root / daemons / ganeti-watcher @ 9ae49f27
History | View | Annotate | Download (12.9 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 | 5a3103e9 | Michael Hanselmann | import re |
33 | a8083063 | Iustin Pop | import time |
34 | a8083063 | Iustin Pop | import fcntl |
35 | a8083063 | Iustin Pop | import errno |
36 | 438b45d4 | Michael Hanselmann | import logging |
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 | 67fe61c4 | Michael Hanselmann | from ganeti import serializer |
42 | 38242904 | Iustin Pop | from ganeti import ssconf |
43 | 89e1fc26 | Iustin Pop | from ganeti import errors |
44 | a8083063 | Iustin Pop | |
45 | a8083063 | Iustin Pop | |
46 | 5a3103e9 | Michael Hanselmann | MAXTRIES = 5 |
47 | 5a3103e9 | Michael Hanselmann | BAD_STATES = ['stopped'] |
48 | 5a3103e9 | Michael Hanselmann | HELPLESS_STATES = ['(node down)'] |
49 | 5a3103e9 | Michael Hanselmann | NOTICE = 'NOTICE' |
50 | 5a3103e9 | Michael Hanselmann | ERROR = 'ERROR' |
51 | 7b195d9b | Michael Hanselmann | KEY_RESTART_COUNT = "restart_count" |
52 | 7b195d9b | Michael Hanselmann | KEY_RESTART_WHEN = "restart_when" |
53 | 7b195d9b | Michael Hanselmann | KEY_BOOT_ID = "bootid" |
54 | 5a3103e9 | Michael Hanselmann | |
55 | 5a3103e9 | Michael Hanselmann | |
56 | 7bca53e4 | Michael Hanselmann | class NotMasterError(errors.GenericError): |
57 | 38242904 | Iustin Pop | """Exception raised when this host is not the master.""" |
58 | a8083063 | Iustin Pop | |
59 | a8083063 | Iustin Pop | |
60 | a8083063 | Iustin Pop | def Indent(s, prefix='| '): |
61 | a8083063 | Iustin Pop | """Indent a piece of text with a given prefix before each line. |
62 | a8083063 | Iustin Pop | |
63 | a8083063 | Iustin Pop | Args: |
64 | a8083063 | Iustin Pop | s: The string to indent |
65 | a8083063 | Iustin Pop | prefix: The string to prepend each line. |
66 | 38242904 | Iustin Pop | |
67 | a8083063 | Iustin Pop | """ |
68 | a8083063 | Iustin Pop | return "%s%s\n" % (prefix, ('\n' + prefix).join(s.splitlines())) |
69 | a8083063 | Iustin Pop | |
70 | a8083063 | Iustin Pop | |
71 | a8083063 | Iustin Pop | def DoCmd(cmd): |
72 | a8083063 | Iustin Pop | """Run a shell command. |
73 | a8083063 | Iustin Pop | |
74 | a8083063 | Iustin Pop | Args: |
75 | a8083063 | Iustin Pop | cmd: the command to run. |
76 | a8083063 | Iustin Pop | |
77 | a8083063 | Iustin Pop | Raises CommandError with verbose commentary on error. |
78 | 38242904 | Iustin Pop | |
79 | a8083063 | Iustin Pop | """ |
80 | a8083063 | Iustin Pop | res = utils.RunCmd(cmd) |
81 | a8083063 | Iustin Pop | |
82 | a8083063 | Iustin Pop | if res.failed: |
83 | 7bca53e4 | Michael Hanselmann | msg = ("Command %s failed:\n%s\nstdout:\n%sstderr:\n%s" % |
84 | 7bca53e4 | Michael Hanselmann | (repr(cmd), |
85 | 7bca53e4 | Michael Hanselmann | Indent(res.fail_reason), |
86 | 7bca53e4 | Michael Hanselmann | Indent(res.stdout), |
87 | 7bca53e4 | Michael Hanselmann | Indent(res.stderr))) |
88 | 7bca53e4 | Michael Hanselmann | raise errors.CommandError(msg) |
89 | a8083063 | Iustin Pop | |
90 | a8083063 | Iustin Pop | return res |
91 | a8083063 | Iustin Pop | |
92 | a8083063 | Iustin Pop | |
93 | fc428e32 | Michael Hanselmann | def LockFile(fd): |
94 | fc428e32 | Michael Hanselmann | """Locks a file using POSIX locks. |
95 | fc428e32 | Michael Hanselmann | |
96 | fc428e32 | Michael Hanselmann | """ |
97 | fc428e32 | Michael Hanselmann | try: |
98 | fc428e32 | Michael Hanselmann | fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB) |
99 | fc428e32 | Michael Hanselmann | except IOError, err: |
100 | fc428e32 | Michael Hanselmann | if err.errno == errno.EAGAIN: |
101 | 7bca53e4 | Michael Hanselmann | raise errors.LockError("File already locked") |
102 | fc428e32 | Michael Hanselmann | raise |
103 | fc428e32 | Michael Hanselmann | |
104 | fc428e32 | Michael Hanselmann | |
105 | 5a3103e9 | Michael Hanselmann | class WatcherState(object): |
106 | a8083063 | Iustin Pop | """Interface to a state file recording restart attempts. |
107 | a8083063 | Iustin Pop | |
108 | a8083063 | Iustin Pop | """ |
109 | a8083063 | Iustin Pop | def __init__(self): |
110 | 5a3103e9 | Michael Hanselmann | """Open, lock, read and parse the file. |
111 | 5a3103e9 | Michael Hanselmann | |
112 | 7bca53e4 | Michael Hanselmann | Raises exception on lock contention. |
113 | 5a3103e9 | Michael Hanselmann | |
114 | 5a3103e9 | Michael Hanselmann | """ |
115 | a8083063 | Iustin Pop | # The two-step dance below is necessary to allow both opening existing |
116 | a8083063 | Iustin Pop | # file read/write and creating if not existing. Vanilla open will truncate |
117 | a8083063 | Iustin Pop | # an existing file -or- allow creating if not existing. |
118 | fc428e32 | Michael Hanselmann | fd = os.open(constants.WATCHER_STATEFILE, os.O_RDWR | os.O_CREAT) |
119 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
120 | a8083063 | Iustin Pop | |
121 | fc428e32 | Michael Hanselmann | LockFile(self.statefile.fileno()) |
122 | a8083063 | Iustin Pop | |
123 | 5a3103e9 | Michael Hanselmann | try: |
124 | b76f660d | Michael Hanselmann | self._data = serializer.Load(self.statefile.read()) |
125 | 5a3103e9 | Michael Hanselmann | except Exception, msg: |
126 | 5a3103e9 | Michael Hanselmann | # Ignore errors while loading the file and treat it as empty |
127 | b76f660d | Michael Hanselmann | self._data = {} |
128 | 438b45d4 | Michael Hanselmann | logging.warning(("Empty or invalid state file. Using defaults." |
129 | 438b45d4 | Michael Hanselmann | " Error message: %s"), msg) |
130 | 5a3103e9 | Michael Hanselmann | |
131 | b76f660d | Michael Hanselmann | if "instance" not in self._data: |
132 | b76f660d | Michael Hanselmann | self._data["instance"] = {} |
133 | b76f660d | Michael Hanselmann | if "node" not in self._data: |
134 | b76f660d | Michael Hanselmann | self._data["node"] = {} |
135 | 5a3103e9 | Michael Hanselmann | |
136 | 2fb96d39 | Michael Hanselmann | self._orig_data = self._data.copy() |
137 | 2fb96d39 | Michael Hanselmann | |
138 | fc428e32 | Michael Hanselmann | def Save(self): |
139 | fc428e32 | Michael Hanselmann | """Save state to file, then unlock and close it. |
140 | 5a3103e9 | Michael Hanselmann | |
141 | 5a3103e9 | Michael Hanselmann | """ |
142 | fc428e32 | Michael Hanselmann | assert self.statefile |
143 | fc428e32 | Michael Hanselmann | |
144 | 2fb96d39 | Michael Hanselmann | if self._orig_data == self._data: |
145 | 2fb96d39 | Michael Hanselmann | logging.debug("Data didn't change, just touching status file") |
146 | 2fb96d39 | Michael Hanselmann | os.utime(constants.WATCHER_STATEFILE, None) |
147 | 2fb96d39 | Michael Hanselmann | return |
148 | 2fb96d39 | Michael Hanselmann | |
149 | fc428e32 | Michael Hanselmann | # We need to make sure the file is locked before renaming it, otherwise |
150 | fc428e32 | Michael Hanselmann | # starting ganeti-watcher again at the same time will create a conflict. |
151 | fc428e32 | Michael Hanselmann | fd = utils.WriteFile(constants.WATCHER_STATEFILE, |
152 | b76f660d | Michael Hanselmann | data=serializer.Dump(self._data), |
153 | fc428e32 | Michael Hanselmann | prewrite=LockFile, close=False) |
154 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
155 | 5a3103e9 | Michael Hanselmann | |
156 | fc428e32 | Michael Hanselmann | def Close(self): |
157 | 5a3103e9 | Michael Hanselmann | """Unlock configuration file and close it. |
158 | 5a3103e9 | Michael Hanselmann | |
159 | 5a3103e9 | Michael Hanselmann | """ |
160 | 5a3103e9 | Michael Hanselmann | assert self.statefile |
161 | 5a3103e9 | Michael Hanselmann | |
162 | fc428e32 | Michael Hanselmann | # Files are automatically unlocked when closing them |
163 | 5a3103e9 | Michael Hanselmann | self.statefile.close() |
164 | 5a3103e9 | Michael Hanselmann | self.statefile = None |
165 | 5a3103e9 | Michael Hanselmann | |
166 | 5a3103e9 | Michael Hanselmann | def GetNodeBootID(self, name): |
167 | 5a3103e9 | Michael Hanselmann | """Returns the last boot ID of a node or None. |
168 | a8083063 | Iustin Pop | |
169 | 5a3103e9 | Michael Hanselmann | """ |
170 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
171 | 5a3103e9 | Michael Hanselmann | |
172 | 7b195d9b | Michael Hanselmann | if name in ndata and KEY_BOOT_ID in ndata[name]: |
173 | 7b195d9b | Michael Hanselmann | return ndata[name][KEY_BOOT_ID] |
174 | 5a3103e9 | Michael Hanselmann | return None |
175 | 5a3103e9 | Michael Hanselmann | |
176 | 5a3103e9 | Michael Hanselmann | def SetNodeBootID(self, name, bootid): |
177 | 5a3103e9 | Michael Hanselmann | """Sets the boot ID of a node. |
178 | 5a3103e9 | Michael Hanselmann | |
179 | 5a3103e9 | Michael Hanselmann | """ |
180 | 5a3103e9 | Michael Hanselmann | assert bootid |
181 | a8083063 | Iustin Pop | |
182 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
183 | a8083063 | Iustin Pop | |
184 | 5a3103e9 | Michael Hanselmann | if name not in ndata: |
185 | 5a3103e9 | Michael Hanselmann | ndata[name] = {} |
186 | 5a3103e9 | Michael Hanselmann | |
187 | 7b195d9b | Michael Hanselmann | ndata[name][KEY_BOOT_ID] = bootid |
188 | 5a3103e9 | Michael Hanselmann | |
189 | 5a3103e9 | Michael Hanselmann | def NumberOfRestartAttempts(self, instance): |
190 | a8083063 | Iustin Pop | """Returns number of previous restart attempts. |
191 | a8083063 | Iustin Pop | |
192 | a8083063 | Iustin Pop | Args: |
193 | a8083063 | Iustin Pop | instance - the instance to look up. |
194 | 38242904 | Iustin Pop | |
195 | a8083063 | Iustin Pop | """ |
196 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
197 | a8083063 | Iustin Pop | |
198 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
199 | 7b195d9b | Michael Hanselmann | return idata[instance.name][KEY_RESTART_COUNT] |
200 | a8083063 | Iustin Pop | |
201 | a8083063 | Iustin Pop | return 0 |
202 | a8083063 | Iustin Pop | |
203 | 5a3103e9 | Michael Hanselmann | def RecordRestartAttempt(self, instance): |
204 | a8083063 | Iustin Pop | """Record a restart attempt. |
205 | a8083063 | Iustin Pop | |
206 | a8083063 | Iustin Pop | Args: |
207 | a8083063 | Iustin Pop | instance - the instance being restarted |
208 | 38242904 | Iustin Pop | |
209 | a8083063 | Iustin Pop | """ |
210 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
211 | a8083063 | Iustin Pop | |
212 | 5a3103e9 | Michael Hanselmann | if instance.name not in idata: |
213 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] = {} |
214 | 5a3103e9 | Michael Hanselmann | else: |
215 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] |
216 | a8083063 | Iustin Pop | |
217 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_WHEN] = time.time() |
218 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
219 | a8083063 | Iustin Pop | |
220 | 5a3103e9 | Michael Hanselmann | def RemoveInstance(self, instance): |
221 | 38242904 | Iustin Pop | """Update state to reflect that a machine is running, i.e. remove record. |
222 | a8083063 | Iustin Pop | |
223 | a8083063 | Iustin Pop | Args: |
224 | a8083063 | Iustin Pop | instance - the instance to remove from books |
225 | a8083063 | Iustin Pop | |
226 | 38242904 | Iustin Pop | This method removes the record for a named instance. |
227 | 38242904 | Iustin Pop | |
228 | a8083063 | Iustin Pop | """ |
229 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
230 | a8083063 | Iustin Pop | |
231 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
232 | 5a3103e9 | Michael Hanselmann | del idata[instance.name] |
233 | a8083063 | Iustin Pop | |
234 | a8083063 | Iustin Pop | |
235 | a8083063 | Iustin Pop | class Instance(object): |
236 | a8083063 | Iustin Pop | """Abstraction for a Virtual Machine instance. |
237 | a8083063 | Iustin Pop | |
238 | a8083063 | Iustin Pop | Methods: |
239 | a8083063 | Iustin Pop | Restart(): issue a command to restart the represented machine. |
240 | 098c0958 | Michael Hanselmann | |
241 | a8083063 | Iustin Pop | """ |
242 | 5a3103e9 | Michael Hanselmann | def __init__(self, name, state, autostart): |
243 | a8083063 | Iustin Pop | self.name = name |
244 | a8083063 | Iustin Pop | self.state = state |
245 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart |
246 | a8083063 | Iustin Pop | |
247 | a8083063 | Iustin Pop | def Restart(self): |
248 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance. |
249 | 3ecf6786 | Iustin Pop | |
250 | 3ecf6786 | Iustin Pop | """ |
251 | a8083063 | Iustin Pop | DoCmd(['gnt-instance', 'startup', '--lock-retries=15', self.name]) |
252 | a8083063 | Iustin Pop | |
253 | 5a3103e9 | Michael Hanselmann | def ActivateDisks(self): |
254 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance. |
255 | 5a3103e9 | Michael Hanselmann | |
256 | 5a3103e9 | Michael Hanselmann | """ |
257 | 5a3103e9 | Michael Hanselmann | DoCmd(['gnt-instance', 'activate-disks', '--lock-retries=15', self.name]) |
258 | 5a3103e9 | Michael Hanselmann | |
259 | a8083063 | Iustin Pop | |
260 | 5a3103e9 | Michael Hanselmann | def _RunListCmd(cmd): |
261 | 5a3103e9 | Michael Hanselmann | """Runs a command and parses its output into lists. |
262 | 38242904 | Iustin Pop | |
263 | a8083063 | Iustin Pop | """ |
264 | 5a3103e9 | Michael Hanselmann | for line in DoCmd(cmd).stdout.splitlines(): |
265 | 5a3103e9 | Michael Hanselmann | yield line.split(':') |
266 | a8083063 | Iustin Pop | |
267 | a8083063 | Iustin Pop | |
268 | 5a3103e9 | Michael Hanselmann | def GetInstanceList(with_secondaries=None): |
269 | 5a3103e9 | Michael Hanselmann | """Get a list of instances on this cluster. |
270 | 5a3103e9 | Michael Hanselmann | |
271 | 5a3103e9 | Michael Hanselmann | """ |
272 | 5a3103e9 | Michael Hanselmann | cmd = ['gnt-instance', 'list', '--lock-retries=15', '--no-headers', |
273 | 5a3103e9 | Michael Hanselmann | '--separator=:'] |
274 | 5a3103e9 | Michael Hanselmann | |
275 | 5a3103e9 | Michael Hanselmann | fields = 'name,oper_state,admin_state' |
276 | a8083063 | Iustin Pop | |
277 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
278 | 5a3103e9 | Michael Hanselmann | fields += ',snodes' |
279 | a8083063 | Iustin Pop | |
280 | 5a3103e9 | Michael Hanselmann | cmd.append('-o') |
281 | 5a3103e9 | Michael Hanselmann | cmd.append(fields) |
282 | 5a3103e9 | Michael Hanselmann | |
283 | 5a3103e9 | Michael Hanselmann | instances = [] |
284 | 5a3103e9 | Michael Hanselmann | for fields in _RunListCmd(cmd): |
285 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
286 | 5a3103e9 | Michael Hanselmann | (name, status, autostart, snodes) = fields |
287 | 5a3103e9 | Michael Hanselmann | |
288 | 5a3103e9 | Michael Hanselmann | if snodes == "-": |
289 | a8083063 | Iustin Pop | continue |
290 | 5a3103e9 | Michael Hanselmann | |
291 | 5a3103e9 | Michael Hanselmann | for node in with_secondaries: |
292 | 5a3103e9 | Michael Hanselmann | if node in snodes.split(','): |
293 | 5a3103e9 | Michael Hanselmann | break |
294 | 5a3103e9 | Michael Hanselmann | else: |
295 | a8083063 | Iustin Pop | continue |
296 | a8083063 | Iustin Pop | |
297 | 5a3103e9 | Michael Hanselmann | else: |
298 | 5a3103e9 | Michael Hanselmann | (name, status, autostart) = fields |
299 | 5a3103e9 | Michael Hanselmann | |
300 | 5a3103e9 | Michael Hanselmann | instances.append(Instance(name, status, autostart != "no")) |
301 | a8083063 | Iustin Pop | |
302 | 5a3103e9 | Michael Hanselmann | return instances |
303 | 5a3103e9 | Michael Hanselmann | |
304 | 5a3103e9 | Michael Hanselmann | |
305 | 5a3103e9 | Michael Hanselmann | def GetNodeBootIDs(): |
306 | 5a3103e9 | Michael Hanselmann | """Get a dict mapping nodes to boot IDs. |
307 | 5a3103e9 | Michael Hanselmann | |
308 | 5a3103e9 | Michael Hanselmann | """ |
309 | 5a3103e9 | Michael Hanselmann | cmd = ['gnt-node', 'list', '--lock-retries=15', '--no-headers', |
310 | 5a3103e9 | Michael Hanselmann | '--separator=:', '-o', 'name,bootid'] |
311 | 5a3103e9 | Michael Hanselmann | |
312 | 5a3103e9 | Michael Hanselmann | ids = {} |
313 | 5a3103e9 | Michael Hanselmann | for fields in _RunListCmd(cmd): |
314 | 5a3103e9 | Michael Hanselmann | (name, bootid) = fields |
315 | 5a3103e9 | Michael Hanselmann | ids[name] = bootid |
316 | 5a3103e9 | Michael Hanselmann | |
317 | 5a3103e9 | Michael Hanselmann | return ids |
318 | a8083063 | Iustin Pop | |
319 | a8083063 | Iustin Pop | |
320 | 5a3103e9 | Michael Hanselmann | class Watcher(object): |
321 | a8083063 | Iustin Pop | """Encapsulate the logic for restarting erronously halted virtual machines. |
322 | a8083063 | Iustin Pop | |
323 | a8083063 | Iustin Pop | The calling program should periodically instantiate me and call Run(). |
324 | a8083063 | Iustin Pop | This will traverse the list of instances, and make up to MAXTRIES attempts |
325 | a8083063 | Iustin Pop | to restart machines that are down. |
326 | 38242904 | Iustin Pop | |
327 | a8083063 | Iustin Pop | """ |
328 | a8083063 | Iustin Pop | def __init__(self): |
329 | 38242904 | Iustin Pop | sstore = ssconf.SimpleStore() |
330 | 38242904 | Iustin Pop | master = sstore.GetMasterNode() |
331 | 89e1fc26 | Iustin Pop | if master != utils.HostInfo().name: |
332 | 3ecf6786 | Iustin Pop | raise NotMasterError("This is not the master node") |
333 | 5a3103e9 | Michael Hanselmann | self.instances = GetInstanceList() |
334 | 5a3103e9 | Michael Hanselmann | self.bootids = GetNodeBootIDs() |
335 | eee1fa2d | Iustin Pop | self.started_instances = set() |
336 | a8083063 | Iustin Pop | |
337 | a8083063 | Iustin Pop | def Run(self): |
338 | 5a3103e9 | Michael Hanselmann | notepad = WatcherState() |
339 | 78f3bd30 | Michael Hanselmann | try: |
340 | 78f3bd30 | Michael Hanselmann | self.CheckInstances(notepad) |
341 | 78f3bd30 | Michael Hanselmann | self.CheckDisks(notepad) |
342 | 78f3bd30 | Michael Hanselmann | self.VerifyDisks() |
343 | 78f3bd30 | Michael Hanselmann | finally: |
344 | 78f3bd30 | Michael Hanselmann | notepad.Save() |
345 | 5a3103e9 | Michael Hanselmann | |
346 | 5a3103e9 | Michael Hanselmann | def CheckDisks(self, notepad): |
347 | 5a3103e9 | Michael Hanselmann | """Check all nodes for restarted ones. |
348 | 38242904 | Iustin Pop | |
349 | a8083063 | Iustin Pop | """ |
350 | 5a3103e9 | Michael Hanselmann | check_nodes = [] |
351 | 5a3103e9 | Michael Hanselmann | for name, id in self.bootids.iteritems(): |
352 | 5a3103e9 | Michael Hanselmann | old = notepad.GetNodeBootID(name) |
353 | 5a3103e9 | Michael Hanselmann | if old != id: |
354 | 5a3103e9 | Michael Hanselmann | # Node's boot ID has changed, proably through a reboot. |
355 | 5a3103e9 | Michael Hanselmann | check_nodes.append(name) |
356 | 5a3103e9 | Michael Hanselmann | |
357 | 5a3103e9 | Michael Hanselmann | if check_nodes: |
358 | 5a3103e9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a |
359 | 5a3103e9 | Michael Hanselmann | # secondary node. |
360 | 5a3103e9 | Michael Hanselmann | for instance in GetInstanceList(with_secondaries=check_nodes): |
361 | 0c0f834d | Iustin Pop | if not instance.autostart: |
362 | 438b45d4 | Michael Hanselmann | logging.info(("Skipping disk activation for non-autostart" |
363 | 438b45d4 | Michael Hanselmann | " instance %s"), instance.name) |
364 | 0c0f834d | Iustin Pop | continue |
365 | eee1fa2d | Iustin Pop | if instance.name in self.started_instances: |
366 | eee1fa2d | Iustin Pop | # we already tried to start the instance, which should have |
367 | eee1fa2d | Iustin Pop | # activated its drives (if they can be at all) |
368 | eee1fa2d | Iustin Pop | continue |
369 | 5a3103e9 | Michael Hanselmann | try: |
370 | 438b45d4 | Michael Hanselmann | logging.info("Activating disks for instance %s", instance.name) |
371 | 5a3103e9 | Michael Hanselmann | instance.ActivateDisks() |
372 | 438b45d4 | Michael Hanselmann | except Error, err: |
373 | 438b45d4 | Michael Hanselmann | logging.error(str(err), exc_info=True) |
374 | 5a3103e9 | Michael Hanselmann | |
375 | 5a3103e9 | Michael Hanselmann | # Keep changed boot IDs |
376 | 5a3103e9 | Michael Hanselmann | for name in check_nodes: |
377 | 5a3103e9 | Michael Hanselmann | notepad.SetNodeBootID(name, self.bootids[name]) |
378 | a8083063 | Iustin Pop | |
379 | 5a3103e9 | Michael Hanselmann | def CheckInstances(self, notepad): |
380 | 5a3103e9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones. |
381 | 5a3103e9 | Michael Hanselmann | |
382 | 5a3103e9 | Michael Hanselmann | """ |
383 | a8083063 | Iustin Pop | for instance in self.instances: |
384 | 5a3103e9 | Michael Hanselmann | # Don't care about manually stopped instances |
385 | 5a3103e9 | Michael Hanselmann | if not instance.autostart: |
386 | 5a3103e9 | Michael Hanselmann | continue |
387 | 5a3103e9 | Michael Hanselmann | |
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 | 438b45d4 | Michael Hanselmann | except Error, err: |
407 | 438b45d4 | Michael Hanselmann | logging.error(str(err), exc_info=True) |
408 | a8083063 | Iustin Pop | |
409 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
410 | a8083063 | Iustin Pop | elif instance.state in HELPLESS_STATES: |
411 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
412 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
413 | a8083063 | Iustin Pop | else: |
414 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
415 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
416 | 438b45d4 | Michael Hanselmann | logging.info("Restart of %s succeeded", instance.name) |
417 | a8083063 | Iustin Pop | |
418 | d2f311db | Iustin Pop | def VerifyDisks(self): |
419 | d2f311db | Iustin Pop | """Run gnt-cluster verify-disks. |
420 | d2f311db | Iustin Pop | |
421 | d2f311db | Iustin Pop | """ |
422 | d2f311db | Iustin Pop | result = DoCmd(['gnt-cluster', 'verify-disks', '--lock-retries=15']) |
423 | d2f311db | Iustin Pop | if result.output: |
424 | 438b45d4 | Michael Hanselmann | logging.info(result.output) |
425 | a8083063 | Iustin Pop | |
426 | a8083063 | Iustin Pop | |
427 | a8083063 | Iustin Pop | def ParseOptions(): |
428 | a8083063 | Iustin Pop | """Parse the command line options. |
429 | a8083063 | Iustin Pop | |
430 | a8083063 | Iustin Pop | Returns: |
431 | a8083063 | Iustin Pop | (options, args) as from OptionParser.parse_args() |
432 | a8083063 | Iustin Pop | |
433 | a8083063 | Iustin Pop | """ |
434 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher", |
435 | a8083063 | Iustin Pop | usage="%prog [-d]", |
436 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" % |
437 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
438 | a8083063 | Iustin Pop | |
439 | a8083063 | Iustin Pop | parser.add_option("-d", "--debug", dest="debug", |
440 | 438b45d4 | Michael Hanselmann | help="Write all messages to stderr", |
441 | a8083063 | Iustin Pop | default=False, action="store_true") |
442 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
443 | a8083063 | Iustin Pop | return options, args |
444 | a8083063 | Iustin Pop | |
445 | a8083063 | Iustin Pop | |
446 | 438b45d4 | Michael Hanselmann | def SetupLogging(debug): |
447 | 438b45d4 | Michael Hanselmann | """Configures the logging module. |
448 | 438b45d4 | Michael Hanselmann | |
449 | 438b45d4 | Michael Hanselmann | """ |
450 | 438b45d4 | Michael Hanselmann | formatter = logging.Formatter("%(asctime)s: %(message)s") |
451 | 438b45d4 | Michael Hanselmann | |
452 | 438b45d4 | Michael Hanselmann | logfile_handler = logging.FileHandler(constants.LOG_WATCHER) |
453 | 438b45d4 | Michael Hanselmann | logfile_handler.setFormatter(formatter) |
454 | 438b45d4 | Michael Hanselmann | logfile_handler.setLevel(logging.INFO) |
455 | 438b45d4 | Michael Hanselmann | |
456 | 438b45d4 | Michael Hanselmann | stderr_handler = logging.StreamHandler() |
457 | 438b45d4 | Michael Hanselmann | stderr_handler.setFormatter(formatter) |
458 | 438b45d4 | Michael Hanselmann | if debug: |
459 | 438b45d4 | Michael Hanselmann | stderr_handler.setLevel(logging.NOTSET) |
460 | 438b45d4 | Michael Hanselmann | else: |
461 | 438b45d4 | Michael Hanselmann | stderr_handler.setLevel(logging.CRITICAL) |
462 | 438b45d4 | Michael Hanselmann | |
463 | 438b45d4 | Michael Hanselmann | root_logger = logging.getLogger("") |
464 | 438b45d4 | Michael Hanselmann | root_logger.setLevel(logging.NOTSET) |
465 | 438b45d4 | Michael Hanselmann | root_logger.addHandler(logfile_handler) |
466 | 438b45d4 | Michael Hanselmann | root_logger.addHandler(stderr_handler) |
467 | 438b45d4 | Michael Hanselmann | |
468 | 438b45d4 | Michael Hanselmann | |
469 | a8083063 | Iustin Pop | def main(): |
470 | a8083063 | Iustin Pop | """Main function. |
471 | a8083063 | Iustin Pop | |
472 | a8083063 | Iustin Pop | """ |
473 | a8083063 | Iustin Pop | options, args = ParseOptions() |
474 | a8083063 | Iustin Pop | |
475 | 438b45d4 | Michael Hanselmann | SetupLogging(options.debug) |
476 | a8083063 | Iustin Pop | |
477 | a8083063 | Iustin Pop | try: |
478 | 781b2b2b | Michael Hanselmann | try: |
479 | 781b2b2b | Michael Hanselmann | watcher = Watcher() |
480 | 781b2b2b | Michael Hanselmann | except errors.ConfigurationError: |
481 | 781b2b2b | Michael Hanselmann | # Just exit if there's no configuration |
482 | 781b2b2b | Michael Hanselmann | sys.exit(constants.EXIT_SUCCESS) |
483 | 5a3103e9 | Michael Hanselmann | watcher.Run() |
484 | 1b052f42 | Michael Hanselmann | except SystemExit: |
485 | 1b052f42 | Michael Hanselmann | raise |
486 | 38242904 | Iustin Pop | except NotMasterError: |
487 | 438b45d4 | Michael Hanselmann | logging.debug("Not master, exiting") |
488 | 38242904 | Iustin Pop | sys.exit(constants.EXIT_NOTMASTER) |
489 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err: |
490 | 438b45d4 | Michael Hanselmann | logging.error("Cannot resolve hostname '%s', exiting.", err.args[0]) |
491 | 89e1fc26 | Iustin Pop | sys.exit(constants.EXIT_NODESETUP_ERROR) |
492 | 438b45d4 | Michael Hanselmann | except Exception, err: |
493 | 438b45d4 | Michael Hanselmann | logging.error(str(err), exc_info=True) |
494 | 438b45d4 | Michael Hanselmann | sys.exit(constants.EXIT_FAILURE) |
495 | a8083063 | Iustin Pop | |
496 | 5a3103e9 | Michael Hanselmann | |
497 | a8083063 | Iustin Pop | if __name__ == '__main__': |
498 | a8083063 | Iustin Pop | main() |