Synnefo » snf-ganeti

#

#

# Copyright (C) 2006, 2007, 2008 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 os

import logging

import threading

import errno

import re

import time

import weakref

from ganeti import constants

from ganeti import serializer

from ganeti import workerpool

from ganeti import opcodes

from ganeti import errors

from ganeti import mcpu

from ganeti import utils

from ganeti import jstore

from ganeti import rpc

JOBQUEUE_THREADS = 25

JOBS_PER_ARCHIVE_DIRECTORY = 10000

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 stop_timestamp: timestamp for the end of the execution

  """

  __slots__ = ["input", "status", "result", "log",

               "start_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.end_timestamp = None

  @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.end_timestamp = state.get("end_timestamp", None)

    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,

      "end_timestamp": self.end_timestamp,

      }

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

  @ivar lock_status: In-memory locking information for debugging

  @ivar change: a Condition variable we use for waiting for job changes

  """

  # pylint: disable-msg=W0212

  __slots__ = ["queue", "id", "ops", "log_serial",

               "received_timestamp", "start_timestamp", "end_timestamp",

               "lock_status", "change",

               "__weakref__"]

  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:

      # TODO: use a better exception

      raise Exception("No opcodes")

    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

    # In-memory attributes

    self.lock_status = None

    # Condition to wait for changes

    self.change = threading.Condition(self.queue._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)

    # In-memory attributes

    obj.lock_status = 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)

    # Condition to wait for changes

    obj.change = threading.Condition(obj.queue._lock)

    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 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 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

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 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.

    """

    self._queue.acquire()

    try:

      assert self._op.status in (constants.OP_STATUS_WAITLOCK,

                                 constants.OP_STATUS_CANCELING)

      # All locks are acquired by now

      self._job.lock_status = None

      # Cancel here if we were asked to

      if self._op.status == constants.OP_STATUS_CANCELING:

        raise CancelJob()

      self._op.status = constants.OP_STATUS_RUNNING

    finally:

      self._queue.release()

  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._queue.acquire()

    try:

      self._job.log_serial += 1

      self._op.log.append((self._job.log_serial, timestamp, log_type, log_msg))

      self._job.change.notifyAll()

    finally:

      self._queue.release()

  def ReportLocks(self, msg):

    """Write locking information to the job.

    Called whenever the LU processor is waiting for a lock or has acquired one.

    """

    # Not getting the queue lock because this is a single assignment

    self._job.lock_status = msg

class _JobQueueWorker(workerpool.BaseWorker):

  """The actual job workers.

  """

  def RunTask(self, job): # pylint: disable-msg=W0221

    """Job executor.

    This functions processes a job. It is closely tied to the _QueuedJob and

    _QueuedOpCode classes.

    @type job: L{_QueuedJob}

    @param job: the job to be processed

    """

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

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

    queue = job.queue

    try:

      try:

        count = len(job.ops)

        for idx, op in enumerate(job.ops):

          op_summary = op.input.Summary()

          if op.status == constants.OP_STATUS_SUCCESS:

            # 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("Op %s/%s: opcode %s already processed, skipping",

                         idx + 1, count, op_summary)

            continue

          try:

            logging.info("Op %s/%s: Starting opcode %s", idx + 1, count,

                         op_summary)

            queue.acquire()

            try:

              if op.status == constants.OP_STATUS_CANCELED:

                raise CancelJob()

              assert op.status == constants.OP_STATUS_QUEUED

              op.status = constants.OP_STATUS_WAITLOCK

              op.result = None

              op.start_timestamp = TimeStampNow()

              if idx == 0: # first opcode

                job.start_timestamp = op.start_timestamp

              queue.UpdateJobUnlocked(job)

              input_opcode = op.input

            finally:

              queue.release()

            # Make sure not to hold queue lock while calling ExecOpCode

            result = proc.ExecOpCode(input_opcode,

                                     _OpExecCallbacks(queue, job, op))

            queue.acquire()

            try:

              op.status = constants.OP_STATUS_SUCCESS

              op.result = result

              op.end_timestamp = TimeStampNow()

              queue.UpdateJobUnlocked(job)

            finally:

              queue.release()

            logging.info("Op %s/%s: Successfully finished opcode %s",

                         idx + 1, count, op_summary)

          except CancelJob:

            # Will be handled further up

            raise

          except Exception, err:

            queue.acquire()

            try:

              try:

                op.status = constants.OP_STATUS_ERROR

                if isinstance(err, errors.GenericError):

                  op.result = errors.EncodeException(err)

                else:

                  op.result = str(err)

                op.end_timestamp = TimeStampNow()

                logging.info("Op %s/%s: Error in opcode %s: %s",

                             idx + 1, count, op_summary, err)

              finally:

                queue.UpdateJobUnlocked(job)

            finally:

              queue.release()

            raise

      except CancelJob:

        queue.acquire()

        try:

          queue.CancelJobUnlocked(job)

        finally:

          queue.release()

      except errors.GenericError, err:

        logging.exception("Ganeti exception")

      except:

        logging.exception("Unhandled exception")

    finally:

      queue.acquire()

      try:

        try:

          job.lock_status = None

          job.end_timestamp = TimeStampNow()

          queue.UpdateJobUnlocked(job)

        finally:

          job_id = job.id

          status = job.CalcStatus()

      finally:

        queue.release()

      logging.info("Finished job %s, status = %s", job_id, status)

class _JobQueueWorkerPool(workerpool.WorkerPool):

  """Simple class implementing a job-processing workerpool.

  """

  def __init__(self, queue):

    super(_JobQueueWorkerPool, self).__init__("JobQueue",

                                              JOBQUEUE_THREADS,

                                              _JobQueueWorker)

    self.queue = queue

def _RequireOpenQueue(fn):

  """Decorator for "public" functions.

  This function should be used for all 'public' functions. That is,

  functions usually called from other classes. Note that this should

  be applied only to methods (not plain functions), since it expects

  that the decorated function is called with a first argument that has

  a '_queue_lock' argument.

  @warning: Use this decorator only after utils.LockedMethod!

  Example::

    @utils.LockedMethod

    @_RequireOpenQueue

    def Example(self):

      pass

  """

  def wrapper(self, *args, **kwargs):

    # pylint: disable-msg=W0212

    assert self._queue_lock is not None, "Queue should be open"

    return fn(self, *args, **kwargs)

  return wrapper

class JobQueue(object):

  """Queue used to manage the jobs.

  @cvar _RE_JOB_FILE: regex matching the valid job file names

  """

  _RE_JOB_FILE = re.compile(r"^job-(%s)$" % constants.JOB_ID_TEMPLATE)

  def __init__(self, context):

    """Constructor for JobQueue.

    The constructor will initialize the job queue object and then

    start loading the current jobs from disk, either for starting them

    (if they were queue) or for aborting them (if they were already

    running).

    @type context: GanetiContext

    @param context: the context object for access to the configuration

        data and other ganeti objects

    """

    self.context = context

    self._memcache = weakref.WeakValueDictionary()

    self._my_hostname = utils.HostInfo().name

    # Locking

    self._lock = threading.Lock()

    self.acquire = self._lock.acquire

    self.release = self._lock.release

    # Initialize

    self._queue_lock = jstore.InitAndVerifyQueue(must_lock=True)

    # Read serial file

    self._last_serial = jstore.ReadSerial()

    assert self._last_serial is not None, ("Serial file was modified between"

                                           " check in jstore and here")

    # Get initial list of nodes

    self._nodes = dict((n.name, n.primary_ip)

                       for n in self.context.cfg.GetAllNodesInfo().values()

                       if n.master_candidate)

    # Remove master node

    try:

      del self._nodes[self._my_hostname]

    except KeyError:

      pass

    # TODO: Check consistency across nodes

    # Setup worker pool

    self._wpool = _JobQueueWorkerPool(self)

    try:

      # We need to lock here because WorkerPool.AddTask() may start a job while

      # we're still doing our work.

      self.acquire()

      try:

        logging.info("Inspecting job queue")

        all_job_ids = self._GetJobIDsUnlocked()

        jobs_count = len(all_job_ids)

        lastinfo = time.time()

        for idx, job_id in enumerate(all_job_ids):

          # Give an update every 1000 jobs or 10 seconds

          if (idx % 1000 == 0 or time.time() >= (lastinfo + 10.0) or

              idx == (jobs_count - 1)):

            logging.info("Job queue inspection: %d/%d (%0.1f %%)",

                         idx, jobs_count - 1, 100.0 * (idx + 1) / jobs_count)

            lastinfo = time.time()

          job = self._LoadJobUnlocked(job_id)

          # a failure in loading the job can cause 'None' to be returned

          if job is None:

            continue

          status = job.CalcStatus()

          if status in (constants.JOB_STATUS_QUEUED, ):

            self._wpool.AddTask(job)

          elif status in (constants.JOB_STATUS_RUNNING,

                          constants.JOB_STATUS_WAITLOCK,

                          constants.JOB_STATUS_CANCELING):

            logging.warning("Unfinished job %s found: %s", job.id, job)

            try:

              job.MarkUnfinishedOps(constants.OP_STATUS_ERROR,

                                    "Unclean master daemon shutdown")

            finally:

              self.UpdateJobUnlocked(job)

        logging.info("Job queue inspection finished")

      finally:

        self.release()

    except:

      self._wpool.TerminateWorkers()

      raise

  @utils.LockedMethod

  @_RequireOpenQueue

  def AddNode(self, node):

    """Register a new node with the queue.

    @type node: L{objects.Node}

    @param node: the node object to be added

    """

    node_name = node.name

    assert node_name != self._my_hostname

    # Clean queue directory on added node

    result = rpc.RpcRunner.call_jobqueue_purge(node_name)

    msg = result.fail_msg

    if msg:

      logging.warning("Cannot cleanup queue directory on node %s: %s",

                      node_name, msg)

    if not node.master_candidate:

      # remove if existing, ignoring errors

      self._nodes.pop(node_name, None)

      # and skip the replication of the job ids

      return

    # Upload the whole queue excluding archived jobs

    files = [self._GetJobPath(job_id) for job_id in self._GetJobIDsUnlocked()]

    # Upload current serial file

    files.append(constants.JOB_QUEUE_SERIAL_FILE)

    for file_name in files:

      # Read file content

      content = utils.ReadFile(file_name)

      result = rpc.RpcRunner.call_jobqueue_update([node_name],

                                                  [node.primary_ip],

                                                  file_name, content)

      msg = result[node_name].fail_msg

      if msg:

        logging.error("Failed to upload file %s to node %s: %s",

                      file_name, node_name, msg)

    self._nodes[node_name] = node.primary_ip

  @utils.LockedMethod

  @_RequireOpenQueue

  def RemoveNode(self, node_name):

    """Callback called when removing nodes from the cluster.

    @type node_name: str

    @param node_name: the name of the node to remove

    """

    try:

      # The queue is removed by the "leave node" RPC call.

      del self._nodes[node_name]

    except KeyError:

      pass

  @staticmethod

  def _CheckRpcResult(result, nodes, failmsg):

    """Verifies the status of an RPC call.

    Since we aim to keep consistency should this node (the current

    master) fail, we will log errors if our rpc fail, and especially

    log the case when more than half of the nodes fails.

    @param result: the data as returned from the rpc call

    @type nodes: list

    @param nodes: the list of nodes we made the call to

    @type failmsg: str

    @param failmsg: the identifier to be used for logging

    """

    failed = []

    success = []

    for node in nodes:

      msg = result[node].fail_msg

      if msg:

        failed.append(node)

        logging.error("RPC call %s (%s) failed on node %s: %s",