root / daemons / ganeti-watcher @ 82d9caef
History | View | Annotate | Download (12.4 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 | b7309a0d | Iustin Pop | HELPLESS_STATES = ['ERROR_nodedown'] |
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 | a8083063 | Iustin Pop | Args: |
66 | a8083063 | Iustin Pop | s: The string to indent |
67 | a8083063 | Iustin Pop | prefix: The string to prepend each line. |
68 | 38242904 | Iustin Pop | |
69 | a8083063 | Iustin Pop | """ |
70 | a8083063 | Iustin Pop | return "%s%s\n" % (prefix, ('\n' + prefix).join(s.splitlines())) |
71 | a8083063 | Iustin Pop | |
72 | a8083063 | Iustin Pop | |
73 | 5a3103e9 | Michael Hanselmann | class WatcherState(object): |
74 | a8083063 | Iustin Pop | """Interface to a state file recording restart attempts. |
75 | a8083063 | Iustin Pop | |
76 | a8083063 | Iustin Pop | """ |
77 | a8083063 | Iustin Pop | def __init__(self): |
78 | 5a3103e9 | Michael Hanselmann | """Open, lock, read and parse the file. |
79 | 5a3103e9 | Michael Hanselmann | |
80 | 7bca53e4 | Michael Hanselmann | Raises exception on lock contention. |
81 | 5a3103e9 | Michael Hanselmann | |
82 | 5a3103e9 | Michael Hanselmann | """ |
83 | a8083063 | Iustin Pop | # The two-step dance below is necessary to allow both opening existing |
84 | a8083063 | Iustin Pop | # file read/write and creating if not existing. Vanilla open will truncate |
85 | a8083063 | Iustin Pop | # an existing file -or- allow creating if not existing. |
86 | fc428e32 | Michael Hanselmann | fd = os.open(constants.WATCHER_STATEFILE, os.O_RDWR | os.O_CREAT) |
87 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
88 | a8083063 | Iustin Pop | |
89 | eb0f0ce0 | Michael Hanselmann | utils.LockFile(self.statefile.fileno()) |
90 | a8083063 | Iustin Pop | |
91 | 5a3103e9 | Michael Hanselmann | try: |
92 | b76f660d | Michael Hanselmann | self._data = serializer.Load(self.statefile.read()) |
93 | 5a3103e9 | Michael Hanselmann | except Exception, msg: |
94 | 5a3103e9 | Michael Hanselmann | # Ignore errors while loading the file and treat it as empty |
95 | b76f660d | Michael Hanselmann | self._data = {} |
96 | 438b45d4 | Michael Hanselmann | logging.warning(("Empty or invalid state file. Using defaults." |
97 | 438b45d4 | Michael Hanselmann | " Error message: %s"), msg) |
98 | 5a3103e9 | Michael Hanselmann | |
99 | b76f660d | Michael Hanselmann | if "instance" not in self._data: |
100 | b76f660d | Michael Hanselmann | self._data["instance"] = {} |
101 | b76f660d | Michael Hanselmann | if "node" not in self._data: |
102 | b76f660d | Michael Hanselmann | self._data["node"] = {} |
103 | 5a3103e9 | Michael Hanselmann | |
104 | 26517d45 | Iustin Pop | self._orig_data = serializer.Dump(self._data) |
105 | 2fb96d39 | Michael Hanselmann | |
106 | fc428e32 | Michael Hanselmann | def Save(self): |
107 | fc428e32 | Michael Hanselmann | """Save state to file, then unlock and close it. |
108 | 5a3103e9 | Michael Hanselmann | |
109 | 5a3103e9 | Michael Hanselmann | """ |
110 | fc428e32 | Michael Hanselmann | assert self.statefile |
111 | fc428e32 | Michael Hanselmann | |
112 | 26517d45 | Iustin Pop | serialized_form = serializer.Dump(self._data) |
113 | 26517d45 | Iustin Pop | if self._orig_data == serialized_form: |
114 | 2fb96d39 | Michael Hanselmann | logging.debug("Data didn't change, just touching status file") |
115 | 2fb96d39 | Michael Hanselmann | os.utime(constants.WATCHER_STATEFILE, None) |
116 | 2fb96d39 | Michael Hanselmann | return |
117 | 2fb96d39 | Michael Hanselmann | |
118 | fc428e32 | Michael Hanselmann | # We need to make sure the file is locked before renaming it, otherwise |
119 | fc428e32 | Michael Hanselmann | # starting ganeti-watcher again at the same time will create a conflict. |
120 | fc428e32 | Michael Hanselmann | fd = utils.WriteFile(constants.WATCHER_STATEFILE, |
121 | 26517d45 | Iustin Pop | data=serialized_form, |
122 | eb0f0ce0 | Michael Hanselmann | prewrite=utils.LockFile, close=False) |
123 | fc428e32 | Michael Hanselmann | self.statefile = os.fdopen(fd, 'w+') |
124 | 5a3103e9 | Michael Hanselmann | |
125 | fc428e32 | Michael Hanselmann | def Close(self): |
126 | 5a3103e9 | Michael Hanselmann | """Unlock configuration file and close it. |
127 | 5a3103e9 | Michael Hanselmann | |
128 | 5a3103e9 | Michael Hanselmann | """ |
129 | 5a3103e9 | Michael Hanselmann | assert self.statefile |
130 | 5a3103e9 | Michael Hanselmann | |
131 | fc428e32 | Michael Hanselmann | # Files are automatically unlocked when closing them |
132 | 5a3103e9 | Michael Hanselmann | self.statefile.close() |
133 | 5a3103e9 | Michael Hanselmann | self.statefile = None |
134 | 5a3103e9 | Michael Hanselmann | |
135 | 5a3103e9 | Michael Hanselmann | def GetNodeBootID(self, name): |
136 | 5a3103e9 | Michael Hanselmann | """Returns the last boot ID of a node or None. |
137 | a8083063 | Iustin Pop | |
138 | 5a3103e9 | Michael Hanselmann | """ |
139 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
140 | 5a3103e9 | Michael Hanselmann | |
141 | 7b195d9b | Michael Hanselmann | if name in ndata and KEY_BOOT_ID in ndata[name]: |
142 | 7b195d9b | Michael Hanselmann | return ndata[name][KEY_BOOT_ID] |
143 | 5a3103e9 | Michael Hanselmann | return None |
144 | 5a3103e9 | Michael Hanselmann | |
145 | 5a3103e9 | Michael Hanselmann | def SetNodeBootID(self, name, bootid): |
146 | 5a3103e9 | Michael Hanselmann | """Sets the boot ID of a node. |
147 | 5a3103e9 | Michael Hanselmann | |
148 | 5a3103e9 | Michael Hanselmann | """ |
149 | 5a3103e9 | Michael Hanselmann | assert bootid |
150 | a8083063 | Iustin Pop | |
151 | b76f660d | Michael Hanselmann | ndata = self._data["node"] |
152 | a8083063 | Iustin Pop | |
153 | 5a3103e9 | Michael Hanselmann | if name not in ndata: |
154 | 5a3103e9 | Michael Hanselmann | ndata[name] = {} |
155 | 5a3103e9 | Michael Hanselmann | |
156 | 7b195d9b | Michael Hanselmann | ndata[name][KEY_BOOT_ID] = bootid |
157 | 5a3103e9 | Michael Hanselmann | |
158 | 5a3103e9 | Michael Hanselmann | def NumberOfRestartAttempts(self, instance): |
159 | a8083063 | Iustin Pop | """Returns number of previous restart attempts. |
160 | a8083063 | Iustin Pop | |
161 | a8083063 | Iustin Pop | Args: |
162 | a8083063 | Iustin Pop | instance - the instance to look up. |
163 | 38242904 | Iustin Pop | |
164 | a8083063 | Iustin Pop | """ |
165 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
166 | a8083063 | Iustin Pop | |
167 | 5a3103e9 | Michael Hanselmann | if instance.name in idata: |
168 | 7b195d9b | Michael Hanselmann | return idata[instance.name][KEY_RESTART_COUNT] |
169 | a8083063 | Iustin Pop | |
170 | a8083063 | Iustin Pop | return 0 |
171 | a8083063 | Iustin Pop | |
172 | 5a3103e9 | Michael Hanselmann | def RecordRestartAttempt(self, instance): |
173 | a8083063 | Iustin Pop | """Record a restart attempt. |
174 | a8083063 | Iustin Pop | |
175 | a8083063 | Iustin Pop | Args: |
176 | a8083063 | Iustin Pop | instance - the instance being restarted |
177 | 38242904 | Iustin Pop | |
178 | a8083063 | Iustin Pop | """ |
179 | b76f660d | Michael Hanselmann | idata = self._data["instance"] |
180 | a8083063 | Iustin Pop | |
181 | 5a3103e9 | Michael Hanselmann | if instance.name not in idata: |
182 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] = {} |
183 | 5a3103e9 | Michael Hanselmann | else: |
184 | 5a3103e9 | Michael Hanselmann | inst = idata[instance.name] |
185 | a8083063 | Iustin Pop | |
186 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_WHEN] = time.time() |
187 | 7b195d9b | Michael Hanselmann | inst[KEY_RESTART_COUNT] = inst.get(KEY_RESTART_COUNT, 0) + 1 |
188 | a8083063 | Iustin Pop | |
189 | 5a3103e9 | Michael Hanselmann | def RemoveInstance(self, instance): |
190 | 38242904 | Iustin Pop | """Update state to reflect that a machine is running, i.e. remove record. |
191 | a8083063 | Iustin Pop | |
192 | a8083063 | Iustin Pop | Args: |
193 | a8083063 | Iustin Pop | instance - the instance to remove from books |
194 | a8083063 | Iustin Pop | |
195 | 38242904 | Iustin Pop | This method removes the record for a named instance. |
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 | Methods: |
208 | a8083063 | Iustin Pop | Restart(): issue a command to restart the represented machine. |
209 | 098c0958 | Michael Hanselmann | |
210 | a8083063 | Iustin Pop | """ |
211 | 5a3103e9 | Michael Hanselmann | def __init__(self, name, state, autostart): |
212 | a8083063 | Iustin Pop | self.name = name |
213 | a8083063 | Iustin Pop | self.state = state |
214 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart |
215 | a8083063 | Iustin Pop | |
216 | a8083063 | Iustin Pop | def Restart(self): |
217 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance. |
218 | 3ecf6786 | Iustin Pop | |
219 | 3ecf6786 | Iustin Pop | """ |
220 | e125c67c | Michael Hanselmann | op = opcodes.OpStartupInstance(instance_name=self.name, |
221 | e125c67c | Michael Hanselmann | force=False, |
222 | e125c67c | Michael Hanselmann | extra_args=None) |
223 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
224 | a8083063 | Iustin Pop | |
225 | 5a3103e9 | Michael Hanselmann | def ActivateDisks(self): |
226 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance. |
227 | 5a3103e9 | Michael Hanselmann | |
228 | 5a3103e9 | Michael Hanselmann | """ |
229 | e125c67c | Michael Hanselmann | op = opcodes.OpActivateInstanceDisks(instance_name=self.name) |
230 | e125c67c | Michael Hanselmann | cli.SubmitOpCode(op, cl=client) |
231 | a8083063 | Iustin Pop | |
232 | a8083063 | Iustin Pop | |
233 | 5a3103e9 | Michael Hanselmann | def GetInstanceList(with_secondaries=None): |
234 | 5a3103e9 | Michael Hanselmann | """Get a list of instances on this cluster. |
235 | 5a3103e9 | Michael Hanselmann | |
236 | 5a3103e9 | Michael Hanselmann | """ |
237 | b7309a0d | Iustin Pop | fields = ["name", "status", "admin_state"] |
238 | a8083063 | Iustin Pop | |
239 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
240 | e125c67c | Michael Hanselmann | fields.append("snodes") |
241 | a8083063 | Iustin Pop | |
242 | e125c67c | Michael Hanselmann | result = client.QueryInstances([], fields) |
243 | 5a3103e9 | Michael Hanselmann | |
244 | 5a3103e9 | Michael Hanselmann | instances = [] |
245 | e125c67c | Michael Hanselmann | for fields in result: |
246 | 5a3103e9 | Michael Hanselmann | if with_secondaries is not None: |
247 | 5a3103e9 | Michael Hanselmann | (name, status, autostart, snodes) = fields |
248 | 5a3103e9 | Michael Hanselmann | |
249 | e125c67c | Michael Hanselmann | if not snodes: |
250 | a8083063 | Iustin Pop | continue |
251 | 5a3103e9 | Michael Hanselmann | |
252 | 5a3103e9 | Michael Hanselmann | for node in with_secondaries: |
253 | e125c67c | Michael Hanselmann | if node in snodes: |
254 | 5a3103e9 | Michael Hanselmann | break |
255 | 5a3103e9 | Michael Hanselmann | else: |
256 | a8083063 | Iustin Pop | continue |
257 | a8083063 | Iustin Pop | |
258 | 5a3103e9 | Michael Hanselmann | else: |
259 | 5a3103e9 | Michael Hanselmann | (name, status, autostart) = fields |
260 | 5a3103e9 | Michael Hanselmann | |
261 | e125c67c | Michael Hanselmann | instances.append(Instance(name, status, autostart)) |
262 | a8083063 | Iustin Pop | |
263 | 5a3103e9 | Michael Hanselmann | return instances |
264 | 5a3103e9 | Michael Hanselmann | |
265 | 5a3103e9 | Michael Hanselmann | |
266 | 5a3103e9 | Michael Hanselmann | def GetNodeBootIDs(): |
267 | 5a3103e9 | Michael Hanselmann | """Get a dict mapping nodes to boot IDs. |
268 | 5a3103e9 | Michael Hanselmann | |
269 | 5a3103e9 | Michael Hanselmann | """ |
270 | e125c67c | Michael Hanselmann | result = client.QueryNodes([], ["name", "bootid"]) |
271 | e125c67c | Michael Hanselmann | return dict([(name, bootid) for name, bootid in result]) |
272 | a8083063 | Iustin Pop | |
273 | a8083063 | Iustin Pop | |
274 | 5a3103e9 | Michael Hanselmann | class Watcher(object): |
275 | a8083063 | Iustin Pop | """Encapsulate the logic for restarting erronously halted virtual machines. |
276 | a8083063 | Iustin Pop | |
277 | a8083063 | Iustin Pop | The calling program should periodically instantiate me and call Run(). |
278 | a8083063 | Iustin Pop | This will traverse the list of instances, and make up to MAXTRIES attempts |
279 | a8083063 | Iustin Pop | to restart machines that are down. |
280 | 38242904 | Iustin Pop | |
281 | a8083063 | Iustin Pop | """ |
282 | a8083063 | Iustin Pop | def __init__(self): |
283 | 2859b87b | Michael Hanselmann | master = client.QueryConfigValues(["master_node"])[0] |
284 | 89e1fc26 | Iustin Pop | if master != utils.HostInfo().name: |
285 | 3ecf6786 | Iustin Pop | raise NotMasterError("This is not the master node") |
286 | 5a3103e9 | Michael Hanselmann | self.instances = GetInstanceList() |
287 | 5a3103e9 | Michael Hanselmann | self.bootids = GetNodeBootIDs() |
288 | eee1fa2d | Iustin Pop | self.started_instances = set() |
289 | a8083063 | Iustin Pop | |
290 | a8083063 | Iustin Pop | def Run(self): |
291 | 5a3103e9 | Michael Hanselmann | notepad = WatcherState() |
292 | 78f3bd30 | Michael Hanselmann | try: |
293 | 78f3bd30 | Michael Hanselmann | self.CheckInstances(notepad) |
294 | 78f3bd30 | Michael Hanselmann | self.CheckDisks(notepad) |
295 | 78f3bd30 | Michael Hanselmann | self.VerifyDisks() |
296 | 78f3bd30 | Michael Hanselmann | finally: |
297 | 78f3bd30 | Michael Hanselmann | notepad.Save() |
298 | 5a3103e9 | Michael Hanselmann | |
299 | 5a3103e9 | Michael Hanselmann | def CheckDisks(self, notepad): |
300 | 5a3103e9 | Michael Hanselmann | """Check all nodes for restarted ones. |
301 | 38242904 | Iustin Pop | |
302 | a8083063 | Iustin Pop | """ |
303 | 5a3103e9 | Michael Hanselmann | check_nodes = [] |
304 | 26517d45 | Iustin Pop | for name, new_id in self.bootids.iteritems(): |
305 | 5a3103e9 | Michael Hanselmann | old = notepad.GetNodeBootID(name) |
306 | 37b77b18 | Iustin Pop | if new_id is None: |
307 | 37b77b18 | Iustin Pop | # Bad node, not returning a boot id |
308 | 37b77b18 | Iustin Pop | logging.debug("Node %s missing boot id, skipping secondary checks", |
309 | 37b77b18 | Iustin Pop | name) |
310 | 37b77b18 | Iustin Pop | continue |
311 | 26517d45 | Iustin Pop | if old != new_id: |
312 | 5a3103e9 | Michael Hanselmann | # Node's boot ID has changed, proably through a reboot. |
313 | 5a3103e9 | Michael Hanselmann | check_nodes.append(name) |
314 | 5a3103e9 | Michael Hanselmann | |
315 | 5a3103e9 | Michael Hanselmann | if check_nodes: |
316 | 5a3103e9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a |
317 | 5a3103e9 | Michael Hanselmann | # secondary node. |
318 | 5a3103e9 | Michael Hanselmann | for instance in GetInstanceList(with_secondaries=check_nodes): |
319 | 0c0f834d | Iustin Pop | if not instance.autostart: |
320 | 438b45d4 | Michael Hanselmann | logging.info(("Skipping disk activation for non-autostart" |
321 | 438b45d4 | Michael Hanselmann | " instance %s"), instance.name) |
322 | 0c0f834d | Iustin Pop | continue |
323 | eee1fa2d | Iustin Pop | if instance.name in self.started_instances: |
324 | eee1fa2d | Iustin Pop | # we already tried to start the instance, which should have |
325 | eee1fa2d | Iustin Pop | # activated its drives (if they can be at all) |
326 | eee1fa2d | Iustin Pop | continue |
327 | 5a3103e9 | Michael Hanselmann | try: |
328 | 438b45d4 | Michael Hanselmann | logging.info("Activating disks for instance %s", instance.name) |
329 | 5a3103e9 | Michael Hanselmann | instance.ActivateDisks() |
330 | b7309a0d | Iustin Pop | except Exception: |
331 | b7309a0d | Iustin Pop | logging.exception("Error while activating disks for instance %s", |
332 | b7309a0d | Iustin Pop | instance.name) |
333 | 5a3103e9 | Michael Hanselmann | |
334 | 5a3103e9 | Michael Hanselmann | # Keep changed boot IDs |
335 | 5a3103e9 | Michael Hanselmann | for name in check_nodes: |
336 | 5a3103e9 | Michael Hanselmann | notepad.SetNodeBootID(name, self.bootids[name]) |
337 | a8083063 | Iustin Pop | |
338 | 5a3103e9 | Michael Hanselmann | def CheckInstances(self, notepad): |
339 | 5a3103e9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones. |
340 | 5a3103e9 | Michael Hanselmann | |
341 | 5a3103e9 | Michael Hanselmann | """ |
342 | a8083063 | Iustin Pop | for instance in self.instances: |
343 | a8083063 | Iustin Pop | if instance.state in BAD_STATES: |
344 | 5a3103e9 | Michael Hanselmann | n = notepad.NumberOfRestartAttempts(instance) |
345 | a8083063 | Iustin Pop | |
346 | a8083063 | Iustin Pop | if n > MAXTRIES: |
347 | a8083063 | Iustin Pop | # stay quiet. |
348 | a8083063 | Iustin Pop | continue |
349 | a8083063 | Iustin Pop | elif n < MAXTRIES: |
350 | a8083063 | Iustin Pop | last = " (Attempt #%d)" % (n + 1) |
351 | a8083063 | Iustin Pop | else: |
352 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
353 | 438b45d4 | Michael Hanselmann | logging.error("Could not restart %s after %d attempts, giving up", |
354 | 438b45d4 | Michael Hanselmann | instance.name, MAXTRIES) |
355 | a8083063 | Iustin Pop | continue |
356 | a8083063 | Iustin Pop | try: |
357 | 438b45d4 | Michael Hanselmann | logging.info("Restarting %s%s", |
358 | 438b45d4 | Michael Hanselmann | instance.name, last) |
359 | a8083063 | Iustin Pop | instance.Restart() |
360 | eee1fa2d | Iustin Pop | self.started_instances.add(instance.name) |
361 | b7309a0d | Iustin Pop | except Exception: |
362 | b7309a0d | Iustin Pop | logging.exception("Erro while restarting instance %s", instance.name) |
363 | a8083063 | Iustin Pop | |
364 | 5a3103e9 | Michael Hanselmann | notepad.RecordRestartAttempt(instance) |
365 | a8083063 | Iustin Pop | elif instance.state in HELPLESS_STATES: |
366 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
367 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
368 | a8083063 | Iustin Pop | else: |
369 | 5a3103e9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(instance): |
370 | 5a3103e9 | Michael Hanselmann | notepad.RemoveInstance(instance) |
371 | 438b45d4 | Michael Hanselmann | logging.info("Restart of %s succeeded", instance.name) |
372 | a8083063 | Iustin Pop | |
373 | b7309a0d | Iustin Pop | @staticmethod |
374 | b7309a0d | Iustin Pop | def VerifyDisks(): |
375 | d2f311db | Iustin Pop | """Run gnt-cluster verify-disks. |
376 | d2f311db | Iustin Pop | |
377 | d2f311db | Iustin Pop | """ |
378 | 5188ab37 | Iustin Pop | op = opcodes.OpVerifyDisks() |
379 | 5188ab37 | Iustin Pop | result = cli.SubmitOpCode(op, cl=client) |
380 | 5188ab37 | Iustin Pop | if not isinstance(result, (tuple, list)): |
381 | 5188ab37 | Iustin Pop | logging.error("Can't get a valid result from verify-disks") |
382 | 5188ab37 | Iustin Pop | return |
383 | 5188ab37 | Iustin Pop | offline_disk_instances = result[2] |
384 | 5188ab37 | Iustin Pop | if not offline_disk_instances: |
385 | 5188ab37 | Iustin Pop | # nothing to do |
386 | 5188ab37 | Iustin Pop | return |
387 | 5188ab37 | Iustin Pop | logging.debug("Will activate disks for instances %s", |
388 | 5188ab37 | Iustin Pop | ", ".join(offline_disk_instances)) |
389 | 5188ab37 | Iustin Pop | # we submit only one job, and wait for it. not optimal, but spams |
390 | 5188ab37 | Iustin Pop | # less the job queue |
391 | 5188ab37 | Iustin Pop | job = [opcodes.OpActivateInstanceDisks(instance_name=name) |
392 | 5188ab37 | Iustin Pop | for name in offline_disk_instances] |
393 | 5188ab37 | Iustin Pop | job_id = cli.SendJob(job, cl=client) |
394 | 5188ab37 | Iustin Pop | |
395 | 5188ab37 | Iustin Pop | cli.PollJob(job_id, cl=client, feedback_fn=logging.debug) |
396 | a8083063 | Iustin Pop | |
397 | a8083063 | Iustin Pop | |
398 | a8083063 | Iustin Pop | def ParseOptions(): |
399 | a8083063 | Iustin Pop | """Parse the command line options. |
400 | a8083063 | Iustin Pop | |
401 | a8083063 | Iustin Pop | Returns: |
402 | a8083063 | Iustin Pop | (options, args) as from OptionParser.parse_args() |
403 | a8083063 | Iustin Pop | |
404 | a8083063 | Iustin Pop | """ |
405 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher", |
406 | a8083063 | Iustin Pop | usage="%prog [-d]", |
407 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" % |
408 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
409 | a8083063 | Iustin Pop | |
410 | a8083063 | Iustin Pop | parser.add_option("-d", "--debug", dest="debug", |
411 | 438b45d4 | Michael Hanselmann | help="Write all messages to stderr", |
412 | a8083063 | Iustin Pop | default=False, action="store_true") |
413 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
414 | a8083063 | Iustin Pop | return options, args |
415 | a8083063 | Iustin Pop | |
416 | a8083063 | Iustin Pop | |
417 | a8083063 | Iustin Pop | def main(): |
418 | a8083063 | Iustin Pop | """Main function. |
419 | a8083063 | Iustin Pop | |
420 | a8083063 | Iustin Pop | """ |
421 | e125c67c | Michael Hanselmann | global client |
422 | e125c67c | Michael Hanselmann | |
423 | a8083063 | Iustin Pop | options, args = ParseOptions() |
424 | a8083063 | Iustin Pop | |
425 | 82d9caef | Iustin Pop | utils.SetupLogging(constants.LOG_WATCHER, debug=options.debug, |
426 | 82d9caef | Iustin Pop | stderr_logging=options.debug) |
427 | a8083063 | Iustin Pop | |
428 | a8083063 | Iustin Pop | try: |
429 | e125c67c | Michael Hanselmann | client = cli.GetClient() |
430 | e125c67c | Michael Hanselmann | |
431 | 781b2b2b | Michael Hanselmann | try: |
432 | 781b2b2b | Michael Hanselmann | watcher = Watcher() |
433 | 781b2b2b | Michael Hanselmann | except errors.ConfigurationError: |
434 | 781b2b2b | Michael Hanselmann | # Just exit if there's no configuration |
435 | 781b2b2b | Michael Hanselmann | sys.exit(constants.EXIT_SUCCESS) |
436 | e125c67c | Michael Hanselmann | |
437 | 5a3103e9 | Michael Hanselmann | watcher.Run() |
438 | 1b052f42 | Michael Hanselmann | except SystemExit: |
439 | 1b052f42 | Michael Hanselmann | raise |
440 | 38242904 | Iustin Pop | except NotMasterError: |
441 | 438b45d4 | Michael Hanselmann | logging.debug("Not master, exiting") |
442 | 38242904 | Iustin Pop | sys.exit(constants.EXIT_NOTMASTER) |
443 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err: |
444 | 438b45d4 | Michael Hanselmann | logging.error("Cannot resolve hostname '%s', exiting.", err.args[0]) |
445 | 89e1fc26 | Iustin Pop | sys.exit(constants.EXIT_NODESETUP_ERROR) |
446 | 438b45d4 | Michael Hanselmann | except Exception, err: |
447 | 438b45d4 | Michael Hanselmann | logging.error(str(err), exc_info=True) |
448 | 438b45d4 | Michael Hanselmann | sys.exit(constants.EXIT_FAILURE) |
449 | a8083063 | Iustin Pop | |
450 | 5a3103e9 | Michael Hanselmann | |
451 | a8083063 | Iustin Pop | if __name__ == '__main__': |
452 | a8083063 | Iustin Pop | main() |