root / daemons / ganeti-watcher @ 26d3fd2f
History | View | Annotate | Download (23.8 kB)
1 | a8083063 | Iustin Pop | #!/usr/bin/python |
---|---|---|---|
2 | a8083063 | Iustin Pop | # |
3 | a8083063 | Iustin Pop | |
4 | 34f06005 | Iustin Pop | # Copyright (C) 2006, 2007, 2008, 2009, 2010 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 | 55c85950 | Iustin Pop | """Tool to restart erroneously 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 | 7260cfbe | Iustin Pop | # pylint: disable-msg=C0103,W0142 |
31 | 7260cfbe | Iustin Pop | |
32 | 7260cfbe | Iustin Pop | # C0103: Invalid name ganeti-watcher |
33 | 7260cfbe | Iustin Pop | |
34 | a8083063 | Iustin Pop | import os |
35 | a8083063 | Iustin Pop | import sys |
36 | a8083063 | Iustin Pop | import time |
37 | 438b45d4 | Michael Hanselmann | import logging |
38 | a8083063 | Iustin Pop | from optparse import OptionParser |
39 | a8083063 | Iustin Pop | |
40 | a8083063 | Iustin Pop | from ganeti import utils |
41 | a8083063 | Iustin Pop | from ganeti import constants |
42 | 67fe61c4 | Michael Hanselmann | from ganeti import serializer |
43 | 89e1fc26 | Iustin Pop | from ganeti import errors |
44 | e125c67c | Michael Hanselmann | from ganeti import opcodes |
45 | e125c67c | Michael Hanselmann | from ganeti import cli |
46 | 7dfb83c2 | Iustin Pop | from ganeti import luxi |
47 | 50273051 | Iustin Pop | from ganeti import ssconf |
48 | 50273051 | Iustin Pop | from ganeti import bdev |
49 | 50273051 | Iustin Pop | from ganeti import hypervisor |
50 | db147305 | Tom Limoncelli | from ganeti import rapi |
51 | 50273051 | Iustin Pop | from ganeti.confd import client as confd_client |
52 | a744b676 | Manuel Franceschini | from ganeti import netutils |
53 | a8083063 | Iustin Pop | |
54 | db147305 | Tom Limoncelli | import ganeti.rapi.client # pylint: disable-msg=W0611 |
55 | db147305 | Tom Limoncelli | |
56 | a8083063 | Iustin Pop | |
57 | 5a3103e9 | Michael Hanselmann | MAXTRIES = 5 |
58 | f5116c87 | Iustin Pop | # Delete any record that is older than 8 hours; this value is based on |
59 | f5116c87 | Iustin Pop | # the fact that the current retry counter is 5, and watcher runs every |
60 | f5116c87 | Iustin Pop | # 5 minutes, so it takes around half an hour to exceed the retry |
61 | f5116c87 | Iustin Pop | # counter, so 8 hours (16*1/2h) seems like a reasonable reset time |
62 | f5116c87 | Iustin Pop | RETRY_EXPIRATION = 8 * 3600 |
63 | b7309a0d | Iustin Pop | BAD_STATES = ['ERROR_down'] |
64 | cbfc4681 | Iustin Pop | HELPLESS_STATES = ['ERROR_nodedown', 'ERROR_nodeoffline'] |
65 | 5a3103e9 | Michael Hanselmann | NOTICE = 'NOTICE' |
66 | 5a3103e9 | Michael Hanselmann | ERROR = 'ERROR' |
67 | 7b195d9b | Michael Hanselmann | KEY_RESTART_COUNT = "restart_count" |
68 | 7b195d9b | Michael Hanselmann | KEY_RESTART_WHEN = "restart_when" |
69 | 7b195d9b | Michael Hanselmann | KEY_BOOT_ID = "bootid" |
70 | 5a3103e9 | Michael Hanselmann | |
71 | 5a3103e9 | Michael Hanselmann | |
72 | e125c67c | Michael Hanselmann | # Global client object |
73 | e125c67c | Michael Hanselmann | client = None |
74 | e125c67c | Michael Hanselmann | |
75 | e125c67c | Michael Hanselmann | |
76 | 7bca53e4 | Michael Hanselmann | class NotMasterError(errors.GenericError): |
77 | 38242904 | Iustin Pop | """Exception raised when this host is not the master.""" |
78 | a8083063 | Iustin Pop | |
79 | a8083063 | Iustin Pop | |
80 | 3753b2cb | Michael Hanselmann | def ShouldPause(): |
81 | 3753b2cb | Michael Hanselmann | """Check whether we should pause. |
82 | 3753b2cb | Michael Hanselmann | |
83 | 3753b2cb | Michael Hanselmann | """ |
84 | 3753b2cb | Michael Hanselmann | return bool(utils.ReadWatcherPauseFile(constants.WATCHER_PAUSEFILE)) |
85 | 3753b2cb | Michael Hanselmann | |
86 | 3753b2cb | Michael Hanselmann | |
87 | f1115454 | Guido Trotter | def StartNodeDaemons(): |
88 | f1115454 | Guido Trotter | """Start all the daemons that should be running on all nodes. |
89 | f1115454 | Guido Trotter | |
90 | f1115454 | Guido Trotter | """ |
91 | 55c85950 | Iustin Pop | # on master or not, try to start the node daemon |
92 | 2826b361 | Guido Trotter | utils.EnsureDaemon(constants.NODED) |
93 | f1115454 | Guido Trotter | # start confd as well. On non candidates it will be in disabled mode. |
94 | 2826b361 | Guido Trotter | utils.EnsureDaemon(constants.CONFD) |
95 | f1115454 | Guido Trotter | |
96 | f1115454 | Guido Trotter | |
97 | 9e289e36 | Guido Trotter | def RunWatcherHooks(): |
98 | 9e289e36 | Guido Trotter | """Run the watcher hooks. |
99 | 9e289e36 | Guido Trotter | |
100 | 9e289e36 | Guido Trotter | """ |
101 | c4feafe8 | Iustin Pop | hooks_dir = utils.PathJoin(constants.HOOKS_BASE_DIR, |
102 | c4feafe8 | Iustin Pop | constants.HOOKS_NAME_WATCHER) |
103 | 10e689d4 | Iustin Pop | if not os.path.isdir(hooks_dir): |
104 | 10e689d4 | Iustin Pop | return |
105 | 9e289e36 | Guido Trotter | |
106 | 9e289e36 | Guido Trotter | try: |
107 | 9e289e36 | Guido Trotter | results = utils.RunParts(hooks_dir) |
108 | 9e289e36 | Guido Trotter | except Exception, msg: # pylint: disable-msg=W0703 |
109 | 9e289e36 | Guido Trotter | logging.critical("RunParts %s failed: %s", hooks_dir, msg) |
110 | 9e289e36 | Guido Trotter | |
111 | 9e289e36 | Guido Trotter | for (relname, status, runresult) in results: |
112 | 9e289e36 | Guido Trotter | if status == constants.RUNPARTS_SKIP: |
113 | 9e289e36 | Guido Trotter | logging.debug("Watcher hook %s: skipped", relname) |
114 | 9e289e36 | Guido Trotter | elif status == constants.RUNPARTS_ERR: |
115 | 9e289e36 | Guido Trotter | logging.warning("Watcher hook %s: error (%s)", relname, runresult) |
116 | 9e289e36 | Guido Trotter | elif status == constants.RUNPARTS_RUN: |
117 | 9e289e36 | Guido Trotter | if runresult.failed: |
118 | 9e289e36 | Guido Trotter | logging.warning("Watcher hook %s: failed (exit: %d) (output: %s)", |
119 | 9e289e36 | Guido Trotter | relname, runresult.exit_code, runresult.output) |
120 | 9e289e36 | Guido Trotter | else: |
121 | 9e289e36 | Guido Trotter | logging.debug("Watcher hook %s: success (output: %s)", relname, |
122 | 9e289e36 | Guido Trotter | runresult.output) |
123 | 9e289e36 | Guido Trotter | |
124 | 001b3825 | Michael Hanselmann | |
125 | 50273051 | Iustin Pop | class NodeMaintenance(object): |
126 | 50273051 | Iustin Pop | """Talks to confd daemons and possible shutdown instances/drbd devices. |
127 | 50273051 | Iustin Pop | |
128 | 50273051 | Iustin Pop | """ |
129 | 50273051 | Iustin Pop | def __init__(self): |
130 | 50273051 | Iustin Pop | self.store_cb = confd_client.StoreResultCallback() |
131 | 50273051 | Iustin Pop | self.filter_cb = confd_client.ConfdFilterCallback(self.store_cb) |
132 | 50273051 | Iustin Pop | self.confd_client = confd_client.GetConfdClient(self.filter_cb) |
133 | 50273051 | Iustin Pop | |
134 | 50273051 | Iustin Pop | @staticmethod |
135 | 50273051 | Iustin Pop | def ShouldRun(): |
136 | 50273051 | Iustin Pop | """Checks whether node maintenance should run. |
137 | 50273051 | Iustin Pop | |
138 | 50273051 | Iustin Pop | """ |
139 | 50273051 | Iustin Pop | try: |
140 | 50273051 | Iustin Pop | return ssconf.SimpleStore().GetMaintainNodeHealth() |
141 | 50273051 | Iustin Pop | except errors.ConfigurationError, err: |
142 | 50273051 | Iustin Pop | logging.error("Configuration error, not activating node maintenance: %s", |
143 | 50273051 | Iustin Pop | err) |
144 | 50273051 | Iustin Pop | return False |
145 | 50273051 | Iustin Pop | |
146 | 50273051 | Iustin Pop | @staticmethod |
147 | 50273051 | Iustin Pop | def GetRunningInstances(): |
148 | 50273051 | Iustin Pop | """Compute list of hypervisor/running instances. |
149 | 50273051 | Iustin Pop | |
150 | 50273051 | Iustin Pop | """ |
151 | 50273051 | Iustin Pop | hyp_list = ssconf.SimpleStore().GetHypervisorList() |
152 | 50273051 | Iustin Pop | results = [] |
153 | 50273051 | Iustin Pop | for hv_name in hyp_list: |
154 | 50273051 | Iustin Pop | try: |
155 | 50273051 | Iustin Pop | hv = hypervisor.GetHypervisor(hv_name) |
156 | 50273051 | Iustin Pop | ilist = hv.ListInstances() |
157 | 50273051 | Iustin Pop | results.extend([(iname, hv_name) for iname in ilist]) |
158 | 50273051 | Iustin Pop | except: # pylint: disable-msg=W0702 |
159 | 50273051 | Iustin Pop | logging.error("Error while listing instances for hypervisor %s", |
160 | 50273051 | Iustin Pop | hv_name, exc_info=True) |
161 | 50273051 | Iustin Pop | return results |
162 | 50273051 | Iustin Pop | |
163 | 50273051 | Iustin Pop | @staticmethod |
164 | 50273051 | Iustin Pop | def GetUsedDRBDs(): |
165 | 50273051 | Iustin Pop | """Get list of used DRBD minors. |
166 | 50273051 | Iustin Pop | |
167 | 50273051 | Iustin Pop | """ |
168 | 50273051 | Iustin Pop | return bdev.DRBD8.GetUsedDevs().keys() |
169 | 50273051 | Iustin Pop | |
170 | 50273051 | Iustin Pop | @classmethod |
171 | 50273051 | Iustin Pop | def DoMaintenance(cls, role): |
172 | 50273051 | Iustin Pop | """Maintain the instance list. |
173 | 50273051 | Iustin Pop | |
174 | 50273051 | Iustin Pop | """ |
175 | 50273051 | Iustin Pop | if role == constants.CONFD_NODE_ROLE_OFFLINE: |
176 | 50273051 | Iustin Pop | inst_running = cls.GetRunningInstances() |
177 | 50273051 | Iustin Pop | cls.ShutdownInstances(inst_running) |
178 | 50273051 | Iustin Pop | drbd_running = cls.GetUsedDRBDs() |
179 | 50273051 | Iustin Pop | cls.ShutdownDRBD(drbd_running) |
180 | 50273051 | Iustin Pop | else: |
181 | 50273051 | Iustin Pop | logging.debug("Not doing anything for role %s", role) |
182 | 50273051 | Iustin Pop | |
183 | 50273051 | Iustin Pop | @staticmethod |
184 | 50273051 | Iustin Pop | def ShutdownInstances(inst_running): |
185 | 50273051 | Iustin Pop | """Shutdown running instances. |
186 | 50273051 | Iustin Pop | |
187 | 50273051 | Iustin Pop | """ |
188 | 50273051 | Iustin Pop | names_running = set([i[0] for i in inst_running]) |
189 | 50273051 | Iustin Pop | if names_running: |
190 | 50273051 | Iustin Pop | logging.info("Following instances should not be running," |
191 | 50273051 | Iustin Pop | " shutting them down: %s", utils.CommaJoin(names_running)) |
192 | 50273051 | Iustin Pop | # this dictionary will collapse duplicate instance names (only |
193 | 50273051 | Iustin Pop | # xen pvm/vhm) into a single key, which is fine |
194 | 50273051 | Iustin Pop | i2h = dict(inst_running) |
195 | 50273051 | Iustin Pop | for name in names_running: |
196 | 50273051 | Iustin Pop | hv_name = i2h[name] |
197 | 50273051 | Iustin Pop | hv = hypervisor.GetHypervisor(hv_name) |
198 | 50273051 | Iustin Pop | hv.StopInstance(None, force=True, name=name) |
199 | 50273051 | Iustin Pop | |
200 | 50273051 | Iustin Pop | @staticmethod |
201 | 50273051 | Iustin Pop | def ShutdownDRBD(drbd_running): |
202 | 50273051 | Iustin Pop | """Shutdown active DRBD devices. |
203 | 50273051 | Iustin Pop | |
204 | 50273051 | Iustin Pop | """ |
205 | 50273051 | Iustin Pop | if drbd_running: |
206 | 50273051 | Iustin Pop | logging.info("Following DRBD minors should not be active," |
207 | 50273051 | Iustin Pop | " shutting them down: %s", utils.CommaJoin(drbd_running)) |
208 | 50273051 | Iustin Pop | for minor in drbd_running: |
209 | 50273051 | Iustin Pop | # pylint: disable-msg=W0212 |
210 | 50273051 | Iustin Pop | # using the private method as is, pending enhancements to the DRBD |
211 | 50273051 | Iustin Pop | # interface |
212 | 50273051 | Iustin Pop | bdev.DRBD8._ShutdownAll(minor) |
213 | 50273051 | Iustin Pop | |
214 | 50273051 | Iustin Pop | def Exec(self): |
215 | 50273051 | Iustin Pop | """Check node status versus cluster desired state. |
216 | 50273051 | Iustin Pop | |
217 | 50273051 | Iustin Pop | """ |
218 | b705c7a6 | Manuel Franceschini | my_name = netutils.Hostname.GetSysName() |
219 | 50273051 | Iustin Pop | req = confd_client.ConfdClientRequest(type= |
220 | 50273051 | Iustin Pop | constants.CONFD_REQ_NODE_ROLE_BYNAME, |
221 | 50273051 | Iustin Pop | query=my_name) |
222 | ebacb943 | Iustin Pop | self.confd_client.SendRequest(req, async=False, coverage=-1) |
223 | 50273051 | Iustin Pop | timed_out, _, _ = self.confd_client.WaitForReply(req.rsalt) |
224 | 50273051 | Iustin Pop | if not timed_out: |
225 | 50273051 | Iustin Pop | # should have a valid response |
226 | 50273051 | Iustin Pop | status, result = self.store_cb.GetResponse(req.rsalt) |
227 | 50273051 | Iustin Pop | assert status, "Missing result but received replies" |
228 | 50273051 | Iustin Pop | if not self.filter_cb.consistent[req.rsalt]: |
229 | 50273051 | Iustin Pop | logging.warning("Inconsistent replies, not doing anything") |
230 | 50273051 | Iustin Pop | return |
231 | 50273051 | Iustin Pop | self.DoMaintenance(result.server_reply.answer) |
232 | 50273051 | Iustin Pop | else: |
233 | 50273051 | Iustin Pop | logging.warning("Confd query timed out, cannot do maintenance actions") |
234 | 50273051 | Iustin Pop | |
235 | 50273051 | Iustin Pop | |
236 | 5a3103e9 | Michael Hanselmann | class WatcherState(object): |
237 | a8083063 | Iustin Pop | """Interface to a state file recording restart attempts. |
238 | a8083063 | Iustin Pop | |
239 | a8083063 | Iustin Pop | """ |
240 | 001b3825 | Michael Hanselmann | def __init__(self, statefile): |
241 | 5a3103e9 | Michael Hanselmann | """Open, lock, read and parse the file. |
242 | 5a3103e9 | Michael Hanselmann | |
243 | 001b3825 | Michael Hanselmann | @type statefile: file |
244 | 001b3825 | Michael Hanselmann | @param statefile: State file object |
245 | 5a3103e9 | Michael Hanselmann | |
246 | 5a3103e9 | Michael Hanselmann | """ |
247 | 001b3825 | Michael Hanselmann | self.statefile = statefile |
248 | a8083063 | Iustin Pop | |
249 | 5a3103e9 | Michael Hanselmann | try: |
250 | 2c404217 | Iustin Pop | state_data = self.statefile.read() |
251 | 2c404217 | Iustin Pop | if not state_data: |
252 | 2c404217 | Iustin Pop | self._data = {} |
253 | 2c404217 | Iustin Pop | else: |
254 | 2c404217 | Iustin Pop | self._data = serializer.Load(state_data) |
255 | 7260cfbe | Iustin Pop | except Exception, msg: # pylint: disable-msg=W0703 |
256 | 5a3103e9 | Michael Hanselmann | # Ignore errors while loading the file and treat it as empty |
257 | b76f660d | Michael Hanselmann | self._data = {} |
258 | 2c404217 | Iustin Pop | logging.warning(("Invalid state file. Using defaults." |
259 | 438b45d4 | Michael Hanselmann | " Error message: %s"), msg) |
260 | 5a3103e9 | Michael Hanselmann | |
261 | b76f660d | Michael Hanselmann | if "instance" not in self._data: |
262 | b76f660d | Michael Hanselmann | self._data["instance"] = {} |
263 | b76f660d | Michael Hanselmann | if "node" not in self._data: |
264 | b76f660d | Michael Hanselmann | self._data["node"] = {} |
265 | 5a3103e9 | Michael Hanselmann | |
266 | 26517d45 | Iustin Pop | self._orig_data = serializer.Dump(self._data) |
267 | 2fb96d39 | Michael Hanselmann | |
268 | fc428e32 | Michael Hanselmann | def Save(self): |
269 | fc428e32 | Michael Hanselmann | """Save state to file, then unlock and close it. |
270 | 5a3103e9 | Michael Hanselmann | |
271 | 5a3103e9 | Michael Hanselmann | """ |
272 | fc428e32 | Michael Hanselmann | assert self.statefile |
273 | fc428e32 | Michael Hanselmann | |
274 | 26517d45 | Iustin Pop | serialized_form = serializer.Dump(self._data) |
275 | 26517d45 | Iustin Pop | if self._orig_data == serialized_form: |
276 | 2fb96d39 | Michael Hanselmann | logging.debug("Data didn't change, just touching status file") |
277 | 2fb96d39 | Michael Hanselmann | os.utime(constants.WATCHER_STATEFILE, None) |
278 | 2fb96d39 | Michael Hanselmann | return |
279 | 2fb96d39 | Michael Hanselmann | |
280 | fc428e32 | Michael Hanselmann | # We need to make sure the file is locked before renaming it, otherwise |
281 | fc428e32 | Michael Hanselmann | # starting ganeti-watcher again at the same time will create a conflict. |
282 | fc428e32 | Michael Hanselmann | fd = utils.WriteFile(constants.WATCHER_STATEFILE, |
283 | 26517d45 | Iustin Pop | data=serialized_form, |
284 | eb0f0ce0 | Michael Hanselmann | prewrite=utils.LockFile, close=False) |
285 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
286 | 5a3103e9 | Michael Hanselmann | |
287 | fc428e32 | Michael Hanselmann | def Close(self): |
288 | 5a3103e9 | Michael Hanselmann | """Unlock configuration file and close it. |
289 | 5a3103e9 | Michael Hanselmann | |
290 | 5a3103e9 | Michael Hanselmann | """ |
291 | 5a3103e9 | Michael Hanselmann | assert self.statefile |
292 | 5a3103e9 | Michael Hanselmann | |
293 | fc428e32 | Michael Hanselmann | # Files are automatically unlocked when closing them |
294 | 5a3103e9 | Michael Hanselmann | self.statefile.close() |
295 | 5a3103e9 | Michael Hanselmann | self.statefile = None |
296 | 5a3103e9 | Michael Hanselmann | |
297 | 5a3103e9 | Michael Hanselmann | def GetNodeBootID(self, name): |
298 | 5a3103e9 | Michael Hanselmann | """Returns the last boot ID of a node or None. |
299 | a8083063 | Iustin Pop | |
300 | 5a3103e9 | Michael Hanselmann | """ |
301 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
302 | 5a3103e9 | Michael Hanselmann | |
303 | 7b195d9b | Michael Hanselmann | if name in ndata and KEY_BOOT_ID in ndata[name]: |
304 | 7b195d9b | Michael Hanselmann | return ndata[name][KEY_BOOT_ID] |
305 | 5a3103e9 | Michael Hanselmann | return None |
306 | 5a3103e9 | Michael Hanselmann | |
307 | 5a3103e9 | Michael Hanselmann | def SetNodeBootID(self, name, bootid): |
308 | 5a3103e9 | Michael Hanselmann | """Sets the boot ID of a node. |
309 | 5a3103e9 | Michael Hanselmann | |
310 | 5a3103e9 | Michael Hanselmann | """ |
311 | 5a3103e9 | Michael Hanselmann | assert bootid |
312 | a8083063 | Iustin Pop | |
313 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
314 | a8083063 | Iustin Pop | |
315 | 5a3103e9 | Michael Hanselmann | if name not in ndata: |
316 | 5a3103e9 | Michael Hanselmann | ndata[name] = {} |
317 | 5a3103e9 | Michael Hanselmann | |
318 | 7b195d9b | Michael Hanselmann | ndata[name][KEY_BOOT_ID] = bootid |
319 | 5a3103e9 | Michael Hanselmann | |
320 | 5a3103e9 | Michael Hanselmann | def NumberOfRestartAttempts(self, instance): |
321 | a8083063 | Iustin Pop | """Returns number of previous restart attempts. |
322 | a8083063 | Iustin Pop | |
323 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
324 | c41eea6e | Iustin Pop | @param instance: the instance to look up |
325 | 38242904 | Iustin Pop | |
326 | a8083063 | Iustin Pop | """ |
327 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
328 | a8083063 | Iustin Pop | |
329 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
330 | 7b195d9b | Michael Hanselmann | return idata[instance.name][KEY_RESTART_COUNT] |
331 | a8083063 | Iustin Pop | |
332 | a8083063 | Iustin Pop | return 0 |
333 | a8083063 | Iustin Pop | |
334 | f5116c87 | Iustin Pop | def MaintainInstanceList(self, instances): |
335 | f5116c87 | Iustin Pop | """Perform maintenance on the recorded instances. |
336 | f5116c87 | Iustin Pop | |
337 | f5116c87 | Iustin Pop | @type instances: list of string |
338 | f5116c87 | Iustin Pop | @param instances: the list of currently existing instances |
339 | f5116c87 | Iustin Pop | |
340 | f5116c87 | Iustin Pop | """ |
341 | f5116c87 | Iustin Pop | idict = self._data["instance"] |
342 | f5116c87 | Iustin Pop | # First, delete obsolete instances |
343 | f5116c87 | Iustin Pop | obsolete_instances = set(idict).difference(instances) |
344 | f5116c87 | Iustin Pop | for inst in obsolete_instances: |
345 | f5116c87 | Iustin Pop | logging.debug("Forgetting obsolete instance %s", inst) |
346 | f5116c87 | Iustin Pop | del idict[inst] |
347 | f5116c87 | Iustin Pop | |
348 | f5116c87 | Iustin Pop | # Second, delete expired records |
349 | f5116c87 | Iustin Pop | earliest = time.time() - RETRY_EXPIRATION |
350 | f5116c87 | Iustin Pop | expired_instances = [i for i in idict |
351 | f5116c87 | Iustin Pop | if idict[i][KEY_RESTART_WHEN] < earliest] |
352 | f5116c87 | Iustin Pop | for inst in expired_instances: |
353 | f5116c87 | Iustin Pop | logging.debug("Expiring record for instance %s", inst) |
354 | f5116c87 | Iustin Pop | del idict[inst] |
355 | f5116c87 | Iustin Pop | |
356 | 5a3103e9 | Michael Hanselmann | def RecordRestartAttempt(self, instance): |
357 | a8083063 | Iustin Pop | """Record a restart attempt. |
358 | a8083063 | Iustin Pop | |
359 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
360 | c41eea6e | Iustin Pop | @param instance: the instance being restarted |
361 | 38242904 | Iustin Pop | |
362 | a8083063 | Iustin Pop | """ |
363 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
364 | a8083063 | Iustin Pop | |
365 | 5a3103e9 | Michael Hanselmann | if instance.name not in idata: |
366 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] = {} |
367 | 5a3103e9 | Michael Hanselmann | else: |
368 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] |
369 | a8083063 | Iustin Pop | |
370 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_WHEN] = time.time() |
371 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
372 | a8083063 | Iustin Pop | |
373 | 5a3103e9 | Michael Hanselmann | def RemoveInstance(self, instance): |
374 | c41eea6e | Iustin Pop | """Update state to reflect that a machine is running. |
375 | a8083063 | Iustin Pop | |
376 | c41eea6e | Iustin Pop | This method removes the record for a named instance (as we only |
377 | c41eea6e | Iustin Pop | track down instances). |
378 | a8083063 | Iustin Pop | |
379 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
380 | c41eea6e | Iustin Pop | @param instance: the instance to remove from books |
381 | 38242904 | Iustin Pop | |
382 | a8083063 | Iustin Pop | """ |
383 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
384 | a8083063 | Iustin Pop | |
385 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
386 | 5a3103e9 | Michael Hanselmann | del idata[instance.name] |
387 | a8083063 | Iustin Pop | |
388 | a8083063 | Iustin Pop | |
389 | a8083063 | Iustin Pop | class Instance(object): |
390 | a8083063 | Iustin Pop | """Abstraction for a Virtual Machine instance. |
391 | a8083063 | Iustin Pop | |
392 | a8083063 | Iustin Pop | """ |
393 | 5a3103e9 | Michael Hanselmann | def __init__(self, name, state, autostart): |
394 | a8083063 | Iustin Pop | self.name = name |
395 | a8083063 | Iustin Pop | self.state = state |
396 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart |
397 | a8083063 | Iustin Pop | |
398 | a8083063 | Iustin Pop | def Restart(self): |
399 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance. |
400 | 3ecf6786 | Iustin Pop | |
401 | 3ecf6786 | Iustin Pop | """ |
402 | 07813a9e | Iustin Pop | op = opcodes.OpStartupInstance(instance_name=self.name, force=False) |
403 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
404 | a8083063 | Iustin Pop | |
405 | 5a3103e9 | Michael Hanselmann | def ActivateDisks(self): |
406 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance. |
407 | 5a3103e9 | Michael Hanselmann | |
408 | 5a3103e9 | Michael Hanselmann | """ |
409 | e125c67c | Michael Hanselmann | op = opcodes.OpActivateInstanceDisks(instance_name=self.name) |
410 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
411 | a8083063 | Iustin Pop | |
412 | a8083063 | Iustin Pop | |
413 | 6dfcc47b | Iustin Pop | def GetClusterData(): |
414 | 5a3103e9 | Michael Hanselmann | """Get a list of instances on this cluster. |
415 | 5a3103e9 | Michael Hanselmann | |
416 | 5a3103e9 | Michael Hanselmann | """ |
417 | 6dfcc47b | Iustin Pop | op1_fields = ["name", "status", "admin_state", "snodes"] |
418 | 6dfcc47b | Iustin Pop | op1 = opcodes.OpQueryInstances(output_fields=op1_fields, names=[], |
419 | 6dfcc47b | Iustin Pop | use_locking=True) |
420 | 6dfcc47b | Iustin Pop | op2_fields = ["name", "bootid", "offline"] |
421 | 6dfcc47b | Iustin Pop | op2 = opcodes.OpQueryNodes(output_fields=op2_fields, names=[], |
422 | 6dfcc47b | Iustin Pop | use_locking=True) |
423 | a8083063 | Iustin Pop | |
424 | 6dfcc47b | Iustin Pop | job_id = client.SubmitJob([op1, op2]) |
425 | a8083063 | Iustin Pop | |
426 | 6dfcc47b | Iustin Pop | all_results = cli.PollJob(job_id, cl=client, feedback_fn=logging.debug) |
427 | 5a3103e9 | Michael Hanselmann | |
428 | 78f44650 | Iustin Pop | logging.debug("Got data from cluster, writing instance status file") |
429 | 78f44650 | Iustin Pop | |
430 | 6dfcc47b | Iustin Pop | result = all_results[0] |
431 | 6dfcc47b | Iustin Pop | smap = {} |
432 | 5a3103e9 | Michael Hanselmann | |
433 | 6dfcc47b | Iustin Pop | instances = {} |
434 | 78f44650 | Iustin Pop | |
435 | 78f44650 | Iustin Pop | # write the upfile |
436 | 78f44650 | Iustin Pop | up_data = "".join(["%s %s\n" % (fields[0], fields[1]) for fields in result]) |
437 | 78f44650 | Iustin Pop | utils.WriteFile(file_name=constants.INSTANCE_UPFILE, data=up_data) |
438 | 78f44650 | Iustin Pop | |
439 | 6dfcc47b | Iustin Pop | for fields in result: |
440 | 6dfcc47b | Iustin Pop | (name, status, autostart, snodes) = fields |
441 | 5a3103e9 | Michael Hanselmann | |
442 | 6dfcc47b | Iustin Pop | # update the secondary node map |
443 | 6dfcc47b | Iustin Pop | for node in snodes: |
444 | 6dfcc47b | Iustin Pop | if node not in smap: |
445 | 6dfcc47b | Iustin Pop | smap[node] = [] |
446 | 6dfcc47b | Iustin Pop | smap[node].append(name) |
447 | a8083063 | Iustin Pop | |
448 | 6dfcc47b | Iustin Pop | instances[name] = Instance(name, status, autostart) |
449 | 5a3103e9 | Michael Hanselmann | |
450 | 6dfcc47b | Iustin Pop | nodes = dict([(name, (bootid, offline)) |
451 | 6dfcc47b | Iustin Pop | for name, bootid, offline in all_results[1]]) |
452 | 5a3103e9 | Michael Hanselmann | |
453 | 6dfcc47b | Iustin Pop | client.ArchiveJob(job_id) |
454 | 5a3103e9 | Michael Hanselmann | |
455 | 6dfcc47b | Iustin Pop | return instances, nodes, smap |
456 | a8083063 | Iustin Pop | |
457 | a8083063 | Iustin Pop | |
458 | 5a3103e9 | Michael Hanselmann | class Watcher(object): |
459 | 55c85950 | Iustin Pop | """Encapsulate the logic for restarting erroneously halted virtual machines. |
460 | a8083063 | Iustin Pop | |
461 | a8083063 | Iustin Pop | The calling program should periodically instantiate me and call Run(). |
462 | a8083063 | Iustin Pop | This will traverse the list of instances, and make up to MAXTRIES attempts |
463 | a8083063 | Iustin Pop | to restart machines that are down. |
464 | 38242904 | Iustin Pop | |
465 | a8083063 | Iustin Pop | """ |
466 | cc962d58 | Iustin Pop | def __init__(self, opts, notepad): |
467 | cc962d58 | Iustin Pop | self.notepad = notepad |
468 | 2859b87b | Michael Hanselmann | master = client.QueryConfigValues(["master_node"])[0] |
469 | b705c7a6 | Manuel Franceschini | if master != netutils.Hostname.GetSysName(): |
470 | 3ecf6786 | Iustin Pop | raise NotMasterError("This is not the master node") |
471 | 24edc6d4 | Iustin Pop | # first archive old jobs |
472 | 24edc6d4 | Iustin Pop | self.ArchiveJobs(opts.job_age) |
473 | 24edc6d4 | Iustin Pop | # and only then submit new ones |
474 | 6dfcc47b | Iustin Pop | self.instances, self.bootids, self.smap = GetClusterData() |
475 | eee1fa2d | Iustin Pop | self.started_instances = set() |
476 | f07521e5 | Iustin Pop | self.opts = opts |
477 | a8083063 | Iustin Pop | |
478 | a8083063 | Iustin Pop | def Run(self): |
479 | cc962d58 | Iustin Pop | """Watcher run sequence. |
480 | cc962d58 | Iustin Pop | |
481 | cc962d58 | Iustin Pop | """ |
482 | cc962d58 | Iustin Pop | notepad = self.notepad |
483 | cc962d58 | Iustin Pop | self.CheckInstances(notepad) |
484 | cc962d58 | Iustin Pop | self.CheckDisks(notepad) |
485 | cc962d58 | Iustin Pop | self.VerifyDisks() |
486 | 5a3103e9 | Michael Hanselmann | |
487 | 24edc6d4 | Iustin Pop | @staticmethod |
488 | 24edc6d4 | Iustin Pop | def ArchiveJobs(age): |
489 | f07521e5 | Iustin Pop | """Archive old jobs. |
490 | f07521e5 | Iustin Pop | |
491 | f07521e5 | Iustin Pop | """ |
492 | f07521e5 | Iustin Pop | arch_count, left_count = client.AutoArchiveJobs(age) |
493 | 07b8a2b5 | Iustin Pop | logging.debug("Archived %s jobs, left %s", arch_count, left_count) |
494 | f07521e5 | Iustin Pop | |
495 | 5a3103e9 | Michael Hanselmann | def CheckDisks(self, notepad): |
496 | 5a3103e9 | Michael Hanselmann | """Check all nodes for restarted ones. |
497 | 38242904 | Iustin Pop | |
498 | a8083063 | Iustin Pop | """ |
499 | 5a3103e9 | Michael Hanselmann | check_nodes = [] |
500 | cbfc4681 | Iustin Pop | for name, (new_id, offline) in self.bootids.iteritems(): |
501 | 5a3103e9 | Michael Hanselmann | old = notepad.GetNodeBootID(name) |
502 | 37b77b18 | Iustin Pop | if new_id is None: |
503 | 37b77b18 | Iustin Pop | # Bad node, not returning a boot id |
504 | cbfc4681 | Iustin Pop | if not offline: |
505 | cbfc4681 | Iustin Pop | logging.debug("Node %s missing boot id, skipping secondary checks", |
506 | cbfc4681 | Iustin Pop | name) |
507 | 37b77b18 | Iustin Pop | continue |
508 | 26517d45 | Iustin Pop | if old != new_id: |
509 | 5a3103e9 | Michael Hanselmann | # Node's boot ID has changed, proably through a reboot. |
510 | 5a3103e9 | Michael Hanselmann | check_nodes.append(name) |
511 | 5a3103e9 | Michael Hanselmann | |
512 | 5a3103e9 | Michael Hanselmann | if check_nodes: |
513 | 5a3103e9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a |
514 | 5a3103e9 | Michael Hanselmann | # secondary node. |
515 | 6dfcc47b | Iustin Pop | for node in check_nodes: |
516 | 6dfcc47b | Iustin Pop | if node not in self.smap: |
517 | eee1fa2d | Iustin Pop | continue |
518 | 6dfcc47b | Iustin Pop | for instance_name in self.smap[node]: |
519 | 6dfcc47b | Iustin Pop | instance = self.instances[instance_name] |
520 | 6dfcc47b | Iustin Pop | if not instance.autostart: |
521 | 6dfcc47b | Iustin Pop | logging.info(("Skipping disk activation for non-autostart" |
522 | 6dfcc47b | Iustin Pop | " instance %s"), instance.name) |
523 | 6dfcc47b | Iustin Pop | continue |
524 | 6dfcc47b | Iustin Pop | if instance.name in self.started_instances: |
525 | 6dfcc47b | Iustin Pop | # we already tried to start the instance, which should have |
526 | 6dfcc47b | Iustin Pop | # activated its drives (if they can be at all) |
527 | 6dfcc47b | Iustin Pop | continue |
528 | 6dfcc47b | Iustin Pop | try: |
529 | 6dfcc47b | Iustin Pop | logging.info("Activating disks for instance %s", instance.name) |
530 | 6dfcc47b | Iustin Pop | instance.ActivateDisks() |
531 | 7260cfbe | Iustin Pop | except Exception: # pylint: disable-msg=W0703 |
532 | 6dfcc47b | Iustin Pop | logging.exception("Error while activating disks for instance %s", |
533 | 6dfcc47b | Iustin Pop | instance.name) |
534 | 5a3103e9 | Michael Hanselmann | |
535 | 5a3103e9 | Michael Hanselmann | # Keep changed boot IDs |
536 | 5a3103e9 | Michael Hanselmann | for name in check_nodes: |
537 | 3448aa22 | Iustin Pop | notepad.SetNodeBootID(name, self.bootids[name][0]) |
538 | a8083063 | Iustin Pop | |
539 | 5a3103e9 | Michael Hanselmann | def CheckInstances(self, notepad): |
540 | 5a3103e9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones. |
541 | 5a3103e9 | Michael Hanselmann | |
542 | 5a3103e9 | Michael Hanselmann | """ |
543 | f5116c87 | Iustin Pop | notepad.MaintainInstanceList(self.instances.keys()) |
544 | f5116c87 | Iustin Pop | |
545 | 6dfcc47b | Iustin Pop | for instance in self.instances.values(): |
546 | a8083063 | Iustin Pop | if instance.state in BAD_STATES: |
547 | 5a3103e9 | Michael Hanselmann | n = notepad.NumberOfRestartAttempts(instance) |
548 | a8083063 | Iustin Pop | |
549 | a8083063 | Iustin Pop | if n > MAXTRIES: |
550 | f5116c87 | Iustin Pop | logging.warning("Not restarting instance %s, retries exhausted", |
551 | f5116c87 | Iustin Pop | instance.name) |
552 | a8083063 | Iustin Pop | continue |
553 | a8083063 | Iustin Pop | elif n < MAXTRIES: |
554 | a8083063 | Iustin Pop | last = " (Attempt #%d)" % (n + 1) |
555 | a8083063 | Iustin Pop | else: |
556 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
557 | 438b45d4 | Michael Hanselmann | logging.error("Could not restart %s after %d attempts, giving up", |
558 | 438b45d4 | Michael Hanselmann | instance.name, MAXTRIES) |
559 | a8083063 | Iustin Pop | continue |
560 | a8083063 | Iustin Pop | try: |
561 | 438b45d4 | Michael Hanselmann | logging.info("Restarting %s%s", |
562 | 438b45d4 | Michael Hanselmann | instance.name, last) |
563 | a8083063 | Iustin Pop | instance.Restart() |
564 | eee1fa2d | Iustin Pop | self.started_instances.add(instance.name) |
565 | 7260cfbe | Iustin Pop | except Exception: # pylint: disable-msg=W0703 |
566 | 4bffa7f7 | Iustin Pop | logging.exception("Error while restarting instance %s", |
567 | 4bffa7f7 | Iustin Pop | instance.name) |
568 | a8083063 | Iustin Pop | |
569 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
570 | a8083063 | Iustin Pop | elif instance.state in HELPLESS_STATES: |
571 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
572 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
573 | a8083063 | Iustin Pop | else: |
574 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
575 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
576 | 438b45d4 | Michael Hanselmann | logging.info("Restart of %s succeeded", instance.name) |
577 | a8083063 | Iustin Pop | |
578 | b7309a0d | Iustin Pop | @staticmethod |
579 | b7309a0d | Iustin Pop | def VerifyDisks(): |
580 | d2f311db | Iustin Pop | """Run gnt-cluster verify-disks. |
581 | d2f311db | Iustin Pop | |
582 | d2f311db | Iustin Pop | """ |
583 | 5188ab37 | Iustin Pop | op = opcodes.OpVerifyDisks() |
584 | 6dfcc47b | Iustin Pop | job_id = client.SubmitJob([op]) |
585 | 6dfcc47b | Iustin Pop | result = cli.PollJob(job_id, cl=client, feedback_fn=logging.debug)[0] |
586 | 6dfcc47b | Iustin Pop | client.ArchiveJob(job_id) |
587 | 5188ab37 | Iustin Pop | if not isinstance(result, (tuple, list)): |
588 | 5188ab37 | Iustin Pop | logging.error("Can't get a valid result from verify-disks") |
589 | 5188ab37 | Iustin Pop | return |
590 | 5188ab37 | Iustin Pop | offline_disk_instances = result[2] |
591 | 5188ab37 | Iustin Pop | if not offline_disk_instances: |
592 | 5188ab37 | Iustin Pop | # nothing to do |
593 | 5188ab37 | Iustin Pop | return |
594 | 5188ab37 | Iustin Pop | logging.debug("Will activate disks for instances %s", |
595 | 1f864b60 | Iustin Pop | utils.CommaJoin(offline_disk_instances)) |
596 | 5188ab37 | Iustin Pop | # we submit only one job, and wait for it. not optimal, but spams |
597 | 5188ab37 | Iustin Pop | # less the job queue |
598 | 5188ab37 | Iustin Pop | job = [opcodes.OpActivateInstanceDisks(instance_name=name) |
599 | 5188ab37 | Iustin Pop | for name in offline_disk_instances] |
600 | 5188ab37 | Iustin Pop | job_id = cli.SendJob(job, cl=client) |
601 | 5188ab37 | Iustin Pop | |
602 | a9105b24 | Michael Hanselmann | try: |
603 | a9105b24 | Michael Hanselmann | cli.PollJob(job_id, cl=client, feedback_fn=logging.debug) |
604 | a9105b24 | Michael Hanselmann | except Exception: # pylint: disable-msg=W0703 |
605 | a9105b24 | Michael Hanselmann | logging.exception("Error while activating disks") |
606 | a8083063 | Iustin Pop | |
607 | a8083063 | Iustin Pop | |
608 | 001b3825 | Michael Hanselmann | def OpenStateFile(path): |
609 | 001b3825 | Michael Hanselmann | """Opens the state file and acquires a lock on it. |
610 | 001b3825 | Michael Hanselmann | |
611 | 001b3825 | Michael Hanselmann | @type path: string |
612 | 001b3825 | Michael Hanselmann | @param path: Path to state file |
613 | 001b3825 | Michael Hanselmann | |
614 | 001b3825 | Michael Hanselmann | """ |
615 | 001b3825 | Michael Hanselmann | # The two-step dance below is necessary to allow both opening existing |
616 | 001b3825 | Michael Hanselmann | # file read/write and creating if not existing. Vanilla open will truncate |
617 | 001b3825 | Michael Hanselmann | # an existing file -or- allow creating if not existing. |
618 | 001b3825 | Michael Hanselmann | statefile_fd = os.open(path, os.O_RDWR | os.O_CREAT) |
619 | 001b3825 | Michael Hanselmann | |
620 | 001b3825 | Michael Hanselmann | # Try to acquire lock on state file. If this fails, another watcher instance |
621 | 001b3825 | Michael Hanselmann | # might already be running or another program is temporarily blocking the |
622 | 001b3825 | Michael Hanselmann | # watcher from running. |
623 | 001b3825 | Michael Hanselmann | try: |
624 | 001b3825 | Michael Hanselmann | utils.LockFile(statefile_fd) |
625 | 001b3825 | Michael Hanselmann | except errors.LockError, err: |
626 | 001b3825 | Michael Hanselmann | logging.error("Can't acquire lock on state file %s: %s", path, err) |
627 | 001b3825 | Michael Hanselmann | return None |
628 | 001b3825 | Michael Hanselmann | |
629 | 001b3825 | Michael Hanselmann | return os.fdopen(statefile_fd, "w+") |
630 | 001b3825 | Michael Hanselmann | |
631 | 001b3825 | Michael Hanselmann | |
632 | db147305 | Tom Limoncelli | def IsRapiResponding(hostname): |
633 | db147305 | Tom Limoncelli | """Connects to RAPI port and does a simple test. |
634 | db147305 | Tom Limoncelli | |
635 | db147305 | Tom Limoncelli | Connects to RAPI port of hostname and does a simple test. At this time, the |
636 | db147305 | Tom Limoncelli | test is GetVersion. |
637 | db147305 | Tom Limoncelli | |
638 | db147305 | Tom Limoncelli | @type hostname: string |
639 | db147305 | Tom Limoncelli | @param hostname: hostname of the node to connect to. |
640 | db147305 | Tom Limoncelli | @rtype: bool |
641 | db147305 | Tom Limoncelli | @return: Whether RAPI is working properly |
642 | db147305 | Tom Limoncelli | |
643 | db147305 | Tom Limoncelli | """ |
644 | 34f06005 | Iustin Pop | curl_config = rapi.client.GenericCurlConfig() |
645 | 2a7c3583 | Michael Hanselmann | rapi_client = rapi.client.GanetiRapiClient(hostname, |
646 | 2a7c3583 | Michael Hanselmann | curl_config_fn=curl_config) |
647 | db147305 | Tom Limoncelli | try: |
648 | db147305 | Tom Limoncelli | master_version = rapi_client.GetVersion() |
649 | db147305 | Tom Limoncelli | except rapi.client.CertificateError, err: |
650 | db147305 | Tom Limoncelli | logging.warning("RAPI Error: CertificateError (%s)", err) |
651 | db147305 | Tom Limoncelli | return False |
652 | db147305 | Tom Limoncelli | except rapi.client.GanetiApiError, err: |
653 | db147305 | Tom Limoncelli | logging.warning("RAPI Error: GanetiApiError (%s)", err) |
654 | db147305 | Tom Limoncelli | return False |
655 | db147305 | Tom Limoncelli | logging.debug("RAPI Result: master_version is %s", master_version) |
656 | db147305 | Tom Limoncelli | return master_version == constants.RAPI_VERSION |
657 | db147305 | Tom Limoncelli | |
658 | db147305 | Tom Limoncelli | |
659 | a8083063 | Iustin Pop | def ParseOptions(): |
660 | a8083063 | Iustin Pop | """Parse the command line options. |
661 | a8083063 | Iustin Pop | |
662 | c41eea6e | Iustin Pop | @return: (options, args) as from OptionParser.parse_args() |
663 | a8083063 | Iustin Pop | |
664 | a8083063 | Iustin Pop | """ |
665 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher", |
666 | a8083063 | Iustin Pop | usage="%prog [-d]", |
667 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" % |
668 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
669 | a8083063 | Iustin Pop | |
670 | 6d4e8ec0 | Iustin Pop | parser.add_option(cli.DEBUG_OPT) |
671 | f07521e5 | Iustin Pop | parser.add_option("-A", "--job-age", dest="job_age", |
672 | f07521e5 | Iustin Pop | help="Autoarchive jobs older than this age (default" |
673 | f07521e5 | Iustin Pop | " 6 hours)", default=6*3600) |
674 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
675 | f07521e5 | Iustin Pop | options.job_age = cli.ParseTimespec(options.job_age) |
676 | a8083063 | Iustin Pop | return options, args |
677 | a8083063 | Iustin Pop | |
678 | a8083063 | Iustin Pop | |
679 | 2a7c3583 | Michael Hanselmann | @rapi.client.UsesRapiClient |
680 | a8083063 | Iustin Pop | def main(): |
681 | a8083063 | Iustin Pop | """Main function. |
682 | a8083063 | Iustin Pop | |
683 | a8083063 | Iustin Pop | """ |
684 | 7260cfbe | Iustin Pop | global client # pylint: disable-msg=W0603 |
685 | e125c67c | Michael Hanselmann | |
686 | f93427cd | Iustin Pop | options, args = ParseOptions() |
687 | f93427cd | Iustin Pop | |
688 | f93427cd | Iustin Pop | if args: # watcher doesn't take any arguments |
689 | f93427cd | Iustin Pop | print >> sys.stderr, ("Usage: %s [-f] " % sys.argv[0]) |
690 | f93427cd | Iustin Pop | sys.exit(constants.EXIT_FAILURE) |
691 | a8083063 | Iustin Pop | |
692 | 82d9caef | Iustin Pop | utils.SetupLogging(constants.LOG_WATCHER, debug=options.debug, |
693 | 82d9caef | Iustin Pop | stderr_logging=options.debug) |
694 | a8083063 | Iustin Pop | |
695 | 3753b2cb | Michael Hanselmann | if ShouldPause(): |
696 | 3753b2cb | Michael Hanselmann | logging.debug("Pause has been set, exiting") |
697 | 3753b2cb | Michael Hanselmann | sys.exit(constants.EXIT_SUCCESS) |
698 | 3753b2cb | Michael Hanselmann | |
699 | 001b3825 | Michael Hanselmann | statefile = OpenStateFile(constants.WATCHER_STATEFILE) |
700 | 001b3825 | Michael Hanselmann | if not statefile: |
701 | 001b3825 | Michael Hanselmann | sys.exit(constants.EXIT_FAILURE) |
702 | 001b3825 | Michael Hanselmann | |
703 | 24edc6d4 | Iustin Pop | update_file = False |
704 | a8083063 | Iustin Pop | try: |
705 | f1115454 | Guido Trotter | StartNodeDaemons() |
706 | 9e289e36 | Guido Trotter | RunWatcherHooks() |
707 | 50273051 | Iustin Pop | # run node maintenance in all cases, even if master, so that old |
708 | 50273051 | Iustin Pop | # masters can be properly cleaned up too |
709 | 50273051 | Iustin Pop | if NodeMaintenance.ShouldRun(): |
710 | 50273051 | Iustin Pop | NodeMaintenance().Exec() |
711 | c4f0219c | Iustin Pop | |
712 | 001b3825 | Michael Hanselmann | notepad = WatcherState(statefile) |
713 | 781b2b2b | Michael Hanselmann | try: |
714 | 2c404217 | Iustin Pop | try: |
715 | 2c404217 | Iustin Pop | client = cli.GetClient() |
716 | 2c404217 | Iustin Pop | except errors.OpPrereqError: |
717 | 2c404217 | Iustin Pop | # this is, from cli.GetClient, a not-master case |
718 | 7dfb83c2 | Iustin Pop | logging.debug("Not on master, exiting") |
719 | 24edc6d4 | Iustin Pop | update_file = True |
720 | 2c404217 | Iustin Pop | sys.exit(constants.EXIT_SUCCESS) |
721 | 7dfb83c2 | Iustin Pop | except luxi.NoMasterError, err: |
722 | 7dfb83c2 | Iustin Pop | logging.warning("Master seems to be down (%s), trying to restart", |
723 | 7dfb83c2 | Iustin Pop | str(err)) |
724 | 2826b361 | Guido Trotter | if not utils.EnsureDaemon(constants.MASTERD): |
725 | 7dfb83c2 | Iustin Pop | logging.critical("Can't start the master, exiting") |
726 | 7dfb83c2 | Iustin Pop | sys.exit(constants.EXIT_FAILURE) |
727 | 7dfb83c2 | Iustin Pop | # else retry the connection |
728 | 7dfb83c2 | Iustin Pop | client = cli.GetClient() |
729 | cc962d58 | Iustin Pop | |
730 | 83052f9e | Guido Trotter | # we are on master now |
731 | 2826b361 | Guido Trotter | utils.EnsureDaemon(constants.RAPI) |
732 | c4f0219c | Iustin Pop | |
733 | db147305 | Tom Limoncelli | # If RAPI isn't responding to queries, try one restart. |
734 | db147305 | Tom Limoncelli | logging.debug("Attempting to talk with RAPI.") |
735 | 9769bb78 | Manuel Franceschini | if not IsRapiResponding(constants.IP4_ADDRESS_LOCALHOST): |
736 | db147305 | Tom Limoncelli | logging.warning("Couldn't get answer from Ganeti RAPI daemon." |
737 | db147305 | Tom Limoncelli | " Restarting Ganeti RAPI.") |
738 | db147305 | Tom Limoncelli | utils.StopDaemon(constants.RAPI) |
739 | db147305 | Tom Limoncelli | utils.EnsureDaemon(constants.RAPI) |
740 | db147305 | Tom Limoncelli | logging.debug("Second attempt to talk with RAPI") |
741 | 9769bb78 | Manuel Franceschini | if not IsRapiResponding(constants.IP4_ADDRESS_LOCALHOST): |
742 | db147305 | Tom Limoncelli | logging.fatal("RAPI is not responding. Please investigate.") |
743 | db147305 | Tom Limoncelli | logging.debug("Successfully talked to RAPI.") |
744 | db147305 | Tom Limoncelli | |
745 | cc962d58 | Iustin Pop | try: |
746 | cc962d58 | Iustin Pop | watcher = Watcher(options, notepad) |
747 | cc962d58 | Iustin Pop | except errors.ConfigurationError: |
748 | cc962d58 | Iustin Pop | # Just exit if there's no configuration |
749 | 24edc6d4 | Iustin Pop | update_file = True |
750 | cc962d58 | Iustin Pop | sys.exit(constants.EXIT_SUCCESS) |
751 | e125c67c | Michael Hanselmann | |
752 | cc962d58 | Iustin Pop | watcher.Run() |
753 | 24edc6d4 | Iustin Pop | update_file = True |
754 | 24edc6d4 | Iustin Pop | |
755 | cc962d58 | Iustin Pop | finally: |
756 | 7dfb83c2 | Iustin Pop | if update_file: |
757 | 7dfb83c2 | Iustin Pop | notepad.Save() |
758 | 7dfb83c2 | Iustin Pop | else: |
759 | 7dfb83c2 | Iustin Pop | logging.debug("Not updating status file due to failure") |
760 | 1b052f42 | Michael Hanselmann | except SystemExit: |
761 | 1b052f42 | Michael Hanselmann | raise |
762 | 38242904 | Iustin Pop | except NotMasterError: |
763 | 438b45d4 | Michael Hanselmann | logging.debug("Not master, exiting") |
764 | 38242904 | Iustin Pop | sys.exit(constants.EXIT_NOTMASTER) |
765 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err: |
766 | 438b45d4 | Michael Hanselmann | logging.error("Cannot resolve hostname '%s', exiting.", err.args[0]) |
767 | 89e1fc26 | Iustin Pop | sys.exit(constants.EXIT_NODESETUP_ERROR) |
768 | 24edc6d4 | Iustin Pop | except errors.JobQueueFull: |
769 | 24edc6d4 | Iustin Pop | logging.error("Job queue is full, can't query cluster state") |
770 | 24edc6d4 | Iustin Pop | except errors.JobQueueDrainError: |
771 | 24edc6d4 | Iustin Pop | logging.error("Job queue is drained, can't maintain cluster state") |
772 | 438b45d4 | Michael Hanselmann | except Exception, err: |
773 | 001b3825 | Michael Hanselmann | logging.exception(str(err)) |
774 | 438b45d4 | Michael Hanselmann | sys.exit(constants.EXIT_FAILURE) |
775 | a8083063 | Iustin Pop | |
776 | 5a3103e9 | Michael Hanselmann | |
777 | a8083063 | Iustin Pop | if __name__ == '__main__': |
778 | a8083063 | Iustin Pop | main() |