root / daemons / ganeti-watcher @ 973d7867
History | View | Annotate | Download (12.2 kB)
1 | a8083063 | Iustin Pop | #!/usr/bin/python |
---|---|---|---|
2 | a8083063 | Iustin Pop | # |
3 | a8083063 | Iustin Pop | |
4 | a8083063 | Iustin Pop | # Copyright (C) 2006, 2007 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 | 5a3103e9 | Michael Hanselmann | import re |
33 | a8083063 | Iustin Pop | import time |
34 | a8083063 | Iustin Pop | import fcntl |
35 | a8083063 | Iustin Pop | import errno |
36 | 5a3103e9 | Michael Hanselmann | import simplejson |
37 | a8083063 | Iustin Pop | from optparse import OptionParser |
38 | a8083063 | Iustin Pop | |
39 | a8083063 | Iustin Pop | from ganeti import utils |
40 | a8083063 | Iustin Pop | from ganeti import constants |
41 | 38242904 | Iustin Pop | from ganeti import ssconf |
42 | 89e1fc26 | Iustin Pop | from ganeti import errors |
43 | a8083063 | Iustin Pop | |
44 | a8083063 | Iustin Pop | |
45 | 5a3103e9 | Michael Hanselmann | MAXTRIES = 5 |
46 | 5a3103e9 | Michael Hanselmann | BAD_STATES = ['stopped'] |
47 | 5a3103e9 | Michael Hanselmann | HELPLESS_STATES = ['(node down)'] |
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 | a8083063 | Iustin Pop | class Error(Exception): |
56 | a8083063 | Iustin Pop | """Generic custom error class.""" |
57 | 38242904 | Iustin Pop | |
58 | 38242904 | Iustin Pop | |
59 | 38242904 | Iustin Pop | class NotMasterError(Error): |
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 | a8083063 | Iustin Pop | Args: |
67 | a8083063 | Iustin Pop | s: The string to indent |
68 | a8083063 | Iustin Pop | prefix: The string to prepend each line. |
69 | 38242904 | Iustin Pop | |
70 | a8083063 | Iustin Pop | """ |
71 | a8083063 | Iustin Pop | return "%s%s\n" % (prefix, ('\n' + prefix).join(s.splitlines())) |
72 | a8083063 | Iustin Pop | |
73 | a8083063 | Iustin Pop | |
74 | a8083063 | Iustin Pop | def DoCmd(cmd): |
75 | a8083063 | Iustin Pop | """Run a shell command. |
76 | a8083063 | Iustin Pop | |
77 | a8083063 | Iustin Pop | Args: |
78 | a8083063 | Iustin Pop | cmd: the command to run. |
79 | a8083063 | Iustin Pop | |
80 | a8083063 | Iustin Pop | Raises CommandError with verbose commentary on error. |
81 | 38242904 | Iustin Pop | |
82 | a8083063 | Iustin Pop | """ |
83 | a8083063 | Iustin Pop | res = utils.RunCmd(cmd) |
84 | a8083063 | Iustin Pop | |
85 | a8083063 | Iustin Pop | if res.failed: |
86 | a8083063 | Iustin Pop | raise Error("Command %s failed:\n%s\nstdout:\n%sstderr:\n%s" % |
87 | a8083063 | Iustin Pop | (repr(cmd), |
88 | a8083063 | Iustin Pop | Indent(res.fail_reason), |
89 | a8083063 | Iustin Pop | Indent(res.stdout), |
90 | a8083063 | Iustin Pop | Indent(res.stderr))) |
91 | a8083063 | Iustin Pop | |
92 | a8083063 | Iustin Pop | return res |
93 | a8083063 | Iustin Pop | |
94 | a8083063 | Iustin Pop | |
95 | 5a3103e9 | Michael Hanselmann | class WatcherState(object): |
96 | a8083063 | Iustin Pop | """Interface to a state file recording restart attempts. |
97 | a8083063 | Iustin Pop | |
98 | a8083063 | Iustin Pop | """ |
99 | a8083063 | Iustin Pop | def __init__(self): |
100 | 5a3103e9 | Michael Hanselmann | """Open, lock, read and parse the file. |
101 | 5a3103e9 | Michael Hanselmann | |
102 | 5a3103e9 | Michael Hanselmann | Raises StandardError on lock contention. |
103 | 5a3103e9 | Michael Hanselmann | |
104 | 5a3103e9 | Michael Hanselmann | """ |
105 | a8083063 | Iustin Pop | # The two-step dance below is necessary to allow both opening existing |
106 | a8083063 | Iustin Pop | # file read/write and creating if not existing. Vanilla open will truncate |
107 | a8083063 | Iustin Pop | # an existing file -or- allow creating if not existing. |
108 | a8083063 | Iustin Pop | f = os.open(constants.WATCHER_STATEFILE, os.O_RDWR | os.O_CREAT) |
109 | a8083063 | Iustin Pop | f = os.fdopen(f, 'w+') |
110 | a8083063 | Iustin Pop | |
111 | a8083063 | Iustin Pop | try: |
112 | a8083063 | Iustin Pop | fcntl.flock(f.fileno(), fcntl.LOCK_EX|fcntl.LOCK_NB) |
113 | a8083063 | Iustin Pop | except IOError, x: |
114 | a8083063 | Iustin Pop | if x.errno == errno.EAGAIN: |
115 | 3ecf6786 | Iustin Pop | raise StandardError("State file already locked") |
116 | a8083063 | Iustin Pop | raise |
117 | a8083063 | Iustin Pop | |
118 | a8083063 | Iustin Pop | self.statefile = f |
119 | a8083063 | Iustin Pop | |
120 | 5a3103e9 | Michael Hanselmann | try: |
121 | 5a3103e9 | Michael Hanselmann | self.data = simplejson.load(self.statefile) |
122 | 5a3103e9 | Michael Hanselmann | except Exception, msg: |
123 | 5a3103e9 | Michael Hanselmann | # Ignore errors while loading the file and treat it as empty |
124 | 5a3103e9 | Michael Hanselmann | self.data = {} |
125 | f4bc1f2c | Michael Hanselmann | sys.stderr.write("Empty or invalid state file." |
126 | f4bc1f2c | Michael Hanselmann | " Using defaults. Error message: %s\n" % msg) |
127 | 5a3103e9 | Michael Hanselmann | |
128 | 5a3103e9 | Michael Hanselmann | if "instance" not in self.data: |
129 | 5a3103e9 | Michael Hanselmann | self.data["instance"] = {} |
130 | 5a3103e9 | Michael Hanselmann | if "node" not in self.data: |
131 | 5a3103e9 | Michael Hanselmann | self.data["node"] = {} |
132 | 5a3103e9 | Michael Hanselmann | |
133 | 5a3103e9 | Michael Hanselmann | def __del__(self): |
134 | 5a3103e9 | Michael Hanselmann | """Called on destruction. |
135 | 5a3103e9 | Michael Hanselmann | |
136 | 5a3103e9 | Michael Hanselmann | """ |
137 | 5a3103e9 | Michael Hanselmann | if self.statefile: |
138 | 5a3103e9 | Michael Hanselmann | self._Close() |
139 | 5a3103e9 | Michael Hanselmann | |
140 | 5a3103e9 | Michael Hanselmann | def _Close(self): |
141 | 5a3103e9 | Michael Hanselmann | """Unlock configuration file and close it. |
142 | 5a3103e9 | Michael Hanselmann | |
143 | 5a3103e9 | Michael Hanselmann | """ |
144 | 5a3103e9 | Michael Hanselmann | assert self.statefile |
145 | 5a3103e9 | Michael Hanselmann | |
146 | 5a3103e9 | Michael Hanselmann | fcntl.flock(self.statefile.fileno(), fcntl.LOCK_UN) |
147 | 5a3103e9 | Michael Hanselmann | |
148 | 5a3103e9 | Michael Hanselmann | self.statefile.close() |
149 | 5a3103e9 | Michael Hanselmann | self.statefile = None |
150 | 5a3103e9 | Michael Hanselmann | |
151 | 5a3103e9 | Michael Hanselmann | def GetNodeBootID(self, name): |
152 | 5a3103e9 | Michael Hanselmann | """Returns the last boot ID of a node or None. |
153 | a8083063 | Iustin Pop | |
154 | 5a3103e9 | Michael Hanselmann | """ |
155 | 5a3103e9 | Michael Hanselmann | ndata = self.data["node"] |
156 | 5a3103e9 | Michael Hanselmann | |
157 | 7b195d9b | Michael Hanselmann | if name in ndata and KEY_BOOT_ID in ndata[name]: |
158 | 7b195d9b | Michael Hanselmann | return ndata[name][KEY_BOOT_ID] |
159 | 5a3103e9 | Michael Hanselmann | return None |
160 | 5a3103e9 | Michael Hanselmann | |
161 | 5a3103e9 | Michael Hanselmann | def SetNodeBootID(self, name, bootid): |
162 | 5a3103e9 | Michael Hanselmann | """Sets the boot ID of a node. |
163 | 5a3103e9 | Michael Hanselmann | |
164 | 5a3103e9 | Michael Hanselmann | """ |
165 | 5a3103e9 | Michael Hanselmann | assert bootid |
166 | a8083063 | Iustin Pop | |
167 | 5a3103e9 | Michael Hanselmann | ndata = self.data["node"] |
168 | a8083063 | Iustin Pop | |
169 | 5a3103e9 | Michael Hanselmann | if name not in ndata: |
170 | 5a3103e9 | Michael Hanselmann | ndata[name] = {} |
171 | 5a3103e9 | Michael Hanselmann | |
172 | 7b195d9b | Michael Hanselmann | ndata[name][KEY_BOOT_ID] = bootid |
173 | 5a3103e9 | Michael Hanselmann | |
174 | 5a3103e9 | Michael Hanselmann | def NumberOfRestartAttempts(self, instance): |
175 | a8083063 | Iustin Pop | """Returns number of previous restart attempts. |
176 | a8083063 | Iustin Pop | |
177 | a8083063 | Iustin Pop | Args: |
178 | a8083063 | Iustin Pop | instance - the instance to look up. |
179 | 38242904 | Iustin Pop | |
180 | a8083063 | Iustin Pop | """ |
181 | 5a3103e9 | Michael Hanselmann | idata = self.data["instance"] |
182 | a8083063 | Iustin Pop | |
183 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
184 | 7b195d9b | Michael Hanselmann | return idata[instance.name][KEY_RESTART_COUNT] |
185 | a8083063 | Iustin Pop | |
186 | a8083063 | Iustin Pop | return 0 |
187 | a8083063 | Iustin Pop | |
188 | 5a3103e9 | Michael Hanselmann | def RecordRestartAttempt(self, instance): |
189 | a8083063 | Iustin Pop | """Record a restart attempt. |
190 | a8083063 | Iustin Pop | |
191 | a8083063 | Iustin Pop | Args: |
192 | a8083063 | Iustin Pop | instance - the instance being restarted |
193 | 38242904 | Iustin Pop | |
194 | a8083063 | Iustin Pop | """ |
195 | 5a3103e9 | Michael Hanselmann | idata = self.data["instance"] |
196 | a8083063 | Iustin Pop | |
197 | 5a3103e9 | Michael Hanselmann | if instance.name not in idata: |
198 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] = {} |
199 | 5a3103e9 | Michael Hanselmann | else: |
200 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] |
201 | a8083063 | Iustin Pop | |
202 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_WHEN] = time.time() |
203 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
204 | a8083063 | Iustin Pop | |
205 | 5a3103e9 | Michael Hanselmann | def RemoveInstance(self, instance): |
206 | 38242904 | Iustin Pop | """Update state to reflect that a machine is running, i.e. remove record. |
207 | a8083063 | Iustin Pop | |
208 | a8083063 | Iustin Pop | Args: |
209 | a8083063 | Iustin Pop | instance - the instance to remove from books |
210 | a8083063 | Iustin Pop | |
211 | 38242904 | Iustin Pop | This method removes the record for a named instance. |
212 | 38242904 | Iustin Pop | |
213 | a8083063 | Iustin Pop | """ |
214 | 5a3103e9 | Michael Hanselmann | idata = self.data["instance"] |
215 | a8083063 | Iustin Pop | |
216 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
217 | 5a3103e9 | Michael Hanselmann | del idata[instance.name] |
218 | a8083063 | Iustin Pop | |
219 | a8083063 | Iustin Pop | def Save(self): |
220 | 5a3103e9 | Michael Hanselmann | """Save state to file, then unlock and close it. |
221 | 38242904 | Iustin Pop | |
222 | a8083063 | Iustin Pop | """ |
223 | a8083063 | Iustin Pop | assert self.statefile |
224 | a8083063 | Iustin Pop | |
225 | a8083063 | Iustin Pop | self.statefile.seek(0) |
226 | a8083063 | Iustin Pop | self.statefile.truncate() |
227 | a8083063 | Iustin Pop | |
228 | 5a3103e9 | Michael Hanselmann | simplejson.dump(self.data, self.statefile) |
229 | a8083063 | Iustin Pop | |
230 | 5a3103e9 | Michael Hanselmann | self._Close() |
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 | Methods: |
237 | a8083063 | Iustin Pop | Restart(): issue a command to restart the represented machine. |
238 | 098c0958 | Michael Hanselmann | |
239 | a8083063 | Iustin Pop | """ |
240 | 5a3103e9 | Michael Hanselmann | def __init__(self, name, state, autostart): |
241 | a8083063 | Iustin Pop | self.name = name |
242 | a8083063 | Iustin Pop | self.state = state |
243 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart |
244 | a8083063 | Iustin Pop | |
245 | a8083063 | Iustin Pop | def Restart(self): |
246 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance. |
247 | 3ecf6786 | Iustin Pop | |
248 | 3ecf6786 | Iustin Pop | """ |
249 | a8083063 | Iustin Pop | DoCmd(['gnt-instance', 'startup', '--lock-retries=15', self.name]) |
250 | a8083063 | Iustin Pop | |
251 | 5a3103e9 | Michael Hanselmann | def ActivateDisks(self): |
252 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance. |
253 | 5a3103e9 | Michael Hanselmann | |
254 | 5a3103e9 | Michael Hanselmann | """ |
255 | 5a3103e9 | Michael Hanselmann | DoCmd(['gnt-instance', 'activate-disks', '--lock-retries=15', self.name]) |
256 | 5a3103e9 | Michael Hanselmann | |
257 | a8083063 | Iustin Pop | |
258 | 5a3103e9 | Michael Hanselmann | def _RunListCmd(cmd): |
259 | 5a3103e9 | Michael Hanselmann | """Runs a command and parses its output into lists. |
260 | 38242904 | Iustin Pop | |
261 | a8083063 | Iustin Pop | """ |
262 | 5a3103e9 | Michael Hanselmann | for line in DoCmd(cmd).stdout.splitlines(): |
263 | 5a3103e9 | Michael Hanselmann | yield line.split(':') |
264 | a8083063 | Iustin Pop | |
265 | a8083063 | Iustin Pop | |
266 | 5a3103e9 | Michael Hanselmann | def GetInstanceList(with_secondaries=None): |
267 | 5a3103e9 | Michael Hanselmann | """Get a list of instances on this cluster. |
268 | 5a3103e9 | Michael Hanselmann | |
269 | 5a3103e9 | Michael Hanselmann | """ |
270 | 5a3103e9 | Michael Hanselmann | cmd = ['gnt-instance', 'list', '--lock-retries=15', '--no-headers', |
271 | 5a3103e9 | Michael Hanselmann | '--separator=:'] |
272 | 5a3103e9 | Michael Hanselmann | |
273 | 5a3103e9 | Michael Hanselmann | fields = 'name,oper_state,admin_state' |
274 | a8083063 | Iustin Pop | |
275 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
276 | 5a3103e9 | Michael Hanselmann | fields += ',snodes' |
277 | a8083063 | Iustin Pop | |
278 | 5a3103e9 | Michael Hanselmann | cmd.append('-o') |
279 | 5a3103e9 | Michael Hanselmann | cmd.append(fields) |
280 | 5a3103e9 | Michael Hanselmann | |
281 | 5a3103e9 | Michael Hanselmann | instances = [] |
282 | 5a3103e9 | Michael Hanselmann | for fields in _RunListCmd(cmd): |
283 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
284 | 5a3103e9 | Michael Hanselmann | (name, status, autostart, snodes) = fields |
285 | 5a3103e9 | Michael Hanselmann | |
286 | 5a3103e9 | Michael Hanselmann | if snodes == "-": |
287 | a8083063 | Iustin Pop | continue |
288 | 5a3103e9 | Michael Hanselmann | |
289 | 5a3103e9 | Michael Hanselmann | for node in with_secondaries: |
290 | 5a3103e9 | Michael Hanselmann | if node in snodes.split(','): |
291 | 5a3103e9 | Michael Hanselmann | break |
292 | 5a3103e9 | Michael Hanselmann | else: |
293 | a8083063 | Iustin Pop | continue |
294 | a8083063 | Iustin Pop | |
295 | 5a3103e9 | Michael Hanselmann | else: |
296 | 5a3103e9 | Michael Hanselmann | (name, status, autostart) = fields |
297 | 5a3103e9 | Michael Hanselmann | |
298 | 5a3103e9 | Michael Hanselmann | instances.append(Instance(name, status, autostart != "no")) |
299 | a8083063 | Iustin Pop | |
300 | 5a3103e9 | Michael Hanselmann | return instances |
301 | 5a3103e9 | Michael Hanselmann | |
302 | 5a3103e9 | Michael Hanselmann | |
303 | 5a3103e9 | Michael Hanselmann | def GetNodeBootIDs(): |
304 | 5a3103e9 | Michael Hanselmann | """Get a dict mapping nodes to boot IDs. |
305 | 5a3103e9 | Michael Hanselmann | |
306 | 5a3103e9 | Michael Hanselmann | """ |
307 | 5a3103e9 | Michael Hanselmann | cmd = ['gnt-node', 'list', '--lock-retries=15', '--no-headers', |
308 | 5a3103e9 | Michael Hanselmann | '--separator=:', '-o', 'name,bootid'] |
309 | 5a3103e9 | Michael Hanselmann | |
310 | 5a3103e9 | Michael Hanselmann | ids = {} |
311 | 5a3103e9 | Michael Hanselmann | for fields in _RunListCmd(cmd): |
312 | 5a3103e9 | Michael Hanselmann | (name, bootid) = fields |
313 | 5a3103e9 | Michael Hanselmann | ids[name] = bootid |
314 | 5a3103e9 | Michael Hanselmann | |
315 | 5a3103e9 | Michael Hanselmann | return ids |
316 | a8083063 | Iustin Pop | |
317 | a8083063 | Iustin Pop | |
318 | a8083063 | Iustin Pop | class Message(object): |
319 | a8083063 | Iustin Pop | """Encapsulation of a notice or error message. |
320 | 38242904 | Iustin Pop | |
321 | a8083063 | Iustin Pop | """ |
322 | a8083063 | Iustin Pop | def __init__(self, level, msg): |
323 | a8083063 | Iustin Pop | self.level = level |
324 | a8083063 | Iustin Pop | self.msg = msg |
325 | a8083063 | Iustin Pop | self.when = time.time() |
326 | a8083063 | Iustin Pop | |
327 | a8083063 | Iustin Pop | def __str__(self): |
328 | a8083063 | Iustin Pop | return self.level + ' ' + time.ctime(self.when) + '\n' + Indent(self.msg) |
329 | a8083063 | Iustin Pop | |
330 | a8083063 | Iustin Pop | |
331 | 5a3103e9 | Michael Hanselmann | class Watcher(object): |
332 | a8083063 | Iustin Pop | """Encapsulate the logic for restarting erronously halted virtual machines. |
333 | a8083063 | Iustin Pop | |
334 | a8083063 | Iustin Pop | The calling program should periodically instantiate me and call Run(). |
335 | a8083063 | Iustin Pop | This will traverse the list of instances, and make up to MAXTRIES attempts |
336 | a8083063 | Iustin Pop | to restart machines that are down. |
337 | 38242904 | Iustin Pop | |
338 | a8083063 | Iustin Pop | """ |
339 | a8083063 | Iustin Pop | def __init__(self): |
340 | 38242904 | Iustin Pop | sstore = ssconf.SimpleStore() |
341 | 38242904 | Iustin Pop | master = sstore.GetMasterNode() |
342 | 89e1fc26 | Iustin Pop | if master != utils.HostInfo().name: |
343 | 3ecf6786 | Iustin Pop | raise NotMasterError("This is not the master node") |
344 | 5a3103e9 | Michael Hanselmann | self.instances = GetInstanceList() |
345 | 5a3103e9 | Michael Hanselmann | self.bootids = GetNodeBootIDs() |
346 | a8083063 | Iustin Pop | self.messages = [] |
347 | a8083063 | Iustin Pop | |
348 | a8083063 | Iustin Pop | def Run(self): |
349 | 5a3103e9 | Michael Hanselmann | notepad = WatcherState() |
350 | 5a3103e9 | Michael Hanselmann | self.CheckInstances(notepad) |
351 | 5a3103e9 | Michael Hanselmann | self.CheckDisks(notepad) |
352 | d2f311db | Iustin Pop | self.VerifyDisks() |
353 | 5a3103e9 | Michael Hanselmann | notepad.Save() |
354 | 5a3103e9 | Michael Hanselmann | |
355 | 5a3103e9 | Michael Hanselmann | def CheckDisks(self, notepad): |
356 | 5a3103e9 | Michael Hanselmann | """Check all nodes for restarted ones. |
357 | 38242904 | Iustin Pop | |
358 | a8083063 | Iustin Pop | """ |
359 | 5a3103e9 | Michael Hanselmann | check_nodes = [] |
360 | 5a3103e9 | Michael Hanselmann | for name, id in self.bootids.iteritems(): |
361 | 5a3103e9 | Michael Hanselmann | old = notepad.GetNodeBootID(name) |
362 | 5a3103e9 | Michael Hanselmann | if old != id: |
363 | 5a3103e9 | Michael Hanselmann | # Node's boot ID has changed, proably through a reboot. |
364 | 5a3103e9 | Michael Hanselmann | check_nodes.append(name) |
365 | 5a3103e9 | Michael Hanselmann | |
366 | 5a3103e9 | Michael Hanselmann | if check_nodes: |
367 | 5a3103e9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a |
368 | 5a3103e9 | Michael Hanselmann | # secondary node. |
369 | 5a3103e9 | Michael Hanselmann | for instance in GetInstanceList(with_secondaries=check_nodes): |
370 | 5a3103e9 | Michael Hanselmann | try: |
371 | f4bc1f2c | Michael Hanselmann | self.messages.append(Message(NOTICE, ("Activating disks for %s." % |
372 | f4bc1f2c | Michael Hanselmann | instance.name))) |
373 | 5a3103e9 | Michael Hanselmann | instance.ActivateDisks() |
374 | 5a3103e9 | Michael Hanselmann | except Error, x: |
375 | 5a3103e9 | Michael Hanselmann | self.messages.append(Message(ERROR, str(x))) |
376 | 5a3103e9 | Michael Hanselmann | |
377 | 5a3103e9 | Michael Hanselmann | # Keep changed boot IDs |
378 | 5a3103e9 | Michael Hanselmann | for name in check_nodes: |
379 | 5a3103e9 | Michael Hanselmann | notepad.SetNodeBootID(name, self.bootids[name]) |
380 | a8083063 | Iustin Pop | |
381 | 5a3103e9 | Michael Hanselmann | def CheckInstances(self, notepad): |
382 | 5a3103e9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones. |
383 | 5a3103e9 | Michael Hanselmann | |
384 | 5a3103e9 | Michael Hanselmann | """ |
385 | a8083063 | Iustin Pop | for instance in self.instances: |
386 | 5a3103e9 | Michael Hanselmann | # Don't care about manually stopped instances |
387 | 5a3103e9 | Michael Hanselmann | if not instance.autostart: |
388 | 5a3103e9 | Michael Hanselmann | continue |
389 | 5a3103e9 | Michael Hanselmann | |
390 | a8083063 | Iustin Pop | if instance.state in BAD_STATES: |
391 | 5a3103e9 | Michael Hanselmann | n = notepad.NumberOfRestartAttempts(instance) |
392 | a8083063 | Iustin Pop | |
393 | a8083063 | Iustin Pop | if n > MAXTRIES: |
394 | a8083063 | Iustin Pop | # stay quiet. |
395 | a8083063 | Iustin Pop | continue |
396 | a8083063 | Iustin Pop | elif n < MAXTRIES: |
397 | a8083063 | Iustin Pop | last = " (Attempt #%d)" % (n + 1) |
398 | a8083063 | Iustin Pop | else: |
399 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
400 | a8083063 | Iustin Pop | self.messages.append(Message(ERROR, "Could not restart %s for %d" |
401 | a8083063 | Iustin Pop | " times, giving up..." % |
402 | a8083063 | Iustin Pop | (instance.name, MAXTRIES))) |
403 | a8083063 | Iustin Pop | continue |
404 | a8083063 | Iustin Pop | try: |
405 | f4bc1f2c | Michael Hanselmann | self.messages.append(Message(NOTICE, ("Restarting %s%s." % |
406 | f4bc1f2c | Michael Hanselmann | (instance.name, last)))) |
407 | a8083063 | Iustin Pop | instance.Restart() |
408 | a8083063 | Iustin Pop | except Error, x: |
409 | a8083063 | Iustin Pop | self.messages.append(Message(ERROR, str(x))) |
410 | a8083063 | Iustin Pop | |
411 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
412 | a8083063 | Iustin Pop | elif instance.state in HELPLESS_STATES: |
413 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
414 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
415 | a8083063 | Iustin Pop | else: |
416 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
417 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
418 | f4bc1f2c | Michael Hanselmann | msg = Message(NOTICE, "Restart of %s succeeded." % instance.name) |
419 | a8083063 | Iustin Pop | self.messages.append(msg) |
420 | a8083063 | Iustin Pop | |
421 | d2f311db | Iustin Pop | def VerifyDisks(self): |
422 | d2f311db | Iustin Pop | """Run gnt-cluster verify-disks. |
423 | d2f311db | Iustin Pop | |
424 | d2f311db | Iustin Pop | """ |
425 | d2f311db | Iustin Pop | result = DoCmd(['gnt-cluster', 'verify-disks', '--lock-retries=15']) |
426 | d2f311db | Iustin Pop | if result.output: |
427 | d2f311db | Iustin Pop | self.messages.append(Message(NOTICE, result.output)) |
428 | d2f311db | Iustin Pop | |
429 | a8083063 | Iustin Pop | def WriteReport(self, logfile): |
430 | 38242904 | Iustin Pop | """Log all messages to file. |
431 | a8083063 | Iustin Pop | |
432 | a8083063 | Iustin Pop | Args: |
433 | a8083063 | Iustin Pop | logfile: file object open for writing (the log file) |
434 | 38242904 | Iustin Pop | |
435 | a8083063 | Iustin Pop | """ |
436 | a8083063 | Iustin Pop | for msg in self.messages: |
437 | a8083063 | Iustin Pop | print >> logfile, str(msg) |
438 | a8083063 | Iustin Pop | |
439 | a8083063 | Iustin Pop | |
440 | a8083063 | Iustin Pop | def ParseOptions(): |
441 | a8083063 | Iustin Pop | """Parse the command line options. |
442 | a8083063 | Iustin Pop | |
443 | a8083063 | Iustin Pop | Returns: |
444 | a8083063 | Iustin Pop | (options, args) as from OptionParser.parse_args() |
445 | a8083063 | Iustin Pop | |
446 | a8083063 | Iustin Pop | """ |
447 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher", |
448 | a8083063 | Iustin Pop | usage="%prog [-d]", |
449 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" % |
450 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
451 | a8083063 | Iustin Pop | |
452 | a8083063 | Iustin Pop | parser.add_option("-d", "--debug", dest="debug", |
453 | a8083063 | Iustin Pop | help="Don't redirect messages to the log file", |
454 | a8083063 | Iustin Pop | default=False, action="store_true") |
455 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
456 | a8083063 | Iustin Pop | return options, args |
457 | a8083063 | Iustin Pop | |
458 | a8083063 | Iustin Pop | |
459 | a8083063 | Iustin Pop | def main(): |
460 | a8083063 | Iustin Pop | """Main function. |
461 | a8083063 | Iustin Pop | |
462 | a8083063 | Iustin Pop | """ |
463 | a8083063 | Iustin Pop | options, args = ParseOptions() |
464 | a8083063 | Iustin Pop | |
465 | a8083063 | Iustin Pop | if not options.debug: |
466 | 5a3103e9 | Michael Hanselmann | sys.stderr = sys.stdout = open(constants.LOG_WATCHER, 'a') |
467 | a8083063 | Iustin Pop | |
468 | a8083063 | Iustin Pop | try: |
469 | 781b2b2b | Michael Hanselmann | try: |
470 | 781b2b2b | Michael Hanselmann | watcher = Watcher() |
471 | 781b2b2b | Michael Hanselmann | except errors.ConfigurationError: |
472 | 781b2b2b | Michael Hanselmann | # Just exit if there's no configuration |
473 | 781b2b2b | Michael Hanselmann | sys.exit(constants.EXIT_SUCCESS) |
474 | 5a3103e9 | Michael Hanselmann | watcher.Run() |
475 | 5a3103e9 | Michael Hanselmann | watcher.WriteReport(sys.stdout) |
476 | 38242904 | Iustin Pop | except NotMasterError: |
477 | 38242904 | Iustin Pop | if options.debug: |
478 | 38242904 | Iustin Pop | sys.stderr.write("Not master, exiting.\n") |
479 | 38242904 | Iustin Pop | sys.exit(constants.EXIT_NOTMASTER) |
480 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err: |
481 | 89e1fc26 | Iustin Pop | sys.stderr.write("Cannot resolve hostname '%s', exiting.\n" % err.args[0]) |
482 | 89e1fc26 | Iustin Pop | sys.exit(constants.EXIT_NODESETUP_ERROR) |
483 | a8083063 | Iustin Pop | except Error, err: |
484 | a8083063 | Iustin Pop | print err |
485 | a8083063 | Iustin Pop | |
486 | 5a3103e9 | Michael Hanselmann | |
487 | a8083063 | Iustin Pop | if __name__ == '__main__': |
488 | a8083063 | Iustin Pop | main() |