jqueue.py 69.7 KB
Newer Older
Iustin Pop's avatar
Iustin Pop committed
1
2
3
#
#

Michael Hanselmann's avatar
Michael Hanselmann committed
4
# Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 Google Inc.
Iustin Pop's avatar
Iustin Pop committed
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
#
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 2 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful, but
# WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
# General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program; if not, write to the Free Software
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
# 02110-1301, USA.


22
23
"""Module implementing the job queue handling.

24
25
26
27
28
Locking: there's a single, large lock in the L{JobQueue} class. It's
used by all other classes in this module.

@var JOBQUEUE_THREADS: the number of worker threads we start for
    processing jobs
29
30

"""
Iustin Pop's avatar
Iustin Pop committed
31

Michael Hanselmann's avatar
Michael Hanselmann committed
32
import logging
33
34
import errno
import re
35
import time
36
import weakref
Michael Hanselmann's avatar
Michael Hanselmann committed
37
import threading
Iustin Pop's avatar
Iustin Pop committed
38

Guido Trotter's avatar
Guido Trotter committed
39
40
41
42
43
44
45
try:
  # pylint: disable-msg=E0611
  from pyinotify import pyinotify
except ImportError:
  import pyinotify

from ganeti import asyncnotifier
Michael Hanselmann's avatar
Michael Hanselmann committed
46
from ganeti import constants
47
from ganeti import serializer
Michael Hanselmann's avatar
Michael Hanselmann committed
48
from ganeti import workerpool
49
from ganeti import locking
50
from ganeti import opcodes
Iustin Pop's avatar
Iustin Pop committed
51
from ganeti import errors
Michael Hanselmann's avatar
Michael Hanselmann committed
52
from ganeti import mcpu
53
from ganeti import utils
54
from ganeti import jstore
55
from ganeti import rpc
56
from ganeti import runtime
57
from ganeti import netutils
58
from ganeti import compat
Michael Hanselmann's avatar
Michael Hanselmann committed
59
from ganeti import ht
Michael Hanselmann's avatar
Michael Hanselmann committed
60

61

62
JOBQUEUE_THREADS = 25
63
JOBS_PER_ARCHIVE_DIRECTORY = 10000
Michael Hanselmann's avatar
Michael Hanselmann committed
64

65
66
67
# member lock names to be passed to @ssynchronized decorator
_LOCK = "_lock"
_QUEUE = "_queue"
68

Iustin Pop's avatar
Iustin Pop committed
69

70
class CancelJob(Exception):
71
72
73
74
75
  """Special exception to cancel a job.

  """


76
def TimeStampNow():
77
78
79
80
81
82
  """Returns the current timestamp.

  @rtype: tuple
  @return: the current time in the (seconds, microseconds) format

  """
83
84
85
  return utils.SplitTime(time.time())


Michael Hanselmann's avatar
Michael Hanselmann committed
86
class _QueuedOpCode(object):
Michael Hanselmann's avatar
Michael Hanselmann committed
87
  """Encapsulates an opcode object.
Michael Hanselmann's avatar
Michael Hanselmann committed
88

89
90
91
92
93
94
  @ivar log: holds the execution log and consists of tuples
  of the form C{(log_serial, timestamp, level, message)}
  @ivar input: the OpCode we encapsulate
  @ivar status: the current status
  @ivar result: the result of the LU execution
  @ivar start_timestamp: timestamp for the start of the execution
Iustin Pop's avatar
Iustin Pop committed
95
  @ivar exec_timestamp: timestamp for the actual LU Exec() function invocation
96
  @ivar stop_timestamp: timestamp for the end of the execution
97

Michael Hanselmann's avatar
Michael Hanselmann committed
98
  """
99
  __slots__ = ["input", "status", "result", "log", "priority",
Iustin Pop's avatar
Iustin Pop committed
100
               "start_timestamp", "exec_timestamp", "end_timestamp",
101
102
               "__weakref__"]

Michael Hanselmann's avatar
Michael Hanselmann committed
103
  def __init__(self, op):
104
105
106
107
108
109
    """Constructor for the _QuededOpCode.

    @type op: L{opcodes.OpCode}
    @param op: the opcode we encapsulate

    """
Michael Hanselmann's avatar
Michael Hanselmann committed
110
111
112
113
    self.input = op
    self.status = constants.OP_STATUS_QUEUED
    self.result = None
    self.log = []
114
    self.start_timestamp = None
Iustin Pop's avatar
Iustin Pop committed
115
    self.exec_timestamp = None
116
    self.end_timestamp = None
117

118
119
120
    # Get initial priority (it might change during the lifetime of this opcode)
    self.priority = getattr(op, "priority", constants.OP_PRIO_DEFAULT)

121
122
  @classmethod
  def Restore(cls, state):
123
124
125
126
127
128
129
130
    """Restore the _QueuedOpCode from the serialized form.

    @type state: dict
    @param state: the serialized state
    @rtype: _QueuedOpCode
    @return: a new _QueuedOpCode instance

    """
Michael Hanselmann's avatar
Michael Hanselmann committed
131
132
133
134
135
    obj = _QueuedOpCode.__new__(cls)
    obj.input = opcodes.OpCode.LoadOpCode(state["input"])
    obj.status = state["status"]
    obj.result = state["result"]
    obj.log = state["log"]
136
    obj.start_timestamp = state.get("start_timestamp", None)
Iustin Pop's avatar
Iustin Pop committed
137
    obj.exec_timestamp = state.get("exec_timestamp", None)
138
    obj.end_timestamp = state.get("end_timestamp", None)
139
    obj.priority = state.get("priority", constants.OP_PRIO_DEFAULT)
140
141
142
    return obj

  def Serialize(self):
143
144
145
146
147
148
    """Serializes this _QueuedOpCode.

    @rtype: dict
    @return: the dictionary holding the serialized state

    """
149
150
151
152
153
    return {
      "input": self.input.__getstate__(),
      "status": self.status,
      "result": self.result,
      "log": self.log,
154
      "start_timestamp": self.start_timestamp,
Iustin Pop's avatar
Iustin Pop committed
155
      "exec_timestamp": self.exec_timestamp,
156
      "end_timestamp": self.end_timestamp,
157
      "priority": self.priority,
158
      }
159

Michael Hanselmann's avatar
Michael Hanselmann committed
160
161
162
163

class _QueuedJob(object):
  """In-memory job representation.

164
165
166
167
168
169
170
171
172
173
174
175
176
  This is what we use to track the user-submitted jobs. Locking must
  be taken care of by users of this class.

  @type queue: L{JobQueue}
  @ivar queue: the parent queue
  @ivar id: the job ID
  @type ops: list
  @ivar ops: the list of _QueuedOpCode that constitute the job
  @type log_serial: int
  @ivar log_serial: holds the index for the next log entry
  @ivar received_timestamp: the timestamp for when the job was received
  @ivar start_timestmap: the timestamp for start of execution
  @ivar end_timestamp: the timestamp for end of execution
177
  @ivar writable: Whether the job is allowed to be modified
Michael Hanselmann's avatar
Michael Hanselmann committed
178
179

  """
Iustin Pop's avatar
Iustin Pop committed
180
  # pylint: disable-msg=W0212
181
  __slots__ = ["queue", "id", "ops", "log_serial", "ops_iter", "cur_opctx",
182
               "received_timestamp", "start_timestamp", "end_timestamp",
183
               "__weakref__", "processor_lock", "writable"]
184

185
  def __init__(self, queue, job_id, ops, writable):
186
187
188
189
190
191
192
193
194
    """Constructor for the _QueuedJob.

    @type queue: L{JobQueue}
    @param queue: our parent queue
    @type job_id: job_id
    @param job_id: our job id
    @type ops: list
    @param ops: the list of opcodes we hold, which will be encapsulated
        in _QueuedOpCodes
195
196
    @type writable: bool
    @param writable: Whether job can be modified
197
198

    """
Michael Hanselmann's avatar
Michael Hanselmann committed
199
    if not ops:
Guido Trotter's avatar
Guido Trotter committed
200
      raise errors.GenericError("A job needs at least one opcode")
Michael Hanselmann's avatar
Michael Hanselmann committed
201

Michael Hanselmann's avatar
Michael Hanselmann committed
202
    self.queue = queue
203
    self.id = job_id
Michael Hanselmann's avatar
Michael Hanselmann committed
204
    self.ops = [_QueuedOpCode(op) for op in ops]
205
    self.log_serial = 0
206
207
208
    self.received_timestamp = TimeStampNow()
    self.start_timestamp = None
    self.end_timestamp = None
209

210
    self._InitInMemory(self, writable)
211
212

  @staticmethod
213
  def _InitInMemory(obj, writable):
214
215
216
    """Initializes in-memory variables.

    """
217
    obj.writable = writable
218
    obj.ops_iter = None
219
    obj.cur_opctx = None
Michael Hanselmann's avatar
Michael Hanselmann committed
220
    obj.processor_lock = threading.Lock()
221

222
223
224
225
226
227
228
  def __repr__(self):
    status = ["%s.%s" % (self.__class__.__module__, self.__class__.__name__),
              "id=%s" % self.id,
              "ops=%s" % ",".join([op.input.Summary() for op in self.ops])]

    return "<%s at %#x>" % (" ".join(status), id(self))

229
  @classmethod
230
  def Restore(cls, queue, state, writable):
231
232
233
234
235
236
    """Restore a _QueuedJob from serialized state:

    @type queue: L{JobQueue}
    @param queue: to which queue the restored job belongs
    @type state: dict
    @param state: the serialized state
237
238
    @type writable: bool
    @param writable: Whether job can be modified
239
240
241
242
    @rtype: _JobQueue
    @return: the restored _JobQueue instance

    """
Michael Hanselmann's avatar
Michael Hanselmann committed
243
244
245
    obj = _QueuedJob.__new__(cls)
    obj.queue = queue
    obj.id = state["id"]
246
247
248
    obj.received_timestamp = state.get("received_timestamp", None)
    obj.start_timestamp = state.get("start_timestamp", None)
    obj.end_timestamp = state.get("end_timestamp", None)
249
250
251
252
253
254
255
256
257

    obj.ops = []
    obj.log_serial = 0
    for op_state in state["ops"]:
      op = _QueuedOpCode.Restore(op_state)
      for log_entry in op.log:
        obj.log_serial = max(obj.log_serial, log_entry[0])
      obj.ops.append(op)

258
    cls._InitInMemory(obj, writable)
259

260
261
262
    return obj

  def Serialize(self):
263
264
265
266
267
268
    """Serialize the _JobQueue instance.

    @rtype: dict
    @return: the serialized state

    """
269
270
    return {
      "id": self.id,
Michael Hanselmann's avatar
Michael Hanselmann committed
271
      "ops": [op.Serialize() for op in self.ops],
272
273
274
      "start_timestamp": self.start_timestamp,
      "end_timestamp": self.end_timestamp,
      "received_timestamp": self.received_timestamp,
275
276
      }

Michael Hanselmann's avatar
Michael Hanselmann committed
277
  def CalcStatus(self):
278
279
280
281
282
283
284
285
286
287
    """Compute the status of this job.

    This function iterates over all the _QueuedOpCodes in the job and
    based on their status, computes the job status.

    The algorithm is:
      - if we find a cancelled, or finished with error, the job
        status will be the same
      - otherwise, the last opcode with the status one of:
          - waitlock
288
          - canceling
289
290
291
292
293
294
295
296
297
298
          - running

        will determine the job status

      - otherwise, it means either all opcodes are queued, or success,
        and the job status will be the same

    @return: the job status

    """
Michael Hanselmann's avatar
Michael Hanselmann committed
299
300
301
    status = constants.JOB_STATUS_QUEUED

    all_success = True
Michael Hanselmann's avatar
Michael Hanselmann committed
302
303
    for op in self.ops:
      if op.status == constants.OP_STATUS_SUCCESS:
Michael Hanselmann's avatar
Michael Hanselmann committed
304
305
306
307
        continue

      all_success = False

Michael Hanselmann's avatar
Michael Hanselmann committed
308
      if op.status == constants.OP_STATUS_QUEUED:
Michael Hanselmann's avatar
Michael Hanselmann committed
309
        pass
Iustin Pop's avatar
Iustin Pop committed
310
311
      elif op.status == constants.OP_STATUS_WAITLOCK:
        status = constants.JOB_STATUS_WAITLOCK
Michael Hanselmann's avatar
Michael Hanselmann committed
312
      elif op.status == constants.OP_STATUS_RUNNING:
Michael Hanselmann's avatar
Michael Hanselmann committed
313
        status = constants.JOB_STATUS_RUNNING
314
315
316
      elif op.status == constants.OP_STATUS_CANCELING:
        status = constants.JOB_STATUS_CANCELING
        break
Michael Hanselmann's avatar
Michael Hanselmann committed
317
      elif op.status == constants.OP_STATUS_ERROR:
318
319
320
        status = constants.JOB_STATUS_ERROR
        # The whole job fails if one opcode failed
        break
Michael Hanselmann's avatar
Michael Hanselmann committed
321
      elif op.status == constants.OP_STATUS_CANCELED:
322
323
        status = constants.OP_STATUS_CANCELED
        break
Michael Hanselmann's avatar
Michael Hanselmann committed
324
325
326
327
328
329

    if all_success:
      status = constants.JOB_STATUS_SUCCESS

    return status

330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
  def CalcPriority(self):
    """Gets the current priority for this job.

    Only unfinished opcodes are considered. When all are done, the default
    priority is used.

    @rtype: int

    """
    priorities = [op.priority for op in self.ops
                  if op.status not in constants.OPS_FINALIZED]

    if not priorities:
      # All opcodes are done, assume default priority
      return constants.OP_PRIO_DEFAULT

    return min(priorities)

348
  def GetLogEntries(self, newer_than):
349
350
351
    """Selectively returns the log entries.

    @type newer_than: None or int
Michael Hanselmann's avatar
Michael Hanselmann committed
352
    @param newer_than: if this is None, return all log entries,
353
354
355
356
357
358
        otherwise return only the log entries with serial higher
        than this value
    @rtype: list
    @return: the list of the log entries selected

    """
359
360
361
362
363
364
365
    if newer_than is None:
      serial = -1
    else:
      serial = newer_than

    entries = []
    for op in self.ops:
366
      entries.extend(filter(lambda entry: entry[0] > serial, op.log))
367
368
369

    return entries

370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
  def GetInfo(self, fields):
    """Returns information about a job.

    @type fields: list
    @param fields: names of fields to return
    @rtype: list
    @return: list with one element for each field
    @raise errors.OpExecError: when an invalid field
        has been passed

    """
    row = []
    for fname in fields:
      if fname == "id":
        row.append(self.id)
      elif fname == "status":
        row.append(self.CalcStatus())
387
388
      elif fname == "priority":
        row.append(self.CalcPriority())
389
390
391
392
393
394
395
396
397
398
399
400
401
402
      elif fname == "ops":
        row.append([op.input.__getstate__() for op in self.ops])
      elif fname == "opresult":
        row.append([op.result for op in self.ops])
      elif fname == "opstatus":
        row.append([op.status for op in self.ops])
      elif fname == "oplog":
        row.append([op.log for op in self.ops])
      elif fname == "opstart":
        row.append([op.start_timestamp for op in self.ops])
      elif fname == "opexec":
        row.append([op.exec_timestamp for op in self.ops])
      elif fname == "opend":
        row.append([op.end_timestamp for op in self.ops])
403
404
      elif fname == "oppriority":
        row.append([op.priority for op in self.ops])
405
406
407
408
409
410
411
412
413
414
415
416
      elif fname == "received_ts":
        row.append(self.received_timestamp)
      elif fname == "start_ts":
        row.append(self.start_timestamp)
      elif fname == "end_ts":
        row.append(self.end_timestamp)
      elif fname == "summary":
        row.append([op.input.Summary() for op in self.ops])
      else:
        raise errors.OpExecError("Invalid self query field '%s'" % fname)
    return row

417
418
419
420
421
422
423
424
425
426
427
  def MarkUnfinishedOps(self, status, result):
    """Mark unfinished opcodes with a given status and result.

    This is an utility function for marking all running or waiting to
    be run opcodes with a given status. Opcodes which are already
    finalised are not changed.

    @param status: a given opcode status
    @param result: the opcode result

    """
428
429
430
431
432
433
434
435
    not_marked = True
    for op in self.ops:
      if op.status in constants.OPS_FINALIZED:
        assert not_marked, "Finalized opcodes found after non-finalized ones"
        continue
      op.status = status
      op.result = result
      not_marked = False
436

437
438
439
440
441
442
  def Finalize(self):
    """Marks the job as finalized.

    """
    self.end_timestamp = TimeStampNow()

443
  def Cancel(self):
444
445
446
447
448
449
450
    """Marks job as canceled/-ing if possible.

    @rtype: tuple; (bool, string)
    @return: Boolean describing whether job was successfully canceled or marked
      as canceling and a text message

    """
451
452
453
454
455
    status = self.CalcStatus()

    if status == constants.JOB_STATUS_QUEUED:
      self.MarkUnfinishedOps(constants.OP_STATUS_CANCELED,
                             "Job canceled by request")
456
      self.Finalize()
457
      return (True, "Job %s canceled" % self.id)
458
459
460
461

    elif status == constants.JOB_STATUS_WAITLOCK:
      # The worker will notice the new status and cancel the job
      self.MarkUnfinishedOps(constants.OP_STATUS_CANCELING, None)
462
      return (True, "Job %s will be canceled" % self.id)
463

464
465
466
    else:
      logging.debug("Job %s is no longer waiting in the queue", self.id)
      return (False, "Job %s is no longer waiting in the queue" % self.id)
467

468

469
class _OpExecCallbacks(mcpu.OpExecCbBase):
470
471
  def __init__(self, queue, job, op):
    """Initializes this class.
472

473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
    @type queue: L{JobQueue}
    @param queue: Job queue
    @type job: L{_QueuedJob}
    @param job: Job object
    @type op: L{_QueuedOpCode}
    @param op: OpCode

    """
    assert queue, "Queue is missing"
    assert job, "Job is missing"
    assert op, "Opcode is missing"

    self._queue = queue
    self._job = job
    self._op = op

489
490
491
492
493
494
495
496
497
  def _CheckCancel(self):
    """Raises an exception to cancel the job if asked to.

    """
    # Cancel here if we were asked to
    if self._op.status == constants.OP_STATUS_CANCELING:
      logging.debug("Canceling opcode")
      raise CancelJob()

498
  @locking.ssynchronized(_QUEUE, shared=1)
499
  def NotifyStart(self):
Iustin Pop's avatar
Iustin Pop committed
500
501
    """Mark the opcode as running, not lock-waiting.

502
503
504
505
    This is called from the mcpu code as a notifier function, when the LU is
    finally about to start the Exec() method. Of course, to have end-user
    visible results, the opcode must be initially (before calling into
    Processor.ExecOpCode) set to OP_STATUS_WAITLOCK.
Iustin Pop's avatar
Iustin Pop committed
506
507

    """
508
    assert self._op in self._job.ops
509
510
    assert self._op.status in (constants.OP_STATUS_WAITLOCK,
                               constants.OP_STATUS_CANCELING)
511

512
    # Cancel here if we were asked to
513
    self._CheckCancel()
514

515
    logging.debug("Opcode is now running")
516

517
518
519
520
521
    self._op.status = constants.OP_STATUS_RUNNING
    self._op.exec_timestamp = TimeStampNow()

    # And finally replicate the job status
    self._queue.UpdateJobUnlocked(self._job)
522

523
  @locking.ssynchronized(_QUEUE, shared=1)
524
525
526
527
528
529
530
531
  def _AppendFeedback(self, timestamp, log_type, log_msg):
    """Internal feedback append function, with locks

    """
    self._job.log_serial += 1
    self._op.log.append((self._job.log_serial, timestamp, log_type, log_msg))
    self._queue.UpdateJobUnlocked(self._job, replicate=False)

532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
  def Feedback(self, *args):
    """Append a log entry.

    """
    assert len(args) < 3

    if len(args) == 1:
      log_type = constants.ELOG_MESSAGE
      log_msg = args[0]
    else:
      (log_type, log_msg) = args

    # The time is split to make serialization easier and not lose
    # precision.
    timestamp = utils.SplitTime(time.time())
547
    self._AppendFeedback(timestamp, log_type, log_msg)
548

549
550
  def CheckCancel(self):
    """Check whether job has been cancelled.
551
552

    """
553
554
555
556
557
558
    assert self._op.status in (constants.OP_STATUS_WAITLOCK,
                               constants.OP_STATUS_CANCELING)

    # Cancel here if we were asked to
    self._CheckCancel()

559
560
561
562
563
564
565
566
567
  def SubmitManyJobs(self, jobs):
    """Submits jobs for processing.

    See L{JobQueue.SubmitManyJobs}.

    """
    # Locking is done in job queue
    return self._queue.SubmitManyJobs(jobs)

568

569
570
571
class _JobChangesChecker(object):
  def __init__(self, fields, prev_job_info, prev_log_serial):
    """Initializes this class.
Guido Trotter's avatar
Guido Trotter committed
572

573
574
575
576
577
578
    @type fields: list of strings
    @param fields: Fields requested by LUXI client
    @type prev_job_info: string
    @param prev_job_info: previous job info, as passed by the LUXI client
    @type prev_log_serial: string
    @param prev_log_serial: previous job serial, as passed by the LUXI client
Guido Trotter's avatar
Guido Trotter committed
579

580
581
582
583
    """
    self._fields = fields
    self._prev_job_info = prev_job_info
    self._prev_log_serial = prev_log_serial
Guido Trotter's avatar
Guido Trotter committed
584

585
586
  def __call__(self, job):
    """Checks whether job has changed.
Guido Trotter's avatar
Guido Trotter committed
587

588
589
    @type job: L{_QueuedJob}
    @param job: Job object
Guido Trotter's avatar
Guido Trotter committed
590
591

    """
592
593
    assert not job.writable, "Expected read-only job"

594
595
596
    status = job.CalcStatus()
    job_info = job.GetInfo(self._fields)
    log_entries = job.GetLogEntries(self._prev_log_serial)
Guido Trotter's avatar
Guido Trotter committed
597
598
599
600
601
602
603
604

    # Serializing and deserializing data can cause type changes (e.g. from
    # tuple to list) or precision loss. We're doing it here so that we get
    # the same modifications as the data received from the client. Without
    # this, the comparison afterwards might fail without the data being
    # significantly different.
    # TODO: we just deserialized from disk, investigate how to make sure that
    # the job info and log entries are compatible to avoid this further step.
605
606
607
608
    # TODO: Doing something like in testutils.py:UnifyValueType might be more
    # efficient, though floats will be tricky
    job_info = serializer.LoadJson(serializer.DumpJson(job_info))
    log_entries = serializer.LoadJson(serializer.DumpJson(log_entries))
Guido Trotter's avatar
Guido Trotter committed
609
610
611

    # Don't even try to wait if the job is no longer running, there will be
    # no changes.
612
613
614
615
616
617
618
    if (status not in (constants.JOB_STATUS_QUEUED,
                       constants.JOB_STATUS_RUNNING,
                       constants.JOB_STATUS_WAITLOCK) or
        job_info != self._prev_job_info or
        (log_entries and self._prev_log_serial != log_entries[0][0])):
      logging.debug("Job %s changed", job.id)
      return (job_info, log_entries)
Guido Trotter's avatar
Guido Trotter committed
619

620
621
622
623
624
625
626
627
628
629
    return None


class _JobFileChangesWaiter(object):
  def __init__(self, filename):
    """Initializes this class.

    @type filename: string
    @param filename: Path to job file
    @raises errors.InotifyError: if the notifier cannot be setup
Guido Trotter's avatar
Guido Trotter committed
630

631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
    """
    self._wm = pyinotify.WatchManager()
    self._inotify_handler = \
      asyncnotifier.SingleFileEventHandler(self._wm, self._OnInotify, filename)
    self._notifier = \
      pyinotify.Notifier(self._wm, default_proc_fun=self._inotify_handler)
    try:
      self._inotify_handler.enable()
    except Exception:
      # pyinotify doesn't close file descriptors automatically
      self._notifier.stop()
      raise

  def _OnInotify(self, notifier_enabled):
    """Callback for inotify.

    """
Guido Trotter's avatar
Guido Trotter committed
648
    if not notifier_enabled:
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
      self._inotify_handler.enable()

  def Wait(self, timeout):
    """Waits for the job file to change.

    @type timeout: float
    @param timeout: Timeout in seconds
    @return: Whether there have been events

    """
    assert timeout >= 0
    have_events = self._notifier.check_events(timeout * 1000)
    if have_events:
      self._notifier.read_events()
    self._notifier.process_events()
    return have_events

  def Close(self):
    """Closes underlying notifier and its file descriptor.

    """
    self._notifier.stop()


class _JobChangesWaiter(object):
  def __init__(self, filename):
    """Initializes this class.

    @type filename: string
    @param filename: Path to job file

    """
    self._filewaiter = None
    self._filename = filename
Guido Trotter's avatar
Guido Trotter committed
683

684
685
  def Wait(self, timeout):
    """Waits for a job to change.
Guido Trotter's avatar
Guido Trotter committed
686

687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
    @type timeout: float
    @param timeout: Timeout in seconds
    @return: Whether there have been events

    """
    if self._filewaiter:
      return self._filewaiter.Wait(timeout)

    # Lazy setup: Avoid inotify setup cost when job file has already changed.
    # If this point is reached, return immediately and let caller check the job
    # file again in case there were changes since the last check. This avoids a
    # race condition.
    self._filewaiter = _JobFileChangesWaiter(self._filename)

    return True

  def Close(self):
    """Closes underlying waiter.

    """
    if self._filewaiter:
      self._filewaiter.Close()


class _WaitForJobChangesHelper(object):
  """Helper class using inotify to wait for changes in a job file.

  This class takes a previous job status and serial, and alerts the client when
  the current job status has changed.

  """
  @staticmethod
  def _CheckForChanges(job_load_fn, check_fn):
    job = job_load_fn()
    if not job:
      raise errors.JobLost()

    result = check_fn(job)
    if result is None:
      raise utils.RetryAgain()

    return result

  def __call__(self, filename, job_load_fn,
               fields, prev_job_info, prev_log_serial, timeout):
    """Waits for changes on a job.

    @type filename: string
    @param filename: File on which to wait for changes
    @type job_load_fn: callable
    @param job_load_fn: Function to load job
    @type fields: list of strings
    @param fields: Which fields to check for changes
    @type prev_job_info: list or None
    @param prev_job_info: Last job information returned
    @type prev_log_serial: int
    @param prev_log_serial: Last job message serial number
    @type timeout: float
    @param timeout: maximum time to wait in seconds

    """
Guido Trotter's avatar
Guido Trotter committed
748
    try:
749
750
751
752
753
754
755
756
757
      check_fn = _JobChangesChecker(fields, prev_job_info, prev_log_serial)
      waiter = _JobChangesWaiter(filename)
      try:
        return utils.Retry(compat.partial(self._CheckForChanges,
                                          job_load_fn, check_fn),
                           utils.RETRY_REMAINING_TIME, timeout,
                           wait_fn=waiter.Wait)
      finally:
        waiter.Close()
Guido Trotter's avatar
Guido Trotter committed
758
759
760
761
762
763
    except (errors.InotifyError, errors.JobLost):
      return None
    except utils.RetryTimeout:
      return constants.JOB_NOTCHANGED


764
765
766
767
768
769
770
771
772
773
774
775
def _EncodeOpError(err):
  """Encodes an error which occurred while processing an opcode.

  """
  if isinstance(err, errors.GenericError):
    to_encode = err
  else:
    to_encode = errors.OpExecError(str(err))

  return errors.EncodeException(to_encode)


776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
class _TimeoutStrategyWrapper:
  def __init__(self, fn):
    """Initializes this class.

    """
    self._fn = fn
    self._next = None

  def _Advance(self):
    """Gets the next timeout if necessary.

    """
    if self._next is None:
      self._next = self._fn()

  def Peek(self):
    """Returns the next timeout.

    """
    self._Advance()
    return self._next

  def Next(self):
    """Returns the current timeout and advances the internal state.

    """
    self._Advance()
    result = self._next
    self._next = None
    return result


808
class _OpExecContext:
809
  def __init__(self, op, index, log_prefix, timeout_strategy_factory):
810
811
812
813
814
815
816
817
    """Initializes this class.

    """
    self.op = op
    self.index = index
    self.log_prefix = log_prefix
    self.summary = op.input.Summary()

Michael Hanselmann's avatar
Michael Hanselmann committed
818
819
820
821
822
823
    # Create local copy to modify
    if getattr(op.input, opcodes.DEPEND_ATTR, None):
      self.jobdeps = op.input.depends[:]
    else:
      self.jobdeps = None

824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
    self._timeout_strategy_factory = timeout_strategy_factory
    self._ResetTimeoutStrategy()

  def _ResetTimeoutStrategy(self):
    """Creates a new timeout strategy.

    """
    self._timeout_strategy = \
      _TimeoutStrategyWrapper(self._timeout_strategy_factory().NextAttempt)

  def CheckPriorityIncrease(self):
    """Checks whether priority can and should be increased.

    Called when locks couldn't be acquired.

    """
    op = self.op

    # Exhausted all retries and next round should not use blocking acquire
    # for locks?
    if (self._timeout_strategy.Peek() is None and
        op.priority > constants.OP_PRIO_HIGHEST):
      logging.debug("Increasing priority")
      op.priority -= 1
      self._ResetTimeoutStrategy()
      return True

    return False

  def GetNextLockTimeout(self):
    """Returns the next lock acquire timeout.

    """
    return self._timeout_strategy.Next()

859

860
class _JobProcessor(object):
861
862
  def __init__(self, queue, opexec_fn, job,
               _timeout_strategy_factory=mcpu.LockAttemptTimeoutStrategy):
863
864
865
866
867
868
    """Initializes this class.

    """
    self.queue = queue
    self.opexec_fn = opexec_fn
    self.job = job
869
    self._timeout_strategy_factory = _timeout_strategy_factory
870
871

  @staticmethod
872
  def _FindNextOpcode(job, timeout_strategy_factory):
873
874
875
876
    """Locates the next opcode to run.

    @type job: L{_QueuedJob}
    @param job: Job object
877
    @param timeout_strategy_factory: Callable to create new timeout strategy
878
879
880
881
882
883

    """
    # Create some sort of a cache to speed up locating next opcode for future
    # lookups
    # TODO: Consider splitting _QueuedJob.ops into two separate lists, one for
    # pending and one for processed ops.
884
885
    if job.ops_iter is None:
      job.ops_iter = enumerate(job.ops)
886
887
888
889

    # Find next opcode to run
    while True:
      try:
890
        (idx, op) = job.ops_iter.next()
891
892
893
894
895
896
897
      except StopIteration:
        raise errors.ProgrammerError("Called for a finished job")

      if op.status == constants.OP_STATUS_RUNNING:
        # Found an opcode already marked as running
        raise errors.ProgrammerError("Called for job marked as running")

898
899
      opctx = _OpExecContext(op, idx, "Op %s/%s" % (idx + 1, len(job.ops)),
                             timeout_strategy_factory)
900

901
902
      if op.status not in constants.OPS_FINALIZED:
        return opctx
903

904
905
906
907
908
909
      # This is a job that was partially completed before master daemon
      # shutdown, so it can be expected that some opcodes are already
      # completed successfully (if any did error out, then the whole job
      # should have been aborted and not resubmitted for processing).
      logging.info("%s: opcode %s already processed, skipping",
                   opctx.log_prefix, opctx.summary)
910
911
912
913
914
915
916
917
918

  @staticmethod
  def _MarkWaitlock(job, op):
    """Marks an opcode as waiting for locks.

    The job's start timestamp is also set if necessary.

    @type job: L{_QueuedJob}
    @param job: Job object
919
920
    @type op: L{_QueuedOpCode}
    @param op: Opcode object
921
922
923

    """
    assert op in job.ops
924
925
926
927
    assert op.status in (constants.OP_STATUS_QUEUED,
                         constants.OP_STATUS_WAITLOCK)

    update = False
928
929

    op.result = None
930
931
932
933
934
935
936
937

    if op.status == constants.OP_STATUS_QUEUED:
      op.status = constants.OP_STATUS_WAITLOCK
      update = True

    if op.start_timestamp is None:
      op.start_timestamp = TimeStampNow()
      update = True
938
939
940

    if job.start_timestamp is None:
      job.start_timestamp = op.start_timestamp
941
942
943
944
945
      update = True

    assert op.status == constants.OP_STATUS_WAITLOCK

    return update
946

Michael Hanselmann's avatar
Michael Hanselmann committed
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
  @staticmethod
  def _CheckDependencies(queue, job, opctx):
    """Checks if an opcode has dependencies and if so, processes them.

    @type queue: L{JobQueue}
    @param queue: Queue object
    @type job: L{_QueuedJob}
    @param job: Job object
    @type opctx: L{_OpExecContext}
    @param opctx: Opcode execution context
    @rtype: bool
    @return: Whether opcode will be re-scheduled by dependency tracker

    """
    op = opctx.op

    result = False

    while opctx.jobdeps:
      (dep_job_id, dep_status) = opctx.jobdeps[0]

      (depresult, depmsg) = queue.depmgr.CheckAndRegister(job, dep_job_id,
                                                          dep_status)
      assert ht.TNonEmptyString(depmsg), "No dependency message"

      logging.info("%s: %s", opctx.log_prefix, depmsg)

      if depresult == _JobDependencyManager.CONTINUE:
        # Remove dependency and continue
        opctx.jobdeps.pop(0)

      elif depresult == _JobDependencyManager.WAIT:
        # Need to wait for notification, dependency tracker will re-add job
        # to workerpool
        result = True
        break

      elif depresult == _JobDependencyManager.CANCEL:
        # Job was cancelled, cancel this job as well
        job.Cancel()
        assert op.status == constants.OP_STATUS_CANCELING
        break

      elif depresult in (_JobDependencyManager.WRONGSTATUS,
                         _JobDependencyManager.ERROR):
        # Job failed or there was an error, this job must fail
        op.status = constants.OP_STATUS_ERROR
        op.result = _EncodeOpError(errors.OpExecError(depmsg))
        break

      else:
        raise errors.ProgrammerError("Unknown dependency result '%s'" %
                                     depresult)

    return result

1003
  def _ExecOpCodeUnlocked(self, opctx):
1004
1005
1006
    """Processes one opcode and returns the result.

    """
1007
1008
    op = opctx.op

1009
1010
    assert op.status == constants.OP_STATUS_WAITLOCK

1011
1012
    timeout = opctx.GetNextLockTimeout()

1013
1014
1015
    try:
      # Make sure not to hold queue lock while calling ExecOpCode
      result = self.opexec_fn(op.input,
1016
                              _OpExecCallbacks(self.queue, self.job, op),
1017
                              timeout=timeout, priority=op.priority)
1018
1019
1020
    except mcpu.LockAcquireTimeout:
      assert timeout is not None, "Received timeout for blocking acquire"
      logging.debug("Couldn't acquire locks in %0.6fs", timeout)
1021
1022
1023
1024
1025
1026
1027
1028

      assert op.status in (constants.OP_STATUS_WAITLOCK,
                           constants.OP_STATUS_CANCELING)

      # Was job cancelled while we were waiting for the lock?
      if op.status == constants.OP_STATUS_CANCELING:
        return (constants.OP_STATUS_CANCELING, None)

1029
1030
      # Stay in waitlock while trying to re-acquire lock
      return (constants.OP_STATUS_WAITLOCK, None)
1031
    except CancelJob:
1032
      logging.exception("%s: Canceling job", opctx.log_prefix)
1033
1034
1035
      assert op.status == constants.OP_STATUS_CANCELING
      return (constants.OP_STATUS_CANCELING, None)
    except Exception, err: # pylint: disable-msg=W0703
1036
1037
      logging.exception("%s: Caught exception in %s",
                        opctx.log_prefix, opctx.summary)
1038
1039
      return (constants.OP_STATUS_ERROR, _EncodeOpError(err))
    else:
1040
1041
      logging.debug("%s: %s successful",
                    opctx.log_prefix, opctx.summary)
1042
1043
      return (constants.OP_STATUS_SUCCESS, result)

1044
  def __call__(self, _nextop_fn=None):
1045
1046
    """Continues execution of a job.

1047
    @param _nextop_fn: Callback function for tests
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
    @rtype: bool
    @return: True if job is finished, False if processor needs to be called
             again

    """
    queue = self.queue
    job = self.job

    logging.debug("Processing job %s", job.id)

    queue.acquire(shared=1)
    try:
      opcount = len(job.ops)

1062
1063
      assert job.writable, "Expected writable job"

1064
1065
1066
1067
      # Don't do anything for finalized jobs
      if job.CalcStatus() in constants.JOBS_FINALIZED:
        return True

1068
1069
1070
      # Is a previous opcode still pending?
      if job.cur_opctx:
        opctx = job.cur_opctx
1071
        job.cur_opctx = None
1072
1073
1074
1075
1076
      else:
        if __debug__ and _nextop_fn:
          _nextop_fn()
        opctx = self._FindNextOpcode(job, self._timeout_strategy_factory)

1077
      op = opctx.op
1078
1079
1080

      # Consistency check
      assert compat.all(i.status in (constants.OP_STATUS_QUEUED,
1081
                                     constants.OP_STATUS_CANCELING)
1082
                        for i in job.ops[opctx.index + 1:])
1083
1084
1085

      assert op.status in (constants.OP_STATUS_QUEUED,
                           constants.OP_STATUS_WAITLOCK,
1086
                           constants.OP_STATUS_CANCELING)
1087

1088
1089
1090
      assert (op.priority <= constants.OP_PRIO_LOWEST and
              op.priority >= constants.OP_PRIO_HIGHEST)

Michael Hanselmann's avatar
Michael Hanselmann committed
1091
1092
      waitjob = None

1093
      if op.status != constants.OP_STATUS_CANCELING:
1094
1095
1096
        assert op.status in (constants.OP_STATUS_QUEUED,
                             constants.OP_STATUS_WAITLOCK)

1097
        # Prepare to start opcode
1098
1099
1100
        if self._MarkWaitlock(job, op):
          # Write to disk
          queue.UpdateJobUnlocked(job)
1101
1102
1103

        assert op.status == constants.OP_STATUS_WAITLOCK
        assert job.CalcStatus() == constants.JOB_STATUS_WAITLOCK
1104
        assert job.start_timestamp and op.start_timestamp
Michael Hanselmann's avatar
Michael Hanselmann committed
1105
1106
1107
1108
        assert waitjob is None

        # Check if waiting for a job is necessary
        waitjob = self._CheckDependencies(queue, job, opctx)
1109

Michael Hanselmann's avatar
Michael Hanselmann committed
1110
1111
1112
        assert op.status in (constants.OP_STATUS_WAITLOCK,
                             constants.OP_STATUS_CANCELING,
                             constants.OP_STATUS_ERROR)
1113

Michael Hanselmann's avatar
Michael Hanselmann committed
1114
1115
1116
1117
        if not (waitjob or op.status in (constants.OP_STATUS_CANCELING,
                                         constants.OP_STATUS_ERROR)):
          logging.info("%s: opcode %s waiting for locks",
                       opctx.log_prefix, opctx.summary)
1118

Michael Hanselmann's avatar
Michael Hanselmann committed
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
          assert not opctx.jobdeps, "Not all dependencies were removed"

          queue.release()
          try:
            (op_status, op_result) = self._ExecOpCodeUnlocked(opctx)
          finally:
            queue.acquire(shared=1)

          op.status = op_status
          op.result = op_result

          assert not waitjob
1131

1132
        if op.status == constants.OP_STATUS_WAITLOCK:
1133
1134
          # Couldn't get locks in time
          assert not op.end_timestamp
1135
        else:
1136
1137
          # Finalize opcode
          op.end_timestamp = TimeStampNow()
1138

1139
1140
1141
1142
1143
          if op.status == constants.OP_STATUS_CANCELING:
            assert not compat.any(i.status != constants.OP_STATUS_CANCELING
                                  for i in job.ops[opctx.index:])
          else:
            assert op.status in constants.OPS_FINALIZED
1144

Michael Hanselmann's avatar
Michael Hanselmann committed
1145
      if op.status == constants.OP_STATUS_WAITLOCK or waitjob:
1146
1147
        finalize = False

Michael Hanselmann's avatar
Michael Hanselmann committed
1148
        if not waitjob and opctx.CheckPriorityIncrease():
1149
1150
          # Priority was changed, need to update on-disk file
          queue.UpdateJobUnlocked(job)
1151

1152
1153
        # Keep around for another round
        job.cur_opctx = opctx
1154

1155
1156
        assert (op.priority <= constants.OP_PRIO_LOWEST and
                op.priority >= constants.OP_PRIO_HIGHEST)
1157

1158
        # In no case must the status be finalized here
1159
        assert job.CalcStatus() == constants.JOB_STATUS_WAITLOCK
1160
1161

      else:
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
        # Ensure all opcodes so far have been successful
        assert (opctx.index == 0 or
                compat.all(i.status == constants.OP_STATUS_SUCCESS
                           for i in job.ops[:opctx.index]))

        # Reset context
        job.cur_opctx = None

        if op.status == constants.OP_STATUS_SUCCESS:
          finalize = False

        elif op.status == constants.OP_STATUS_ERROR:
          # Ensure failed opcode has an exception as its result
          assert errors.GetEncodedError(job.ops[opctx.index].result)

          to_encode = errors.OpExecError("Preceding opcode failed")
          job.MarkUnfinishedOps(constants.OP_STATUS_ERROR,
                                _EncodeOpError(to_encode))
          finalize = True
1181

1182
1183
1184
1185
          # Consistency check
          assert compat.all(i.status == constants.OP_STATUS_ERROR and
                            errors.GetEncodedError(i.result)
                            for i in job.ops[opctx.index:])
1186

1187
1188
1189
1190
1191
1192
1193
1194
        elif op.status == constants.OP_STATUS_CANCELING:
          job.MarkUnfinishedOps(constants.OP_STATUS_CANCELED,
                                "Job canceled by request")
          finalize = True

        else:
          raise errors.ProgrammerError("Unknown status '%s'" % op.status)

1195
1196
1197
1198
1199
        if opctx.index == (opcount - 1):
          # Finalize on last opcode
          finalize = True

        if finalize:
1200
          # All opcodes have been run, finalize job
1201
          job.Finalize()
1202
1203
1204
1205

        # Write to disk. If the job status is final, this is the final write
        # allowed. Once the file has been written, it can be archived anytime.
        queue.UpdateJobUnlocked(job)
1206

Michael Hanselmann's avatar
Michael Hanselmann committed
1207
1208
        assert not waitjob

1209
        if finalize:
1210
          logging.info("Finished job %s, status = %s", job.id, job.CalcStatus())
Michael Hanselmann's avatar
Michael Hanselmann committed
1211
1212
          # TODO: Check locking
          queue.depmgr.NotifyWaiters(job.id)
1213
          return True
1214

Michael Hanselmann's avatar
Michael Hanselmann committed
1215
1216
1217
      assert not waitjob or queue.depmgr.JobWaiting(job)

      return bool(waitjob)
1218
    finally:
1219
      assert job.writable, "Job became read-only while being processed"
1220
1221
1222
      queue.release()


1223
1224
1225
1226
class _JobQueueWorker(workerpool.BaseWorker):
  """The actual job workers.

  """
Iustin Pop's avatar
Iustin Pop committed
1227
  def RunTask(self, job): # pylint: disable-msg=W0221
Michael Hanselmann's avatar
Michael Hanselmann committed
1228
1229
    """Job executor.

1230
1231
1232
    @type job: L{_QueuedJob}
    @param job: the job to be processed

Michael Hanselmann's avatar
Michael Hanselmann committed
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
    """
    # Ensure only one worker is active on a single job. If a job registers for
    # a dependency job, and the other job notifies before the first worker is
    # done, the job can end up in the tasklist more than once.
    job.processor_lock.acquire()
    try:
      return self._RunTaskInner(job)
    finally:
      job.processor_lock.release()

  def _RunTaskInner(self, job):
    """Executes a job.

    Must be called with per-job lock acquired.

Michael Hanselmann's avatar
Michael Hanselmann committed
1248
    """
1249
1250
1251
    queue = job.queue
    assert queue == self.pool.queue

1252
1253
    setname_fn = lambda op: self.SetTaskName(self._GetWorkerName(job, op))
    setname_fn(None)
1254

1255
1256
    proc = mcpu.Processor(queue.context, job.id)

1257
1258
1259
1260
1261
    # Create wrapper for setting thread name
    wrap_execop_fn = compat.partial(self._WrapExecOpCode, setname_fn,
                                    proc.ExecOpCode)

    if not _JobProcessor(queue, wrap_execop_fn, job)():
1262
      # Schedule again
1263
      raise workerpool.DeferTask(priority=job.CalcPriority())
Michael Hanselmann's avatar
Michael Hanselmann committed
1264

1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
  @staticmethod
  def _WrapExecOpCode(setname_fn, execop_fn, op, *args, **kwargs):
    """Updates the worker thread name to include a short summary of the opcode.

    @param setname_fn: Callable setting worker thread name
    @param execop_fn: Callable for executing opcode (usually
                      L{mcpu.Processor.ExecOpCode})

    """
    setname_fn(op)
    try:
      return execop_fn(op, *args, **kwargs)
    finally:
      setname_fn(None)

  @staticmethod
  def _GetWorkerName(job, op):
    """Sets the worker thread name.

    @type job: L{_QueuedJob}
    @type op: L{opcodes.OpCode}

    """
    parts = ["Job%s" % job.id]

    if op:
      parts.append(op.TinySummary())

    return "/".join(parts)

Michael Hanselmann's avatar
Michael Hanselmann committed
1295
1296

class _JobQueueWorkerPool(workerpool.WorkerPool):
1297
1298
1299
  """Simple class implementing a job-processing workerpool.

  """
1300
  def __init__(self, queue):
1301
    super(_JobQueueWorkerPool, self).__init__("Jq",
1302
                                              JOBQUEUE_THREADS,
Michael Hanselmann's avatar
Michael Hanselmann committed
1303
                                              _JobQueueWorker)
1304
    self.queue = queue