+ counter = itertools.count()
+ try:
+ check_fn = _JobChangesChecker(fields, prev_job_info, prev_log_serial)
+ waiter = _waiter_cls(filename)
+ try:
+ return utils.Retry(compat.partial(self._CheckForChanges,
+ counter, job_load_fn, check_fn),
+ utils.RETRY_REMAINING_TIME, timeout,
+ wait_fn=waiter.Wait)
+ finally:
+ waiter.Close()
+ except 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_base.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):
+ (DEFER,
+ WAITDEP,
+ FINISHED) = range(1, 4)
+
+ 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_WAITING)
+
+ update = False
+
+ op.result = None
+
+ if op.status == constants.OP_STATUS_QUEUED:
+ op.status = constants.OP_STATUS_WAITING
+ 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_WAITING
+
+ 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_WAITING
+
+ 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)
+ 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_WAITING,
+ 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)
+
+ # Queue is shutting down, return to queued
+ if not self.queue.AcceptingJobsUnlocked():
+ return (constants.OP_STATUS_QUEUED, None)
+
+ # Stay in waitlock while trying to re-acquire lock
+ return (constants.OP_STATUS_WAITING, 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 QueueShutdown:
+ logging.exception("%s: Queue is shutting down", opctx.log_prefix)
+
+ assert op.status == constants.OP_STATUS_WAITING
+
+ # Job hadn't been started yet, so it should return to the queue
+ return (constants.OP_STATUS_QUEUED, None)
+
+ except Exception, err: # pylint: disable=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
+ @return: C{FINISHED} if job is fully processed, C{DEFER} if the job should
+ be deferred and C{WAITDEP} if the dependency manager
+ (L{_JobDependencyManager}) will re-schedule the job when appropriate
+
+ """
+ queue = self.queue
+ job = self.job
+
+ logging.debug("Processing job %s", job.id)
+
+ queue.acquire(shared=1)
+ try:
+ opcount = len(job.ops)
+
+ assert job.writable, "Expected writable job"
+
+ # Don't do anything for finalized jobs
+ if job.CalcStatus() in constants.JOBS_FINALIZED:
+ return self.FINISHED
+
+ # 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_WAITING,
+ 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_WAITING)
+
+ # Prepare to start opcode
+ if self._MarkWaitlock(job, op):
+ # Write to disk
+ queue.UpdateJobUnlocked(job)
+
+ assert op.status == constants.OP_STATUS_WAITING
+ assert job.CalcStatus() == constants.JOB_STATUS_WAITING
+ 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_WAITING,
+ 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 in (constants.OP_STATUS_WAITING,
+ constants.OP_STATUS_QUEUED):
+ # waiting: Couldn't get locks in time
+ # queued: Queue is shutting down
+ 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_QUEUED:
+ # Queue is shutting down
+ assert not waitjob
+
+ finalize = False
+
+ # Reset context
+ job.cur_opctx = None
+
+ # In no case must the status be finalized here
+ assert job.CalcStatus() == constants.JOB_STATUS_QUEUED
+
+ elif op.status == constants.OP_STATUS_WAITING 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_WAITING
+
+ 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())
+ return self.FINISHED
+
+ assert not waitjob or queue.depmgr.JobWaiting(job)
+
+ if waitjob:
+ return self.WAITDEP
+ else:
+ return self.DEFER
+ finally:
+ assert job.writable, "Job became read-only while being processed"
+ queue.release()
+
+
+def _EvaluateJobProcessorResult(depmgr, job, result):
+ """Looks at a result from L{_JobProcessor} for a job.
+
+ To be used in a L{_JobQueueWorker}.
+
+ """
+ if result == _JobProcessor.FINISHED:
+ # Notify waiting jobs
+ depmgr.NotifyWaiters(job.id)
+
+ elif result == _JobProcessor.DEFER:
+ # Schedule again
+ raise workerpool.DeferTask(priority=job.CalcPriority())
+
+ elif result == _JobProcessor.WAITDEP:
+ # No-op, dependency manager will re-schedule
+ pass
+
+ else:
+ raise errors.ProgrammerError("Job processor returned unknown status %s" %
+ (result, ))
+
+
+class _JobQueueWorker(workerpool.BaseWorker):
+ """The actual job workers.
+
+ """
+ def RunTask(self, job): # pylint: disable=W0221
+ """Job executor.
+
+ @type job: L{_QueuedJob}
+ @param job: the job to be processed
+
+ """
+ assert job.writable, "Expected writable job"
+
+ # 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)
+
+ _EvaluateJobProcessorResult(queue.depmgr, job,
+ _JobProcessor(queue, wrap_execop_fn, job)())
+
+ @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)
+
+ 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("JobDepMgr")
+
+ @locking.ssynchronized(_LOCK, shared=1)
+ def GetLockInfo(self, requested): # pylint: disable=W0613
+ """Retrieves information about waiting jobs.
+
+ @type requested: set
+ @param requested: Requested information, see C{query.LQ_*}
+
+ """
+ # No need to sort here, that's being done by the lock manager and query
+ # library. There are no priorities for notifying jobs, hence all show up as
+ # one item under "pending".
+ return [("job/%s" % job_id, None, None,
+ [("job", [job.id for job in waiters])])
+ for job_id, waiters in self._waiters.items()
+ if waiters]
+
+ @locking.ssynchronized(_LOCK, shared=1)
+ def JobWaiting(self, job):
+ """Checks if a job is waiting.
+
+ """
+ return compat.any(job in jobs
+ for jobs in self._waiters.values())
+
+ @locking.ssynchronized(_LOCK)
+ def CheckAndRegister(self, job, dep_job_id, dep_status):
+ """Checks if a dependency job has the requested status.
+
+ If the other job is not yet in a finalized status, the calling job will be
+ notified (re-added to the workerpool) at a later point.
+
+ @type job: L{_QueuedJob}
+ @param job: Job object
+ @type dep_job_id: int
+ @param dep_job_id: ID of dependency job
+ @type dep_status: list
+ @param dep_status: Required status
+
+ """
+ assert ht.TJobId(job.id)
+ assert ht.TJobId(dep_job_id)
+ assert ht.TListOf(ht.TElemOf(constants.JOBS_FINALIZED))(dep_status)
+
+ if job.id == dep_job_id:
+ return (self.ERROR, "Job can't depend on itself")
+
+ # Get status of dependency job
+ try:
+ status = self._getstatus_fn(dep_job_id)
+ except errors.JobLost, err:
+ return (self.ERROR, "Dependency error: %s" % err)
+
+ assert status in constants.JOB_STATUS_ALL
+
+ job_id_waiters = self._waiters.setdefault(dep_job_id, set())
+
+ if status not in constants.JOBS_FINALIZED:
+ # Register for notification and wait for job to finish
+ job_id_waiters.add(job)
+ return (self.WAIT,
+ "Need to wait for job %s, wanted status '%s'" %
+ (dep_job_id, dep_status))
+
+ # Remove from waiters list
+ if job in job_id_waiters:
+ job_id_waiters.remove(job)
+
+ if (status == constants.JOB_STATUS_CANCELED and
+ constants.JOB_STATUS_CANCELED not in dep_status):
+ return (self.CANCEL, "Dependency job %s was cancelled" % dep_job_id)
+
+ elif not dep_status or status in dep_status:
+ return (self.CONTINUE,
+ "Dependency job %s finished with status '%s'" %
+ (dep_job_id, status))
+
+ else:
+ return (self.WRONGSTATUS,
+ "Dependency job %s finished with status '%s',"
+ " not one of '%s' as required" %
+ (dep_job_id, status, utils.CommaJoin(dep_status)))
+
+ def _RemoveEmptyWaitersUnlocked(self):
+ """Remove all jobs without actual waiters.
+
+ """
+ for job_id in [job_id for (job_id, waiters) in self._waiters.items()
+ if not waiters]:
+ del self._waiters[job_id]
+
+ def NotifyWaiters(self, job_id):
+ """Notifies all jobs waiting for a certain job ID.
+
+ @attention: Do not call until L{CheckAndRegister} returned a status other
+ than C{WAITDEP} for C{job_id}, or behaviour is undefined
+ @type job_id: int
+ @param job_id: Job ID
+
+ """
+ assert ht.TJobId(job_id)
+
+ self._lock.acquire()
+ try:
+ self._RemoveEmptyWaitersUnlocked()
+
+ jobs = self._waiters.pop(job_id, None)
+ finally:
+ self._lock.release()
+
+ if jobs:
+ # Re-add jobs to workerpool
+ logging.debug("Re-adding %s jobs which were waiting for job %s",
+ len(jobs), job_id)
+ self._enqueue_fn(jobs)
+
+
+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_filelock' argument.
+
+ @warning: Use this decorator only after locking.ssynchronized
+
+ Example::
+ @locking.ssynchronized(_LOCK)
+ @_RequireOpenQueue
+ def Example(self):
+ pass
+
+ """
+ def wrapper(self, *args, **kwargs):
+ # pylint: disable=W0212
+ assert self._queue_filelock is not None, "Queue should be open"
+ return fn(self, *args, **kwargs)
+ return wrapper
+
+
+def _RequireNonDrainedQueue(fn):
+ """Decorator checking for a non-drained queue.
+
+ To be used with functions submitting new jobs.
+
+ """
+ def wrapper(self, *args, **kwargs):
+ """Wrapper function.
+
+ @raise errors.JobQueueDrainError: if the job queue is marked for draining
+
+ """
+ # Ok when sharing the big job queue lock, as the drain file is created when
+ # the lock is exclusive.
+ # Needs access to protected member, pylint: disable=W0212
+ if self._drained:
+ raise errors.JobQueueDrainError("Job queue is drained, refusing job")
+
+ if not self._accepting_jobs:
+ raise errors.JobQueueError("Job queue is shutting down, refusing job")
+
+ return fn(self, *args, **kwargs)
+ return wrapper
+
+
+class JobQueue(object):
+ """Queue used to manage the jobs.
+
+ """
+ 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 = netutils.Hostname.GetSysName()
+
+ # The Big JobQueue lock. If a code block or method acquires it in shared
+ # mode safe it must guarantee concurrency with all the code acquiring it in
+ # shared mode, including itself. In order not to acquire it at all
+ # concurrency must be guaranteed with all code acquiring it in shared mode
+ # and all code acquiring it exclusively.
+ self._lock = locking.SharedLock("JobQueue")
+
+ self.acquire = self._lock.acquire
+ self.release = self._lock.release
+
+ # Accept jobs by default
+ self._accepting_jobs = True
+
+ # Initialize the queue, and acquire the filelock.
+ # This ensures no other process is working on the job queue.
+ self._queue_filelock = 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
+ self._nodes.pop(self._my_hostname, None)
+
+ # TODO: Check consistency across nodes
+
+ self._queue_size = None
+ self._UpdateQueueSizeUnlocked()
+ assert ht.TInt(self._queue_size)
+ self._drained = jstore.CheckDrainFlag()
+
+ # Job dependencies
+ self.depmgr = _JobDependencyManager(self._GetJobStatusForDependencies,
+ self._EnqueueJobs)
+ self.context.glm.AddToLockMonitor(self.depmgr)
+
+ # Setup worker pool
+ self._wpool = _JobQueueWorkerPool(self)
+ try:
+ self._InspectQueue()
+ except:
+ self._wpool.TerminateWorkers()
+ raise
+
+ @locking.ssynchronized(_LOCK)
+ @_RequireOpenQueue
+ def _InspectQueue(self):
+ """Loads the whole job queue and resumes unfinished jobs.
+
+ This function needs the lock here because WorkerPool.AddTask() may start a
+ job while we're still doing our work.
+
+ """
+ logging.info("Inspecting job queue")
+
+ restartjobs = []
+
+ 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 == constants.JOB_STATUS_QUEUED:
+ restartjobs.append(job)
+
+ elif status in (constants.JOB_STATUS_RUNNING,
+ constants.JOB_STATUS_WAITING,
+ constants.JOB_STATUS_CANCELING):
+ logging.warning("Unfinished job %s found: %s", job.id, job)
+
+ if status == constants.JOB_STATUS_WAITING:
+ # Restart job
+ job.MarkUnfinishedOps(constants.OP_STATUS_QUEUED, None)
+ restartjobs.append(job)
+ else:
+ job.MarkUnfinishedOps(constants.OP_STATUS_ERROR,
+ "Unclean master daemon shutdown")
+ job.Finalize()
+
+ self.UpdateJobUnlocked(job)
+
+ if restartjobs:
+ logging.info("Restarting %s jobs", len(restartjobs))
+ self._EnqueueJobsUnlocked(restartjobs)
+
+ logging.info("Job queue inspection finished")
+
+ def _GetRpc(self, address_list):
+ """Gets RPC runner with context.
+
+ """
+ return rpc.JobQueueRunner(self.context, address_list)
+
+ @locking.ssynchronized(_LOCK)
+ @_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 = self._GetRpc(None).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(pathutils.JOB_QUEUE_SERIAL_FILE)
+
+ # Static address list
+ addrs = [node.primary_ip]
+
+ for file_name in files:
+ # Read file content
+ content = utils.ReadFile(file_name)
+
+ result = _CallJqUpdate(self._GetRpc(addrs), [node_name],
+ 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)
+
+ # Set queue drained flag
+ result = \
+ self._GetRpc(addrs).call_jobqueue_set_drain_flag([node_name],
+ self._drained)
+ msg = result[node_name].fail_msg
+ if msg:
+ logging.error("Failed to set queue drained flag on node %s: %s",
+ node_name, msg)
+
+ self._nodes[node_name] = node.primary_ip
+
+ @locking.ssynchronized(_LOCK)
+ @_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
+
+ """
+ self._nodes.pop(node_name, None)
+
+ @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",
+ result[node].call, failmsg, node, msg)
+ else:
+ success.append(node)
+
+ # +1 for the master node
+ if (len(success) + 1) < len(failed):
+ # TODO: Handle failing nodes
+ logging.error("More than half of the nodes failed")
+
+ def _GetNodeIp(self):
+ """Helper for returning the node name/ip list.
+
+ @rtype: (list, list)
+ @return: a tuple of two lists, the first one with the node
+ names and the second one with the node addresses
+
+ """
+ # TODO: Change to "tuple(map(list, zip(*self._nodes.items())))"?
+ name_list = self._nodes.keys()
+ addr_list = [self._nodes[name] for name in name_list]
+ return name_list, addr_list
+
+ def _UpdateJobQueueFile(self, file_name, data, replicate):
+ """Writes a file locally and then replicates it to all nodes.
+
+ This function will replace the contents of a file on the local
+ node and then replicate it to all the other nodes we have.
+
+ @type file_name: str
+ @param file_name: the path of the file to be replicated
+ @type data: str
+ @param data: the new contents of the file
+ @type replicate: boolean
+ @param replicate: whether to spread the changes to the remote nodes
+
+ """
+ getents = runtime.GetEnts()
+ utils.WriteFile(file_name, data=data, uid=getents.masterd_uid,
+ gid=getents.daemons_gid,
+ mode=constants.JOB_QUEUE_FILES_PERMS)
+
+ if replicate:
+ names, addrs = self._GetNodeIp()
+ result = _CallJqUpdate(self._GetRpc(addrs), names, file_name, data)
+ self._CheckRpcResult(result, self._nodes, "Updating %s" % file_name)
+
+ def _RenameFilesUnlocked(self, rename):
+ """Renames a file locally and then replicate the change.
+
+ This function will rename a file in the local queue directory
+ and then replicate this rename to all the other nodes we have.
+
+ @type rename: list of (old, new)
+ @param rename: List containing tuples mapping old to new names
+
+ """
+ # Rename them locally
+ for old, new in rename:
+ utils.RenameFile(old, new, mkdir=True)
+
+ # ... and on all nodes
+ names, addrs = self._GetNodeIp()
+ result = self._GetRpc(addrs).call_jobqueue_rename(names, rename)
+ self._CheckRpcResult(result, self._nodes, "Renaming files (%r)" % rename)
+
+ def _NewSerialsUnlocked(self, count):
+ """Generates a new job identifier.
+
+ Job identifiers are unique during the lifetime of a cluster.