063 | Iustin Pop | |
242 | a8083063 | Iustin Pop | def Restart(self): |
243 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance. |
244 | 3ecf6786 | Iustin Pop | |
245 | 3ecf6786 | Iustin Pop | """ |
246 | 07813a9e | Iustin Pop | op = opcodes.OpStartupInstance(instance_name=self.name, force=False) |
247 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
248 | a8083063 | Iustin Pop | |
249 | 5a3103e9 | Michael Hanselmann | def ActivateDisks(self): |
250 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance. |
251 | 5a3103e9 | Michael Hanselmann | |
252 | 5a3103e9 | Michael Hanselmann | """ |
253 | e125c67c | Michael Hanselmann | op = opcodes.OpActivateInstanceDisks(instance_name=self.name) |
254 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
255 | a8083063 | Iustin Pop | |
256 | a8083063 | Iustin Pop | |
257 | 6dfcc47b | Iustin Pop | def GetClusterData(): |
258 | 5a3103e9 | Michael Hanselmann | """Get a list of instances on this cluster. |
259 | 5a3103e9 | Michael Hanselmann | |
260 | 5a3103e9 | Michael Hanselmann | """ |
261 | 6dfcc47b | Iustin Pop | op1_fields = ["name", "status", "admin_state", "snodes"] |
262 | 6dfcc47b | Iustin Pop | op1 = opcodes.OpQueryInstances(output_fields=op1_fields, names=[], |
263 | 6dfcc47b | Iustin Pop | use_locking=True) |
264 | 6dfcc47b | Iustin Pop | op2_fields = ["name", "bootid", "offline"] |
265 | 6dfcc47b | Iustin Pop | op2 = opcodes.OpQueryNodes(output_fields=op2_fields, names=[], |
266 | 6dfcc47b | Iustin Pop | use_locking=True) |
267 | a8083063 | Iustin Pop | |
268 | 6dfcc47b | Iustin Pop | job_id = client.SubmitJob([op1, op2]) |
269 | a8083063 | Iustin Pop | |
270 | 6dfcc47b | Iustin Pop | all_results = cli.PollJob(job_id, cl=client, feedback_fn=logging.debug) |
271 | 5a3103e9 | Michael Hanselmann | |
272 | 78f44650 | Iustin Pop | logging.debug("Got data from cluster, writing instance status file") |
273 | 78f44650 | Iustin Pop | |
274 | 6dfcc47b | Iustin Pop | result = all_results[0] |
275 | 6dfcc47b | Iustin Pop | smap = {} |
276 | 5a3103e9 | Michael Hanselmann | |
277 | 6dfcc47b | Iustin Pop | instances = {} |
278 | 78f44650 | Iustin Pop | |
279 | 78f44650 | Iustin Pop | # write the upfile |
280 | 78f44650 | Iustin Pop | up_data = "".join(["%s %s\n" % (fields[0], fields[1]) for fields in result]) |
281 | 78f44650 | Iustin Pop | utils.WriteFile(file_name=constants.INSTANCE_UPFILE, data=up_data) |
282 | 78f44650 | Iustin Pop | |
283 | 6dfcc47b | Iustin Pop | for fields in result: |
284 | 6dfcc47b | Iustin Pop | (name, status, autostart, snodes) = fields |
285 | 5a3103e9 | Michael Hanselmann | |
286 | 6dfcc47b | Iustin Pop | # update the secondary node map |
287 | 6dfcc47b | Iustin Pop | for node in snodes: |
288 | 6dfcc47b | Iustin Pop | if node not in smap: |
289 | 6dfcc47b | Iustin Pop | smap[node] = [] |
290 | 6dfcc47b | Iustin Pop | smap[node].append(name) |
291 | a8083063 | Iustin Pop | |
292 | 6dfcc47b | Iustin Pop | instances[name] = Instance(name, status, autostart) |
293 | 5a3103e9 | Michael Hanselmann | |
294 | 6dfcc47b | Iustin Pop | nodes = dict([(name, (bootid, offline)) |
295 | 6dfcc47b | Iustin Pop | for name, bootid, offline in all_results[1]]) |
296 | 5a3103e9 | Michael Hanselmann | |
297 | 6dfcc47b | Iustin Pop | client.ArchiveJob(job_id) |
298 | 5a3103e9 | Michael Hanselmann | |
299 | 6dfcc47b | Iustin Pop | return instances, nodes, smap |
300 | a8083063 | Iustin Pop | |
301 | a8083063 | Iustin Pop | |
302 | 5a3103e9 | Michael Hanselmann | class Watcher(object): |
303 | a8083063 | Iustin Pop | """Encapsulate the logic for restarting erronously halted virtual machines. |
304 | a8083063 | Iustin Pop | |
305 | a8083063 | Iustin Pop | The calling program should periodically instantiate me and call Run(). |
306 | a8083063 | Iustin Pop | This will traverse the list of instances, and make up to MAXTRIES attempts |
307 | a8083063 | Iustin Pop | to restart machines that are down. |
308 | 38242904 | Iustin Pop | |
309 | a8083063 | Iustin Pop | """ |
310 | cc962d58 | Iustin Pop | def __init__(self, opts, notepad): |
311 | cc962d58 | Iustin Pop | self.notepad = notepad |
312 | 2859b87b | Michael Hanselmann | master = client.QueryConfigValues(["master_node"])[0] |
313 | 89e1fc26 | Iustin Pop | if master != utils.HostInfo().name: |
314 | 3ecf6786 | Iustin Pop | raise NotMasterError("This is not the master node") |
315 | 24edc6d4 | Iustin Pop | # first archive old jobs |
316 | 24edc6d4 | Iustin Pop | self.ArchiveJobs(opts.job_age) |
317 | 24edc6d4 | Iustin Pop | # and only then submit new ones |
318 | 6dfcc47b | Iustin Pop | self.instances, self.bootids, self.smap = GetClusterData() |
319 | eee1fa2d | Iustin Pop | self.started_instances = set() |
320 | f07521e5 | Iustin Pop | self.opts = opts |
321 | a8083063 | Iustin Pop | |
322 | a8083063 | Iustin Pop | def Run(self): |
323 | cc962d58 | Iustin Pop | """Watcher run sequence. |
324 | cc962d58 | Iustin Pop | |
325 | cc962d58 | Iustin Pop | """ |
326 | cc962d58 | Iustin Pop | notepad = self.notepad |
327 | cc962d58 | Iustin Pop | self.CheckInstances(notepad) |
328 | cc962d58 | Iustin Pop | self.CheckDisks(notepad) |
329 | cc962d58 | Iustin Pop | self.VerifyDisks() |
330 | 5a3103e9 | Michael Hanselmann | |
331 | 24edc6d4 | Iustin Pop | @staticmethod |
332 | 24edc6d4 | Iustin Pop | def ArchiveJobs(age): |
333 | f07521e5 | Iustin Pop | """Archive old jobs. |
334 | f07521e5 | Iustin Pop | |
335 | f07521e5 | Iustin Pop | """ |
336 | f07521e5 | Iustin Pop | arch_count, left_count = client.AutoArchiveJobs(age) |
337 | f07521e5 | Iustin Pop | logging.debug("Archived %s jobs, left %s" % (arch_count, left_count)) |
338 | f07521e5 | Iustin Pop | |
339 | 5a3103e9 | Michael Hanselmann | def CheckDisks(self, notepad): |
340 | 5a3103e9 | Michael Hanselmann | """Check all nodes for restarted ones. |
341 | 38242904 | Iustin Pop | |
342 | a8083063 | Iustin Pop | """ |
343 | 5a3103e9 | Michael Hanselmann | check_nodes = [] |
344 | cbfc4681 | Iustin Pop | for name, (new_id, offline) in self.bootids.iteritems(): |
345 | 5a3103e9 | Michael Hanselmann | old = notepad.GetNodeBootID(name) |
346 | 37b77b18 | Iustin Pop | if new_id is None: |
347 | 37b77b18 | Iustin Pop | # Bad node, not returning a boot id |
348 | cbfc4681 | Iustin Pop | if not offline: |
349 | cbfc4681 | Iustin Pop | logging.debug("Node %s missing boot id, skipping secondary checks", |
350 | cbfc4681 | Iustin Pop | name) |
351 | 37b77b18 | Iustin Pop | continue |
352 | 26517d45 | Iustin Pop | if old != new_id: |
353 | 5a3103e9 | Michael Hanselmann | # Node's boot ID has changed, proably through a reboot. |
354 | 5a3103e9 | Michael Hanselmann | check_nodes.append(name) |
355 | 5a3103e9 | Michael Hanselmann | |
356 | 5a3103e9 | Michael Hanselmann | if check_nodes: |
357 | 5a3103e9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a |
358 | 5a3103e9 | Michael Hanselmann | # secondary node. |
359 | 6dfcc47b | Iustin Pop | for node in check_nodes: |
360 | 6dfcc47b | Iustin Pop | if node not in self.smap: |
361 | eee1fa2d | Iustin Pop | continue |
362 | 6dfcc47b | Iustin Pop | for instance_name in self.smap[node]: |
363 | 6dfcc47b | Iustin Pop | instance = self.instances[instance_name] |
364 | 6dfcc47b | Iustin Pop | if not instance.autostart: |
365 | 6dfcc47b | Iustin Pop | logging.info(("Skipping disk activation for non-autostart" |
366 | 6dfcc47b | Iustin Pop | " instance %s"), instance.name) |
367 | 6dfcc47b | Iustin Pop | continue |
368 | 6dfcc47b | Iustin Pop | if instance.name in self.started_instances: |
369 | 6dfcc47b | Iustin Pop | # we already tried to start the instance, which should have |
370 | 6dfcc47b | Iustin Pop | # activated its drives (if they can be at all) |
371 | 6dfcc47b | Iustin Pop | continue |
372 | 6dfcc47b | Iustin Pop | try: |
373 | 6dfcc47b | Iustin Pop | logging.info("Activating disks for instance %s", instance.name) |
374 | 6dfcc47b | Iustin Pop | instance.ActivateDisks() |
375 | 6dfcc47b | Iustin Pop | except Exception: |
376 | 6dfcc47b | Iustin Pop | logging.exception("Error while activating disks for instance %s", |
377 | 6dfcc47b | Iustin Pop | instance.name) |
378 | 5a3103e9 | Michael Hanselmann | |
379 | 5a3103e9 | Michael Hanselmann | # Keep changed boot IDs |
380 | 5a3103e9 | Michael Hanselmann | for name in check_nodes: |
381 | 3448aa22 | Iustin Pop | notepad.SetNodeBootID(name, self.bootids[name][0]) |
382 | a8083063 | Iustin Pop | |
383 | 5a3103e9 | Michael Hanselmann | def CheckInstances(self, notepad): |
384 | 5a3103e9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones. |
385 | 5a3103e9 | Michael Hanselmann | |
386 | 5a3103e9 | Michael Hanselmann | """ |
387 | 6dfcc47b | Iustin Pop | for instance in self.instances.values(): |
388 | a8083063 | Iustin Pop | if instance.state in BAD_STATES: |
389 | 5a3103e9 | Michael Hanselmann | n = notepad.NumberOfRestartAttempts(instance) |
390 | a8083063 | Iustin Pop | |
391 | a8083063 | Iustin Pop | if n > MAXTRIES: |
392 | a8083063 | Iustin Pop | # stay quiet. |
393 | a8083063 | Iustin Pop | continue |
394 | a8083063 | Iustin Pop | elif n < MAXTRIES: |
395 | a8083063 | Iustin Pop | last = " (Attempt #%d)" % (n + 1) |
396 | a8083063 | Iustin Pop | else: |
397 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
398 | 438b45d4 | Michael Hanselmann | logging.error("Could not restart %s after %d attempts, giving up", |
399 | 438b45d4 | Michael Hanselmann | instance.name, MAXTRIES) |
400 | a8083063 | Iustin Pop | continue |
401 | a8083063 | Iustin Pop | try: |
402 | 438b45d4 | Michael Hanselmann | logging.info("Restarting %s%s", |
403 | 438b45d4 | Michael Hanselmann | instance.name, last) |
404 | a8083063 | Iustin Pop | instance.Restart() |
405 | eee1fa2d | Iustin Pop | self.started_instances.add(instance.name) |
406 | b7309a0d | Iustin Pop | except Exception: |
407 | 4bffa7f7 | Iustin Pop | logging.exception("Error while restarting instance %s", |
408 | 4bffa7f7 | Iustin Pop | instance.name) |
409 | a8083063 | Iustin Pop | |
410 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
411 | a8083063 | Iustin Pop | elif instance.state in HELPLESS_STATES: |
412 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
413 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
414 | a8083063 | Iustin Pop | else: |
415 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
416 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
417 | 438b45d4 | Michael Hanselmann | logging.info("Restart of %s succeeded", instance.name) |
418 | a8083063 | Iustin Pop | |
419 | b7309a0d | Iustin Pop | @staticmethod |
420 | b7309a0d | Iustin Pop | def VerifyDisks(): |
421 | d2f311db | Iustin Pop | """Run gnt-cluster verify-disks. |
422 | d2f311db | Iustin Pop | |
423 | d2f311db | Iustin Pop | """ |
424 | 5188ab37 | Iustin Pop | op = opcodes.OpVerifyDisks() |
425 | 6dfcc47b | Iustin Pop | job_id = client.SubmitJob([op]) |
426 | 6dfcc47b | Iustin Pop | result = cli.PollJob(job_id, cl=client, feedback_fn=logging.debug)[0] |
427 | 6dfcc47b | Iustin Pop | client.ArchiveJob(job_id) |
428 | 5188ab37 | Iustin Pop | if not isinstance(result, (tuple, list)): |
429 | 5188ab37 | Iustin Pop | logging.error("Can't get a valid result from verify-disks") |
430 | 5188ab37 | Iustin Pop | return |
431 | 5188ab37 | Iustin Pop | offline_disk_instances = result[2] |
432 | 5188ab37 | Iustin Pop | if not offline_disk_instances: |
433 | 5188ab37 | Iustin Pop | # nothing to do |
434 | 5188ab37 | Iustin Pop | return |
435 | 5188ab37 | Iustin Pop | logging.debug("Will activate disks for instances %s", |
436 | 5188ab37 | Iustin Pop | ", ".join(offline_disk_instances)) |
437 | 5188ab37 | Iustin Pop | # we submit only one job, and wait for it. not optimal, but spams |
438 | 5188ab37 | Iustin Pop | # less the job queue |
439 | 5188ab37 | Iustin Pop | job = [opcodes.OpActivateInstanceDisks(instance_name=name) |
440 | 5188ab37 | Iustin Pop | for name in offline_disk_instances] |
441 | 5188ab37 | Iustin Pop | job_id = cli.SendJob(job, cl=client) |
442 | 5188ab37 | Iustin Pop | |
443 | 5188ab37 | Iustin Pop | cli.PollJob(job_id, cl=client, feedback_fn=logging.debug) |
444 | a8083063 | Iustin Pop | |
445 | a8083063 | Iustin Pop | |
446 | a8083063 | Iustin Pop | def ParseOptions(): |
447 | a8083063 | Iustin Pop | """Parse the command line options. |
448 | a8083063 | Iustin Pop | |
449 | c41eea6e | Iustin Pop | @return: (options, args) as from OptionParser.parse_args() |
450 | a8083063 | Iustin Pop | |
451 | a8083063 | Iustin Pop | """ |
452 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher", |
453 | a8083063 | Iustin Pop | usage="%prog [-d]", |
454 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" % |
455 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
456 | a8083063 | Iustin Pop | |
457 | a8083063 | Iustin Pop | parser.add_option("-d", "--debug", dest="debug", |
458 | 438b45d4 | Michael Hanselmann | help="Write all messages to stderr", |
459 | a8083063 | Iustin Pop | default=False, action="store_true") |
460 | f07521e5 | Iustin Pop | parser.add_option("-A", "--job-age", dest="job_age", |
461 | f07521e5 | Iustin Pop | help="Autoarchive jobs older than this age (default" |
462 | f07521e5 | Iustin Pop | " 6 hours)", default=6*3600) |
463 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
464 | f07521e5 | Iustin Pop | options.job_age = cli.ParseTimespec(options.job_age) |
465 | a8083063 | Iustin Pop | return options, args |
466 | a8083063 | Iustin Pop | |
467 | a8083063 | Iustin Pop | |
468 | a8083063 | Iustin Pop | def main(): |
469 | a8083063 | Iustin Pop | """Main function. |
470 | a8083063 | Iustin Pop | |
471 | a8083063 | Iustin Pop | """ |
472 | e125c67c | Michael Hanselmann | global client |
473 | e125c67c | Michael Hanselmann | |
474 | a8083063 | Iustin Pop | options, args = ParseOptions() |
475 | a8083063 | Iustin Pop | |
476 | 82d9caef | Iustin Pop | utils.SetupLogging(constants.LOG_WATCHER, debug=options.debug, |
477 | 82d9caef | Iustin Pop | stderr_logging=options.debug) |
478 | a8083063 | Iustin Pop | |
479 | 24edc6d4 | Iustin Pop | update_file = False |
480 | a8083063 | Iustin Pop | try: |
481 | 83052f9e | Guido Trotter | # on master or not, try to start the node dameon |
482 | 83052f9e | Guido Trotter | EnsureDaemon(constants.NODED) |
483 | c4f0219c | Iustin Pop | |
484 | cc962d58 | Iustin Pop | notepad = WatcherState() |
485 | 781b2b2b | Michael Hanselmann | try: |
486 | 2c404217 | Iustin Pop | try: |
487 | 2c404217 | Iustin Pop | client = cli.GetClient() |
488 | 2c404217 | Iustin Pop | except errors.OpPrereqError: |
489 | 2c404217 | Iustin Pop | # this is, from cli.GetClient, a not-master case |
490 | 7dfb83c2 | Iustin Pop | logging.debug("Not on master, exiting") |
491 | 24edc6d4 | Iustin Pop | update_file = True |
492 | 2c404217 | Iustin Pop | sys.exit(constants.EXIT_SUCCESS) |
493 | 7dfb83c2 | Iustin Pop | except luxi.NoMasterError, err: |
494 | 7dfb83c2 | Iustin Pop | logging.warning("Master seems to be down (%s), trying to restart", |
495 | 7dfb83c2 | Iustin Pop | str(err)) |
496 | 7dfb83c2 | Iustin Pop | if not StartMaster(): |
497 | 7dfb83c2 | Iustin Pop | logging.critical("Can't start the master, exiting") |
498 | 7dfb83c2 | Iustin Pop | sys.exit(constants.EXIT_FAILURE) |
499 | 7dfb83c2 | Iustin Pop | # else retry the connection |
500 | 7dfb83c2 | Iustin Pop | client = cli.GetClient() |
501 | cc962d58 | Iustin Pop | |
502 | 83052f9e | Guido Trotter | # we are on master now |
503 | 83052f9e | Guido Trotter | EnsureDaemon(constants.RAPI) |
504 | c4f0219c | Iustin Pop | |
505 | cc962d58 | Iustin Pop | try: |
506 | cc962d58 | Iustin Pop | watcher = Watcher(options, notepad) |
507 | cc962d58 | Iustin Pop | except errors.ConfigurationError: |
508 | cc962d58 | Iustin Pop | # Just exit if there's no configuration |
509 | 24edc6d4 | Iustin Pop | update_file = True |
510 | cc962d58 | Iustin Pop | sys.exit(constants.EXIT_SUCCESS) |
511 | e125c67c | Michael Hanselmann | |
512 | cc962d58 | Iustin Pop | watcher.Run() |
513 | 24edc6d4 | Iustin Pop | update_file = True |
514 | 24edc6d4 | Iustin Pop | |
515 | cc962d58 | Iustin Pop | finally: |
516 | 7dfb83c2 | Iustin Pop | if update_file: |
517 | 7dfb83c2 | Iustin Pop | notepad.Save() |
518 | 7dfb83c2 | Iustin Pop | else: |
519 | 7dfb83c2 | Iustin Pop | logging.debug("Not updating status file due to failure") |
520 | 1b052f42 | Michael Hanselmann | except SystemExit: |
521 | 1b052f42 | Michael Hanselmann | raise |
522 | 38242904 | Iustin Pop | except NotMasterError: |
523 | 438b45d4 | Michael Hanselmann | logging.debug("Not master, exiting") |
524 | 38242904 | Iustin Pop | sys.exit(constants.EXIT_NOTMASTER) |
525 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err: |
526 | 438b45d4 | Michael Hanselmann | logging.error("Cannot resolve hostname '%s', exiting.", err.args[0]) |
527 | 89e1fc26 | Iustin Pop | sys.exit(constants.EXIT_NODESETUP_ERROR) |
528 | 24edc6d4 | Iustin Pop | except errors.JobQueueFull: |
529 | 24edc6d4 | Iustin Pop | logging.error("Job queue is full, can't query cluster state") |
530 | 24edc6d4 | Iustin Pop | except errors.JobQueueDrainError: |
531 | 24edc6d4 | Iustin Pop | logging.error("Job queue is drained, can't maintain cluster state") |
532 | 438b45d4 | Michael Hanselmann | except Exception, err: |
533 | 438b45d4 | Michael Hanselmann | logging.error(str(err), exc_info=True) |
534 | 438b45d4 | Michael Hanselmann | sys.exit(constants.EXIT_FAILURE) |
535 | a8083063 | Iustin Pop | |
536 | 5a3103e9 | Michael Hanselmann | |
537 | a8083063 | Iustin Pop | if __name__ == '__main__': |
538 | a8083063 | Iustin Pop | main() |