root / lib / jqueue.py @ 8062638d
History | View | Annotate | Download (48.9 kB)
1 |
#
|
---|---|
2 |
#
|
3 |
|
4 |
# Copyright (C) 2006, 2007, 2008, 2009, 2010 Google Inc.
|
5 |
#
|
6 |
# This program is free software; you can redistribute it and/or modify
|
7 |
# it under the terms of the GNU General Public License as published by
|
8 |
# the Free Software Foundation; either version 2 of the License, or
|
9 |
# (at your option) any later version.
|
10 |
#
|
11 |
# This program is distributed in the hope that it will be useful, but
|
12 |
# WITHOUT ANY WARRANTY; without even the implied warranty of
|
13 |
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
14 |
# General Public License for more details.
|
15 |
#
|
16 |
# You should have received a copy of the GNU General Public License
|
17 |
# along with this program; if not, write to the Free Software
|
18 |
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
|
19 |
# 02110-1301, USA.
|
20 |
|
21 |
|
22 |
"""Module implementing the job queue handling.
|
23 |
|
24 |
Locking: there's a single, large lock in the L{JobQueue} class. It's
|
25 |
used by all other classes in this module.
|
26 |
|
27 |
@var JOBQUEUE_THREADS: the number of worker threads we start for
|
28 |
processing jobs
|
29 |
|
30 |
"""
|
31 |
|
32 |
import os |
33 |
import logging |
34 |
import errno |
35 |
import re |
36 |
import time |
37 |
import weakref |
38 |
|
39 |
try:
|
40 |
# pylint: disable-msg=E0611
|
41 |
from pyinotify import pyinotify |
42 |
except ImportError: |
43 |
import pyinotify |
44 |
|
45 |
from ganeti import asyncnotifier |
46 |
from ganeti import constants |
47 |
from ganeti import serializer |
48 |
from ganeti import workerpool |
49 |
from ganeti import locking |
50 |
from ganeti import opcodes |
51 |
from ganeti import errors |
52 |
from ganeti import mcpu |
53 |
from ganeti import utils |
54 |
from ganeti import jstore |
55 |
from ganeti import rpc |
56 |
from ganeti import netutils |
57 |
from ganeti import compat |
58 |
|
59 |
|
60 |
JOBQUEUE_THREADS = 25
|
61 |
JOBS_PER_ARCHIVE_DIRECTORY = 10000
|
62 |
|
63 |
# member lock names to be passed to @ssynchronized decorator
|
64 |
_LOCK = "_lock"
|
65 |
_QUEUE = "_queue"
|
66 |
|
67 |
|
68 |
class CancelJob(Exception): |
69 |
"""Special exception to cancel a job.
|
70 |
|
71 |
"""
|
72 |
|
73 |
|
74 |
def TimeStampNow(): |
75 |
"""Returns the current timestamp.
|
76 |
|
77 |
@rtype: tuple
|
78 |
@return: the current time in the (seconds, microseconds) format
|
79 |
|
80 |
"""
|
81 |
return utils.SplitTime(time.time())
|
82 |
|
83 |
|
84 |
class _QueuedOpCode(object): |
85 |
"""Encapsulates an opcode object.
|
86 |
|
87 |
@ivar log: holds the execution log and consists of tuples
|
88 |
of the form C{(log_serial, timestamp, level, message)}
|
89 |
@ivar input: the OpCode we encapsulate
|
90 |
@ivar status: the current status
|
91 |
@ivar result: the result of the LU execution
|
92 |
@ivar start_timestamp: timestamp for the start of the execution
|
93 |
@ivar exec_timestamp: timestamp for the actual LU Exec() function invocation
|
94 |
@ivar stop_timestamp: timestamp for the end of the execution
|
95 |
|
96 |
"""
|
97 |
__slots__ = ["input", "status", "result", "log", |
98 |
"start_timestamp", "exec_timestamp", "end_timestamp", |
99 |
"__weakref__"]
|
100 |
|
101 |
def __init__(self, op): |
102 |
"""Constructor for the _QuededOpCode.
|
103 |
|
104 |
@type op: L{opcodes.OpCode}
|
105 |
@param op: the opcode we encapsulate
|
106 |
|
107 |
"""
|
108 |
self.input = op
|
109 |
self.status = constants.OP_STATUS_QUEUED
|
110 |
self.result = None |
111 |
self.log = []
|
112 |
self.start_timestamp = None |
113 |
self.exec_timestamp = None |
114 |
self.end_timestamp = None |
115 |
|
116 |
@classmethod
|
117 |
def Restore(cls, state): |
118 |
"""Restore the _QueuedOpCode from the serialized form.
|
119 |
|
120 |
@type state: dict
|
121 |
@param state: the serialized state
|
122 |
@rtype: _QueuedOpCode
|
123 |
@return: a new _QueuedOpCode instance
|
124 |
|
125 |
"""
|
126 |
obj = _QueuedOpCode.__new__(cls) |
127 |
obj.input = opcodes.OpCode.LoadOpCode(state["input"])
|
128 |
obj.status = state["status"]
|
129 |
obj.result = state["result"]
|
130 |
obj.log = state["log"]
|
131 |
obj.start_timestamp = state.get("start_timestamp", None) |
132 |
obj.exec_timestamp = state.get("exec_timestamp", None) |
133 |
obj.end_timestamp = state.get("end_timestamp", None) |
134 |
return obj
|
135 |
|
136 |
def Serialize(self): |
137 |
"""Serializes this _QueuedOpCode.
|
138 |
|
139 |
@rtype: dict
|
140 |
@return: the dictionary holding the serialized state
|
141 |
|
142 |
"""
|
143 |
return {
|
144 |
"input": self.input.__getstate__(), |
145 |
"status": self.status, |
146 |
"result": self.result, |
147 |
"log": self.log, |
148 |
"start_timestamp": self.start_timestamp, |
149 |
"exec_timestamp": self.exec_timestamp, |
150 |
"end_timestamp": self.end_timestamp, |
151 |
} |
152 |
|
153 |
|
154 |
class _QueuedJob(object): |
155 |
"""In-memory job representation.
|
156 |
|
157 |
This is what we use to track the user-submitted jobs. Locking must
|
158 |
be taken care of by users of this class.
|
159 |
|
160 |
@type queue: L{JobQueue}
|
161 |
@ivar queue: the parent queue
|
162 |
@ivar id: the job ID
|
163 |
@type ops: list
|
164 |
@ivar ops: the list of _QueuedOpCode that constitute the job
|
165 |
@type log_serial: int
|
166 |
@ivar log_serial: holds the index for the next log entry
|
167 |
@ivar received_timestamp: the timestamp for when the job was received
|
168 |
@ivar start_timestmap: the timestamp for start of execution
|
169 |
@ivar end_timestamp: the timestamp for end of execution
|
170 |
|
171 |
"""
|
172 |
# pylint: disable-msg=W0212
|
173 |
__slots__ = ["queue", "id", "ops", "log_serial", |
174 |
"received_timestamp", "start_timestamp", "end_timestamp", |
175 |
"__weakref__"]
|
176 |
|
177 |
def __init__(self, queue, job_id, ops): |
178 |
"""Constructor for the _QueuedJob.
|
179 |
|
180 |
@type queue: L{JobQueue}
|
181 |
@param queue: our parent queue
|
182 |
@type job_id: job_id
|
183 |
@param job_id: our job id
|
184 |
@type ops: list
|
185 |
@param ops: the list of opcodes we hold, which will be encapsulated
|
186 |
in _QueuedOpCodes
|
187 |
|
188 |
"""
|
189 |
if not ops: |
190 |
raise errors.GenericError("A job needs at least one opcode") |
191 |
|
192 |
self.queue = queue
|
193 |
self.id = job_id
|
194 |
self.ops = [_QueuedOpCode(op) for op in ops] |
195 |
self.log_serial = 0 |
196 |
self.received_timestamp = TimeStampNow()
|
197 |
self.start_timestamp = None |
198 |
self.end_timestamp = None |
199 |
|
200 |
def __repr__(self): |
201 |
status = ["%s.%s" % (self.__class__.__module__, self.__class__.__name__), |
202 |
"id=%s" % self.id, |
203 |
"ops=%s" % ",".join([op.input.Summary() for op in self.ops])] |
204 |
|
205 |
return "<%s at %#x>" % (" ".join(status), id(self)) |
206 |
|
207 |
@classmethod
|
208 |
def Restore(cls, queue, state): |
209 |
"""Restore a _QueuedJob from serialized state:
|
210 |
|
211 |
@type queue: L{JobQueue}
|
212 |
@param queue: to which queue the restored job belongs
|
213 |
@type state: dict
|
214 |
@param state: the serialized state
|
215 |
@rtype: _JobQueue
|
216 |
@return: the restored _JobQueue instance
|
217 |
|
218 |
"""
|
219 |
obj = _QueuedJob.__new__(cls) |
220 |
obj.queue = queue |
221 |
obj.id = state["id"]
|
222 |
obj.received_timestamp = state.get("received_timestamp", None) |
223 |
obj.start_timestamp = state.get("start_timestamp", None) |
224 |
obj.end_timestamp = state.get("end_timestamp", None) |
225 |
|
226 |
obj.ops = [] |
227 |
obj.log_serial = 0
|
228 |
for op_state in state["ops"]: |
229 |
op = _QueuedOpCode.Restore(op_state) |
230 |
for log_entry in op.log: |
231 |
obj.log_serial = max(obj.log_serial, log_entry[0]) |
232 |
obj.ops.append(op) |
233 |
|
234 |
return obj
|
235 |
|
236 |
def Serialize(self): |
237 |
"""Serialize the _JobQueue instance.
|
238 |
|
239 |
@rtype: dict
|
240 |
@return: the serialized state
|
241 |
|
242 |
"""
|
243 |
return {
|
244 |
"id": self.id, |
245 |
"ops": [op.Serialize() for op in self.ops], |
246 |
"start_timestamp": self.start_timestamp, |
247 |
"end_timestamp": self.end_timestamp, |
248 |
"received_timestamp": self.received_timestamp, |
249 |
} |
250 |
|
251 |
def CalcStatus(self): |
252 |
"""Compute the status of this job.
|
253 |
|
254 |
This function iterates over all the _QueuedOpCodes in the job and
|
255 |
based on their status, computes the job status.
|
256 |
|
257 |
The algorithm is:
|
258 |
- if we find a cancelled, or finished with error, the job
|
259 |
status will be the same
|
260 |
- otherwise, the last opcode with the status one of:
|
261 |
- waitlock
|
262 |
- canceling
|
263 |
- running
|
264 |
|
265 |
will determine the job status
|
266 |
|
267 |
- otherwise, it means either all opcodes are queued, or success,
|
268 |
and the job status will be the same
|
269 |
|
270 |
@return: the job status
|
271 |
|
272 |
"""
|
273 |
status = constants.JOB_STATUS_QUEUED |
274 |
|
275 |
all_success = True
|
276 |
for op in self.ops: |
277 |
if op.status == constants.OP_STATUS_SUCCESS:
|
278 |
continue
|
279 |
|
280 |
all_success = False
|
281 |
|
282 |
if op.status == constants.OP_STATUS_QUEUED:
|
283 |
pass
|
284 |
elif op.status == constants.OP_STATUS_WAITLOCK:
|
285 |
status = constants.JOB_STATUS_WAITLOCK |
286 |
elif op.status == constants.OP_STATUS_RUNNING:
|
287 |
status = constants.JOB_STATUS_RUNNING |
288 |
elif op.status == constants.OP_STATUS_CANCELING:
|
289 |
status = constants.JOB_STATUS_CANCELING |
290 |
break
|
291 |
elif op.status == constants.OP_STATUS_ERROR:
|
292 |
status = constants.JOB_STATUS_ERROR |
293 |
# The whole job fails if one opcode failed
|
294 |
break
|
295 |
elif op.status == constants.OP_STATUS_CANCELED:
|
296 |
status = constants.OP_STATUS_CANCELED |
297 |
break
|
298 |
|
299 |
if all_success:
|
300 |
status = constants.JOB_STATUS_SUCCESS |
301 |
|
302 |
return status
|
303 |
|
304 |
def GetLogEntries(self, newer_than): |
305 |
"""Selectively returns the log entries.
|
306 |
|
307 |
@type newer_than: None or int
|
308 |
@param newer_than: if this is None, return all log entries,
|
309 |
otherwise return only the log entries with serial higher
|
310 |
than this value
|
311 |
@rtype: list
|
312 |
@return: the list of the log entries selected
|
313 |
|
314 |
"""
|
315 |
if newer_than is None: |
316 |
serial = -1
|
317 |
else:
|
318 |
serial = newer_than |
319 |
|
320 |
entries = [] |
321 |
for op in self.ops: |
322 |
entries.extend(filter(lambda entry: entry[0] > serial, op.log)) |
323 |
|
324 |
return entries
|
325 |
|
326 |
def GetInfo(self, fields): |
327 |
"""Returns information about a job.
|
328 |
|
329 |
@type fields: list
|
330 |
@param fields: names of fields to return
|
331 |
@rtype: list
|
332 |
@return: list with one element for each field
|
333 |
@raise errors.OpExecError: when an invalid field
|
334 |
has been passed
|
335 |
|
336 |
"""
|
337 |
row = [] |
338 |
for fname in fields: |
339 |
if fname == "id": |
340 |
row.append(self.id)
|
341 |
elif fname == "status": |
342 |
row.append(self.CalcStatus())
|
343 |
elif fname == "ops": |
344 |
row.append([op.input.__getstate__() for op in self.ops]) |
345 |
elif fname == "opresult": |
346 |
row.append([op.result for op in self.ops]) |
347 |
elif fname == "opstatus": |
348 |
row.append([op.status for op in self.ops]) |
349 |
elif fname == "oplog": |
350 |
row.append([op.log for op in self.ops]) |
351 |
elif fname == "opstart": |
352 |
row.append([op.start_timestamp for op in self.ops]) |
353 |
elif fname == "opexec": |
354 |
row.append([op.exec_timestamp for op in self.ops]) |
355 |
elif fname == "opend": |
356 |
row.append([op.end_timestamp for op in self.ops]) |
357 |
elif fname == "received_ts": |
358 |
row.append(self.received_timestamp)
|
359 |
elif fname == "start_ts": |
360 |
row.append(self.start_timestamp)
|
361 |
elif fname == "end_ts": |
362 |
row.append(self.end_timestamp)
|
363 |
elif fname == "summary": |
364 |
row.append([op.input.Summary() for op in self.ops]) |
365 |
else:
|
366 |
raise errors.OpExecError("Invalid self query field '%s'" % fname) |
367 |
return row
|
368 |
|
369 |
def MarkUnfinishedOps(self, status, result): |
370 |
"""Mark unfinished opcodes with a given status and result.
|
371 |
|
372 |
This is an utility function for marking all running or waiting to
|
373 |
be run opcodes with a given status. Opcodes which are already
|
374 |
finalised are not changed.
|
375 |
|
376 |
@param status: a given opcode status
|
377 |
@param result: the opcode result
|
378 |
|
379 |
"""
|
380 |
try:
|
381 |
not_marked = True
|
382 |
for op in self.ops: |
383 |
if op.status in constants.OPS_FINALIZED: |
384 |
assert not_marked, "Finalized opcodes found after non-finalized ones" |
385 |
continue
|
386 |
op.status = status |
387 |
op.result = result |
388 |
not_marked = False
|
389 |
finally:
|
390 |
self.queue.UpdateJobUnlocked(self) |
391 |
|
392 |
|
393 |
class _OpExecCallbacks(mcpu.OpExecCbBase): |
394 |
def __init__(self, queue, job, op): |
395 |
"""Initializes this class.
|
396 |
|
397 |
@type queue: L{JobQueue}
|
398 |
@param queue: Job queue
|
399 |
@type job: L{_QueuedJob}
|
400 |
@param job: Job object
|
401 |
@type op: L{_QueuedOpCode}
|
402 |
@param op: OpCode
|
403 |
|
404 |
"""
|
405 |
assert queue, "Queue is missing" |
406 |
assert job, "Job is missing" |
407 |
assert op, "Opcode is missing" |
408 |
|
409 |
self._queue = queue
|
410 |
self._job = job
|
411 |
self._op = op
|
412 |
|
413 |
def _CheckCancel(self): |
414 |
"""Raises an exception to cancel the job if asked to.
|
415 |
|
416 |
"""
|
417 |
# Cancel here if we were asked to
|
418 |
if self._op.status == constants.OP_STATUS_CANCELING: |
419 |
logging.debug("Canceling opcode")
|
420 |
raise CancelJob()
|
421 |
|
422 |
@locking.ssynchronized(_QUEUE, shared=1) |
423 |
def NotifyStart(self): |
424 |
"""Mark the opcode as running, not lock-waiting.
|
425 |
|
426 |
This is called from the mcpu code as a notifier function, when the LU is
|
427 |
finally about to start the Exec() method. Of course, to have end-user
|
428 |
visible results, the opcode must be initially (before calling into
|
429 |
Processor.ExecOpCode) set to OP_STATUS_WAITLOCK.
|
430 |
|
431 |
"""
|
432 |
assert self._op in self._job.ops |
433 |
assert self._op.status in (constants.OP_STATUS_WAITLOCK, |
434 |
constants.OP_STATUS_CANCELING) |
435 |
|
436 |
# Cancel here if we were asked to
|
437 |
self._CheckCancel()
|
438 |
|
439 |
logging.debug("Opcode is now running")
|
440 |
|
441 |
self._op.status = constants.OP_STATUS_RUNNING
|
442 |
self._op.exec_timestamp = TimeStampNow()
|
443 |
|
444 |
# And finally replicate the job status
|
445 |
self._queue.UpdateJobUnlocked(self._job) |
446 |
|
447 |
@locking.ssynchronized(_QUEUE, shared=1) |
448 |
def _AppendFeedback(self, timestamp, log_type, log_msg): |
449 |
"""Internal feedback append function, with locks
|
450 |
|
451 |
"""
|
452 |
self._job.log_serial += 1 |
453 |
self._op.log.append((self._job.log_serial, timestamp, log_type, log_msg)) |
454 |
self._queue.UpdateJobUnlocked(self._job, replicate=False) |
455 |
|
456 |
def Feedback(self, *args): |
457 |
"""Append a log entry.
|
458 |
|
459 |
"""
|
460 |
assert len(args) < 3 |
461 |
|
462 |
if len(args) == 1: |
463 |
log_type = constants.ELOG_MESSAGE |
464 |
log_msg = args[0]
|
465 |
else:
|
466 |
(log_type, log_msg) = args |
467 |
|
468 |
# The time is split to make serialization easier and not lose
|
469 |
# precision.
|
470 |
timestamp = utils.SplitTime(time.time()) |
471 |
self._AppendFeedback(timestamp, log_type, log_msg)
|
472 |
|
473 |
def ReportLocks(self, msg): |
474 |
"""Write locking information to the job.
|
475 |
|
476 |
Called whenever the LU processor is waiting for a lock or has acquired one.
|
477 |
|
478 |
"""
|
479 |
assert self._op.status in (constants.OP_STATUS_WAITLOCK, |
480 |
constants.OP_STATUS_CANCELING) |
481 |
|
482 |
# Cancel here if we were asked to
|
483 |
self._CheckCancel()
|
484 |
|
485 |
|
486 |
class _JobChangesChecker(object): |
487 |
def __init__(self, fields, prev_job_info, prev_log_serial): |
488 |
"""Initializes this class.
|
489 |
|
490 |
@type fields: list of strings
|
491 |
@param fields: Fields requested by LUXI client
|
492 |
@type prev_job_info: string
|
493 |
@param prev_job_info: previous job info, as passed by the LUXI client
|
494 |
@type prev_log_serial: string
|
495 |
@param prev_log_serial: previous job serial, as passed by the LUXI client
|
496 |
|
497 |
"""
|
498 |
self._fields = fields
|
499 |
self._prev_job_info = prev_job_info
|
500 |
self._prev_log_serial = prev_log_serial
|
501 |
|
502 |
def __call__(self, job): |
503 |
"""Checks whether job has changed.
|
504 |
|
505 |
@type job: L{_QueuedJob}
|
506 |
@param job: Job object
|
507 |
|
508 |
"""
|
509 |
status = job.CalcStatus() |
510 |
job_info = job.GetInfo(self._fields)
|
511 |
log_entries = job.GetLogEntries(self._prev_log_serial)
|
512 |
|
513 |
# Serializing and deserializing data can cause type changes (e.g. from
|
514 |
# tuple to list) or precision loss. We're doing it here so that we get
|
515 |
# the same modifications as the data received from the client. Without
|
516 |
# this, the comparison afterwards might fail without the data being
|
517 |
# significantly different.
|
518 |
# TODO: we just deserialized from disk, investigate how to make sure that
|
519 |
# the job info and log entries are compatible to avoid this further step.
|
520 |
# TODO: Doing something like in testutils.py:UnifyValueType might be more
|
521 |
# efficient, though floats will be tricky
|
522 |
job_info = serializer.LoadJson(serializer.DumpJson(job_info)) |
523 |
log_entries = serializer.LoadJson(serializer.DumpJson(log_entries)) |
524 |
|
525 |
# Don't even try to wait if the job is no longer running, there will be
|
526 |
# no changes.
|
527 |
if (status not in (constants.JOB_STATUS_QUEUED, |
528 |
constants.JOB_STATUS_RUNNING, |
529 |
constants.JOB_STATUS_WAITLOCK) or
|
530 |
job_info != self._prev_job_info or |
531 |
(log_entries and self._prev_log_serial != log_entries[0][0])): |
532 |
logging.debug("Job %s changed", job.id)
|
533 |
return (job_info, log_entries)
|
534 |
|
535 |
return None |
536 |
|
537 |
|
538 |
class _JobFileChangesWaiter(object): |
539 |
def __init__(self, filename): |
540 |
"""Initializes this class.
|
541 |
|
542 |
@type filename: string
|
543 |
@param filename: Path to job file
|
544 |
@raises errors.InotifyError: if the notifier cannot be setup
|
545 |
|
546 |
"""
|
547 |
self._wm = pyinotify.WatchManager()
|
548 |
self._inotify_handler = \
|
549 |
asyncnotifier.SingleFileEventHandler(self._wm, self._OnInotify, filename) |
550 |
self._notifier = \
|
551 |
pyinotify.Notifier(self._wm, default_proc_fun=self._inotify_handler) |
552 |
try:
|
553 |
self._inotify_handler.enable()
|
554 |
except Exception: |
555 |
# pyinotify doesn't close file descriptors automatically
|
556 |
self._notifier.stop()
|
557 |
raise
|
558 |
|
559 |
def _OnInotify(self, notifier_enabled): |
560 |
"""Callback for inotify.
|
561 |
|
562 |
"""
|
563 |
if not notifier_enabled: |
564 |
self._inotify_handler.enable()
|
565 |
|
566 |
def Wait(self, timeout): |
567 |
"""Waits for the job file to change.
|
568 |
|
569 |
@type timeout: float
|
570 |
@param timeout: Timeout in seconds
|
571 |
@return: Whether there have been events
|
572 |
|
573 |
"""
|
574 |
assert timeout >= 0 |
575 |
have_events = self._notifier.check_events(timeout * 1000) |
576 |
if have_events:
|
577 |
self._notifier.read_events()
|
578 |
self._notifier.process_events()
|
579 |
return have_events
|
580 |
|
581 |
def Close(self): |
582 |
"""Closes underlying notifier and its file descriptor.
|
583 |
|
584 |
"""
|
585 |
self._notifier.stop()
|
586 |
|
587 |
|
588 |
class _JobChangesWaiter(object): |
589 |
def __init__(self, filename): |
590 |
"""Initializes this class.
|
591 |
|
592 |
@type filename: string
|
593 |
@param filename: Path to job file
|
594 |
|
595 |
"""
|
596 |
self._filewaiter = None |
597 |
self._filename = filename
|
598 |
|
599 |
def Wait(self, timeout): |
600 |
"""Waits for a job to change.
|
601 |
|
602 |
@type timeout: float
|
603 |
@param timeout: Timeout in seconds
|
604 |
@return: Whether there have been events
|
605 |
|
606 |
"""
|
607 |
if self._filewaiter: |
608 |
return self._filewaiter.Wait(timeout) |
609 |
|
610 |
# Lazy setup: Avoid inotify setup cost when job file has already changed.
|
611 |
# If this point is reached, return immediately and let caller check the job
|
612 |
# file again in case there were changes since the last check. This avoids a
|
613 |
# race condition.
|
614 |
self._filewaiter = _JobFileChangesWaiter(self._filename) |
615 |
|
616 |
return True |
617 |
|
618 |
def Close(self): |
619 |
"""Closes underlying waiter.
|
620 |
|
621 |
"""
|
622 |
if self._filewaiter: |
623 |
self._filewaiter.Close()
|
624 |
|
625 |
|
626 |
class _WaitForJobChangesHelper(object): |
627 |
"""Helper class using inotify to wait for changes in a job file.
|
628 |
|
629 |
This class takes a previous job status and serial, and alerts the client when
|
630 |
the current job status has changed.
|
631 |
|
632 |
"""
|
633 |
@staticmethod
|
634 |
def _CheckForChanges(job_load_fn, check_fn): |
635 |
job = job_load_fn() |
636 |
if not job: |
637 |
raise errors.JobLost()
|
638 |
|
639 |
result = check_fn(job) |
640 |
if result is None: |
641 |
raise utils.RetryAgain()
|
642 |
|
643 |
return result
|
644 |
|
645 |
def __call__(self, filename, job_load_fn, |
646 |
fields, prev_job_info, prev_log_serial, timeout): |
647 |
"""Waits for changes on a job.
|
648 |
|
649 |
@type filename: string
|
650 |
@param filename: File on which to wait for changes
|
651 |
@type job_load_fn: callable
|
652 |
@param job_load_fn: Function to load job
|
653 |
@type fields: list of strings
|
654 |
@param fields: Which fields to check for changes
|
655 |
@type prev_job_info: list or None
|
656 |
@param prev_job_info: Last job information returned
|
657 |
@type prev_log_serial: int
|
658 |
@param prev_log_serial: Last job message serial number
|
659 |
@type timeout: float
|
660 |
@param timeout: maximum time to wait in seconds
|
661 |
|
662 |
"""
|
663 |
try:
|
664 |
check_fn = _JobChangesChecker(fields, prev_job_info, prev_log_serial) |
665 |
waiter = _JobChangesWaiter(filename) |
666 |
try:
|
667 |
return utils.Retry(compat.partial(self._CheckForChanges, |
668 |
job_load_fn, check_fn), |
669 |
utils.RETRY_REMAINING_TIME, timeout, |
670 |
wait_fn=waiter.Wait) |
671 |
finally:
|
672 |
waiter.Close() |
673 |
except (errors.InotifyError, errors.JobLost):
|
674 |
return None |
675 |
except utils.RetryTimeout:
|
676 |
return constants.JOB_NOTCHANGED
|
677 |
|
678 |
|
679 |
class _JobQueueWorker(workerpool.BaseWorker): |
680 |
"""The actual job workers.
|
681 |
|
682 |
"""
|
683 |
def RunTask(self, job): # pylint: disable-msg=W0221 |
684 |
"""Job executor.
|
685 |
|
686 |
This functions processes a job. It is closely tied to the _QueuedJob and
|
687 |
_QueuedOpCode classes.
|
688 |
|
689 |
@type job: L{_QueuedJob}
|
690 |
@param job: the job to be processed
|
691 |
|
692 |
"""
|
693 |
self.SetTaskName("Job%s" % job.id) |
694 |
|
695 |
logging.info("Processing job %s", job.id)
|
696 |
proc = mcpu.Processor(self.pool.queue.context, job.id)
|
697 |
queue = job.queue |
698 |
try:
|
699 |
try:
|
700 |
count = len(job.ops)
|
701 |
for idx, op in enumerate(job.ops): |
702 |
op_summary = op.input.Summary() |
703 |
if op.status == constants.OP_STATUS_SUCCESS:
|
704 |
# this is a job that was partially completed before master
|
705 |
# daemon shutdown, so it can be expected that some opcodes
|
706 |
# are already completed successfully (if any did error
|
707 |
# out, then the whole job should have been aborted and not
|
708 |
# resubmitted for processing)
|
709 |
logging.info("Op %s/%s: opcode %s already processed, skipping",
|
710 |
idx + 1, count, op_summary)
|
711 |
continue
|
712 |
try:
|
713 |
logging.info("Op %s/%s: Starting opcode %s", idx + 1, count, |
714 |
op_summary) |
715 |
|
716 |
queue.acquire(shared=1)
|
717 |
try:
|
718 |
if op.status == constants.OP_STATUS_CANCELED:
|
719 |
logging.debug("Canceling opcode")
|
720 |
raise CancelJob()
|
721 |
assert op.status == constants.OP_STATUS_QUEUED
|
722 |
logging.debug("Opcode %s/%s waiting for locks",
|
723 |
idx + 1, count)
|
724 |
op.status = constants.OP_STATUS_WAITLOCK |
725 |
op.result = None
|
726 |
op.start_timestamp = TimeStampNow() |
727 |
if idx == 0: # first opcode |
728 |
job.start_timestamp = op.start_timestamp |
729 |
queue.UpdateJobUnlocked(job) |
730 |
|
731 |
input_opcode = op.input |
732 |
finally:
|
733 |
queue.release() |
734 |
|
735 |
# Make sure not to hold queue lock while calling ExecOpCode
|
736 |
result = proc.ExecOpCode(input_opcode, |
737 |
_OpExecCallbacks(queue, job, op)) |
738 |
|
739 |
queue.acquire(shared=1)
|
740 |
try:
|
741 |
logging.debug("Opcode %s/%s succeeded", idx + 1, count) |
742 |
op.status = constants.OP_STATUS_SUCCESS |
743 |
op.result = result |
744 |
op.end_timestamp = TimeStampNow() |
745 |
if idx == count - 1: |
746 |
job.end_timestamp = TimeStampNow() |
747 |
|
748 |
# Consistency check
|
749 |
assert compat.all(i.status == constants.OP_STATUS_SUCCESS
|
750 |
for i in job.ops) |
751 |
|
752 |
queue.UpdateJobUnlocked(job) |
753 |
finally:
|
754 |
queue.release() |
755 |
|
756 |
logging.info("Op %s/%s: Successfully finished opcode %s",
|
757 |
idx + 1, count, op_summary)
|
758 |
except CancelJob:
|
759 |
# Will be handled further up
|
760 |
raise
|
761 |
except Exception, err: |
762 |
queue.acquire(shared=1)
|
763 |
try:
|
764 |
try:
|
765 |
logging.debug("Opcode %s/%s failed", idx + 1, count) |
766 |
op.status = constants.OP_STATUS_ERROR |
767 |
if isinstance(err, errors.GenericError): |
768 |
to_encode = err |
769 |
else:
|
770 |
to_encode = errors.OpExecError(str(err))
|
771 |
op.result = errors.EncodeException(to_encode) |
772 |
op.end_timestamp = TimeStampNow() |
773 |
logging.info("Op %s/%s: Error in opcode %s: %s",
|
774 |
idx + 1, count, op_summary, err)
|
775 |
|
776 |
to_encode = errors.OpExecError("Preceding opcode failed")
|
777 |
job.MarkUnfinishedOps(constants.OP_STATUS_ERROR, |
778 |
errors.EncodeException(to_encode)) |
779 |
|
780 |
# Consistency check
|
781 |
assert compat.all(i.status == constants.OP_STATUS_SUCCESS
|
782 |
for i in job.ops[:idx]) |
783 |
assert compat.all(i.status == constants.OP_STATUS_ERROR and |
784 |
errors.GetEncodedError(i.result) |
785 |
for i in job.ops[idx:]) |
786 |
finally:
|
787 |
job.end_timestamp = TimeStampNow() |
788 |
queue.UpdateJobUnlocked(job) |
789 |
finally:
|
790 |
queue.release() |
791 |
raise
|
792 |
|
793 |
except CancelJob:
|
794 |
queue.acquire(shared=1)
|
795 |
try:
|
796 |
job.MarkUnfinishedOps(constants.OP_STATUS_CANCELED, |
797 |
"Job canceled by request")
|
798 |
job.end_timestamp = TimeStampNow() |
799 |
queue.UpdateJobUnlocked(job) |
800 |
finally:
|
801 |
queue.release() |
802 |
except errors.GenericError, err:
|
803 |
logging.exception("Ganeti exception")
|
804 |
except:
|
805 |
logging.exception("Unhandled exception")
|
806 |
finally:
|
807 |
status = job.CalcStatus() |
808 |
logging.info("Finished job %s, status = %s", job.id, status)
|
809 |
|
810 |
|
811 |
class _JobQueueWorkerPool(workerpool.WorkerPool): |
812 |
"""Simple class implementing a job-processing workerpool.
|
813 |
|
814 |
"""
|
815 |
def __init__(self, queue): |
816 |
super(_JobQueueWorkerPool, self).__init__("JobQueue", |
817 |
JOBQUEUE_THREADS, |
818 |
_JobQueueWorker) |
819 |
self.queue = queue
|
820 |
|
821 |
|
822 |
def _RequireOpenQueue(fn): |
823 |
"""Decorator for "public" functions.
|
824 |
|
825 |
This function should be used for all 'public' functions. That is,
|
826 |
functions usually called from other classes. Note that this should
|
827 |
be applied only to methods (not plain functions), since it expects
|
828 |
that the decorated function is called with a first argument that has
|
829 |
a '_queue_filelock' argument.
|
830 |
|
831 |
@warning: Use this decorator only after locking.ssynchronized
|
832 |
|
833 |
Example::
|
834 |
@locking.ssynchronized(_LOCK)
|
835 |
@_RequireOpenQueue
|
836 |
def Example(self):
|
837 |
pass
|
838 |
|
839 |
"""
|
840 |
def wrapper(self, *args, **kwargs): |
841 |
# pylint: disable-msg=W0212
|
842 |
assert self._queue_filelock is not None, "Queue should be open" |
843 |
return fn(self, *args, **kwargs) |
844 |
return wrapper
|
845 |
|
846 |
|
847 |
class JobQueue(object): |
848 |
"""Queue used to manage the jobs.
|
849 |
|
850 |
@cvar _RE_JOB_FILE: regex matching the valid job file names
|
851 |
|
852 |
"""
|
853 |
_RE_JOB_FILE = re.compile(r"^job-(%s)$" % constants.JOB_ID_TEMPLATE)
|
854 |
|
855 |
def __init__(self, context): |
856 |
"""Constructor for JobQueue.
|
857 |
|
858 |
The constructor will initialize the job queue object and then
|
859 |
start loading the current jobs from disk, either for starting them
|
860 |
(if they were queue) or for aborting them (if they were already
|
861 |
running).
|
862 |
|
863 |
@type context: GanetiContext
|
864 |
@param context: the context object for access to the configuration
|
865 |
data and other ganeti objects
|
866 |
|
867 |
"""
|
868 |
self.context = context
|
869 |
self._memcache = weakref.WeakValueDictionary()
|
870 |
self._my_hostname = netutils.Hostname.GetSysName()
|
871 |
|
872 |
# The Big JobQueue lock. If a code block or method acquires it in shared
|
873 |
# mode safe it must guarantee concurrency with all the code acquiring it in
|
874 |
# shared mode, including itself. In order not to acquire it at all
|
875 |
# concurrency must be guaranteed with all code acquiring it in shared mode
|
876 |
# and all code acquiring it exclusively.
|
877 |
self._lock = locking.SharedLock("JobQueue") |
878 |
|
879 |
self.acquire = self._lock.acquire |
880 |
self.release = self._lock.release |
881 |
|
882 |
# Initialize the queue, and acquire the filelock.
|
883 |
# This ensures no other process is working on the job queue.
|
884 |
self._queue_filelock = jstore.InitAndVerifyQueue(must_lock=True) |
885 |
|
886 |
# Read serial file
|
887 |
self._last_serial = jstore.ReadSerial()
|
888 |
assert self._last_serial is not None, ("Serial file was modified between" |
889 |
" check in jstore and here")
|
890 |
|
891 |
# Get initial list of nodes
|
892 |
self._nodes = dict((n.name, n.primary_ip) |
893 |
for n in self.context.cfg.GetAllNodesInfo().values() |
894 |
if n.master_candidate)
|
895 |
|
896 |
# Remove master node
|
897 |
self._nodes.pop(self._my_hostname, None) |
898 |
|
899 |
# TODO: Check consistency across nodes
|
900 |
|
901 |
self._queue_size = 0 |
902 |
self._UpdateQueueSizeUnlocked()
|
903 |
self._drained = self._IsQueueMarkedDrain() |
904 |
|
905 |
# Setup worker pool
|
906 |
self._wpool = _JobQueueWorkerPool(self) |
907 |
try:
|
908 |
# We need to lock here because WorkerPool.AddTask() may start a job while
|
909 |
# we're still doing our work.
|
910 |
self.acquire()
|
911 |
try:
|
912 |
logging.info("Inspecting job queue")
|
913 |
|
914 |
all_job_ids = self._GetJobIDsUnlocked()
|
915 |
jobs_count = len(all_job_ids)
|
916 |
lastinfo = time.time() |
917 |
for idx, job_id in enumerate(all_job_ids): |
918 |
# Give an update every 1000 jobs or 10 seconds
|
919 |
if (idx % 1000 == 0 or time.time() >= (lastinfo + 10.0) or |
920 |
idx == (jobs_count - 1)):
|
921 |
logging.info("Job queue inspection: %d/%d (%0.1f %%)",
|
922 |
idx, jobs_count - 1, 100.0 * (idx + 1) / jobs_count) |
923 |
lastinfo = time.time() |
924 |
|
925 |
job = self._LoadJobUnlocked(job_id)
|
926 |
|
927 |
# a failure in loading the job can cause 'None' to be returned
|
928 |
if job is None: |
929 |
continue
|
930 |
|
931 |
status = job.CalcStatus() |
932 |
|
933 |
if status in (constants.JOB_STATUS_QUEUED, ): |
934 |
self._wpool.AddTask((job, ))
|
935 |
|
936 |
elif status in (constants.JOB_STATUS_RUNNING, |
937 |
constants.JOB_STATUS_WAITLOCK, |
938 |
constants.JOB_STATUS_CANCELING): |
939 |
logging.warning("Unfinished job %s found: %s", job.id, job)
|
940 |
job.MarkUnfinishedOps(constants.OP_STATUS_ERROR, |
941 |
"Unclean master daemon shutdown")
|
942 |
|
943 |
logging.info("Job queue inspection finished")
|
944 |
finally:
|
945 |
self.release()
|
946 |
except:
|
947 |
self._wpool.TerminateWorkers()
|
948 |
raise
|
949 |
|
950 |
@locking.ssynchronized(_LOCK)
|
951 |
@_RequireOpenQueue
|
952 |
def AddNode(self, node): |
953 |
"""Register a new node with the queue.
|
954 |
|
955 |
@type node: L{objects.Node}
|
956 |
@param node: the node object to be added
|
957 |
|
958 |
"""
|
959 |
node_name = node.name |
960 |
assert node_name != self._my_hostname |
961 |
|
962 |
# Clean queue directory on added node
|
963 |
result = rpc.RpcRunner.call_jobqueue_purge(node_name) |
964 |
msg = result.fail_msg |
965 |
if msg:
|
966 |
logging.warning("Cannot cleanup queue directory on node %s: %s",
|
967 |
node_name, msg) |
968 |
|
969 |
if not node.master_candidate: |
970 |
# remove if existing, ignoring errors
|
971 |
self._nodes.pop(node_name, None) |
972 |
# and skip the replication of the job ids
|
973 |
return
|
974 |
|
975 |
# Upload the whole queue excluding archived jobs
|
976 |
files = [self._GetJobPath(job_id) for job_id in self._GetJobIDsUnlocked()] |
977 |
|
978 |
# Upload current serial file
|
979 |
files.append(constants.JOB_QUEUE_SERIAL_FILE) |
980 |
|
981 |
for file_name in files: |
982 |
# Read file content
|
983 |
content = utils.ReadFile(file_name) |
984 |
|
985 |
result = rpc.RpcRunner.call_jobqueue_update([node_name], |
986 |
[node.primary_ip], |
987 |
file_name, content) |
988 |
msg = result[node_name].fail_msg |
989 |
if msg:
|
990 |
logging.error("Failed to upload file %s to node %s: %s",
|
991 |
file_name, node_name, msg) |
992 |
|
993 |
self._nodes[node_name] = node.primary_ip
|
994 |
|
995 |
@locking.ssynchronized(_LOCK)
|
996 |
@_RequireOpenQueue
|
997 |
def RemoveNode(self, node_name): |
998 |
"""Callback called when removing nodes from the cluster.
|
999 |
|
1000 |
@type node_name: str
|
1001 |
@param node_name: the name of the node to remove
|
1002 |
|
1003 |
"""
|
1004 |
self._nodes.pop(node_name, None) |
1005 |
|
1006 |
@staticmethod
|
1007 |
def _CheckRpcResult(result, nodes, failmsg): |
1008 |
"""Verifies the status of an RPC call.
|
1009 |
|
1010 |
Since we aim to keep consistency should this node (the current
|
1011 |
master) fail, we will log errors if our rpc fail, and especially
|
1012 |
log the case when more than half of the nodes fails.
|
1013 |
|
1014 |
@param result: the data as returned from the rpc call
|
1015 |
@type nodes: list
|
1016 |
@param nodes: the list of nodes we made the call to
|
1017 |
@type failmsg: str
|
1018 |
@param failmsg: the identifier to be used for logging
|
1019 |
|
1020 |
"""
|
1021 |
failed = [] |
1022 |
success = [] |
1023 |
|
1024 |
for node in nodes: |
1025 |
msg = result[node].fail_msg |
1026 |
if msg:
|
1027 |
failed.append(node) |
1028 |
logging.error("RPC call %s (%s) failed on node %s: %s",
|
1029 |
result[node].call, failmsg, node, msg) |
1030 |
else:
|
1031 |
success.append(node) |
1032 |
|
1033 |
# +1 for the master node
|
1034 |
if (len(success) + 1) < len(failed): |
1035 |
# TODO: Handle failing nodes
|
1036 |
logging.error("More than half of the nodes failed")
|
1037 |
|
1038 |
def _GetNodeIp(self): |
1039 |
"""Helper for returning the node name/ip list.
|
1040 |
|
1041 |
@rtype: (list, list)
|
1042 |
@return: a tuple of two lists, the first one with the node
|
1043 |
names and the second one with the node addresses
|
1044 |
|
1045 |
"""
|
1046 |
# TODO: Change to "tuple(map(list, zip(*self._nodes.items())))"?
|
1047 |
name_list = self._nodes.keys()
|
1048 |
addr_list = [self._nodes[name] for name in name_list] |
1049 |
return name_list, addr_list
|
1050 |
|
1051 |
def _UpdateJobQueueFile(self, file_name, data, replicate): |
1052 |
"""Writes a file locally and then replicates it to all nodes.
|
1053 |
|
1054 |
This function will replace the contents of a file on the local
|
1055 |
node and then replicate it to all the other nodes we have.
|
1056 |
|
1057 |
@type file_name: str
|
1058 |
@param file_name: the path of the file to be replicated
|
1059 |
@type data: str
|
1060 |
@param data: the new contents of the file
|
1061 |
@type replicate: boolean
|
1062 |
@param replicate: whether to spread the changes to the remote nodes
|
1063 |
|
1064 |
"""
|
1065 |
utils.WriteFile(file_name, data=data) |
1066 |
|
1067 |
if replicate:
|
1068 |
names, addrs = self._GetNodeIp()
|
1069 |
result = rpc.RpcRunner.call_jobqueue_update(names, addrs, file_name, data) |
1070 |
self._CheckRpcResult(result, self._nodes, "Updating %s" % file_name) |
1071 |
|
1072 |
def _RenameFilesUnlocked(self, rename): |
1073 |
"""Renames a file locally and then replicate the change.
|
1074 |
|
1075 |
This function will rename a file in the local queue directory
|
1076 |
and then replicate this rename to all the other nodes we have.
|
1077 |
|
1078 |
@type rename: list of (old, new)
|
1079 |
@param rename: List containing tuples mapping old to new names
|
1080 |
|
1081 |
"""
|
1082 |
# Rename them locally
|
1083 |
for old, new in rename: |
1084 |
utils.RenameFile(old, new, mkdir=True)
|
1085 |
|
1086 |
# ... and on all nodes
|
1087 |
names, addrs = self._GetNodeIp()
|
1088 |
result = rpc.RpcRunner.call_jobqueue_rename(names, addrs, rename) |
1089 |
self._CheckRpcResult(result, self._nodes, "Renaming files (%r)" % rename) |
1090 |
|
1091 |
@staticmethod
|
1092 |
def _FormatJobID(job_id): |
1093 |
"""Convert a job ID to string format.
|
1094 |
|
1095 |
Currently this just does C{str(job_id)} after performing some
|
1096 |
checks, but if we want to change the job id format this will
|
1097 |
abstract this change.
|
1098 |
|
1099 |
@type job_id: int or long
|
1100 |
@param job_id: the numeric job id
|
1101 |
@rtype: str
|
1102 |
@return: the formatted job id
|
1103 |
|
1104 |
"""
|
1105 |
if not isinstance(job_id, (int, long)): |
1106 |
raise errors.ProgrammerError("Job ID '%s' not numeric" % job_id) |
1107 |
if job_id < 0: |
1108 |
raise errors.ProgrammerError("Job ID %s is negative" % job_id) |
1109 |
|
1110 |
return str(job_id) |
1111 |
|
1112 |
@classmethod
|
1113 |
def _GetArchiveDirectory(cls, job_id): |
1114 |
"""Returns the archive directory for a job.
|
1115 |
|
1116 |
@type job_id: str
|
1117 |
@param job_id: Job identifier
|
1118 |
@rtype: str
|
1119 |
@return: Directory name
|
1120 |
|
1121 |
"""
|
1122 |
return str(int(job_id) / JOBS_PER_ARCHIVE_DIRECTORY) |
1123 |
|
1124 |
def _NewSerialsUnlocked(self, count): |
1125 |
"""Generates a new job identifier.
|
1126 |
|
1127 |
Job identifiers are unique during the lifetime of a cluster.
|
1128 |
|
1129 |
@type count: integer
|
1130 |
@param count: how many serials to return
|
1131 |
@rtype: str
|
1132 |
@return: a string representing the job identifier.
|
1133 |
|
1134 |
"""
|
1135 |
assert count > 0 |
1136 |
# New number
|
1137 |
serial = self._last_serial + count
|
1138 |
|
1139 |
# Write to file
|
1140 |
self._UpdateJobQueueFile(constants.JOB_QUEUE_SERIAL_FILE,
|
1141 |
"%s\n" % serial, True) |
1142 |
|
1143 |
result = [self._FormatJobID(v)
|
1144 |
for v in range(self._last_serial, serial + 1)] |
1145 |
# Keep it only if we were able to write the file
|
1146 |
self._last_serial = serial
|
1147 |
|
1148 |
return result
|
1149 |
|
1150 |
@staticmethod
|
1151 |
def _GetJobPath(job_id): |
1152 |
"""Returns the job file for a given job id.
|
1153 |
|
1154 |
@type job_id: str
|
1155 |
@param job_id: the job identifier
|
1156 |
@rtype: str
|
1157 |
@return: the path to the job file
|
1158 |
|
1159 |
"""
|
1160 |
return utils.PathJoin(constants.QUEUE_DIR, "job-%s" % job_id) |
1161 |
|
1162 |
@classmethod
|
1163 |
def _GetArchivedJobPath(cls, job_id): |
1164 |
"""Returns the archived job file for a give job id.
|
1165 |
|
1166 |
@type job_id: str
|
1167 |
@param job_id: the job identifier
|
1168 |
@rtype: str
|
1169 |
@return: the path to the archived job file
|
1170 |
|
1171 |
"""
|
1172 |
return utils.PathJoin(constants.JOB_QUEUE_ARCHIVE_DIR,
|
1173 |
cls._GetArchiveDirectory(job_id), "job-%s" % job_id)
|
1174 |
|
1175 |
def _GetJobIDsUnlocked(self, sort=True): |
1176 |
"""Return all known job IDs.
|
1177 |
|
1178 |
The method only looks at disk because it's a requirement that all
|
1179 |
jobs are present on disk (so in the _memcache we don't have any
|
1180 |
extra IDs).
|
1181 |
|
1182 |
@type sort: boolean
|
1183 |
@param sort: perform sorting on the returned job ids
|
1184 |
@rtype: list
|
1185 |
@return: the list of job IDs
|
1186 |
|
1187 |
"""
|
1188 |
jlist = [] |
1189 |
for filename in utils.ListVisibleFiles(constants.QUEUE_DIR): |
1190 |
m = self._RE_JOB_FILE.match(filename)
|
1191 |
if m:
|
1192 |
jlist.append(m.group(1))
|
1193 |
if sort:
|
1194 |
jlist = utils.NiceSort(jlist) |
1195 |
return jlist
|
1196 |
|
1197 |
def _LoadJobUnlocked(self, job_id): |
1198 |
"""Loads a job from the disk or memory.
|
1199 |
|
1200 |
Given a job id, this will return the cached job object if
|
1201 |
existing, or try to load the job from the disk. If loading from
|
1202 |
disk, it will also add the job to the cache.
|
1203 |
|
1204 |
@param job_id: the job id
|
1205 |
@rtype: L{_QueuedJob} or None
|
1206 |
@return: either None or the job object
|
1207 |
|
1208 |
"""
|
1209 |
job = self._memcache.get(job_id, None) |
1210 |
if job:
|
1211 |
logging.debug("Found job %s in memcache", job_id)
|
1212 |
return job
|
1213 |
|
1214 |
try:
|
1215 |
job = self._LoadJobFromDisk(job_id)
|
1216 |
if job is None: |
1217 |
return job
|
1218 |
except errors.JobFileCorrupted:
|
1219 |
old_path = self._GetJobPath(job_id)
|
1220 |
new_path = self._GetArchivedJobPath(job_id)
|
1221 |
if old_path == new_path:
|
1222 |
# job already archived (future case)
|
1223 |
logging.exception("Can't parse job %s", job_id)
|
1224 |
else:
|
1225 |
# non-archived case
|
1226 |
logging.exception("Can't parse job %s, will archive.", job_id)
|
1227 |
self._RenameFilesUnlocked([(old_path, new_path)])
|
1228 |
return None |
1229 |
|
1230 |
self._memcache[job_id] = job
|
1231 |
logging.debug("Added job %s to the cache", job_id)
|
1232 |
return job
|
1233 |
|
1234 |
def _LoadJobFromDisk(self, job_id): |
1235 |
"""Load the given job file from disk.
|
1236 |
|
1237 |
Given a job file, read, load and restore it in a _QueuedJob format.
|
1238 |
|
1239 |
@type job_id: string
|
1240 |
@param job_id: job identifier
|
1241 |
@rtype: L{_QueuedJob} or None
|
1242 |
@return: either None or the job object
|
1243 |
|
1244 |
"""
|
1245 |
filepath = self._GetJobPath(job_id)
|
1246 |
logging.debug("Loading job from %s", filepath)
|
1247 |
try:
|
1248 |
raw_data = utils.ReadFile(filepath) |
1249 |
except EnvironmentError, err: |
1250 |
if err.errno in (errno.ENOENT, ): |
1251 |
return None |
1252 |
raise
|
1253 |
|
1254 |
try:
|
1255 |
data = serializer.LoadJson(raw_data) |
1256 |
job = _QueuedJob.Restore(self, data)
|
1257 |
except Exception, err: # pylint: disable-msg=W0703 |
1258 |
raise errors.JobFileCorrupted(err)
|
1259 |
|
1260 |
return job
|
1261 |
|
1262 |
def SafeLoadJobFromDisk(self, job_id): |
1263 |
"""Load the given job file from disk.
|
1264 |
|
1265 |
Given a job file, read, load and restore it in a _QueuedJob format.
|
1266 |
In case of error reading the job, it gets returned as None, and the
|
1267 |
exception is logged.
|
1268 |
|
1269 |
@type job_id: string
|
1270 |
@param job_id: job identifier
|
1271 |
@rtype: L{_QueuedJob} or None
|
1272 |
@return: either None or the job object
|
1273 |
|
1274 |
"""
|
1275 |
try:
|
1276 |
return self._LoadJobFromDisk(job_id) |
1277 |
except (errors.JobFileCorrupted, EnvironmentError): |
1278 |
logging.exception("Can't load/parse job %s", job_id)
|
1279 |
return None |
1280 |
|
1281 |
@staticmethod
|
1282 |
def _IsQueueMarkedDrain(): |
1283 |
"""Check if the queue is marked from drain.
|
1284 |
|
1285 |
This currently uses the queue drain file, which makes it a
|
1286 |
per-node flag. In the future this can be moved to the config file.
|
1287 |
|
1288 |
@rtype: boolean
|
1289 |
@return: True of the job queue is marked for draining
|
1290 |
|
1291 |
"""
|
1292 |
return os.path.exists(constants.JOB_QUEUE_DRAIN_FILE)
|
1293 |
|
1294 |
def _UpdateQueueSizeUnlocked(self): |
1295 |
"""Update the queue size.
|
1296 |
|
1297 |
"""
|
1298 |
self._queue_size = len(self._GetJobIDsUnlocked(sort=False)) |
1299 |
|
1300 |
@locking.ssynchronized(_LOCK)
|
1301 |
@_RequireOpenQueue
|
1302 |
def SetDrainFlag(self, drain_flag): |
1303 |
"""Sets the drain flag for the queue.
|
1304 |
|
1305 |
@type drain_flag: boolean
|
1306 |
@param drain_flag: Whether to set or unset the drain flag
|
1307 |
|
1308 |
"""
|
1309 |
if drain_flag:
|
1310 |
utils.WriteFile(constants.JOB_QUEUE_DRAIN_FILE, data="", close=True) |
1311 |
else:
|
1312 |
utils.RemoveFile(constants.JOB_QUEUE_DRAIN_FILE) |
1313 |
|
1314 |
self._drained = drain_flag
|
1315 |
|
1316 |
return True |
1317 |
|
1318 |
@_RequireOpenQueue
|
1319 |
def _SubmitJobUnlocked(self, job_id, ops): |
1320 |
"""Create and store a new job.
|
1321 |
|
1322 |
This enters the job into our job queue and also puts it on the new
|
1323 |
queue, in order for it to be picked up by the queue processors.
|
1324 |
|
1325 |
@type job_id: job ID
|
1326 |
@param job_id: the job ID for the new job
|
1327 |
@type ops: list
|
1328 |
@param ops: The list of OpCodes that will become the new job.
|
1329 |
@rtype: L{_QueuedJob}
|
1330 |
@return: the job object to be queued
|
1331 |
@raise errors.JobQueueDrainError: if the job queue is marked for draining
|
1332 |
@raise errors.JobQueueFull: if the job queue has too many jobs in it
|
1333 |
|
1334 |
"""
|
1335 |
# Ok when sharing the big job queue lock, as the drain file is created when
|
1336 |
# the lock is exclusive.
|
1337 |
if self._drained: |
1338 |
raise errors.JobQueueDrainError("Job queue is drained, refusing job") |
1339 |
|
1340 |
if self._queue_size >= constants.JOB_QUEUE_SIZE_HARD_LIMIT: |
1341 |
raise errors.JobQueueFull()
|
1342 |
|
1343 |
job = _QueuedJob(self, job_id, ops)
|
1344 |
|
1345 |
# Write to disk
|
1346 |
self.UpdateJobUnlocked(job)
|
1347 |
|
1348 |
self._queue_size += 1 |
1349 |
|
1350 |
logging.debug("Adding new job %s to the cache", job_id)
|
1351 |
self._memcache[job_id] = job
|
1352 |
|
1353 |
return job
|
1354 |
|
1355 |
@locking.ssynchronized(_LOCK)
|
1356 |
@_RequireOpenQueue
|
1357 |
def SubmitJob(self, ops): |
1358 |
"""Create and store a new job.
|
1359 |
|
1360 |
@see: L{_SubmitJobUnlocked}
|
1361 |
|
1362 |
"""
|
1363 |
job_id = self._NewSerialsUnlocked(1)[0] |
1364 |
self._wpool.AddTask((self._SubmitJobUnlocked(job_id, ops), )) |
1365 |
return job_id
|
1366 |
|
1367 |
@locking.ssynchronized(_LOCK)
|
1368 |
@_RequireOpenQueue
|
1369 |
def SubmitManyJobs(self, jobs): |
1370 |
"""Create and store multiple jobs.
|
1371 |
|
1372 |
@see: L{_SubmitJobUnlocked}
|
1373 |
|
1374 |
"""
|
1375 |
results = [] |
1376 |
tasks = [] |
1377 |
all_job_ids = self._NewSerialsUnlocked(len(jobs)) |
1378 |
for job_id, ops in zip(all_job_ids, jobs): |
1379 |
try:
|
1380 |
tasks.append((self._SubmitJobUnlocked(job_id, ops), ))
|
1381 |
status = True
|
1382 |
data = job_id |
1383 |
except errors.GenericError, err:
|
1384 |
data = str(err)
|
1385 |
status = False
|
1386 |
results.append((status, data)) |
1387 |
self._wpool.AddManyTasks(tasks)
|
1388 |
|
1389 |
return results
|
1390 |
|
1391 |
@_RequireOpenQueue
|
1392 |
def UpdateJobUnlocked(self, job, replicate=True): |
1393 |
"""Update a job's on disk storage.
|
1394 |
|
1395 |
After a job has been modified, this function needs to be called in
|
1396 |
order to write the changes to disk and replicate them to the other
|
1397 |
nodes.
|
1398 |
|
1399 |
@type job: L{_QueuedJob}
|
1400 |
@param job: the changed job
|
1401 |
@type replicate: boolean
|
1402 |
@param replicate: whether to replicate the change to remote nodes
|
1403 |
|
1404 |
"""
|
1405 |
filename = self._GetJobPath(job.id)
|
1406 |
data = serializer.DumpJson(job.Serialize(), indent=False)
|
1407 |
logging.debug("Writing job %s to %s", job.id, filename)
|
1408 |
self._UpdateJobQueueFile(filename, data, replicate)
|
1409 |
|
1410 |
def WaitForJobChanges(self, job_id, fields, prev_job_info, prev_log_serial, |
1411 |
timeout): |
1412 |
"""Waits for changes in a job.
|
1413 |
|
1414 |
@type job_id: string
|
1415 |
@param job_id: Job identifier
|
1416 |
@type fields: list of strings
|
1417 |
@param fields: Which fields to check for changes
|
1418 |
@type prev_job_info: list or None
|
1419 |
@param prev_job_info: Last job information returned
|
1420 |
@type prev_log_serial: int
|
1421 |
@param prev_log_serial: Last job message serial number
|
1422 |
@type timeout: float
|
1423 |
@param timeout: maximum time to wait in seconds
|
1424 |
@rtype: tuple (job info, log entries)
|
1425 |
@return: a tuple of the job information as required via
|
1426 |
the fields parameter, and the log entries as a list
|
1427 |
|
1428 |
if the job has not changed and the timeout has expired,
|
1429 |
we instead return a special value,
|
1430 |
L{constants.JOB_NOTCHANGED}, which should be interpreted
|
1431 |
as such by the clients
|
1432 |
|
1433 |
"""
|
1434 |
load_fn = compat.partial(self.SafeLoadJobFromDisk, job_id)
|
1435 |
|
1436 |
helper = _WaitForJobChangesHelper() |
1437 |
|
1438 |
return helper(self._GetJobPath(job_id), load_fn, |
1439 |
fields, prev_job_info, prev_log_serial, timeout) |
1440 |
|
1441 |
@locking.ssynchronized(_LOCK)
|
1442 |
@_RequireOpenQueue
|
1443 |
def CancelJob(self, job_id): |
1444 |
"""Cancels a job.
|
1445 |
|
1446 |
This will only succeed if the job has not started yet.
|
1447 |
|
1448 |
@type job_id: string
|
1449 |
@param job_id: job ID of job to be cancelled.
|
1450 |
|
1451 |
"""
|
1452 |
logging.info("Cancelling job %s", job_id)
|
1453 |
|
1454 |
job = self._LoadJobUnlocked(job_id)
|
1455 |
if not job: |
1456 |
logging.debug("Job %s not found", job_id)
|
1457 |
return (False, "Job %s not found" % job_id) |
1458 |
|
1459 |
job_status = job.CalcStatus() |
1460 |
|
1461 |
if job_status not in (constants.JOB_STATUS_QUEUED, |
1462 |
constants.JOB_STATUS_WAITLOCK): |
1463 |
logging.debug("Job %s is no longer waiting in the queue", job.id)
|
1464 |
return (False, "Job %s is no longer waiting in the queue" % job.id) |
1465 |
|
1466 |
if job_status == constants.JOB_STATUS_QUEUED:
|
1467 |
job.MarkUnfinishedOps(constants.OP_STATUS_CANCELED, |
1468 |
"Job canceled by request")
|
1469 |
return (True, "Job %s canceled" % job.id) |
1470 |
|
1471 |
elif job_status == constants.JOB_STATUS_WAITLOCK:
|
1472 |
# The worker will notice the new status and cancel the job
|
1473 |
job.MarkUnfinishedOps(constants.OP_STATUS_CANCELING, None)
|
1474 |
return (True, "Job %s will be canceled" % job.id) |
1475 |
|
1476 |
@_RequireOpenQueue
|
1477 |
def _ArchiveJobsUnlocked(self, jobs): |
1478 |
"""Archives jobs.
|
1479 |
|
1480 |
@type jobs: list of L{_QueuedJob}
|
1481 |
@param jobs: Job objects
|
1482 |
@rtype: int
|
1483 |
@return: Number of archived jobs
|
1484 |
|
1485 |
"""
|
1486 |
archive_jobs = [] |
1487 |
rename_files = [] |
1488 |
for job in jobs: |
1489 |
if job.CalcStatus() not in constants.JOBS_FINALIZED: |
1490 |
logging.debug("Job %s is not yet done", job.id)
|
1491 |
continue
|
1492 |
|
1493 |
archive_jobs.append(job) |
1494 |
|
1495 |
old = self._GetJobPath(job.id)
|
1496 |
new = self._GetArchivedJobPath(job.id)
|
1497 |
rename_files.append((old, new)) |
1498 |
|
1499 |
# TODO: What if 1..n files fail to rename?
|
1500 |
self._RenameFilesUnlocked(rename_files)
|
1501 |
|
1502 |
logging.debug("Successfully archived job(s) %s",
|
1503 |
utils.CommaJoin(job.id for job in archive_jobs)) |
1504 |
|
1505 |
# Since we haven't quite checked, above, if we succeeded or failed renaming
|
1506 |
# the files, we update the cached queue size from the filesystem. When we
|
1507 |
# get around to fix the TODO: above, we can use the number of actually
|
1508 |
# archived jobs to fix this.
|
1509 |
self._UpdateQueueSizeUnlocked()
|
1510 |
return len(archive_jobs) |
1511 |
|
1512 |
@locking.ssynchronized(_LOCK)
|
1513 |
@_RequireOpenQueue
|
1514 |
def ArchiveJob(self, job_id): |
1515 |
"""Archives a job.
|
1516 |
|
1517 |
This is just a wrapper over L{_ArchiveJobsUnlocked}.
|
1518 |
|
1519 |
@type job_id: string
|
1520 |
@param job_id: Job ID of job to be archived.
|
1521 |
@rtype: bool
|
1522 |
@return: Whether job was archived
|
1523 |
|
1524 |
"""
|
1525 |
logging.info("Archiving job %s", job_id)
|
1526 |
|
1527 |
job = self._LoadJobUnlocked(job_id)
|
1528 |
if not job: |
1529 |
logging.debug("Job %s not found", job_id)
|
1530 |
return False |
1531 |
|
1532 |
return self._ArchiveJobsUnlocked([job]) == 1 |
1533 |
|
1534 |
@locking.ssynchronized(_LOCK)
|
1535 |
@_RequireOpenQueue
|
1536 |
def AutoArchiveJobs(self, age, timeout): |
1537 |
"""Archives all jobs based on age.
|
1538 |
|
1539 |
The method will archive all jobs which are older than the age
|
1540 |
parameter. For jobs that don't have an end timestamp, the start
|
1541 |
timestamp will be considered. The special '-1' age will cause
|
1542 |
archival of all jobs (that are not running or queued).
|
1543 |
|
1544 |
@type age: int
|
1545 |
@param age: the minimum age in seconds
|
1546 |
|
1547 |
"""
|
1548 |
logging.info("Archiving jobs with age more than %s seconds", age)
|
1549 |
|
1550 |
now = time.time() |
1551 |
end_time = now + timeout |
1552 |
archived_count = 0
|
1553 |
last_touched = 0
|
1554 |
|
1555 |
all_job_ids = self._GetJobIDsUnlocked()
|
1556 |
pending = [] |
1557 |
for idx, job_id in enumerate(all_job_ids): |
1558 |
last_touched = idx + 1
|
1559 |
|
1560 |
# Not optimal because jobs could be pending
|
1561 |
# TODO: Measure average duration for job archival and take number of
|
1562 |
# pending jobs into account.
|
1563 |
if time.time() > end_time:
|
1564 |
break
|
1565 |
|
1566 |
# Returns None if the job failed to load
|
1567 |
job = self._LoadJobUnlocked(job_id)
|
1568 |
if job:
|
1569 |
if job.end_timestamp is None: |
1570 |
if job.start_timestamp is None: |
1571 |
job_age = job.received_timestamp |
1572 |
else:
|
1573 |
job_age = job.start_timestamp |
1574 |
else:
|
1575 |
job_age = job.end_timestamp |
1576 |
|
1577 |
if age == -1 or now - job_age[0] > age: |
1578 |
pending.append(job) |
1579 |
|
1580 |
# Archive 10 jobs at a time
|
1581 |
if len(pending) >= 10: |
1582 |
archived_count += self._ArchiveJobsUnlocked(pending)
|
1583 |
pending = [] |
1584 |
|
1585 |
if pending:
|
1586 |
archived_count += self._ArchiveJobsUnlocked(pending)
|
1587 |
|
1588 |
return (archived_count, len(all_job_ids) - last_touched) |
1589 |
|
1590 |
def QueryJobs(self, job_ids, fields): |
1591 |
"""Returns a list of jobs in queue.
|
1592 |
|
1593 |
@type job_ids: list
|
1594 |
@param job_ids: sequence of job identifiers or None for all
|
1595 |
@type fields: list
|
1596 |
@param fields: names of fields to return
|
1597 |
@rtype: list
|
1598 |
@return: list one element per job, each element being list with
|
1599 |
the requested fields
|
1600 |
|
1601 |
"""
|
1602 |
jobs = [] |
1603 |
list_all = False
|
1604 |
if not job_ids: |
1605 |
# Since files are added to/removed from the queue atomically, there's no
|
1606 |
# risk of getting the job ids in an inconsistent state.
|
1607 |
job_ids = self._GetJobIDsUnlocked()
|
1608 |
list_all = True
|
1609 |
|
1610 |
for job_id in job_ids: |
1611 |
job = self.SafeLoadJobFromDisk(job_id)
|
1612 |
if job is not None: |
1613 |
jobs.append(job.GetInfo(fields)) |
1614 |
elif not list_all: |
1615 |
jobs.append(None)
|
1616 |
|
1617 |
return jobs
|
1618 |
|
1619 |
@locking.ssynchronized(_LOCK)
|
1620 |
@_RequireOpenQueue
|
1621 |
def Shutdown(self): |
1622 |
"""Stops the job queue.
|
1623 |
|
1624 |
This shutdowns all the worker threads an closes the queue.
|
1625 |
|
1626 |
"""
|
1627 |
self._wpool.TerminateWorkers()
|
1628 |
|
1629 |
self._queue_filelock.Close()
|
1630 |
self._queue_filelock = None |