root / lib / watcher / __init__.py @ a24aed2a
History | View | Annotate | Download (24.1 kB)
1 | 9f4bb951 | Michael Hanselmann | #
|
---|---|---|---|
2 | a8083063 | Iustin Pop | #
|
3 | a8083063 | Iustin Pop | |
4 | f2af0bec | Iustin Pop | # Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 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 | a8083063 | Iustin Pop | import os |
31 | cfcc79c6 | Michael Hanselmann | import os.path |
32 | a8083063 | Iustin Pop | import sys |
33 | a8083063 | Iustin Pop | import time |
34 | 438b45d4 | Michael Hanselmann | import logging |
35 | 16e0b9c9 | Michael Hanselmann | import operator |
36 | 9bb69bb5 | Michael Hanselmann | import errno |
37 | a8083063 | Iustin Pop | from optparse import OptionParser |
38 | a8083063 | Iustin Pop | |
39 | a8083063 | Iustin Pop | from ganeti import utils |
40 | a8083063 | Iustin Pop | from ganeti import constants |
41 | 83e5e26f | René Nussbaumer | from ganeti import compat |
42 | 89e1fc26 | Iustin Pop | from ganeti import errors |
43 | e125c67c | Michael Hanselmann | from ganeti import opcodes |
44 | e125c67c | Michael Hanselmann | from ganeti import cli |
45 | 7dfb83c2 | Iustin Pop | from ganeti import luxi |
46 | db147305 | Tom Limoncelli | from ganeti import rapi |
47 | a744b676 | Manuel Franceschini | from ganeti import netutils |
48 | 16e0b9c9 | Michael Hanselmann | from ganeti import qlang |
49 | 16e0b9c9 | Michael Hanselmann | from ganeti import objects |
50 | 16e0b9c9 | Michael Hanselmann | from ganeti import ssconf |
51 | 16e0b9c9 | Michael Hanselmann | from ganeti import ht |
52 | a8083063 | Iustin Pop | |
53 | b459a848 | Andrea Spadaccini | import ganeti.rapi.client # pylint: disable=W0611 |
54 | adf6301e | Michael Hanselmann | |
55 | adf6301e | Michael Hanselmann | from ganeti.watcher import nodemaint |
56 | adf6301e | Michael Hanselmann | from ganeti.watcher import state |
57 | db147305 | Tom Limoncelli | |
58 | a8083063 | Iustin Pop | |
59 | 5a3103e9 | Michael Hanselmann | MAXTRIES = 5
|
60 | 0cc9e018 | Michael Hanselmann | BAD_STATES = frozenset([
|
61 | 0cc9e018 | Michael Hanselmann | constants.INSTST_ERRORDOWN, |
62 | 0cc9e018 | Michael Hanselmann | ]) |
63 | 0cc9e018 | Michael Hanselmann | HELPLESS_STATES = frozenset([
|
64 | 0cc9e018 | Michael Hanselmann | constants.INSTST_NODEDOWN, |
65 | 0cc9e018 | Michael Hanselmann | constants.INSTST_NODEOFFLINE, |
66 | 0cc9e018 | Michael Hanselmann | ]) |
67 | 0cc9e018 | Michael Hanselmann | NOTICE = "NOTICE"
|
68 | 0cc9e018 | Michael Hanselmann | ERROR = "ERROR"
|
69 | e125c67c | Michael Hanselmann | |
70 | 16e0b9c9 | Michael Hanselmann | #: Number of seconds to wait between starting child processes for node groups
|
71 | 16e0b9c9 | Michael Hanselmann | CHILD_PROCESS_DELAY = 1.0
|
72 | 16e0b9c9 | Michael Hanselmann | |
73 | 9bb69bb5 | Michael Hanselmann | #: How many seconds to wait for instance status file lock
|
74 | 9bb69bb5 | Michael Hanselmann | INSTANCE_STATUS_LOCK_TIMEOUT = 10.0
|
75 | 9bb69bb5 | Michael Hanselmann | |
76 | e125c67c | Michael Hanselmann | |
77 | 7bca53e4 | Michael Hanselmann | class NotMasterError(errors.GenericError): |
78 | 38242904 | Iustin Pop | """Exception raised when this host is not the master."""
|
79 | a8083063 | Iustin Pop | |
80 | a8083063 | Iustin Pop | |
81 | 3753b2cb | Michael Hanselmann | def ShouldPause(): |
82 | 3753b2cb | Michael Hanselmann | """Check whether we should pause.
|
83 | 3753b2cb | Michael Hanselmann |
|
84 | 3753b2cb | Michael Hanselmann | """
|
85 | 3753b2cb | Michael Hanselmann | return bool(utils.ReadWatcherPauseFile(constants.WATCHER_PAUSEFILE)) |
86 | 3753b2cb | Michael Hanselmann | |
87 | 3753b2cb | Michael Hanselmann | |
88 | f1115454 | Guido Trotter | def StartNodeDaemons(): |
89 | f1115454 | Guido Trotter | """Start all the daemons that should be running on all nodes.
|
90 | f1115454 | Guido Trotter |
|
91 | f1115454 | Guido Trotter | """
|
92 | 55c85950 | Iustin Pop | # on master or not, try to start the node daemon
|
93 | 2826b361 | Guido Trotter | utils.EnsureDaemon(constants.NODED) |
94 | f1115454 | Guido Trotter | # start confd as well. On non candidates it will be in disabled mode.
|
95 | aa224134 | Iustin Pop | if constants.ENABLE_CONFD:
|
96 | aa224134 | Iustin Pop | utils.EnsureDaemon(constants.CONFD) |
97 | f1115454 | Guido Trotter | |
98 | f1115454 | Guido Trotter | |
99 | 9e289e36 | Guido Trotter | def RunWatcherHooks(): |
100 | 9e289e36 | Guido Trotter | """Run the watcher hooks.
|
101 | 9e289e36 | Guido Trotter |
|
102 | 9e289e36 | Guido Trotter | """
|
103 | c4feafe8 | Iustin Pop | hooks_dir = utils.PathJoin(constants.HOOKS_BASE_DIR, |
104 | c4feafe8 | Iustin Pop | constants.HOOKS_NAME_WATCHER) |
105 | 10e689d4 | Iustin Pop | if not os.path.isdir(hooks_dir): |
106 | 10e689d4 | Iustin Pop | return
|
107 | 9e289e36 | Guido Trotter | |
108 | 9e289e36 | Guido Trotter | try:
|
109 | 9e289e36 | Guido Trotter | results = utils.RunParts(hooks_dir) |
110 | 17385bd2 | Andrea Spadaccini | except Exception, err: # pylint: disable=W0703 |
111 | 17385bd2 | Andrea Spadaccini | logging.exception("RunParts %s failed: %s", hooks_dir, err)
|
112 | a0aa6b49 | Michael Hanselmann | return
|
113 | 9e289e36 | Guido Trotter | |
114 | 9e289e36 | Guido Trotter | for (relname, status, runresult) in results: |
115 | 9e289e36 | Guido Trotter | if status == constants.RUNPARTS_SKIP:
|
116 | 9e289e36 | Guido Trotter | logging.debug("Watcher hook %s: skipped", relname)
|
117 | 9e289e36 | Guido Trotter | elif status == constants.RUNPARTS_ERR:
|
118 | 9e289e36 | Guido Trotter | logging.warning("Watcher hook %s: error (%s)", relname, runresult)
|
119 | 9e289e36 | Guido Trotter | elif status == constants.RUNPARTS_RUN:
|
120 | 9e289e36 | Guido Trotter | if runresult.failed:
|
121 | 9e289e36 | Guido Trotter | logging.warning("Watcher hook %s: failed (exit: %d) (output: %s)",
|
122 | 9e289e36 | Guido Trotter | relname, runresult.exit_code, runresult.output) |
123 | 9e289e36 | Guido Trotter | else:
|
124 | 9e289e36 | Guido Trotter | logging.debug("Watcher hook %s: success (output: %s)", relname,
|
125 | 9e289e36 | Guido Trotter | runresult.output) |
126 | 013ce4ae | Michael Hanselmann | else:
|
127 | 013ce4ae | Michael Hanselmann | raise errors.ProgrammerError("Unknown status %s returned by RunParts", |
128 | 013ce4ae | Michael Hanselmann | status) |
129 | 9e289e36 | Guido Trotter | |
130 | 001b3825 | Michael Hanselmann | |
131 | a8083063 | Iustin Pop | class Instance(object): |
132 | a8083063 | Iustin Pop | """Abstraction for a Virtual Machine instance.
|
133 | a8083063 | Iustin Pop |
|
134 | a8083063 | Iustin Pop | """
|
135 | adf6301e | Michael Hanselmann | def __init__(self, name, status, autostart, snodes): |
136 | a8083063 | Iustin Pop | self.name = name
|
137 | adf6301e | Michael Hanselmann | self.status = status
|
138 | 5a3103e9 | Michael Hanselmann | self.autostart = autostart
|
139 | 83e5e26f | René Nussbaumer | self.snodes = snodes
|
140 | a8083063 | Iustin Pop | |
141 | 16e0b9c9 | Michael Hanselmann | def Restart(self, cl): |
142 | 3ecf6786 | Iustin Pop | """Encapsulates the start of an instance.
|
143 | 3ecf6786 | Iustin Pop |
|
144 | 3ecf6786 | Iustin Pop | """
|
145 | c873d91c | Iustin Pop | op = opcodes.OpInstanceStartup(instance_name=self.name, force=False) |
146 | 16e0b9c9 | Michael Hanselmann | cli.SubmitOpCode(op, cl=cl) |
147 | a8083063 | Iustin Pop | |
148 | 16e0b9c9 | Michael Hanselmann | def ActivateDisks(self, cl): |
149 | 5a3103e9 | Michael Hanselmann | """Encapsulates the activation of all disks of an instance.
|
150 | 5a3103e9 | Michael Hanselmann |
|
151 | 5a3103e9 | Michael Hanselmann | """
|
152 | 83f5d475 | Iustin Pop | op = opcodes.OpInstanceActivateDisks(instance_name=self.name)
|
153 | 16e0b9c9 | Michael Hanselmann | cli.SubmitOpCode(op, cl=cl) |
154 | a8083063 | Iustin Pop | |
155 | a8083063 | Iustin Pop | |
156 | 16e0b9c9 | Michael Hanselmann | class Node: |
157 | 16e0b9c9 | Michael Hanselmann | """Data container representing cluster node.
|
158 | 5a3103e9 | Michael Hanselmann |
|
159 | 5a3103e9 | Michael Hanselmann | """
|
160 | 16e0b9c9 | Michael Hanselmann | def __init__(self, name, bootid, offline, secondaries): |
161 | 16e0b9c9 | Michael Hanselmann | """Initializes this class.
|
162 | a8083063 | Iustin Pop |
|
163 | 16e0b9c9 | Michael Hanselmann | """
|
164 | 16e0b9c9 | Michael Hanselmann | self.name = name
|
165 | 16e0b9c9 | Michael Hanselmann | self.bootid = bootid
|
166 | 16e0b9c9 | Michael Hanselmann | self.offline = offline
|
167 | 16e0b9c9 | Michael Hanselmann | self.secondaries = secondaries
|
168 | 5a3103e9 | Michael Hanselmann | |
169 | 78f44650 | Iustin Pop | |
170 | 16e0b9c9 | Michael Hanselmann | def _CheckInstances(cl, notepad, instances): |
171 | 16e0b9c9 | Michael Hanselmann | """Make a pass over the list of instances, restarting downed ones.
|
172 | 5a3103e9 | Michael Hanselmann |
|
173 | 16e0b9c9 | Michael Hanselmann | """
|
174 | 16e0b9c9 | Michael Hanselmann | notepad.MaintainInstanceList(instances.keys()) |
175 | 78f44650 | Iustin Pop | |
176 | 16e0b9c9 | Michael Hanselmann | started = set()
|
177 | 78f44650 | Iustin Pop | |
178 | 16e0b9c9 | Michael Hanselmann | for inst in instances.values(): |
179 | 16e0b9c9 | Michael Hanselmann | if inst.status in BAD_STATES: |
180 | 16e0b9c9 | Michael Hanselmann | n = notepad.NumberOfRestartAttempts(inst.name) |
181 | 5a3103e9 | Michael Hanselmann | |
182 | 16e0b9c9 | Michael Hanselmann | if n > MAXTRIES:
|
183 | 16e0b9c9 | Michael Hanselmann | logging.warning("Not restarting instance '%s', retries exhausted",
|
184 | 16e0b9c9 | Michael Hanselmann | inst.name) |
185 | 16e0b9c9 | Michael Hanselmann | continue
|
186 | a8083063 | Iustin Pop | |
187 | 16e0b9c9 | Michael Hanselmann | if n == MAXTRIES:
|
188 | 16e0b9c9 | Michael Hanselmann | notepad.RecordRestartAttempt(inst.name) |
189 | 16e0b9c9 | Michael Hanselmann | logging.error("Could not restart instance '%s' after %s attempts,"
|
190 | 16e0b9c9 | Michael Hanselmann | " giving up", inst.name, MAXTRIES)
|
191 | 16e0b9c9 | Michael Hanselmann | continue
|
192 | 5a3103e9 | Michael Hanselmann | |
193 | 16e0b9c9 | Michael Hanselmann | try:
|
194 | 16e0b9c9 | Michael Hanselmann | logging.info("Restarting instance '%s' (attempt #%s)",
|
195 | 16e0b9c9 | Michael Hanselmann | inst.name, n + 1)
|
196 | 16e0b9c9 | Michael Hanselmann | inst.Restart(cl) |
197 | b459a848 | Andrea Spadaccini | except Exception: # pylint: disable=W0703 |
198 | 16e0b9c9 | Michael Hanselmann | logging.exception("Error while restarting instance '%s'", inst.name)
|
199 | 16e0b9c9 | Michael Hanselmann | else:
|
200 | 16e0b9c9 | Michael Hanselmann | started.add(inst.name) |
201 | 5a3103e9 | Michael Hanselmann | |
202 | 16e0b9c9 | Michael Hanselmann | notepad.RecordRestartAttempt(inst.name) |
203 | 5a3103e9 | Michael Hanselmann | |
204 | 16e0b9c9 | Michael Hanselmann | else:
|
205 | 16e0b9c9 | Michael Hanselmann | if notepad.NumberOfRestartAttempts(inst.name):
|
206 | 16e0b9c9 | Michael Hanselmann | notepad.RemoveInstance(inst.name) |
207 | 16e0b9c9 | Michael Hanselmann | if inst.status not in HELPLESS_STATES: |
208 | 16e0b9c9 | Michael Hanselmann | logging.info("Restart of instance '%s' succeeded", inst.name)
|
209 | a8083063 | Iustin Pop | |
210 | 16e0b9c9 | Michael Hanselmann | return started
|
211 | a8083063 | Iustin Pop | |
212 | a8083063 | Iustin Pop | |
213 | 16e0b9c9 | Michael Hanselmann | def _CheckDisks(cl, notepad, nodes, instances, started): |
214 | 16e0b9c9 | Michael Hanselmann | """Check all nodes for restarted ones.
|
215 | 38242904 | Iustin Pop |
|
216 | a8083063 | Iustin Pop | """
|
217 | 16e0b9c9 | Michael Hanselmann | check_nodes = [] |
218 | 16e0b9c9 | Michael Hanselmann | |
219 | 16e0b9c9 | Michael Hanselmann | for node in nodes.values(): |
220 | 16e0b9c9 | Michael Hanselmann | old = notepad.GetNodeBootID(node.name) |
221 | 16e0b9c9 | Michael Hanselmann | if not node.bootid: |
222 | 16e0b9c9 | Michael Hanselmann | # Bad node, not returning a boot id
|
223 | 16e0b9c9 | Michael Hanselmann | if not node.offline: |
224 | 16e0b9c9 | Michael Hanselmann | logging.debug("Node '%s' missing boot ID, skipping secondary checks",
|
225 | 16e0b9c9 | Michael Hanselmann | node.name) |
226 | 16e0b9c9 | Michael Hanselmann | continue
|
227 | 16e0b9c9 | Michael Hanselmann | |
228 | 16e0b9c9 | Michael Hanselmann | if old != node.bootid:
|
229 | 16e0b9c9 | Michael Hanselmann | # Node's boot ID has changed, probably through a reboot
|
230 | 16e0b9c9 | Michael Hanselmann | check_nodes.append(node) |
231 | 16e0b9c9 | Michael Hanselmann | |
232 | 16e0b9c9 | Michael Hanselmann | if check_nodes:
|
233 | 16e0b9c9 | Michael Hanselmann | # Activate disks for all instances with any of the checked nodes as a
|
234 | 16e0b9c9 | Michael Hanselmann | # secondary node.
|
235 | 16e0b9c9 | Michael Hanselmann | for node in check_nodes: |
236 | 16e0b9c9 | Michael Hanselmann | for instance_name in node.secondaries: |
237 | 16e0b9c9 | Michael Hanselmann | try:
|
238 | 16e0b9c9 | Michael Hanselmann | inst = instances[instance_name] |
239 | 16e0b9c9 | Michael Hanselmann | except KeyError: |
240 | 16e0b9c9 | Michael Hanselmann | logging.info("Can't find instance '%s', maybe it was ignored",
|
241 | 16e0b9c9 | Michael Hanselmann | instance_name) |
242 | eee1fa2d | Iustin Pop | continue
|
243 | a8083063 | Iustin Pop | |
244 | 16e0b9c9 | Michael Hanselmann | if not inst.autostart: |
245 | 16e0b9c9 | Michael Hanselmann | logging.info("Skipping disk activation for non-autostart"
|
246 | 16e0b9c9 | Michael Hanselmann | " instance '%s'", inst.name)
|
247 | a8083063 | Iustin Pop | continue
|
248 | 16e0b9c9 | Michael Hanselmann | |
249 | 16e0b9c9 | Michael Hanselmann | if inst.name in started: |
250 | 16e0b9c9 | Michael Hanselmann | # we already tried to start the instance, which should have
|
251 | 16e0b9c9 | Michael Hanselmann | # activated its drives (if they can be at all)
|
252 | 16e0b9c9 | Michael Hanselmann | logging.debug("Skipping disk activation for instance '%s' as"
|
253 | 16e0b9c9 | Michael Hanselmann | " it was already started", inst.name)
|
254 | a8083063 | Iustin Pop | continue
|
255 | 16e0b9c9 | Michael Hanselmann | |
256 | a8083063 | Iustin Pop | try:
|
257 | 16e0b9c9 | Michael Hanselmann | logging.info("Activating disks for instance '%s'", inst.name)
|
258 | 16e0b9c9 | Michael Hanselmann | inst.ActivateDisks(cl) |
259 | b459a848 | Andrea Spadaccini | except Exception: # pylint: disable=W0703 |
260 | 16e0b9c9 | Michael Hanselmann | logging.exception("Error while activating disks for instance '%s'",
|
261 | 16e0b9c9 | Michael Hanselmann | inst.name) |
262 | a8083063 | Iustin Pop | |
263 | 16e0b9c9 | Michael Hanselmann | # Keep changed boot IDs
|
264 | 16e0b9c9 | Michael Hanselmann | for node in check_nodes: |
265 | 16e0b9c9 | Michael Hanselmann | notepad.SetNodeBootID(node.name, node.bootid) |
266 | a8083063 | Iustin Pop | |
267 | 83e5e26f | René Nussbaumer | |
268 | 16e0b9c9 | Michael Hanselmann | def _CheckForOfflineNodes(nodes, instance): |
269 | 16e0b9c9 | Michael Hanselmann | """Checks if given instances has any secondary in offline status.
|
270 | ae1a845c | Michael Hanselmann |
|
271 | 16e0b9c9 | Michael Hanselmann | @param instance: The instance object
|
272 | 16e0b9c9 | Michael Hanselmann | @return: True if any of the secondary is offline, False otherwise
|
273 | ae1a845c | Michael Hanselmann |
|
274 | 16e0b9c9 | Michael Hanselmann | """
|
275 | 16e0b9c9 | Michael Hanselmann | return compat.any(nodes[node_name].offline for node_name in instance.snodes) |
276 | ae1a845c | Michael Hanselmann | |
277 | ae1a845c | Michael Hanselmann | |
278 | 16e0b9c9 | Michael Hanselmann | def _VerifyDisks(cl, uuid, nodes, instances): |
279 | 16e0b9c9 | Michael Hanselmann | """Run a per-group "gnt-cluster verify-disks".
|
280 | ae1a845c | Michael Hanselmann |
|
281 | 16e0b9c9 | Michael Hanselmann | """
|
282 | 16e0b9c9 | Michael Hanselmann | job_id = cl.SubmitJob([opcodes.OpGroupVerifyDisks(group_name=uuid)]) |
283 | 16e0b9c9 | Michael Hanselmann | ((_, offline_disk_instances, _), ) = \ |
284 | 16e0b9c9 | Michael Hanselmann | cli.PollJob(job_id, cl=cl, feedback_fn=logging.debug) |
285 | 16e0b9c9 | Michael Hanselmann | cl.ArchiveJob(job_id) |
286 | ae1a845c | Michael Hanselmann | |
287 | 16e0b9c9 | Michael Hanselmann | if not offline_disk_instances: |
288 | 16e0b9c9 | Michael Hanselmann | # nothing to do
|
289 | 16e0b9c9 | Michael Hanselmann | logging.debug("Verify-disks reported no offline disks, nothing to do")
|
290 | 16e0b9c9 | Michael Hanselmann | return
|
291 | ae1a845c | Michael Hanselmann | |
292 | 16e0b9c9 | Michael Hanselmann | logging.debug("Will activate disks for instance(s) %s",
|
293 | 16e0b9c9 | Michael Hanselmann | utils.CommaJoin(offline_disk_instances)) |
294 | ae1a845c | Michael Hanselmann | |
295 | 16e0b9c9 | Michael Hanselmann | # We submit only one job, and wait for it. Not optimal, but this puts less
|
296 | 16e0b9c9 | Michael Hanselmann | # load on the job queue.
|
297 | 16e0b9c9 | Michael Hanselmann | job = [] |
298 | 16e0b9c9 | Michael Hanselmann | for name in offline_disk_instances: |
299 | 16e0b9c9 | Michael Hanselmann | try:
|
300 | 16e0b9c9 | Michael Hanselmann | inst = instances[name] |
301 | 16e0b9c9 | Michael Hanselmann | except KeyError: |
302 | 16e0b9c9 | Michael Hanselmann | logging.info("Can't find instance '%s', maybe it was ignored", name)
|
303 | 16e0b9c9 | Michael Hanselmann | continue
|
304 | ae1a845c | Michael Hanselmann | |
305 | 16e0b9c9 | Michael Hanselmann | if inst.status in HELPLESS_STATES or _CheckForOfflineNodes(nodes, inst): |
306 | 16e0b9c9 | Michael Hanselmann | logging.info("Skipping instance '%s' because it is in a helpless state or"
|
307 | 16e0b9c9 | Michael Hanselmann | " has offline secondaries", name)
|
308 | 16e0b9c9 | Michael Hanselmann | continue
|
309 | ae1a845c | Michael Hanselmann | |
310 | 16e0b9c9 | Michael Hanselmann | job.append(opcodes.OpInstanceActivateDisks(instance_name=name)) |
311 | 5188ab37 | Iustin Pop | |
312 | 16e0b9c9 | Michael Hanselmann | if job:
|
313 | 16e0b9c9 | Michael Hanselmann | job_id = cli.SendJob(job, cl=cl) |
314 | 83e5e26f | René Nussbaumer | |
315 | 16e0b9c9 | Michael Hanselmann | try:
|
316 | 16e0b9c9 | Michael Hanselmann | cli.PollJob(job_id, cl=cl, feedback_fn=logging.debug) |
317 | b459a848 | Andrea Spadaccini | except Exception: # pylint: disable=W0703 |
318 | 16e0b9c9 | Michael Hanselmann | logging.exception("Error while activating disks")
|
319 | a8083063 | Iustin Pop | |
320 | a8083063 | Iustin Pop | |
321 | db147305 | Tom Limoncelli | def IsRapiResponding(hostname): |
322 | db147305 | Tom Limoncelli | """Connects to RAPI port and does a simple test.
|
323 | db147305 | Tom Limoncelli |
|
324 | db147305 | Tom Limoncelli | Connects to RAPI port of hostname and does a simple test. At this time, the
|
325 | db147305 | Tom Limoncelli | test is GetVersion.
|
326 | db147305 | Tom Limoncelli |
|
327 | db147305 | Tom Limoncelli | @type hostname: string
|
328 | db147305 | Tom Limoncelli | @param hostname: hostname of the node to connect to.
|
329 | db147305 | Tom Limoncelli | @rtype: bool
|
330 | db147305 | Tom Limoncelli | @return: Whether RAPI is working properly
|
331 | db147305 | Tom Limoncelli |
|
332 | db147305 | Tom Limoncelli | """
|
333 | 34f06005 | Iustin Pop | curl_config = rapi.client.GenericCurlConfig() |
334 | 2a7c3583 | Michael Hanselmann | rapi_client = rapi.client.GanetiRapiClient(hostname, |
335 | 2a7c3583 | Michael Hanselmann | curl_config_fn=curl_config) |
336 | db147305 | Tom Limoncelli | try:
|
337 | db147305 | Tom Limoncelli | master_version = rapi_client.GetVersion() |
338 | db147305 | Tom Limoncelli | except rapi.client.CertificateError, err:
|
339 | d7c42723 | Michael Hanselmann | logging.warning("RAPI certificate error: %s", err)
|
340 | db147305 | Tom Limoncelli | return False |
341 | db147305 | Tom Limoncelli | except rapi.client.GanetiApiError, err:
|
342 | d7c42723 | Michael Hanselmann | logging.warning("RAPI error: %s", err)
|
343 | db147305 | Tom Limoncelli | return False |
344 | d7c42723 | Michael Hanselmann | else:
|
345 | d7c42723 | Michael Hanselmann | logging.debug("Reported RAPI version %s", master_version)
|
346 | d7c42723 | Michael Hanselmann | return master_version == constants.RAPI_VERSION
|
347 | db147305 | Tom Limoncelli | |
348 | db147305 | Tom Limoncelli | |
349 | a8083063 | Iustin Pop | def ParseOptions(): |
350 | a8083063 | Iustin Pop | """Parse the command line options.
|
351 | a8083063 | Iustin Pop |
|
352 | c41eea6e | Iustin Pop | @return: (options, args) as from OptionParser.parse_args()
|
353 | a8083063 | Iustin Pop |
|
354 | a8083063 | Iustin Pop | """
|
355 | a8083063 | Iustin Pop | parser = OptionParser(description="Ganeti cluster watcher",
|
356 | a8083063 | Iustin Pop | usage="%prog [-d]",
|
357 | a8083063 | Iustin Pop | version="%%prog (ganeti) %s" %
|
358 | a8083063 | Iustin Pop | constants.RELEASE_VERSION) |
359 | a8083063 | Iustin Pop | |
360 | 6d4e8ec0 | Iustin Pop | parser.add_option(cli.DEBUG_OPT) |
361 | 16e0b9c9 | Michael Hanselmann | parser.add_option(cli.NODEGROUP_OPT) |
362 | f0a80b01 | Michael Hanselmann | parser.add_option("-A", "--job-age", dest="job_age", default=6 * 3600, |
363 | f07521e5 | Iustin Pop | help="Autoarchive jobs older than this age (default"
|
364 | f0a80b01 | Michael Hanselmann | " 6 hours)")
|
365 | 46c8a6ab | Iustin Pop | parser.add_option("--ignore-pause", dest="ignore_pause", default=False, |
366 | 46c8a6ab | Iustin Pop | action="store_true", help="Ignore cluster pause setting") |
367 | 5f01e6ad | Michael Hanselmann | parser.add_option("--wait-children", dest="wait_children", |
368 | 16e0b9c9 | Michael Hanselmann | action="store_true", help="Wait for child processes") |
369 | 5f01e6ad | Michael Hanselmann | parser.add_option("--no-wait-children", dest="wait_children", |
370 | 5f01e6ad | Michael Hanselmann | action="store_false", help="Don't wait for child processes") |
371 | 5f01e6ad | Michael Hanselmann | # See optparse documentation for why default values are not set by options
|
372 | 5f01e6ad | Michael Hanselmann | parser.set_defaults(wait_children=True)
|
373 | a8083063 | Iustin Pop | options, args = parser.parse_args() |
374 | f07521e5 | Iustin Pop | options.job_age = cli.ParseTimespec(options.job_age) |
375 | f0a80b01 | Michael Hanselmann | |
376 | f0a80b01 | Michael Hanselmann | if args:
|
377 | f0a80b01 | Michael Hanselmann | parser.error("No arguments expected")
|
378 | f0a80b01 | Michael Hanselmann | |
379 | f0a80b01 | Michael Hanselmann | return (options, args)
|
380 | a8083063 | Iustin Pop | |
381 | a8083063 | Iustin Pop | |
382 | 9bb69bb5 | Michael Hanselmann | def _WriteInstanceStatus(filename, data): |
383 | 9bb69bb5 | Michael Hanselmann | """Writes the per-group instance status file.
|
384 | 9bb69bb5 | Michael Hanselmann |
|
385 | 9bb69bb5 | Michael Hanselmann | The entries are sorted.
|
386 | 8f07dc0d | Michael Hanselmann |
|
387 | 9bb69bb5 | Michael Hanselmann | @type filename: string
|
388 | 9bb69bb5 | Michael Hanselmann | @param filename: Path to instance status file
|
389 | 9bb69bb5 | Michael Hanselmann | @type data: list of tuple; (instance name as string, status as string)
|
390 | 9bb69bb5 | Michael Hanselmann | @param data: Instance name and status
|
391 | 8f07dc0d | Michael Hanselmann |
|
392 | 8f07dc0d | Michael Hanselmann | """
|
393 | 9bb69bb5 | Michael Hanselmann | logging.debug("Updating instance status file '%s' with %s instances",
|
394 | 9bb69bb5 | Michael Hanselmann | filename, len(data))
|
395 | 8f07dc0d | Michael Hanselmann | |
396 | 9bb69bb5 | Michael Hanselmann | utils.WriteFile(filename, |
397 | 9bb69bb5 | Michael Hanselmann | data="".join(map(compat.partial(operator.mod, "%s %s\n"), |
398 | 9bb69bb5 | Michael Hanselmann | sorted(data))))
|
399 | 9bb69bb5 | Michael Hanselmann | |
400 | 9bb69bb5 | Michael Hanselmann | |
401 | 9bb69bb5 | Michael Hanselmann | def _UpdateInstanceStatus(filename, instances): |
402 | 9bb69bb5 | Michael Hanselmann | """Writes an instance status file from L{Instance} objects.
|
403 | 9bb69bb5 | Michael Hanselmann |
|
404 | 9bb69bb5 | Michael Hanselmann | @type filename: string
|
405 | 9bb69bb5 | Michael Hanselmann | @param filename: Path to status file
|
406 | 9bb69bb5 | Michael Hanselmann | @type instances: list of L{Instance}
|
407 | 9bb69bb5 | Michael Hanselmann |
|
408 | 9bb69bb5 | Michael Hanselmann | """
|
409 | 9bb69bb5 | Michael Hanselmann | _WriteInstanceStatus(filename, [(inst.name, inst.status) |
410 | 9bb69bb5 | Michael Hanselmann | for inst in instances]) |
411 | 9bb69bb5 | Michael Hanselmann | |
412 | 9bb69bb5 | Michael Hanselmann | |
413 | 9bb69bb5 | Michael Hanselmann | def _ReadInstanceStatus(filename): |
414 | 9bb69bb5 | Michael Hanselmann | """Reads an instance status file.
|
415 | 9bb69bb5 | Michael Hanselmann |
|
416 | 9bb69bb5 | Michael Hanselmann | @type filename: string
|
417 | 9bb69bb5 | Michael Hanselmann | @param filename: Path to status file
|
418 | 9bb69bb5 | Michael Hanselmann | @rtype: tuple; (None or number, list of lists containing instance name and
|
419 | 9bb69bb5 | Michael Hanselmann | status)
|
420 | 9bb69bb5 | Michael Hanselmann | @return: File's mtime and instance status contained in the file; mtime is
|
421 | 9bb69bb5 | Michael Hanselmann | C{None} if file can't be read
|
422 | 9bb69bb5 | Michael Hanselmann |
|
423 | 9bb69bb5 | Michael Hanselmann | """
|
424 | 9bb69bb5 | Michael Hanselmann | logging.debug("Reading per-group instance status from '%s'", filename)
|
425 | 9bb69bb5 | Michael Hanselmann | |
426 | 2635bb04 | Michael Hanselmann | statcb = utils.FileStatHelper() |
427 | 9bb69bb5 | Michael Hanselmann | try:
|
428 | 9bb69bb5 | Michael Hanselmann | content = utils.ReadFile(filename, preread=statcb) |
429 | 9bb69bb5 | Michael Hanselmann | except EnvironmentError, err: |
430 | 9bb69bb5 | Michael Hanselmann | if err.errno == errno.ENOENT:
|
431 | 9bb69bb5 | Michael Hanselmann | logging.error("Can't read '%s', does not exist (yet)", filename)
|
432 | 9bb69bb5 | Michael Hanselmann | else:
|
433 | 9bb69bb5 | Michael Hanselmann | logging.exception("Unable to read '%s', ignoring", filename)
|
434 | 9bb69bb5 | Michael Hanselmann | return (None, None) |
435 | 9bb69bb5 | Michael Hanselmann | else:
|
436 | 6f9e71bb | Michael Hanselmann | return (statcb.st.st_mtime, [line.split(None, 1) |
437 | 9bb69bb5 | Michael Hanselmann | for line in content.splitlines()]) |
438 | 9bb69bb5 | Michael Hanselmann | |
439 | 9bb69bb5 | Michael Hanselmann | |
440 | 9bb69bb5 | Michael Hanselmann | def _MergeInstanceStatus(filename, pergroup_filename, groups): |
441 | 9bb69bb5 | Michael Hanselmann | """Merges all per-group instance status files into a global one.
|
442 | 9bb69bb5 | Michael Hanselmann |
|
443 | 9bb69bb5 | Michael Hanselmann | @type filename: string
|
444 | 9bb69bb5 | Michael Hanselmann | @param filename: Path to global instance status file
|
445 | 9bb69bb5 | Michael Hanselmann | @type pergroup_filename: string
|
446 | 9bb69bb5 | Michael Hanselmann | @param pergroup_filename: Path to per-group status files, must contain "%s"
|
447 | 9bb69bb5 | Michael Hanselmann | to be replaced with group UUID
|
448 | 9bb69bb5 | Michael Hanselmann | @type groups: sequence
|
449 | 9bb69bb5 | Michael Hanselmann | @param groups: UUIDs of known groups
|
450 | 9bb69bb5 | Michael Hanselmann |
|
451 | 9bb69bb5 | Michael Hanselmann | """
|
452 | 9bb69bb5 | Michael Hanselmann | # Lock global status file in exclusive mode
|
453 | 9bb69bb5 | Michael Hanselmann | lock = utils.FileLock.Open(filename) |
454 | 9bb69bb5 | Michael Hanselmann | try:
|
455 | 9bb69bb5 | Michael Hanselmann | lock.Exclusive(blocking=True, timeout=INSTANCE_STATUS_LOCK_TIMEOUT)
|
456 | 9bb69bb5 | Michael Hanselmann | except errors.LockError, err:
|
457 | 9bb69bb5 | Michael Hanselmann | # All per-group processes will lock and update the file. None of them
|
458 | 9bb69bb5 | Michael Hanselmann | # should take longer than 10 seconds (the value of
|
459 | 9bb69bb5 | Michael Hanselmann | # INSTANCE_STATUS_LOCK_TIMEOUT).
|
460 | 9bb69bb5 | Michael Hanselmann | logging.error("Can't acquire lock on instance status file '%s', not"
|
461 | 9bb69bb5 | Michael Hanselmann | " updating: %s", filename, err)
|
462 | 9bb69bb5 | Michael Hanselmann | return
|
463 | 9bb69bb5 | Michael Hanselmann | |
464 | 9bb69bb5 | Michael Hanselmann | logging.debug("Acquired exclusive lock on '%s'", filename)
|
465 | 9bb69bb5 | Michael Hanselmann | |
466 | 9bb69bb5 | Michael Hanselmann | data = {} |
467 | 9bb69bb5 | Michael Hanselmann | |
468 | 9bb69bb5 | Michael Hanselmann | # Load instance status from all groups
|
469 | 9bb69bb5 | Michael Hanselmann | for group_uuid in groups: |
470 | 9bb69bb5 | Michael Hanselmann | (mtime, instdata) = _ReadInstanceStatus(pergroup_filename % group_uuid) |
471 | 9bb69bb5 | Michael Hanselmann | |
472 | 9bb69bb5 | Michael Hanselmann | if mtime is not None: |
473 | 9bb69bb5 | Michael Hanselmann | for (instance_name, status) in instdata: |
474 | 9bb69bb5 | Michael Hanselmann | data.setdefault(instance_name, []).append((mtime, status)) |
475 | 9bb69bb5 | Michael Hanselmann | |
476 | 9bb69bb5 | Michael Hanselmann | # Select last update based on file mtime
|
477 | 9bb69bb5 | Michael Hanselmann | inststatus = [(instance_name, sorted(status, reverse=True)[0][1]) |
478 | 9bb69bb5 | Michael Hanselmann | for (instance_name, status) in data.items()] |
479 | 9bb69bb5 | Michael Hanselmann | |
480 | 9bb69bb5 | Michael Hanselmann | # Write the global status file. Don't touch file after it's been
|
481 | 9bb69bb5 | Michael Hanselmann | # updated--there is no lock anymore.
|
482 | 9bb69bb5 | Michael Hanselmann | _WriteInstanceStatus(filename, inststatus) |
483 | 8f07dc0d | Michael Hanselmann | |
484 | 8f07dc0d | Michael Hanselmann | |
485 | 16e0b9c9 | Michael Hanselmann | def GetLuxiClient(try_restart): |
486 | 16e0b9c9 | Michael Hanselmann | """Tries to connect to the master daemon.
|
487 | 16e0b9c9 | Michael Hanselmann |
|
488 | 16e0b9c9 | Michael Hanselmann | @type try_restart: bool
|
489 | 16e0b9c9 | Michael Hanselmann | @param try_restart: Whether to attempt to restart the master daemon
|
490 | 16e0b9c9 | Michael Hanselmann |
|
491 | 16e0b9c9 | Michael Hanselmann | """
|
492 | 16e0b9c9 | Michael Hanselmann | try:
|
493 | 16e0b9c9 | Michael Hanselmann | return cli.GetClient()
|
494 | 16e0b9c9 | Michael Hanselmann | except errors.OpPrereqError, err:
|
495 | 16e0b9c9 | Michael Hanselmann | # this is, from cli.GetClient, a not-master case
|
496 | 16e0b9c9 | Michael Hanselmann | raise NotMasterError("Not on master node (%s)" % err) |
497 | 16e0b9c9 | Michael Hanselmann | |
498 | 16e0b9c9 | Michael Hanselmann | except luxi.NoMasterError, err:
|
499 | 16e0b9c9 | Michael Hanselmann | if not try_restart: |
500 | 16e0b9c9 | Michael Hanselmann | raise
|
501 | 16e0b9c9 | Michael Hanselmann | |
502 | 16e0b9c9 | Michael Hanselmann | logging.warning("Master daemon seems to be down (%s), trying to restart",
|
503 | 16e0b9c9 | Michael Hanselmann | err) |
504 | 16e0b9c9 | Michael Hanselmann | |
505 | 16e0b9c9 | Michael Hanselmann | if not utils.EnsureDaemon(constants.MASTERD): |
506 | 16e0b9c9 | Michael Hanselmann | raise errors.GenericError("Can't start the master daemon") |
507 | 16e0b9c9 | Michael Hanselmann | |
508 | 16e0b9c9 | Michael Hanselmann | # Retry the connection
|
509 | 16e0b9c9 | Michael Hanselmann | return cli.GetClient()
|
510 | 16e0b9c9 | Michael Hanselmann | |
511 | 16e0b9c9 | Michael Hanselmann | |
512 | 16e0b9c9 | Michael Hanselmann | def _StartGroupChildren(cl, wait): |
513 | 16e0b9c9 | Michael Hanselmann | """Starts a new instance of the watcher for every node group.
|
514 | 16e0b9c9 | Michael Hanselmann |
|
515 | 16e0b9c9 | Michael Hanselmann | """
|
516 | 16e0b9c9 | Michael Hanselmann | assert not compat.any(arg.startswith(cli.NODEGROUP_OPT_NAME) |
517 | 16e0b9c9 | Michael Hanselmann | for arg in sys.argv) |
518 | 16e0b9c9 | Michael Hanselmann | |
519 | 16e0b9c9 | Michael Hanselmann | result = cl.QueryGroups([], ["name", "uuid"], False) |
520 | 16e0b9c9 | Michael Hanselmann | |
521 | 16e0b9c9 | Michael Hanselmann | children = [] |
522 | 16e0b9c9 | Michael Hanselmann | |
523 | 16e0b9c9 | Michael Hanselmann | for (idx, (name, uuid)) in enumerate(result): |
524 | 16e0b9c9 | Michael Hanselmann | args = sys.argv + [cli.NODEGROUP_OPT_NAME, uuid] |
525 | 16e0b9c9 | Michael Hanselmann | |
526 | 16e0b9c9 | Michael Hanselmann | if idx > 0: |
527 | 16e0b9c9 | Michael Hanselmann | # Let's not kill the system
|
528 | 16e0b9c9 | Michael Hanselmann | time.sleep(CHILD_PROCESS_DELAY) |
529 | 16e0b9c9 | Michael Hanselmann | |
530 | 16e0b9c9 | Michael Hanselmann | logging.debug("Spawning child for group '%s' (%s), arguments %s",
|
531 | 16e0b9c9 | Michael Hanselmann | name, uuid, args) |
532 | 16e0b9c9 | Michael Hanselmann | |
533 | 16e0b9c9 | Michael Hanselmann | try:
|
534 | 16e0b9c9 | Michael Hanselmann | # TODO: Should utils.StartDaemon be used instead?
|
535 | 16e0b9c9 | Michael Hanselmann | pid = os.spawnv(os.P_NOWAIT, args[0], args)
|
536 | b459a848 | Andrea Spadaccini | except Exception: # pylint: disable=W0703 |
537 | 16e0b9c9 | Michael Hanselmann | logging.exception("Failed to start child for group '%s' (%s)",
|
538 | 16e0b9c9 | Michael Hanselmann | name, uuid) |
539 | 16e0b9c9 | Michael Hanselmann | else:
|
540 | 16e0b9c9 | Michael Hanselmann | logging.debug("Started with PID %s", pid)
|
541 | 16e0b9c9 | Michael Hanselmann | children.append(pid) |
542 | 16e0b9c9 | Michael Hanselmann | |
543 | 16e0b9c9 | Michael Hanselmann | if wait:
|
544 | 16e0b9c9 | Michael Hanselmann | for pid in children: |
545 | 16e0b9c9 | Michael Hanselmann | logging.debug("Waiting for child PID %s", pid)
|
546 | 16e0b9c9 | Michael Hanselmann | try:
|
547 | 16e0b9c9 | Michael Hanselmann | result = utils.RetryOnSignal(os.waitpid, pid, 0)
|
548 | 16e0b9c9 | Michael Hanselmann | except EnvironmentError, err: |
549 | 16e0b9c9 | Michael Hanselmann | result = str(err)
|
550 | 16e0b9c9 | Michael Hanselmann | |
551 | 16e0b9c9 | Michael Hanselmann | logging.debug("Child PID %s exited with status %s", pid, result)
|
552 | 16e0b9c9 | Michael Hanselmann | |
553 | 16e0b9c9 | Michael Hanselmann | |
554 | 16e0b9c9 | Michael Hanselmann | def _ArchiveJobs(cl, age): |
555 | 16e0b9c9 | Michael Hanselmann | """Archives old jobs.
|
556 | 16e0b9c9 | Michael Hanselmann |
|
557 | 16e0b9c9 | Michael Hanselmann | """
|
558 | 16e0b9c9 | Michael Hanselmann | (arch_count, left_count) = cl.AutoArchiveJobs(age) |
559 | 16e0b9c9 | Michael Hanselmann | logging.debug("Archived %s jobs, left %s", arch_count, left_count)
|
560 | 16e0b9c9 | Michael Hanselmann | |
561 | 16e0b9c9 | Michael Hanselmann | |
562 | 16e0b9c9 | Michael Hanselmann | def _CheckMaster(cl): |
563 | 16e0b9c9 | Michael Hanselmann | """Ensures current host is master node.
|
564 | 16e0b9c9 | Michael Hanselmann |
|
565 | 16e0b9c9 | Michael Hanselmann | """
|
566 | 16e0b9c9 | Michael Hanselmann | (master, ) = cl.QueryConfigValues(["master_node"])
|
567 | 16e0b9c9 | Michael Hanselmann | if master != netutils.Hostname.GetSysName():
|
568 | 16e0b9c9 | Michael Hanselmann | raise NotMasterError("This is not the master node") |
569 | 16e0b9c9 | Michael Hanselmann | |
570 | 16e0b9c9 | Michael Hanselmann | |
571 | 2a7c3583 | Michael Hanselmann | @rapi.client.UsesRapiClient
|
572 | 16e0b9c9 | Michael Hanselmann | def _GlobalWatcher(opts): |
573 | 16e0b9c9 | Michael Hanselmann | """Main function for global watcher.
|
574 | 16e0b9c9 | Michael Hanselmann |
|
575 | 16e0b9c9 | Michael Hanselmann | At the end child processes are spawned for every node group.
|
576 | 16e0b9c9 | Michael Hanselmann |
|
577 | 16e0b9c9 | Michael Hanselmann | """
|
578 | 16e0b9c9 | Michael Hanselmann | StartNodeDaemons() |
579 | 16e0b9c9 | Michael Hanselmann | RunWatcherHooks() |
580 | 16e0b9c9 | Michael Hanselmann | |
581 | 16e0b9c9 | Michael Hanselmann | # Run node maintenance in all cases, even if master, so that old masters can
|
582 | 16e0b9c9 | Michael Hanselmann | # be properly cleaned up
|
583 | b459a848 | Andrea Spadaccini | if nodemaint.NodeMaintenance.ShouldRun(): # pylint: disable=E0602 |
584 | b459a848 | Andrea Spadaccini | nodemaint.NodeMaintenance().Exec() # pylint: disable=E0602
|
585 | 16e0b9c9 | Michael Hanselmann | |
586 | 16e0b9c9 | Michael Hanselmann | try:
|
587 | 16e0b9c9 | Michael Hanselmann | client = GetLuxiClient(True)
|
588 | 16e0b9c9 | Michael Hanselmann | except NotMasterError:
|
589 | 16e0b9c9 | Michael Hanselmann | # Don't proceed on non-master nodes
|
590 | 16e0b9c9 | Michael Hanselmann | return constants.EXIT_SUCCESS
|
591 | 16e0b9c9 | Michael Hanselmann | |
592 | 16e0b9c9 | Michael Hanselmann | # we are on master now
|
593 | 16e0b9c9 | Michael Hanselmann | utils.EnsureDaemon(constants.RAPI) |
594 | 16e0b9c9 | Michael Hanselmann | |
595 | 16e0b9c9 | Michael Hanselmann | # If RAPI isn't responding to queries, try one restart
|
596 | 16e0b9c9 | Michael Hanselmann | logging.debug("Attempting to talk to remote API on %s",
|
597 | 16e0b9c9 | Michael Hanselmann | constants.IP4_ADDRESS_LOCALHOST) |
598 | 16e0b9c9 | Michael Hanselmann | if not IsRapiResponding(constants.IP4_ADDRESS_LOCALHOST): |
599 | 16e0b9c9 | Michael Hanselmann | logging.warning("Couldn't get answer from remote API, restaring daemon")
|
600 | 16e0b9c9 | Michael Hanselmann | utils.StopDaemon(constants.RAPI) |
601 | 16e0b9c9 | Michael Hanselmann | utils.EnsureDaemon(constants.RAPI) |
602 | 16e0b9c9 | Michael Hanselmann | logging.debug("Second attempt to talk to remote API")
|
603 | 16e0b9c9 | Michael Hanselmann | if not IsRapiResponding(constants.IP4_ADDRESS_LOCALHOST): |
604 | 16e0b9c9 | Michael Hanselmann | logging.fatal("RAPI is not responding")
|
605 | 16e0b9c9 | Michael Hanselmann | logging.debug("Successfully talked to remote API")
|
606 | 16e0b9c9 | Michael Hanselmann | |
607 | 16e0b9c9 | Michael Hanselmann | _CheckMaster(client) |
608 | 16e0b9c9 | Michael Hanselmann | _ArchiveJobs(client, opts.job_age) |
609 | 16e0b9c9 | Michael Hanselmann | |
610 | 16e0b9c9 | Michael Hanselmann | # Spawn child processes for all node groups
|
611 | 16e0b9c9 | Michael Hanselmann | _StartGroupChildren(client, opts.wait_children) |
612 | 16e0b9c9 | Michael Hanselmann | |
613 | 16e0b9c9 | Michael Hanselmann | return constants.EXIT_SUCCESS
|
614 | 16e0b9c9 | Michael Hanselmann | |
615 | 16e0b9c9 | Michael Hanselmann | |
616 | 16e0b9c9 | Michael Hanselmann | def _GetGroupData(cl, uuid): |
617 | 16e0b9c9 | Michael Hanselmann | """Retrieves instances and nodes per node group.
|
618 | 16e0b9c9 | Michael Hanselmann |
|
619 | 16e0b9c9 | Michael Hanselmann | """
|
620 | 16e0b9c9 | Michael Hanselmann | job = [ |
621 | 16e0b9c9 | Michael Hanselmann | # Get all primary instances in group
|
622 | 16e0b9c9 | Michael Hanselmann | opcodes.OpQuery(what=constants.QR_INSTANCE, |
623 | 16e0b9c9 | Michael Hanselmann | fields=["name", "status", "admin_state", "snodes", |
624 | 16e0b9c9 | Michael Hanselmann | "pnode.group.uuid", "snodes.group.uuid"], |
625 | 2e5c33db | Iustin Pop | qfilter=[qlang.OP_EQUAL, "pnode.group.uuid", uuid],
|
626 | 5bfb1134 | Michael Hanselmann | use_locking=True),
|
627 | 16e0b9c9 | Michael Hanselmann | |
628 | 16e0b9c9 | Michael Hanselmann | # Get all nodes in group
|
629 | 16e0b9c9 | Michael Hanselmann | opcodes.OpQuery(what=constants.QR_NODE, |
630 | 16e0b9c9 | Michael Hanselmann | fields=["name", "bootid", "offline"], |
631 | 2e5c33db | Iustin Pop | qfilter=[qlang.OP_EQUAL, "group.uuid", uuid],
|
632 | 5bfb1134 | Michael Hanselmann | use_locking=True),
|
633 | 16e0b9c9 | Michael Hanselmann | ] |
634 | 16e0b9c9 | Michael Hanselmann | |
635 | 16e0b9c9 | Michael Hanselmann | job_id = cl.SubmitJob(job) |
636 | 16e0b9c9 | Michael Hanselmann | results = map(objects.QueryResponse.FromDict,
|
637 | 16e0b9c9 | Michael Hanselmann | cli.PollJob(job_id, cl=cl, feedback_fn=logging.debug)) |
638 | 16e0b9c9 | Michael Hanselmann | cl.ArchiveJob(job_id) |
639 | 16e0b9c9 | Michael Hanselmann | |
640 | 16e0b9c9 | Michael Hanselmann | results_data = map(operator.attrgetter("data"), results) |
641 | 16e0b9c9 | Michael Hanselmann | |
642 | 16e0b9c9 | Michael Hanselmann | # Ensure results are tuples with two values
|
643 | 16e0b9c9 | Michael Hanselmann | assert compat.all(map(ht.TListOf(ht.TListOf(ht.TIsLength(2))), results_data)) |
644 | 16e0b9c9 | Michael Hanselmann | |
645 | 16e0b9c9 | Michael Hanselmann | # Extract values ignoring result status
|
646 | 16e0b9c9 | Michael Hanselmann | (raw_instances, raw_nodes) = [[map(compat.snd, values)
|
647 | 16e0b9c9 | Michael Hanselmann | for values in res] |
648 | 16e0b9c9 | Michael Hanselmann | for res in results_data] |
649 | 16e0b9c9 | Michael Hanselmann | |
650 | 16e0b9c9 | Michael Hanselmann | secondaries = {} |
651 | 16e0b9c9 | Michael Hanselmann | instances = [] |
652 | 16e0b9c9 | Michael Hanselmann | |
653 | 16e0b9c9 | Michael Hanselmann | # Load all instances
|
654 | 16e0b9c9 | Michael Hanselmann | for (name, status, autostart, snodes, pnode_group_uuid,
|
655 | 16e0b9c9 | Michael Hanselmann | snodes_group_uuid) in raw_instances:
|
656 | 16e0b9c9 | Michael Hanselmann | if snodes and set([pnode_group_uuid]) != set(snodes_group_uuid): |
657 | 16e0b9c9 | Michael Hanselmann | logging.error("Ignoring split instance '%s', primary group %s, secondary"
|
658 | 16e0b9c9 | Michael Hanselmann | " groups %s", name, pnode_group_uuid,
|
659 | 16e0b9c9 | Michael Hanselmann | utils.CommaJoin(snodes_group_uuid)) |
660 | 16e0b9c9 | Michael Hanselmann | else:
|
661 | 16e0b9c9 | Michael Hanselmann | instances.append(Instance(name, status, autostart, snodes)) |
662 | 16e0b9c9 | Michael Hanselmann | |
663 | 16e0b9c9 | Michael Hanselmann | for node in snodes: |
664 | 16e0b9c9 | Michael Hanselmann | secondaries.setdefault(node, set()).add(name)
|
665 | 16e0b9c9 | Michael Hanselmann | |
666 | 16e0b9c9 | Michael Hanselmann | # Load all nodes
|
667 | 16e0b9c9 | Michael Hanselmann | nodes = [Node(name, bootid, offline, secondaries.get(name, set()))
|
668 | 16e0b9c9 | Michael Hanselmann | for (name, bootid, offline) in raw_nodes] |
669 | 16e0b9c9 | Michael Hanselmann | |
670 | 16e0b9c9 | Michael Hanselmann | return (dict((node.name, node) for node in nodes), |
671 | 16e0b9c9 | Michael Hanselmann | dict((inst.name, inst) for inst in instances)) |
672 | 16e0b9c9 | Michael Hanselmann | |
673 | 16e0b9c9 | Michael Hanselmann | |
674 | 9bb69bb5 | Michael Hanselmann | def _LoadKnownGroups(): |
675 | 9bb69bb5 | Michael Hanselmann | """Returns a list of all node groups known by L{ssconf}.
|
676 | 16e0b9c9 | Michael Hanselmann |
|
677 | 16e0b9c9 | Michael Hanselmann | """
|
678 | 16e0b9c9 | Michael Hanselmann | groups = ssconf.SimpleStore().GetNodegroupList() |
679 | 16e0b9c9 | Michael Hanselmann | |
680 | 9bb69bb5 | Michael Hanselmann | result = list(line.split(None, 1)[0] for line in groups |
681 | 9bb69bb5 | Michael Hanselmann | if line.strip())
|
682 | 9bb69bb5 | Michael Hanselmann | |
683 | 9bb69bb5 | Michael Hanselmann | if not compat.all(map(utils.UUID_RE.match, result)): |
684 | 9bb69bb5 | Michael Hanselmann | raise errors.GenericError("Ssconf contains invalid group UUID") |
685 | 9bb69bb5 | Michael Hanselmann | |
686 | 9bb69bb5 | Michael Hanselmann | return result
|
687 | 16e0b9c9 | Michael Hanselmann | |
688 | 16e0b9c9 | Michael Hanselmann | |
689 | 16e0b9c9 | Michael Hanselmann | def _GroupWatcher(opts): |
690 | 16e0b9c9 | Michael Hanselmann | """Main function for per-group watcher process.
|
691 | 16e0b9c9 | Michael Hanselmann |
|
692 | 16e0b9c9 | Michael Hanselmann | """
|
693 | 16e0b9c9 | Michael Hanselmann | group_uuid = opts.nodegroup.lower() |
694 | 16e0b9c9 | Michael Hanselmann | |
695 | 16e0b9c9 | Michael Hanselmann | if not utils.UUID_RE.match(group_uuid): |
696 | 16e0b9c9 | Michael Hanselmann | raise errors.GenericError("Node group parameter (%s) must be given a UUID," |
697 | 16e0b9c9 | Michael Hanselmann | " got '%s'" %
|
698 | 16e0b9c9 | Michael Hanselmann | (cli.NODEGROUP_OPT_NAME, group_uuid)) |
699 | 16e0b9c9 | Michael Hanselmann | |
700 | 16e0b9c9 | Michael Hanselmann | logging.info("Watcher for node group '%s'", group_uuid)
|
701 | 16e0b9c9 | Michael Hanselmann | |
702 | 9bb69bb5 | Michael Hanselmann | known_groups = _LoadKnownGroups() |
703 | 9bb69bb5 | Michael Hanselmann | |
704 | 16e0b9c9 | Michael Hanselmann | # Check if node group is known
|
705 | 9bb69bb5 | Michael Hanselmann | if group_uuid not in known_groups: |
706 | 16e0b9c9 | Michael Hanselmann | raise errors.GenericError("Node group '%s' is not known by ssconf" % |
707 | 16e0b9c9 | Michael Hanselmann | group_uuid) |
708 | 16e0b9c9 | Michael Hanselmann | |
709 | 9bb69bb5 | Michael Hanselmann | # Group UUID has been verified and should not contain any dangerous characters
|
710 | 16e0b9c9 | Michael Hanselmann | state_path = constants.WATCHER_GROUP_STATE_FILE % group_uuid |
711 | 9bb69bb5 | Michael Hanselmann | inst_status_path = constants.WATCHER_GROUP_INSTANCE_STATUS_FILE % group_uuid |
712 | 16e0b9c9 | Michael Hanselmann | |
713 | 16e0b9c9 | Michael Hanselmann | logging.debug("Using state file %s", state_path)
|
714 | 16e0b9c9 | Michael Hanselmann | |
715 | 16e0b9c9 | Michael Hanselmann | # Global watcher
|
716 | b459a848 | Andrea Spadaccini | statefile = state.OpenStateFile(state_path) # pylint: disable=E0602
|
717 | 16e0b9c9 | Michael Hanselmann | if not statefile: |
718 | 16e0b9c9 | Michael Hanselmann | return constants.EXIT_FAILURE
|
719 | 16e0b9c9 | Michael Hanselmann | |
720 | b459a848 | Andrea Spadaccini | notepad = state.WatcherState(statefile) # pylint: disable=E0602
|
721 | 16e0b9c9 | Michael Hanselmann | try:
|
722 | 16e0b9c9 | Michael Hanselmann | # Connect to master daemon
|
723 | 16e0b9c9 | Michael Hanselmann | client = GetLuxiClient(False)
|
724 | 16e0b9c9 | Michael Hanselmann | |
725 | 16e0b9c9 | Michael Hanselmann | _CheckMaster(client) |
726 | 16e0b9c9 | Michael Hanselmann | |
727 | 16e0b9c9 | Michael Hanselmann | (nodes, instances) = _GetGroupData(client, group_uuid) |
728 | 16e0b9c9 | Michael Hanselmann | |
729 | 9bb69bb5 | Michael Hanselmann | # Update per-group instance status file
|
730 | 9bb69bb5 | Michael Hanselmann | _UpdateInstanceStatus(inst_status_path, instances.values()) |
731 | 9bb69bb5 | Michael Hanselmann | |
732 | 9bb69bb5 | Michael Hanselmann | _MergeInstanceStatus(constants.INSTANCE_STATUS_FILE, |
733 | 9bb69bb5 | Michael Hanselmann | constants.WATCHER_GROUP_INSTANCE_STATUS_FILE, |
734 | 9bb69bb5 | Michael Hanselmann | known_groups) |
735 | 9bb69bb5 | Michael Hanselmann | |
736 | 16e0b9c9 | Michael Hanselmann | started = _CheckInstances(client, notepad, instances) |
737 | 16e0b9c9 | Michael Hanselmann | _CheckDisks(client, notepad, nodes, instances, started) |
738 | 16e0b9c9 | Michael Hanselmann | _VerifyDisks(client, group_uuid, nodes, instances) |
739 | 16e0b9c9 | Michael Hanselmann | except Exception, err: |
740 | 16e0b9c9 | Michael Hanselmann | logging.info("Not updating status file due to failure: %s", err)
|
741 | 16e0b9c9 | Michael Hanselmann | raise
|
742 | 16e0b9c9 | Michael Hanselmann | else:
|
743 | 16e0b9c9 | Michael Hanselmann | # Save changes for next run
|
744 | 16e0b9c9 | Michael Hanselmann | notepad.Save(state_path) |
745 | 16e0b9c9 | Michael Hanselmann | |
746 | 16e0b9c9 | Michael Hanselmann | return constants.EXIT_SUCCESS
|
747 | 16e0b9c9 | Michael Hanselmann | |
748 | 16e0b9c9 | Michael Hanselmann | |
749 | 9f4bb951 | Michael Hanselmann | def Main(): |
750 | a8083063 | Iustin Pop | """Main function.
|
751 | a8083063 | Iustin Pop |
|
752 | a8083063 | Iustin Pop | """
|
753 | f0a80b01 | Michael Hanselmann | (options, _) = ParseOptions() |
754 | a8083063 | Iustin Pop | |
755 | cfcc79c6 | Michael Hanselmann | utils.SetupLogging(constants.LOG_WATCHER, sys.argv[0],
|
756 | cfcc79c6 | Michael Hanselmann | debug=options.debug, stderr_logging=options.debug) |
757 | a8083063 | Iustin Pop | |
758 | 46c8a6ab | Iustin Pop | if ShouldPause() and not options.ignore_pause: |
759 | 3753b2cb | Michael Hanselmann | logging.debug("Pause has been set, exiting")
|
760 | 9f4bb951 | Michael Hanselmann | return constants.EXIT_SUCCESS
|
761 | 3753b2cb | Michael Hanselmann | |
762 | 16e0b9c9 | Michael Hanselmann | # Try to acquire global watcher lock in shared mode
|
763 | 16e0b9c9 | Michael Hanselmann | lock = utils.FileLock.Open(constants.WATCHER_LOCK_FILE) |
764 | a8083063 | Iustin Pop | try:
|
765 | 16e0b9c9 | Michael Hanselmann | lock.Shared(blocking=False)
|
766 | 16e0b9c9 | Michael Hanselmann | except (EnvironmentError, errors.LockError), err: |
767 | 16e0b9c9 | Michael Hanselmann | logging.error("Can't acquire lock on %s: %s",
|
768 | 16e0b9c9 | Michael Hanselmann | constants.WATCHER_LOCK_FILE, err) |
769 | 16e0b9c9 | Michael Hanselmann | return constants.EXIT_SUCCESS
|
770 | db147305 | Tom Limoncelli | |
771 | 16e0b9c9 | Michael Hanselmann | if options.nodegroup is None: |
772 | 16e0b9c9 | Michael Hanselmann | fn = _GlobalWatcher |
773 | 16e0b9c9 | Michael Hanselmann | else:
|
774 | 16e0b9c9 | Michael Hanselmann | # Per-nodegroup watcher
|
775 | 16e0b9c9 | Michael Hanselmann | fn = _GroupWatcher |
776 | 16e0b9c9 | Michael Hanselmann | |
777 | 16e0b9c9 | Michael Hanselmann | try:
|
778 | 16e0b9c9 | Michael Hanselmann | return fn(options)
|
779 | 16e0b9c9 | Michael Hanselmann | except (SystemExit, KeyboardInterrupt): |
780 | 1b052f42 | Michael Hanselmann | raise
|
781 | 38242904 | Iustin Pop | except NotMasterError:
|
782 | 438b45d4 | Michael Hanselmann | logging.debug("Not master, exiting")
|
783 | 9f4bb951 | Michael Hanselmann | return constants.EXIT_NOTMASTER
|
784 | 89e1fc26 | Iustin Pop | except errors.ResolverError, err:
|
785 | 013ce4ae | Michael Hanselmann | logging.error("Cannot resolve hostname '%s', exiting", err.args[0]) |
786 | 9f4bb951 | Michael Hanselmann | return constants.EXIT_NODESETUP_ERROR
|
787 | 24edc6d4 | Iustin Pop | except errors.JobQueueFull:
|
788 | 24edc6d4 | Iustin Pop | logging.error("Job queue is full, can't query cluster state")
|
789 | 24edc6d4 | Iustin Pop | except errors.JobQueueDrainError:
|
790 | 24edc6d4 | Iustin Pop | logging.error("Job queue is drained, can't maintain cluster state")
|
791 | 438b45d4 | Michael Hanselmann | except Exception, err: |
792 | 001b3825 | Michael Hanselmann | logging.exception(str(err))
|
793 | 9f4bb951 | Michael Hanselmann | return constants.EXIT_FAILURE
|
794 | 5a3103e9 | Michael Hanselmann | |
795 | 9f4bb951 | Michael Hanselmann | return constants.EXIT_SUCCESS |