root / daemons / ganeti-watcher @ 4bffa7f7
History | View | Annotate | Download (12.5 kB)
1 | a8083063 | Iustin Pop | #!/usr/bin/python |
---|---|---|---|
2 | a8083063 | Iustin Pop | # |
3 | a8083063 | Iustin Pop | |
4 | fc428e32 | Michael Hanselmann | # Copyright (C) 2006, 2007, 2008 Google Inc. |
5 | a8083063 | Iustin Pop | # |
6 | a8083063 | Iustin Pop | # This program is free software; you can redistribute it and/or modify |
7 | a8083063 | Iustin Pop | # it under the terms of the GNU General Public License as published by |
8 | a8083063 | Iustin Pop | # the Free Software Foundation; either version 2 of the License, or |
9 | a8083063 | Iustin Pop | # (at your option) any later version. |
10 | a8083063 | Iustin Pop | # |
11 | a8083063 | Iustin Pop | # This program is distributed in the hope that it will be useful, but |
12 | a8083063 | Iustin Pop | # WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | a8083063 | Iustin Pop | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 | a8083063 | Iustin Pop | # General Public License for more details. |
15 | a8083063 | Iustin Pop | # |
16 | a8083063 | Iustin Pop | # You should have received a copy of the GNU General Public License |
17 | a8083063 | Iustin Pop | # along with this program; if not, write to the Free Software |
18 | a8083063 | Iustin Pop | # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA |
19 | a8083063 | Iustin Pop | # 02110-1301, USA. |
20 | a8083063 | Iustin Pop | |
21 | a8083063 | Iustin Pop | |
22 | a8083063 | Iustin Pop | """Tool to restart erronously downed virtual machines. |
23 | a8083063 | Iustin Pop | |
24 | a8083063 | Iustin Pop | This program and set of classes implement a watchdog to restart |
25 | a8083063 | Iustin Pop | virtual machines in a Ganeti cluster that have crashed or been killed |
26 | a8083063 | Iustin Pop | by a node reboot. Run from cron or similar. |
27 | a8083063 | Iustin Pop | |
28 | 5a3103e9 | Michael Hanselmann | """ |
29 | a8083063 | Iustin Pop | |
30 | a8083063 | Iustin Pop | import os |
31 | a8083063 | Iustin Pop | import sys |
32 | a8083063 | Iustin Pop | import time |
33 | 438b45d4 | Michael Hanselmann | import logging |
34 | a8083063 | Iustin Pop | from optparse import OptionParser |
35 | a8083063 | Iustin Pop | |
36 | a8083063 | Iustin Pop | from ganeti import utils |
37 | a8083063 | Iustin Pop | from ganeti import constants |
38 | 67fe61c4 | Michael Hanselmann | from ganeti import serializer |
39 | 89e1fc26 | Iustin Pop | from ganeti import errors |
40 | e125c67c | Michael Hanselmann | from ganeti import opcodes |
41 | e125c67c | Michael Hanselmann | from ganeti import cli |
42 | a8083063 | Iustin Pop | |
43 | a8083063 | Iustin Pop | |
44 | 5a3103e9 | Michael Hanselmann | MAXTRIES = 5 |
45 | b7309a0d | Iustin Pop | BAD_STATES = ['ERROR_down'] |
46 | cbfc4681 | Iustin Pop | HELPLESS_STATES = ['ERROR_nodedown', 'ERROR_nodeoffline'] |
47 | 5a3103e9 | Michael Hanselmann | NOTICE = 'NOTICE' |
48 | 5a3103e9 | Michael Hanselmann | ERROR = 'ERROR' |
49 | 7b195d9b | Michael Hanselmann | KEY_RESTART_COUNT = "restart_count" |
50 | 7b195d9b | Michael Hanselmann | KEY_RESTART_WHEN = "restart_when" |
51 | 7b195d9b | Michael Hanselmann | KEY_BOOT_ID = "bootid" |
52 | 5a3103e9 | Michael Hanselmann | |
53 | 5a3103e9 | Michael Hanselmann | |
54 | e125c67c | Michael Hanselmann | # Global client object |
55 | e125c67c | Michael Hanselmann | client = None |
56 | e125c67c | Michael Hanselmann | |
57 | e125c67c | Michael Hanselmann | |
58 | 7bca53e4 | Michael Hanselmann | class NotMasterError(errors.GenericError): |
59 | 38242904 | Iustin Pop | """Exception raised when this host is not the master.""" |
60 | a8083063 | Iustin Pop | |
61 | a8083063 | Iustin Pop | |
62 | a8083063 | Iustin Pop | def Indent(s, prefix='| '): |
63 | a8083063 | Iustin Pop | """Indent a piece of text with a given prefix before each line. |
64 | a8083063 | Iustin Pop | |
65 | c41eea6e | Iustin Pop | @param s: the string to indent |
66 | c41eea6e | Iustin Pop | @param prefix: the string to prepend each line |
67 | 38242904 | Iustin Pop | |
68 | a8083063 | Iustin Pop | """ |
69 | a8083063 | Iustin Pop | return "%s%s\n" % (prefix, ('\n' + prefix).join(s.splitlines())) |
70 | a8083063 | Iustin Pop | |
71 | a8083063 | Iustin Pop | |
72 | 5a3103e9 | Michael Hanselmann | class WatcherState(object): |
73 | a8083063 | Iustin Pop | """Interface to a state file recording restart attempts. |
74 | a8083063 | Iustin Pop | |
75 | a8083063 | Iustin Pop | """ |
76 | a8083063 | Iustin Pop | def __init__(self): |
77 | 5a3103e9 | Michael Hanselmann | """Open, lock, read and parse the file. |
78 | 5a3103e9 | Michael Hanselmann | |
79 | 7bca53e4 | Michael Hanselmann | Raises exception on lock contention. |
80 | 5a3103e9 | Michael Hanselmann | |
81 | 5a3103e9 | Michael Hanselmann | """ |
82 | a8083063 | Iustin Pop | # The two-step dance below is necessary to allow both opening existing |
83 | a8083063 | Iustin Pop | # file read/write and creating if not existing. Vanilla open will truncate |
84 | a8083063 | Iustin Pop | # an existing file -or- allow creating if not existing. |
85 | fc428e32 | Michael Hanselmann | fd = os.open(constants.WATCHER_STATEFILE, os.O_RDWR | os.O_CREAT) |
86 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
87 | a8083063 | Iustin Pop | |
88 | eb0f0ce0 | Michael Hanselmann | utils.LockFile(self.statefile.fileno()) |
89 | a8083063 | Iustin Pop | |
90 | 5a3103e9 | Michael Hanselmann | try: |
91 | b76f660d | Michael Hanselmann | self._data = serializer.Load(self.statefile.read()) |
92 | 5a3103e9 | Michael Hanselmann | except Exception, msg: |
93 | 5a3103e9 | Michael Hanselmann | # Ignore errors while loading the file and treat it as empty |
94 | b76f660d | Michael Hanselmann | self._data = {} |
95 | 438b45d4 | Michael Hanselmann | logging.warning(("Empty or invalid state file. Using defaults." |
96 | 438b45d4 | Michael Hanselmann | " Error message: %s"), msg) |
97 | 5a3103e9 | Michael Hanselmann | |
98 | b76f660d | Michael Hanselmann | if "instance" not in self._data: |
99 | b76f660d | Michael Hanselmann | self._data["instance"] = {} |
100 | b76f660d | Michael Hanselmann | if "node" not in self._data: |
101 | b76f660d | Michael Hanselmann | self._data["node"] = {} |
102 | 5a3103e9 | Michael Hanselmann | |
103 | 26517d45 | Iustin Pop | self._orig_data = serializer.Dump(self._data) |
104 | 2fb96d39 | Michael Hanselmann | |
105 | fc428e32 | Michael Hanselmann | def Save(self): |
106 | fc428e32 | Michael Hanselmann | """Save state to file, then unlock and close it. |
107 | 5a3103e9 | Michael Hanselmann | |
108 | 5a3103e9 | Michael Hanselmann | """ |
109 | fc428e32 | Michael Hanselmann | assert self.statefile |
110 | fc428e32 | Michael Hanselmann | |
111 | 26517d45 | Iustin Pop | serialized_form = serializer.Dump(self._data) |
112 | 26517d45 | Iustin Pop | if self._orig_data == serialized_form: |
113 | 2fb96d39 | Michael Hanselmann | logging.debug("Data didn't change, just touching status file") |
114 | 2fb96d39 | Michael Hanselmann | os.utime(constants.WATCHER_STATEFILE, None) |
115 | 2fb96d39 | Michael Hanselmann | return |
116 | 2fb96d39 | Michael Hanselmann | |
117 | fc428e32 | Michael Hanselmann | # We need to make sure the file is locked before renaming it, otherwise |
118 | fc428e32 | Michael Hanselmann | # starting ganeti-watcher again at the same time will create a conflict. |
119 | fc428e32 | Michael Hanselmann | fd = utils.WriteFile(constants.WATCHER_STATEFILE, |
120 | 26517d45 | Iustin Pop | data=serialized_form, |
121 | eb0f0ce0 | Michael Hanselmann | prewrite=utils.LockFile, close=False) |
122 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
123 | 5a3103e9 | Michael Hanselmann | |
124 | fc428e32 | Michael Hanselmann | def Close(self): |
125 | 5a3103e9 | Michael Hanselmann | """Unlock configuration file and close it. |
126 | 5a3103e9 | Michael Hanselmann | |
127 | 5a3103e9 | Michael Hanselmann | """ |
128 | 5a3103e9 | Michael Hanselmann | assert self.statefile |
129 | 5a3103e9 | Michael Hanselmann | |
130 | fc428e32 | Michael Hanselmann | # Files are automatically unlocked when closing them |
131 | 5a3103e9 | Michael Hanselmann | self.statefile.close() |
132 | 5a3103e9 | Michael Hanselmann | self.statefile = None |
133 | 5a3103e9 | Michael Hanselmann | |
134 | 5a3103e9 | Michael Hanselmann | def GetNodeBootID(self, name): |
135 | 5a3103e9 | Michael Hanselmann | """Returns the last boot ID of a node or None. |
136 | a8083063 | Iustin Pop | |
137 | 5a3103e9 | Michael Hanselmann | """ |
138 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
139 | 5a3103e9 | Michael Hanselmann | |
140 | 7b195d9b | Michael Hanselmann | if name in ndata and KEY_BOOT_ID in ndata[name]: |
141 | 7b195d9b | Michael Hanselmann | return ndata[name][KEY_BOOT_ID] |
142 | 5a3103e9 | Michael Hanselmann | return None |
143 | 5a3103e9 | Michael Hanselmann | |
144 | 5a3103e9 | Michael Hanselmann | def SetNodeBootID(self, name, bootid): |
145 | 5a3103e9 | Michael Hanselmann | """Sets the boot ID of a node. |
146 | 5a3103e9 | Michael Hanselmann | |
147 | 5a3103e9 | Michael Hanselmann | """ |
148 | 5a3103e9 | Michael Hanselmann | assert bootid |
149 | a8083063 | Iustin Pop | |
150 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
151 | a8083063 | Iustin Pop | |
152 | 5a3103e9 | Michael Hanselmann | if name not in ndata: |
153 | 5a3103e9 | Michael Hanselmann | ndata[name] = {} |
154 | 5a3103e9 | Michael Hanselmann | |
155 | 7b195d9b | Michael Hanselmann | ndata[name][KEY_BOOT_ID] = bootid |
156 | 5a3103e9 | Michael Hanselmann | |
157 | 5a3103e9 | Michael Hanselmann | def NumberOfRestartAttempts(self, instance): |
158 | a8083063 | Iustin Pop | """Returns number of previous restart attempts. |
159 | a8083063 | Iustin Pop | |
160 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
161 | c41eea6e | Iustin Pop | @param instance: the instance to look up |
162 | 38242904 | Iustin Pop | |
163 | a8083063 | Iustin Pop | """ |
164 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
165 | a8083063 | Iustin Pop | |
166 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
167 | 7b195d9b | Michael Hanselmann | return idata[instance.name][KEY_RESTART_COUNT] |
168 | a8083063 | Iustin Pop | |
169 | a8083063 | Iustin Pop | return 0 |
170 | a8083063 | Iustin Pop | |
171 | 5a3103e9 | Michael Hanselmann | def RecordRestartAttempt(self, instance): |
172 | a8083063 | Iustin Pop | """Record a restart attempt. |
173 | a8083063 | Iustin Pop | |
174 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
175 | c41eea6e | Iustin Pop | @param instance: the instance being restarted |
176 | 38242904 | Iustin Pop | |
177 | a8083063 | Iustin Pop | """ |
178 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
179 | a8083063 | Iustin Pop | |
180 | 5a3103e9 | Michael Hanselmann | if instance.name not in idata: |
181 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] = {} |
182 | 5a3103e9 | Michael Hanselmann | else: |
183 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] |
184 | a8083063 | Iustin Pop | |
185 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_WHEN] = time.time() |
186 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
187 | a8083063 | Iustin Pop | |
188 | 5a3103e9 | Michael Hanselmann | def RemoveInstance(self, instance): |
189 | c41eea6e | Iustin Pop | """Update state to reflect that a machine is running. |
190 | a8083063 | Iustin Pop | |
191 | c41eea6e | Iustin Pop | This method removes the record for a named instance (as we only |
192 | c41eea6e | Iustin Pop | track down instances). |
193 | a8083063 | Iustin Pop | |
194 | c41eea6e | Iustin Pop | @type instance: L{Instance} |
195 | c41eea6e | Iustin Pop | @param instance: the instance to remove from books |
196 | 38242904 | Iustin Pop | |
197 | a8083063 | Iustin Pop | """ |
198 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
199 | a8083063 | Iustin Pop | |
200 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
201 | 5a3103e9 | Michael Hanselmann | del idata[instance.name] |
202 | a8083063 | Iustin Pop | |
203 | a8083063 | Iustin Pop | |
204 | a8083063 | Iustin Pop | class Instance(object): |
205 | a8083063 | Iustin Pop | """Abstraction for a Virtual Machine instance. |
206 | a8083063 | Iustin Pop | |
207 | a8083063 | Iustin Pop | """ |
208 | 5a3103e9 | Michael Hanselmann | def __init__(self, name, state, autostart): |
209 | a8083063 | Iustin Pop | self.name = name |
210 | a8083063 | Iustin Pop | self.state = state |
211 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart |
212 | a8083063 | Iustin Pop | |
213 | a8083063 | Iustin Pop | def Restart(self): |
214 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance. |
215 | 3ecf6786 | Iustin Pop | |
216 | 3ecf6786 | Iustin Pop | """ |
217 | e125c67c | Michael Hanselmann | op = opcodes.OpStartupInstance(instance_name=self.name, |
218 | e125c67c | Michael Hanselmann | force=False, |
219 | e125c67c | Michael Hanselmann | extra_args=None) |
220 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
221 | a8083063 | Iustin Pop | |
222 | 5a3103e9 | Michael Hanselmann | def ActivateDisks(self): |
223 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance. |
224 | 5a3103e9 | Michael Hanselmann | |
225 | 5a3103e9 | Michael Hanselmann | """ |
226 | e125c67c | Michael Hanselmann | op = opcodes.OpActivateInstanceDisks(instance_name=self.name) |
227 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
228 | a8083063 | Iustin Pop | |
229 | a8083063 | Iustin Pop | |
230 | 5a3103e9 | Michael Hanselmann | def GetInstanceList(with_secondaries=None): |
231 | 5a3103e9 | Michael Hanselmann | """Get a list of instances on this cluster. |
232 | 5a3103e9 | Michael Hanselmann | |
233 | 5a3103e9 | Michael Hanselmann | """ |
234 | b7309a0d | Iustin Pop | fields = ["name", "status", "admin_state"] |
235 | a8083063 | Iustin Pop | |
236 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
237 | e125c67c | Michael Hanselmann | fields.append("snodes") |
238 | a8083063 | Iustin Pop | |
239 | e125c67c | Michael Hanselmann | result = client.QueryInstances([], fields) |
240 | 5a3103e9 | Michael Hanselmann | |
241 | 5a3103e9 | Michael Hanselmann | instances = [] |
242 | e125c67c | Michael Hanselmann | for fields in result: |
243 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
244 | 5a3103e9 | Michael Hanselmann | (name, status, autostart, snodes) = fields |
245 | 5a3103e9 | Michael Hanselmann | |
246 | e125c67c | Michael Hanselmann | if not snodes: |
247 | a8083063 | Iustin Pop | continue |
248 | 5a3103e9 | Michael Hanselmann | |
249 | 5a3103e9 | Michael Hanselmann | for node in with_secondaries: |
250 | e125c67c | Michael Hanselmann | if node in snodes: |
251 | 5a3103e9 | Michael Hanselmann | break |
252 | 5a3103e9 | Michael Hanselmann | else: |
253 | a8083063 | Iustin Pop | continue |
254 | a8083063 | Iustin Pop | |
255 | 5a3103e9 | Michael Hanselmann | else: |
256 | 5a3103e9 | Michael Hanselmann | (name, status, autostart) = fields |
257 | 5a3103e9 | Michael Hanselmann | |
258 | e125c67c | Michael Hanselmann | instances.append(Instance(name, status, autostart)) |
259 | a8083063 | Iustin Pop | |
260 | 5a3103e9 | Michael Hanselmann | return instances |
261 | 5a3103e9 | Michael Hanselmann | |
262 | 5a3103e9 | Michael Hanselmann | |
263 | 5a3103e9 | Michael Hanselmann | def GetNodeBootIDs(): |
264 | 5a3103e9 | Michael Hanselmann | """Get a dict mapping nodes to boot IDs. |
265 | 5a3103e9 | Michael Hanselmann | |
266 | 5a3103e9 | Michael Hanselmann | """ |
267 | cbfc4681 | Iustin Pop | result = client.QueryNodes([], ["name", "bootid", "offline"]) |
268 | cbfc4681 | Iustin Pop | return dict([(name, (bootid, offline)) for name, bootid, offline in result]) |
269 | a8083063 | Iustin Pop | |
270 | a8083063 | Iustin Pop | |
271 | 5a3103e9 | Michael Hanselmann | class Watcher(object): |
272 | a8083063 | Iustin Pop | """Encapsulate the logic for restarting erronously halted virtual machines. |
273 | a8083063 | Iustin Pop | |
274 | a8083063 | Iustin Pop | The calling program should periodically instantiate me and call Run(). |
275 | a8083063 | Iustin Pop | This will traverse the list of instances, and make up to MAXTRIES attempts |
276 | a8083063 | Iustin Pop | to restart machines that are down. |
277 | 38242904 | Iustin Pop | |
278 | a8083063 | Iustin Pop | """ |
279 | a8083063 | Iustin Pop | def __init__(self): |
280 | 2859b87b | Michael Hanselmann | master = client.QueryConfigValues(["master_node"])[0] |
281 | 89e1fc26 | Iustin Pop | if master != utils.HostInfo().name: |
282 | 3ecf6786 | Iustin Pop | raise NotMasterError("This is not the master node") |
283 | 5a3103e9 | Michael Hanselmann | self.instances = GetInstanceList() |
284 | 5a3103e9 | Michael Hanselmann | self.bootids = GetNodeBootIDs() |
285 | eee1fa2d | Iustin Pop | self.started_instances = set() |
286 | a8083063 | Iustin Pop | |
287 | a8083063 | Iustin Pop | def Run(self): |
288 | 5a3103e9 | Michael Hanselmann | notepad = WatcherState() |
289 | 78f3bd30 | Michael Hanselmann | try: |
290 | 78f3bd30 | Michael Hanselmann | self.CheckInstances(notepad) |
291 | 78f3bd30 | Michael Hanselmann | self.CheckDisks(notepad) |
292 | 78f3bd30 | Michael Hanselmann | self.VerifyDisks() |
293 | 78f3bd30 | Michael Hanselmann | finally: |
294 | 78f3bd30 | Michael Hanselmann | notepad.Save() |
295 | 5a3103e9 | Michael Hanselmann | |
296 | 5a3103e9 | Michael Hanselmann | def CheckDisks(self, notepad): |
297 | 5a3103e9 | Michael Hanselmann | """Check all nodes for restarted ones. |
298 | 38242904 | Iustin Pop | |
299 | a8083063 | Iustin Pop | """ |
300 | 5a3103e9 | Michael Hanselmann | check_nodes = [] |
301 | cbfc4681 | Iustin Pop | for name, (new_id, offline) in self.bootids.iteritems(): |
302 | 5a3103e9 | Michael Hanselmann | old = notepad.GetNodeBootID(name) |
303 | 37b77b18 | Iustin Pop | if new_id is None: |
304 | 37b77b18 | Iustin Pop | # Bad node, not returning a boot id |
305 | cbfc4681 | Iustin Pop | if not offline: |
306 | cbfc4681 | Iustin Pop | logging.debug("Node %s missing boot id, skipping secondary checks", |
307 | cbfc4681 | Iustin Pop | name) |
308 | 37b77b18 | Iustin Pop | continue |
309 | 26517d45 | Iustin Pop | if old != new_id: |
310 | 5a3103e9 | Michael Hanselmann | # Node's boot ID has changed, proably through a reboot. |
311 | 5a3103e9 | Michael Hanselmann | check_nodes.append(name) |
312 | 5a3103e9 | Michael Hanselmann | |
313 | 5a3103e9 | Michael Hanselmann | if check_nodes: |
314 | 5a3103e9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a |
315 | 5a3103e9 | Michael Hanselmann | # secondary node. |
316 | 5a3103e9 | Michael Hanselmann | for instance in GetInstanceList(with_secondaries=check_nodes): |
317 | 0c0f834d | Iustin Pop | if not instance.autostart: |
318 | 438b45d4 | Michael Hanselmann | logging.info(("Skipping disk activation for non-autostart" |
319 | 438b45d4 | Michael Hanselmann | " instance %s"), instance.name) |
320 | 0c0f834d | Iustin Pop | continue |
321 | eee1fa2d | Iustin Pop | if instance.name in self.started_instances: |
322 | eee1fa2d | Iustin Pop | # we already tried to start the instance, which should have |
323 | eee1fa2d | Iustin Pop | # activated its drives (if they can be at all) |
324 | eee1fa2d | Iustin Pop | continue |
325 | 5a3103e9 | Michael Hanselmann | try: |
326 | 438b45d4 | Michael Hanselmann | logging.info("Activating disks for instance %s", instance.name) |
327 | 5a3103e9 | Michael Hanselmann | instance.ActivateDisks() |
328 | b7309a0d | Iustin Pop | except Exception: |
329 | b7309a0d | Iustin Pop | logging.exception("Error while activating disks for instance %s", |
330 | b7309a0d | Iustin Pop | instance.name) |
331 | 5a3103e9 | Michael Hanselmann | |
332 | 5a3103e9 | Michael Hanselmann | # Keep changed boot IDs |
333 | 5a3103e9 | Michael Hanselmann | for name in check_nodes: |
334 | 5a3103e9 | Michael Hanselmann | notepad.SetNodeBootID(name, self.bootids[name]) |
335 | a8083063 | Iustin Pop | |
336 | 5a3103e9 | Michael Hanselmann | def CheckInstances(self, notepad): |
337 | 5a3103e9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones. |
338 | 5a3103e9 | Michael Hanselmann | |
339 | 5a3103e9 | Michael Hanselmann | """ |
340 | a8083063 | Iustin Pop | for instance in self.instances: |
341 | a8083063 | Iustin Pop | if instance.state in BAD_STATES: |
342 | 5a3103e9 | Michael Hanselmann | n = notepad.NumberOfRestartAttempts(instance) |
343 | a8083063 | Iustin Pop | |
344 | a8083063 | Iustin Pop | if n > MAXTRIES: |
345 | a8083063 | Iustin Pop | # stay quiet. |
346 | a8083063 | Iustin Pop | continue |
347 | a8083063 | Iustin Pop | elif n < MAXTRIES: |
348 | a8083063 | Iustin Pop | last = " (Attempt #%d)" % (n + 1) |
349 | a8083063 | Iustin Pop | else: |
350 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
351 | 438b45d4 | Michael Hanselmann | logging.error("Could not restart %s after %d attempts, giving up", |
352 | 438b45d4 | Michael Hanselmann | instance.name, MAXTRIES) |
353 | a8083063 | Iustin Pop | continue |
354 | a8083063 | Iustin Pop | try: |
355 | 438b45d4 | Michael Hanselmann | logging.info("Restarting %s%s", |
356 | 438b45d4 | Michael Hanselmann | instance.name, last) |
357 | a8083063 | Iustin Pop | instance.Restart() |
358 | eee1fa2d | Iustin Pop | self.started_instances.add(instance.name) |
359 | b7309a0d | Iustin Pop | except Exception: |
360 | 4bffa7f7 | Iustin Pop | logging.exception("Error while restarting instance %s", |
361 | 4bffa7f7 | Iustin Pop | instance.name) |
362 | a8083063 | Iustin Pop | |
363 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
364 | a8083063 | Iustin Pop | elif instance.state in HELPLESS_STATES: |
365 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
366 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
367 | a8083063 | Iustin Pop | else: |
368 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
369 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
370 | 438b45d4 | Michael Hanselmann | logging.info("Restart of %s succeeded", instance.name) |
371 | a8083063 | Iustin Pop | |
372 | b7309a0d | Iustin Pop | @staticmethod |
373 | b7309a0d | Iustin Pop | def VerifyDisks(): |
374 | d2f311db | Iustin Pop | """Run gnt-cluster verify-disks. |
375 | d2f311db | Iustin Pop | |
376 | d2f311db | Iustin Pop | """ |
377 | 5188ab37 | Iustin Pop | op = opcodes.OpVerifyDisks() |
378 | 5188ab37 | Iustin Pop | result = cli.SubmitOpCode(op, cl=client) |
379 | 5188ab37 | Iustin Pop | if not isinstance(result, (tuple, list)): |
380 | 5188ab37 | Iustin Pop | logging.error("Can't get a valid result from verify-disks") |
381 | 5188ab37 | Iustin Pop | return |
382 | 5188ab37 | Iustin Pop | offline_disk_instances = result[2] |
383 | 5188ab37 | Iustin Pop | if not offline_disk_instances: |
384 | 5188ab37 | Iustin Pop | # nothing to do |
385 | 5188ab37 | Iustin Pop | return |
386 | 5188ab37 | Iustin Pop | logging.debug("Will activate disks for instances %s", |
387 | 5188ab37 | Iustin Pop | ", ".join(offline_disk_instances)) |
388 | 5188ab37 | Iustin Pop | # we submit only one job, and wait for it. not optimal, but spams |
389 | 5188ab37 | Iustin Pop | # less the job queue |
390 | 5188ab37 | Iustin Pop | job = [opcodes.OpActivateInstanceDisks(instance_name=name) |
391 | 5188ab37 | Iustin Pop | for name in offline_disk_instances] |
392 | 5188ab37 | Iustin Pop | job_id = cli.SendJob(job, cl=client) |
393 | 5188ab37 | Iustin Pop | |
394 | 5188ab37 | Iustin Pop | cli.PollJob(job_id, cl=client, feedback_fn=logging.debug) |
395 | a8083063 | Iustin Pop | |
396 | a8083063 | Iustin Pop | |
397 | a8083063 | Iustin Pop | def ParseOptions(): |
398 | a8083063 | Iustin Pop | """Parse the command line options. |
399 | a8083063 | Iustin Pop | |
400 | c41eea6e | Iustin Pop | @return: (options, args) as from OptionParser.parse_args() |
401 | a8083063 | Iustin Pop | |
402 | a8083063 | Iustin Pop | """ |
403 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher", |
404 | a8083063 | Iustin Pop | usage="%prog [-d]", |
405 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" % |
406 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
407 | a8083063 | Iustin Pop | |
408 | a8083063 | Iustin Pop | parser.add_option("-d", "--debug", dest="debug", |
409 | 438b45d4 | Michael Hanselmann | help="Write all messages to stderr", |
410 | a8083063 | Iustin Pop | default=False, action="store_true") |
411 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
412 | a8083063 | Iustin Pop | return options, args |
413 | a8083063 | Iustin Pop | |
414 | a8083063 | Iustin Pop | |
415 | a8083063 | Iustin Pop | def main(): |
416 | a8083063 | Iustin Pop | """Main function. |
417 | a8083063 | Iustin Pop | |
418 | a8083063 | Iustin Pop | """ |
419 | e125c67c | Michael Hanselmann | global client |
420 | e125c67c | Michael Hanselmann | |
421 | a8083063 | Iustin Pop | options, args = ParseOptions() |
422 | a8083063 | Iustin Pop | |
423 | 82d9caef | Iustin Pop | utils.SetupLogging(constants.LOG_WATCHER, debug=options.debug, |
424 | 82d9caef | Iustin Pop | stderr_logging=options.debug) |
425 | a8083063 | Iustin Pop | |
426 | a8083063 | Iustin Pop | try: |
427 | e125c67c | Michael Hanselmann | client = cli.GetClient() |
428 | e125c67c | Michael Hanselmann | |
429 | 781b2b2b | Michael Hanselmann | try: |
430 | 781b2b2b | Michael Hanselmann | watcher = Watcher() |
431 | 781b2b2b | Michael Hanselmann | except errors.ConfigurationError: |
432 | 781b2b2b | Michael Hanselmann | # Just exit if there's no configuration |
433 | 781b2b2b | Michael Hanselmann | sys.exit(constants.EXIT_SUCCESS) |
434 | e125c67c | Michael Hanselmann | |
435 | 5a3103e9 | Michael Hanselmann | watcher.Run() |
436 | 1b052f42 | Michael Hanselmann | except SystemExit: |
437 | 1b052f42 | Michael Hanselmann | raise |
438 | 38242904 | Iustin Pop | except NotMasterError: |
439 | 438b45d4 | Michael Hanselmann | logging.debug("Not master, exiting") |
440 | 38242904 | Iustin Pop | sys.exit(constants.EXIT_NOTMASTER) |
441 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err: |
442 | 438b45d4 | Michael Hanselmann | logging.error("Cannot resolve hostname '%s', exiting.", err.args[0]) |
443 | 89e1fc26 | Iustin Pop | sys.exit(constants.EXIT_NODESETUP_ERROR) |
444 | 438b45d4 | Michael Hanselmann | except Exception, err: |
445 | 438b45d4 | Michael Hanselmann | logging.error(str(err), exc_info=True) |
446 | 438b45d4 | Michael Hanselmann | sys.exit(constants.EXIT_FAILURE) |
447 | a8083063 | Iustin Pop | |
448 | 5a3103e9 | Michael Hanselmann | |
449 | a8083063 | Iustin Pop | if __name__ == '__main__': |
450 | a8083063 | Iustin Pop | main() |