jqueue.py

#
#

# Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 Google Inc.
#
# 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.


"""Module implementing the job queue handling.

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

"""

import logging
import errno
import re
import time
import weakref
import threading

try:
  # pylint: disable-msg=E0611
  from pyinotify import pyinotify
except ImportError:
  import pyinotify

from ganeti import asyncnotifier
from ganeti import constants
from ganeti import serializer
from ganeti import workerpool
from ganeti import locking
from ganeti import opcodes
from ganeti import errors
from ganeti import mcpu
from ganeti import utils
from ganeti import jstore
from ganeti import rpc
from ganeti import runtime
from ganeti import netutils
from ganeti import compat
from ganeti import ht


JOBQUEUE_THREADS = 25
JOBS_PER_ARCHIVE_DIRECTORY = 10000

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


class CancelJob(Exception):
  """Special exception to cancel a job.

  """


def TimeStampNow():
  """Returns the current timestamp.

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

  """
  return utils.SplitTime(time.time())


class _QueuedOpCode(object):
  """Encapsulates an opcode object.

  @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
  @ivar exec_timestamp: timestamp for the actual LU Exec() function invocation
  @ivar stop_timestamp: timestamp for the end of the execution

  """
  __slots__ = ["input", "status", "result", "log", "priority",
               "start_timestamp", "exec_timestamp", "end_timestamp",
               "__weakref__"]

  def __init__(self, op):
    """Constructor for the _QuededOpCode.

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

    """
    self.input = op
    self.status = constants.OP_STATUS_QUEUED
    self.result = None
    self.log = []
    self.start_timestamp = None
    self.exec_timestamp = None
    self.end_timestamp = None

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

  @classmethod
  def Restore(cls, state):
    """Restore the _QueuedOpCode from the serialized form.

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

    """
    obj = _QueuedOpCode.__new__(cls)
    obj.input = opcodes.OpCode.LoadOpCode(state["input"])
    obj.status = state["status"]
    obj.result = state["result"]
    obj.log = state["log"]
    obj.start_timestamp = state.get("start_timestamp", None)
    obj.exec_timestamp = state.get("exec_timestamp", None)
    obj.end_timestamp = state.get("end_timestamp", None)
    obj.priority = state.get("priority", constants.OP_PRIO_DEFAULT)
    return obj

  def Serialize(self):
    """Serializes this _QueuedOpCode.

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

    """
    return {
      "input": self.input.__getstate__(),
      "status": self.status,
      "result": self.result,
      "log": self.log,
      "start_timestamp": self.start_timestamp,
      "exec_timestamp": self.exec_timestamp,
      "end_timestamp": self.end_timestamp,
      "priority": self.priority,
      }


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

  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

  """
  # pylint: disable-msg=W0212
  __slots__ = ["queue", "id", "ops", "log_serial", "ops_iter", "cur_opctx",
               "received_timestamp", "start_timestamp", "end_timestamp",
               "__weakref__", "processor_lock"]

  def __init__(self, queue, job_id, ops):
    """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

    """
    if not ops:
      raise errors.GenericError("A job needs at least one opcode")

    self.queue = queue
    self.id = job_id
    self.ops = [_QueuedOpCode(op) for op in ops]
    self.log_serial = 0
    self.received_timestamp = TimeStampNow()
    self.start_timestamp = None
    self.end_timestamp = None

    self._InitInMemory(self)

  @staticmethod
  def _InitInMemory(obj):
    """Initializes in-memory variables.

    """
    obj.ops_iter = None
    obj.cur_opctx = None
    obj.processor_lock = threading.Lock()

  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))

  @classmethod
  def Restore(cls, queue, state):
    """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
    @rtype: _JobQueue
    @return: the restored _JobQueue instance

    """
    obj = _QueuedJob.__new__(cls)
    obj.queue = queue
    obj.id = state["id"]
    obj.received_timestamp = state.get("received_timestamp", None)
    obj.start_timestamp = state.get("start_timestamp", None)
    obj.end_timestamp = state.get("end_timestamp", None)

    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)

    cls._InitInMemory(obj)

    return obj

  def Serialize(self):
    """Serialize the _JobQueue instance.

    @rtype: dict
    @return: the serialized state

    """
    return {
      "id": self.id,
      "ops": [op.Serialize() for op in self.ops],
      "start_timestamp": self.start_timestamp,
      "end_timestamp": self.end_timestamp,
      "received_timestamp": self.received_timestamp,
      }

  def CalcStatus(self):
    """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
          - canceling
          - 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

    """
    status = constants.JOB_STATUS_QUEUED

    all_success = True
    for op in self.ops:
      if op.status == constants.OP_STATUS_SUCCESS:
        continue

      all_success = False

      if op.status == constants.OP_STATUS_QUEUED:
        pass
      elif op.status == constants.OP_STATUS_WAITLOCK:
        status = constants.JOB_STATUS_WAITLOCK
      elif op.status == constants.OP_STATUS_RUNNING:
        status = constants.JOB_STATUS_RUNNING
      elif op.status == constants.OP_STATUS_CANCELING:
        status = constants.JOB_STATUS_CANCELING
        break
      elif op.status == constants.OP_STATUS_ERROR:
        status = constants.JOB_STATUS_ERROR
        # The whole job fails if one opcode failed
        break
      elif op.status == constants.OP_STATUS_CANCELED:
        status = constants.OP_STATUS_CANCELED
        break

    if all_success:
      status = constants.JOB_STATUS_SUCCESS

    return status

  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)

  def GetLogEntries(self, newer_than):
    """Selectively returns the log entries.

    @type newer_than: None or int
    @param newer_than: if this is None, return all log entries,
        otherwise return only the log entries with serial higher
        than this value
    @rtype: list
    @return: the list of the log entries selected

    """
    if newer_than is None:
      serial = -1
    else:
      serial = newer_than

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

    return entries

  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())
      elif fname == "priority":
        row.append(self.CalcPriority())
      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])
      elif fname == "oppriority":
        row.append([op.priority for op in self.ops])
      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

  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

    """
    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

  def Finalize(self):
    """Marks the job as finalized.

    """
    self.end_timestamp = TimeStampNow()

  def Cancel(self):
    """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

    """
    status = self.CalcStatus()

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

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

    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)


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

    @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

  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()

  @locking.ssynchronized(_QUEUE, shared=1)
  def NotifyStart(self):
    """Mark the opcode as running, not lock-waiting.

    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.

    """
    assert self._op in self._job.ops
    assert self._op.status in (constants.OP_STATUS_WAITLOCK,
                               constants.OP_STATUS_CANCELING)

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

    logging.debug("Opcode is now running")

    self._op.status = constants.OP_STATUS_RUNNING
    self._op.exec_timestamp = TimeStampNow()

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

  @locking.ssynchronized(_QUEUE, shared=1)
  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)

  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())
    self._AppendFeedback(timestamp, log_type, log_msg)

  def CheckCancel(self):
    """Check whether job has been cancelled.

    """
    assert self._op.status in (constants.OP_STATUS_WAITLOCK,
                               constants.OP_STATUS_CANCELING)

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

  def SubmitManyJobs(self, jobs):
    """Submits jobs for processing.

    See L{JobQueue.SubmitManyJobs}.

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


class _JobChangesChecker(object):
  def __init__(self, fields, prev_job_info, prev_log_serial):
    """Initializes this class.

    @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

    """
    self._fields = fields
    self._prev_job_info = prev_job_info
    self._prev_log_serial = prev_log_serial

  def __call__(self, job):
    """Checks whether job has changed.

    @type job: L{_QueuedJob}
    @param job: Job object

    """
    status = job.CalcStatus()
    job_info = job.GetInfo(self._fields)
    log_entries = job.GetLogEntries(self._prev_log_serial)

    # 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.
    # 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))

    # Don't even try to wait if the job is no longer running, there will be
    # no changes.
    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)

    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

    """
    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.

    """
    if not notifier_enabled:
      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

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

    @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

    """
    try:
      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()
    except (errors.InotifyError, errors.JobLost):
      return None
    except utils.RetryTimeout:
      return constants.JOB_NOTCHANGED


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)


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


class _OpExecContext:
  def __init__(self, op, index, log_prefix, timeout_strategy_factory):
    """Initializes this class.

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

    # Create local copy to modify
    if getattr(op.input, opcodes.DEPEND_ATTR, None):
      self.jobdeps = op.input.depends[:]
    else:
      self.jobdeps = None

    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()


class _JobProcessor(object):
  def __init__(self, queue, opexec_fn, job,
               _timeout_strategy_factory=mcpu.LockAttemptTimeoutStrategy):
    """Initializes this class.

    """
    self.queue = queue
    self.opexec_fn = opexec_fn
    self.job = job
    self._timeout_strategy_factory = _timeout_strategy_factory

  @staticmethod
  def _FindNextOpcode(job, timeout_strategy_factory):
    """Locates the next opcode to run.

    @type job: L{_QueuedJob}
    @param job: Job object
    @param timeout_strategy_factory: Callable to create new timeout strategy

    """
    # 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.
    if job.ops_iter is None:
      job.ops_iter = enumerate(job.ops)

    # Find next opcode to run
    while True:
      try:
        (idx, op) = job.ops_iter.next()
      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")

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

      if op.status not in constants.OPS_FINALIZED:
        return opctx

      # 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)

  @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
    @type op: L{_QueuedOpCode}
    @param op: Opcode object

    """
    assert op in job.ops
    assert op.status in (constants.OP_STATUS_QUEUED,
                         constants.OP_STATUS_WAITLOCK)

    update = False

    op.result = None

    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

    if job.start_timestamp is None:
      job.start_timestamp = op.start_timestamp
      update = True

    assert op.status == constants.OP_STATUS_WAITLOCK

    return update

  @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

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

    """
    op = opctx.op

    assert op.status == constants.OP_STATUS_WAITLOCK

    timeout = opctx.GetNextLockTimeout()

    try:
      # Make sure not to hold queue lock while calling ExecOpCode
      result = self.opexec_fn(op.input,
                              _OpExecCallbacks(self.queue, self.job, op),
                              timeout=timeout, priority=op.priority)
    except mcpu.LockAcquireTimeout:
      assert timeout is not None, "Received timeout for blocking acquire"
      logging.debug("Couldn't acquire locks in %0.6fs", timeout)

      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)

      # Stay in waitlock while trying to re-acquire lock
      return (constants.OP_STATUS_WAITLOCK, None)
    except CancelJob:
      logging.exception("%s: Canceling job", opctx.log_prefix)
      assert op.status == constants.OP_STATUS_CANCELING
      return (constants.OP_STATUS_CANCELING, None)
    except Exception, err: # pylint: disable-msg=W0703
      logging.exception("%s: Caught exception in %s",
                        opctx.log_prefix, opctx.summary)
      return (constants.OP_STATUS_ERROR, _EncodeOpError(err))
    else:
      logging.debug("%s: %s successful",
                    opctx.log_prefix, opctx.summary)
      return (constants.OP_STATUS_SUCCESS, result)

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

    @param _nextop_fn: Callback function for tests
    @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)

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

      # Is a previous opcode still pending?
      if job.cur_opctx:
        opctx = job.cur_opctx
        job.cur_opctx = None
      else:
        if __debug__ and _nextop_fn:
          _nextop_fn()
        opctx = self._FindNextOpcode(job, self._timeout_strategy_factory)

      op = opctx.op

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

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

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

      waitjob = None

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

        # Prepare to start opcode
        if self._MarkWaitlock(job, op):
          # Write to disk
          queue.UpdateJobUnlocked(job)

        assert op.status == constants.OP_STATUS_WAITLOCK
        assert job.CalcStatus() == constants.JOB_STATUS_WAITLOCK
        assert job.start_timestamp and op.start_timestamp
        assert waitjob is None

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

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

        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)

          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

        if op.status == constants.OP_STATUS_WAITLOCK:
          # Couldn't get locks in time
          assert not op.end_timestamp
        else:
          # Finalize opcode
          op.end_timestamp = TimeStampNow()

          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

      if op.status == constants.OP_STATUS_WAITLOCK or waitjob:
        finalize = False

        if not waitjob and opctx.CheckPriorityIncrease():
          # Priority was changed, need to update on-disk file
          queue.UpdateJobUnlocked(job)

        # Keep around for another round
        job.cur_opctx = opctx

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

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

      else:
        # 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

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

        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)

        if opctx.index == (opcount - 1):
          # Finalize on last opcode
          finalize = True

        if finalize:
          # All opcodes have been run, finalize job
          job.Finalize()

        # 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)

        assert not waitjob

        if finalize:
          logging.info("Finished job %s, status = %s", job.id, job.CalcStatus())
          # TODO: Check locking
          queue.depmgr.NotifyWaiters(job.id)
          return True

      assert not waitjob or queue.depmgr.JobWaiting(job)

      return bool(waitjob)
    finally:
      queue.release()


class _JobQueueWorker(workerpool.BaseWorker):
  """The actual job workers.

  """
  def RunTask(self, job): # pylint: disable-msg=W0221
    """Job executor.

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

    """
    # 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.

    """
    queue = job.queue
    assert queue == self.pool.queue

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

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

    # 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)():
      # Schedule again
      raise workerpool.DeferTask(priority=job.CalcPriority())

  @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)


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

  """
  def __init__(self, queue):
    super(_JobQueueWorkerPool, self).__init__("Jq",
                                              JOBQUEUE_THREADS,
                                              _JobQueueWorker)
    self.queue = queue


class _JobDependencyManager:
  """Keeps track of job dependencies.

  """
  (WAIT,
   ERROR,
   CANCEL,
   CONTINUE,
   WRONGSTATUS) = range(1, 6)

  # TODO: Export waiter information to lock monitor

  def __init__(self, getstatus_fn, enqueue_fn):
    """Initializes this class.

    """
    self._getstatus_fn = getstatus_fn
    self._enqueue_fn = enqueue_fn

    self._waiters = {}
    self._lock = locking.SharedLock(