celery/celery/concurrency/asynpool.py

# -*- coding: utf-8 -*-
"""Version of multiprocessing.Pool using Async I/O.

.. note::

    This module will be moved soon, so don't use it directly.

This is a non-blocking version of :class:`multiprocessing.Pool`.

This code deals with three major challenges:

#. Starting up child processes and keeping them running.
#. Sending jobs to the processes and receiving results back.
#. Safely shutting down this system.
"""
from __future__ import absolute_import, unicode_literals

import errno
import gc
import os
import select
import socket
import struct
import sys
import time
from collections import deque, namedtuple
from io import BytesIO
from numbers import Integral
from pickle import HIGHEST_PROTOCOL
from time import sleep
from weakref import WeakValueDictionary, ref

from billiard import pool as _pool
from billiard.compat import buf_t, isblocking, setblocking
from billiard.pool import ACK, NACK, RUN, TERMINATE, WorkersJoined
from billiard.queues import _SimpleQueue
from kombu.async import ERR, WRITE
from kombu.serialization import pickle as _pickle
from kombu.utils.eventio import SELECT_BAD_FD
from kombu.utils.functional import fxrange
from vine import promise

from celery.five import Counter, items, values
from celery.utils.functional import noop
from celery.utils.log import get_logger
from celery.worker import state as worker_state

# pylint: disable=redefined-outer-name
# We cache globals and attribute lookups, so disable this warning.

try:
    from _billiard import read as __read__
    from struct import unpack_from as _unpack_from
    memoryview = memoryview
    readcanbuf = True

    if sys.version_info[0] == 2 and sys.version_info < (2, 7, 6):

        def unpack_from(fmt, view, _unpack_from=_unpack_from):  # noqa
            return _unpack_from(fmt, view.tobytes())  # <- memoryview
    else:
        # unpack_from supports memoryview in 2.7.6 and 3.3+
        unpack_from = _unpack_from  # noqa

except (ImportError, NameError):  # pragma: no cover

    def __read__(fd, buf, size, read=os.read):  # noqa
        chunk = read(fd, size)
        n = len(chunk)
        if n != 0:
            buf.write(chunk)
        return n
    readcanbuf = False  # noqa

    def unpack_from(fmt, iobuf, unpack=struct.unpack):  # noqa
        return unpack(fmt, iobuf.getvalue())  # <-- BytesIO

__all__ = ('AsynPool',)

logger = get_logger(__name__)
error, debug = logger.error, logger.debug

UNAVAIL = frozenset({errno.EAGAIN, errno.EINTR})

#: Constant sent by child process when started (ready to accept work)
WORKER_UP = 15

#: A process must've started before this timeout (in secs.) expires.
PROC_ALIVE_TIMEOUT = 4.0

SCHED_STRATEGY_FCFS = 1
SCHED_STRATEGY_FAIR = 4

SCHED_STRATEGIES = {
    None: SCHED_STRATEGY_FAIR,
    'fast': SCHED_STRATEGY_FCFS,
    'fcfs': SCHED_STRATEGY_FCFS,
    'fair': SCHED_STRATEGY_FAIR,
}
SCHED_STRATEGY_TO_NAME = {v: k for k, v in SCHED_STRATEGIES.items()}

Ack = namedtuple('Ack', ('id', 'fd', 'payload'))


def gen_not_started(gen):
    """Return true if generator is not started."""
    # gi_frame is None when generator stopped.
    return gen.gi_frame and gen.gi_frame.f_lasti == -1


def _get_job_writer(job):
    try:
        writer = job._writer
    except AttributeError:
        pass
    else:
        return writer()  # is a weakref


if hasattr(select, 'poll'):
    def _select_imp(readers=None, writers=None, err=None, timeout=0,
                    poll=select.poll, POLLIN=select.POLLIN,
                    POLLOUT=select.POLLOUT, POLLERR=select.POLLERR):
        poller = poll()
        register = poller.register

        if readers:
            [register(fd, POLLIN) for fd in readers]
        if writers:
            [register(fd, POLLOUT) for fd in writers]
        if err:
            [register(fd, POLLERR) for fd in err]

        R, W = set(), set()
        timeout = 0 if timeout and timeout < 0 else round(timeout * 1e3)
        events = poller.poll(timeout)
        for fd, event in events:
            if not isinstance(fd, Integral):
                fd = fd.fileno()
            if event & POLLIN:
                R.add(fd)
            if event & POLLOUT:
                W.add(fd)
            if event & POLLERR:
                R.add(fd)
        return R, W, 0
else:
    def _select_imp(readers=None, writers=None, err=None, timeout=0):
        r, w, e = select.select(readers, writers, err, timeout)
        if e:
            r = list(set(r) | set(e))
        return r, w, 0


def _select(readers=None, writers=None, err=None, timeout=0,
            poll=_select_imp):
    """Simple wrapper to :class:`~select.select`, using :`~select.poll`.

    Arguments:
        readers (Set[Fd]): Set of reader fds to test if readable.
        writers (Set[Fd]): Set of writer fds to test if writable.
        err (Set[Fd]): Set of fds to test for error condition.

    All fd sets passed must be mutable as this function
    will remove non-working fds from them, this also means
    the caller must make sure there are still fds in the sets
    before calling us again.

    Returns:
        Tuple[Set, Set, Set]: of ``(readable, writable, again)``, where
        ``readable`` is a set of fds that have data available for read,
        ``writable`` is a set of fds that's ready to be written to
        and ``again`` is a flag that if set means the caller must
        throw away the result and call us again.
    """
    readers = set() if readers is None else readers
    writers = set() if writers is None else writers
    err = set() if err is None else err
    try:
        return poll(readers, writers, err, timeout)
    except (select.error, socket.error) as exc:
        if exc.errno == errno.EINTR:
            return set(), set(), 1
        elif exc.errno in SELECT_BAD_FD:
            for fd in readers | writers | err:
                try:
                    select.select([fd], [], [], 0)
                except (select.error, socket.error) as exc:
                    if getattr(exc, 'errno', None) not in SELECT_BAD_FD:
                        raise
                    readers.discard(fd)
                    writers.discard(fd)
                    err.discard(fd)
            return set(), set(), 1
        else:
            raise


class Worker(_pool.Worker):
    """Pool worker process."""

    def on_loop_start(self, pid):
        # our version sends a WORKER_UP message when the process is ready
        # to accept work, this will tell the parent that the inqueue fd
        # is writable.
        self.outq.put((WORKER_UP, (pid,)))


class ResultHandler(_pool.ResultHandler):
    """Handles messages from the pool processes."""

    def __init__(self, *args, **kwargs):
        self.fileno_to_outq = kwargs.pop('fileno_to_outq')
        self.on_process_alive = kwargs.pop('on_process_alive')
        super(ResultHandler, self).__init__(*args, **kwargs)
        # add our custom message handler
        self.state_handlers[WORKER_UP] = self.on_process_alive

    def _recv_message(self, add_reader, fd, callback,
                      __read__=__read__, readcanbuf=readcanbuf,
                      BytesIO=BytesIO, unpack_from=unpack_from,
                      load=_pickle.load):
        Hr = Br = 0
        if readcanbuf:
            buf = bytearray(4)
            bufv = memoryview(buf)
        else:
            buf = bufv = BytesIO()
        # header

        while Hr < 4:
            try:
                n = __read__(
                    fd, bufv[Hr:] if readcanbuf else bufv, 4 - Hr,
                )
            except OSError as exc:
                if exc.errno not in UNAVAIL:
                    raise
                yield
            else:
                if n == 0:
                    raise (OSError('End of file during message') if Hr
                           else EOFError())
                Hr += n

        body_size, = unpack_from(b'>i', bufv)
        if readcanbuf:
            buf = bytearray(body_size)
            bufv = memoryview(buf)
        else:
            buf = bufv = BytesIO()

        while Br < body_size:
            try:
                n = __read__(
                    fd, bufv[Br:] if readcanbuf else bufv, body_size - Br,
                )
            except OSError as exc:
                if exc.errno not in UNAVAIL:
                    raise
                yield
            else:
                if n == 0:
                    raise (OSError('End of file during message') if Br
                           else EOFError())
                Br += n
        add_reader(fd, self.handle_event, fd)
        if readcanbuf:
            message = load(BytesIO(bufv))
        else:
            bufv.seek(0)
            message = load(bufv)
        if message:
            callback(message)

    def _make_process_result(self, hub):
        """Coroutine reading messages from the pool processes."""
        fileno_to_outq = self.fileno_to_outq
        on_state_change = self.on_state_change
        add_reader = hub.add_reader
        remove_reader = hub.remove_reader
        recv_message = self._recv_message

        def on_result_readable(fileno):
            try:
                fileno_to_outq[fileno]
            except KeyError:  # process gone
                return remove_reader(fileno)
            it = recv_message(add_reader, fileno, on_state_change)
            try:
                next(it)
            except StopIteration:
                pass
            except (IOError, OSError, EOFError):
                remove_reader(fileno)
            else:
                add_reader(fileno, it)
        return on_result_readable

    def register_with_event_loop(self, hub):
        self.handle_event = self._make_process_result(hub)

    def handle_event(self, *args):
        # pylint: disable=method-hidden
        #   register_with_event_loop overrides this
        raise RuntimeError('Not registered with event loop')

    def on_stop_not_started(self):
        # This is always used, since we do not start any threads.
        cache = self.cache
        check_timeouts = self.check_timeouts
        fileno_to_outq = self.fileno_to_outq
        on_state_change = self.on_state_change
        join_exited_workers = self.join_exited_workers

        # flush the processes outqueues until they've all terminated.
        outqueues = set(fileno_to_outq)
        while cache and outqueues and self._state != TERMINATE:
            if check_timeouts is not None:
                # make sure tasks with a time limit will time out.
                check_timeouts()
            # cannot iterate and remove at the same time
            pending_remove_fd = set()
            for fd in outqueues:
                self._flush_outqueue(
                    fd, pending_remove_fd.add, fileno_to_outq,
                    on_state_change,
                )
                try:
                    join_exited_workers(shutdown=True)
                except WorkersJoined:
                    return debug('result handler: all workers terminated')
            outqueues.difference_update(pending_remove_fd)

    def _flush_outqueue(self, fd, remove, process_index, on_state_change):
        try:
            proc = process_index[fd]
        except KeyError:
            # process already found terminated
            # this means its outqueue has already been processed
            # by the worker lost handler.
            return remove(fd)

        reader = proc.outq._reader
        try:
            setblocking(reader, 1)
        except (OSError, IOError):
            return remove(fd)
        try:
            if reader.poll(0):
                task = reader.recv()
            else:
                task = None
                sleep(0.5)
        except (IOError, EOFError):
            return remove(fd)
        else:
            if task:
                on_state_change(task)
        finally:
            try:
                setblocking(reader, 0)
            except (OSError, IOError):
                return remove(fd)


class AsynPool(_pool.Pool):
    """AsyncIO Pool (no threads)."""

    ResultHandler = ResultHandler
    Worker = Worker

    def WorkerProcess(self, worker):
        worker = super(AsynPool, self).WorkerProcess(worker)
        worker.dead = False
        return worker

    def __init__(self, processes=None, synack=False,
                 sched_strategy=None, *args, **kwargs):
        self.sched_strategy = SCHED_STRATEGIES.get(sched_strategy,
                                                   sched_strategy)
        processes = self.cpu_count() if processes is None else processes
        self.synack = synack
        # create queue-pairs for all our processes in advance.
        self._queues = {
            self.create_process_queues(): None for _ in range(processes)
        }

        # inqueue fileno -> process mapping
        self._fileno_to_inq = {}
        # outqueue fileno -> process mapping
        self._fileno_to_outq = {}
        # synqueue fileno -> process mapping
        self._fileno_to_synq = {}

        # We keep track of processes that haven't yet
        # sent a WORKER_UP message.  If a process fails to send
        # this message within proc_up_timeout we terminate it
        # and hope the next process will recover.
        self._proc_alive_timeout = PROC_ALIVE_TIMEOUT
        self._waiting_to_start = set()

        # denormalized set of all inqueues.
        self._all_inqueues = set()

        # Set of fds being written to (busy)
        self._active_writes = set()

        # Set of active co-routines currently writing jobs.
        self._active_writers = set()

        # Set of fds that are busy (executing task)
        self._busy_workers = set()
        self._mark_worker_as_available = self._busy_workers.discard

        # Holds jobs waiting to be written to child processes.
        self.outbound_buffer = deque()

        self.write_stats = Counter()

        super(AsynPool, self).__init__(processes, *args, **kwargs)

        for proc in self._pool:
            # create initial mappings, these will be updated
            # as processes are recycled, or found lost elsewhere.
            self._fileno_to_outq[proc.outqR_fd] = proc
            self._fileno_to_synq[proc.synqW_fd] = proc

        self.on_soft_timeout = getattr(
            self._timeout_handler, 'on_soft_timeout', noop,
        )
        self.on_hard_timeout = getattr(
            self._timeout_handler, 'on_hard_timeout', noop,
        )

    def _create_worker_process(self, i):
        gc.collect()  # Issue #2927
        return super(AsynPool, self)._create_worker_process(i)

    def _event_process_exit(self, hub, proc):
        # This method is called whenever the process sentinel is readable.
        self._untrack_child_process(proc, hub)
        self.maintain_pool()

    def _track_child_process(self, proc, hub):
        try:
            fd = proc._sentinel_poll
        except AttributeError:
            # we need to duplicate the fd here to carefully
            # control when the fd is removed from the process table,
            # as once the original fd is closed we cannot unregister
            # the fd from epoll(7) anymore, causing a 100% CPU poll loop.
            fd = proc._sentinel_poll = os.dup(proc._popen.sentinel)
        hub.add_reader(fd, self._event_process_exit, hub, proc)

    def _untrack_child_process(self, proc, hub):
        if proc._sentinel_poll is not None:
            fd, proc._sentinel_poll = proc._sentinel_poll, None
            hub.remove(fd)
            os.close(fd)

    def register_with_event_loop(self, hub):
        """Register the async pool with the current event loop."""
        self._result_handler.register_with_event_loop(hub)
        self.handle_result_event = self._result_handler.handle_event
        self._create_timelimit_handlers(hub)
        self._create_process_handlers(hub)
        self._create_write_handlers(hub)

        # Add handler for when a process exits (calls maintain_pool)
        [self._track_child_process(w, hub) for w in self._pool]
        # Handle_result_event is called whenever one of the
        # result queues are readable.
        [hub.add_reader(fd, self.handle_result_event, fd)
         for fd in self._fileno_to_outq]

        # Timers include calling maintain_pool at a regular interval
        # to be certain processes are restarted.
        for handler, interval in items(self.timers):
            hub.call_repeatedly(interval, handler)

        hub.on_tick.add(self.on_poll_start)

    def _create_timelimit_handlers(self, hub):
        """Create handlers used to implement time limits."""
        call_later = hub.call_later
        trefs = self._tref_for_id = WeakValueDictionary()

        def on_timeout_set(R, soft, hard):
            if soft:
                trefs[R._job] = call_later(
                    soft, self._on_soft_timeout, R._job, soft, hard, hub,
                )
            elif hard:
                trefs[R._job] = call_later(
                    hard, self._on_hard_timeout, R._job,
                )
        self.on_timeout_set = on_timeout_set

        def _discard_tref(job):
            try:
                tref = trefs.pop(job)
                tref.cancel()
                del tref
            except (KeyError, AttributeError):
                pass  # out of scope
        self._discard_tref = _discard_tref

        def on_timeout_cancel(R):
            _discard_tref(R._job)
        self.on_timeout_cancel = on_timeout_cancel

    def _on_soft_timeout(self, job, soft, hard, hub):
        # only used by async pool.
        if hard:
            self._tref_for_id[job] = hub.call_later(
                hard - soft, self._on_hard_timeout, job,
            )
        try:
            result = self._cache[job]
        except KeyError:
            pass  # job ready
        else:
            self.on_soft_timeout(result)
        finally:
            if not hard:
                # remove tref
                self._discard_tref(job)

    def _on_hard_timeout(self, job):
        # only used by async pool.
        try:
            result = self._cache[job]
        except KeyError:
            pass  # job ready
        else:
            self.on_hard_timeout(result)
        finally:
            # remove tref
            self._discard_tref(job)

    def on_job_ready(self, job, i, obj, inqW_fd):
        self._mark_worker_as_available(inqW_fd)

    def _create_process_handlers(self, hub):
        """Create handlers called on process up/down, etc."""
        add_reader, remove_reader, remove_writer = (
            hub.add_reader, hub.remove_reader, hub.remove_writer,
        )
        cache = self._cache
        all_inqueues = self._all_inqueues
        fileno_to_inq = self._fileno_to_inq
        fileno_to_outq = self._fileno_to_outq
        fileno_to_synq = self._fileno_to_synq
        busy_workers = self._busy_workers
        handle_result_event = self.handle_result_event
        process_flush_queues = self.process_flush_queues
        waiting_to_start = self._waiting_to_start

        def verify_process_alive(proc):
            proc = proc()  # is a weakref
            if (proc is not None and proc._is_alive() and
                    proc in waiting_to_start):
                assert proc.outqR_fd in fileno_to_outq
                assert fileno_to_outq[proc.outqR_fd] is proc
                assert proc.outqR_fd in hub.readers
                error('Timed out waiting for UP message from %r', proc)
                os.kill(proc.pid, 9)

        def on_process_up(proc):
            """Called when a process has started."""
            # If we got the same fd as a previous process then we'll also
            # receive jobs in the old buffer, so we need to reset the
            # job._write_to and job._scheduled_for attributes used to recover
            # message boundaries when processes exit.
            infd = proc.inqW_fd
            for job in values(cache):
                if job._write_to and job._write_to.inqW_fd == infd:
                    job._write_to = proc
                if job._scheduled_for and job._scheduled_for.inqW_fd == infd:
                    job._scheduled_for = proc
            fileno_to_outq[proc.outqR_fd] = proc

            # maintain_pool is called whenever a process exits.
            self._track_child_process(proc, hub)

            assert not isblocking(proc.outq._reader)

            # handle_result_event is called when the processes outqueue is
            # readable.
            add_reader(proc.outqR_fd, handle_result_event, proc.outqR_fd)

            waiting_to_start.add(proc)
            hub.call_later(
                self._proc_alive_timeout, verify_process_alive, ref(proc),
            )

        self.on_process_up = on_process_up

        def _remove_from_index(obj, proc, index, remove_fun, callback=None):
            # this remove the file descriptors for a process from
            # the indices.  we have to make sure we don't overwrite
            # another processes fds, as the fds may be reused.
            try:
                fd = obj.fileno()
            except (IOError, OSError):
                return

            try:
                if index[fd] is proc:
                    # fd hasn't been reused so we can remove it from index.
                    index.pop(fd, None)
            except KeyError:
                pass
            else:
                remove_fun(fd)
                if callback is not None:
                    callback(fd)
            return fd

        def on_process_down(proc):
            """Called when a worker process exits."""
            if getattr(proc, 'dead', None):
                return
            process_flush_queues(proc)
            _remove_from_index(
                proc.outq._reader, proc, fileno_to_outq, remove_reader,
            )
            if proc.synq:
                _remove_from_index(
                    proc.synq._writer, proc, fileno_to_synq, remove_writer,
                )
            inq = _remove_from_index(
                proc.inq._writer, proc, fileno_to_inq, remove_writer,
                callback=all_inqueues.discard,
            )
            if inq:
                busy_workers.discard(inq)
            self._untrack_child_process(proc, hub)
            waiting_to_start.discard(proc)
            self._active_writes.discard(proc.inqW_fd)
            remove_writer(proc.inq._writer)
            remove_reader(proc.outq._reader)
            if proc.synqR_fd:
                remove_reader(proc.synq._reader)
            if proc.synqW_fd:
                self._active_writes.discard(proc.synqW_fd)
                remove_reader(proc.synq._writer)
        self.on_process_down = on_process_down

    def _create_write_handlers(self, hub,
                               pack=struct.pack, dumps=_pickle.dumps,
                               protocol=HIGHEST_PROTOCOL):
        """Create handlers used to write data to child processes."""
        fileno_to_inq = self._fileno_to_inq
        fileno_to_synq = self._fileno_to_synq
        outbound = self.outbound_buffer
        pop_message = outbound.popleft
        put_message = outbound.append
        all_inqueues = self._all_inqueues
        active_writes = self._active_writes
        active_writers = self._active_writers
        busy_workers = self._busy_workers
        diff = all_inqueues.difference
        add_writer = hub.add_writer
        hub_add, hub_remove = hub.add, hub.remove
        mark_write_fd_as_active = active_writes.add
        mark_write_gen_as_active = active_writers.add
        mark_worker_as_busy = busy_workers.add
        write_generator_done = active_writers.discard
        get_job = self._cache.__getitem__
        write_stats = self.write_stats
        is_fair_strategy = self.sched_strategy == SCHED_STRATEGY_FAIR
        revoked_tasks = worker_state.revoked
        getpid = os.getpid

        precalc = {ACK: self._create_payload(ACK, (0,)),
                   NACK: self._create_payload(NACK, (0,))}

        def _put_back(job, _time=time.time):
            # puts back at the end of the queue
            if job._terminated is not None or \
                    job.correlation_id in revoked_tasks:
                if not job._accepted:
                    job._ack(None, _time(), getpid(), None)
                job._set_terminated(job._terminated)
            else:
                # XXX linear lookup, should find a better way,
                # but this happens rarely and is here to protect against races.
                if job not in outbound:
                    outbound.appendleft(job)
        self._put_back = _put_back

        # called for every event loop iteration, and if there
        # are messages pending this will schedule writing one message
        # by registering the 'schedule_writes' function for all currently
        # inactive inqueues (not already being written to)

        # consolidate means the event loop will merge them
        # and call the callback once with the list writable fds as
        # argument.  Using this means we minimize the risk of having
        # the same fd receive every task if the pipe read buffer is not
        # full.
        if is_fair_strategy:

            def on_poll_start():
                if outbound and len(busy_workers) < len(all_inqueues):
                    #  print('ALL: %r ACTIVE: %r' % (len(all_inqueues),
                    #                                len(active_writes)))
                    inactive = diff(active_writes)
                    [hub_add(fd, None, WRITE | ERR, consolidate=True)
                     for fd in inactive]
                else:
                    [hub_remove(fd) for fd in diff(active_writes)]
        else:
            def on_poll_start():  # noqa
                if outbound:
                    [hub_add(fd, None, WRITE | ERR, consolidate=True)
                     for fd in diff(active_writes)]
                else:
                    [hub_remove(fd) for fd in diff(active_writes)]
        self.on_poll_start = on_poll_start

        def on_inqueue_close(fd, proc):
            # Makes sure the fd is removed from tracking when
            # the connection is closed, this is essential as fds may be reused.
            busy_workers.discard(fd)
            try:
                if fileno_to_inq[fd] is proc:
                    fileno_to_inq.pop(fd, None)
                    active_writes.discard(fd)
                    all_inqueues.discard(fd)
                    hub_remove(fd)
            except KeyError:
                pass
        self.on_inqueue_close = on_inqueue_close

        def schedule_writes(ready_fds, total_write_count=[0]):
            # Schedule write operation to ready file descriptor.
            # The file descriptor is writable, but that does not
            # mean the process is currently reading from the socket.
            # The socket is buffered so writable simply means that
            # the buffer can accept at least 1 byte of data.

            # This means we have to cycle between the ready fds.
            # the first version used shuffle, but this version
            # using `total_writes % ready_fds` is about 30% faster
            # with many processes, and also leans more towards fairness
            # in write stats when used with many processes
            # [XXX On macOS, this may vary depending
            # on event loop implementation (i.e, select/poll vs epoll), so
            # have to test further]
            num_ready = len(ready_fds)

            for _ in range(num_ready):
                ready_fd = ready_fds[total_write_count[0] % num_ready]
                total_write_count[0] += 1
                if ready_fd in active_writes:
                    # already writing to this fd
                    continue
                if is_fair_strategy and ready_fd in busy_workers:
                    # worker is already busy with another task
                    continue
                if ready_fd not in all_inqueues:
                    hub_remove(ready_fd)
                    continue
                try:
                    job = pop_message()
                except IndexError:
                    # no more messages, remove all inactive fds from the hub.
                    # this is important since the fds are always writable
                    # as long as there's 1 byte left in the buffer, and so
                    # this may create a spinloop where the event loop
                    # always wakes up.
                    for inqfd in diff(active_writes):
                        hub_remove(inqfd)
                    break

                else:
                    if not job._accepted:  # job not accepted by another worker
                        try:
                            # keep track of what process the write operation
                            # was scheduled for.
                            proc = job._scheduled_for = fileno_to_inq[ready_fd]
                        except KeyError:
                            # write was scheduled for this fd but the process
                            # has since exited and the message must be sent to
                            # another process.
                            put_message(job)
                            continue
                        cor = _write_job(proc, ready_fd, job)
                        job._writer = ref(cor)
                        mark_write_gen_as_active(cor)
                        mark_write_fd_as_active(ready_fd)
                        mark_worker_as_busy(ready_fd)

                        # Try to write immediately, in case there's an error.
                        try:
                            next(cor)
                        except StopIteration:
                            pass
                        except OSError as exc:
                            if exc.errno != errno.EBADF:
                                raise
                        else:
                            add_writer(ready_fd, cor)
        hub.consolidate_callback = schedule_writes

        def send_job(tup):
            # Schedule writing job request for when one of the process
            # inqueues are writable.
            body = dumps(tup, protocol=protocol)
            body_size = len(body)
            header = pack(b'>I', body_size)
            # index 1,0 is the job ID.
            job = get_job(tup[1][0])
            job._payload = buf_t(header), buf_t(body), body_size
            put_message(job)
        self._quick_put = send_job

        def on_not_recovering(proc, fd, job, exc):
            logger.exception(
                'Process inqueue damaged: %r %r: %r', proc, proc.exitcode, exc)
            if proc._is_alive():
                proc.terminate()
            hub.remove(fd)
            self._put_back(job)

        def _write_job(proc, fd, job):
            # writes job to the worker process.
            # Operation must complete if more than one byte of data
            # was written.  If the broker connection is lost
            # and no data was written the operation shall be canceled.
            header, body, body_size = job._payload
            errors = 0
            try:
                # job result keeps track of what process the job is sent to.
                job._write_to = proc
                send = proc.send_job_offset

                Hw = Bw = 0
                # write header
                while Hw < 4:
                    try:
                        Hw += send(header, Hw)
                    except Exception as exc:  # pylint: disable=broad-except
                        if getattr(exc, 'errno', None) not in UNAVAIL:
                            raise
                        # suspend until more data
                        errors += 1
                        if errors > 100:
                            on_not_recovering(proc, fd, job, exc)
                            raise StopIteration()
                        yield
                    else:
                        errors = 0

                # write body
                while Bw < body_size:
                    try:
                        Bw += send(body, Bw)
                    except Exception as exc:  # pylint: disable=broad-except
                        if getattr(exc, 'errno', None) not in UNAVAIL:
                            raise
                        # suspend until more data
                        errors += 1
                        if errors > 100:
                            on_not_recovering(proc, fd, job, exc)
                            raise StopIteration()
                        yield
                    else:
                        errors = 0
            finally:
                hub_remove(fd)
                write_stats[proc.index] += 1
                # message written, so this fd is now available
                active_writes.discard(fd)
                write_generator_done(job._writer())  # is a weakref

        def send_ack(response, pid, job, fd):
            # Only used when synack is enabled.
            # Schedule writing ack response for when the fd is writable.
            msg = Ack(job, fd, precalc[response])
            callback = promise(write_generator_done)
            cor = _write_ack(fd, msg, callback=callback)
            mark_write_gen_as_active(cor)
            mark_write_fd_as_active(fd)
            callback.args = (cor,)
            add_writer(fd, cor)
        self.send_ack = send_ack

        def _write_ack(fd, ack, callback=None):
            # writes ack back to the worker if synack enabled.
            # this operation *MUST* complete, otherwise
            # the worker process will hang waiting for the ack.
            header, body, body_size = ack[2]
            try:
                try:
                    proc = fileno_to_synq[fd]
                except KeyError:
                    # process died, we can safely discard the ack at this
                    # point.
                    raise StopIteration()
                send = proc.send_syn_offset

                Hw = Bw = 0
                # write header
                while Hw < 4:
                    try:
                        Hw += send(header, Hw)
                    except Exception as exc:  # pylint: disable=broad-except
                        if getattr(exc, 'errno', None) not in UNAVAIL:
                            raise
                        yield

                # write body
                while Bw < body_size:
                    try:
                        Bw += send(body, Bw)
                    except Exception as exc:  # pylint: disable=broad-except
                        if getattr(exc, 'errno', None) not in UNAVAIL:
                            raise
                        # suspend until more data
                        yield
            finally:
                if callback:
                    callback()
                # message written, so this fd is now available
                active_writes.discard(fd)

    def flush(self):
        if self._state == TERMINATE:
            return
        # cancel all tasks that haven't been accepted so that NACK is sent.
        for job in values(self._cache):
            if not job._accepted:
                job._cancel()

        # clear the outgoing buffer as the tasks will be redelivered by
        # the broker anyway.
        if self.outbound_buffer:
            self.outbound_buffer.clear()

        self.maintain_pool()

        try:
            # ...but we must continue writing the payloads we already started
            # to keep message boundaries.
            # The messages may be NACK'ed later if synack is enabled.
            if self._state == RUN:
                # flush outgoing buffers
                intervals = fxrange(0.01, 0.1, 0.01, repeatlast=True)
                owned_by = {}
                for job in values(self._cache):
                    writer = _get_job_writer(job)
                    if writer is not None:
                        owned_by[writer] = job

                while self._active_writers:
                    writers = list(self._active_writers)
                    for gen in writers:
                        if (gen.__name__ == '_write_job' and
                                gen_not_started(gen)):
                            # hasn't started writing the job so can
                            # discard the task, but we must also remove
                            # it from the Pool._cache.
                            try:
                                job = owned_by[gen]
                            except KeyError:
                                pass
                            else:
                                # removes from Pool._cache
                                job.discard()
                            self._active_writers.discard(gen)
                        else:
                            try:
                                job = owned_by[gen]
                            except KeyError:
                                pass
                            else:
                                job_proc = job._write_to
                                if job_proc._is_alive():
                                    self._flush_writer(job_proc, gen)
                    # workers may have exited in the meantime.
                    self.maintain_pool()
                    sleep(next(intervals))  # don't busyloop
        finally:
            self.outbound_buffer.clear()
            self._active_writers.clear()
            self._active_writes.clear()
            self._busy_workers.clear()

    def _flush_writer(self, proc, writer):
        fds = {proc.inq._writer}
        try:
            while fds:
                if not proc._is_alive():
                    break  # process exited
                readable, writable, again = _select(
                    writers=fds, err=fds, timeout=0.5,
                )
                if not again and (writable or readable):
                    try:
                        next(writer)
                    except (StopIteration, OSError, IOError, EOFError):
                        break
        finally:
            self._active_writers.discard(writer)

    def get_process_queues(self):
        """Get queues for a new process.

        Here we'll find an unused slot, as there should always
        be one available when we start a new process.
        """
        return next(q for q, owner in items(self._queues)
                    if owner is None)

    def on_grow(self, n):
        """Grow the pool by ``n`` proceses."""
        diff = max(self._processes - len(self._queues), 0)
        if diff:
            self._queues.update({
                self.create_process_queues(): None for _ in range(diff)
            })

    def on_shrink(self, n):
        """Shrink the pool by ``n`` processes."""
        pass

    def create_process_queues(self):
        """Create new in, out, etc. queues, returned as a tuple."""
        # NOTE: Pipes must be set O_NONBLOCK at creation time (the original
        # fd), otherwise it won't be possible to change the flags until
        # there's an actual reader/writer on the other side.
        inq = _SimpleQueue(wnonblock=True)
        outq = _SimpleQueue(rnonblock=True)
        synq = None
        assert isblocking(inq._reader)
        assert not isblocking(inq._writer)
        assert not isblocking(outq._reader)
        assert isblocking(outq._writer)
        if self.synack:
            synq = _SimpleQueue(wnonblock=True)
            assert isblocking(synq._reader)
            assert not isblocking(synq._writer)
        return inq, outq, synq

    def on_process_alive(self, pid):
        """Called when reciving the :const:`WORKER_UP` message.

        Marks the process as ready to receive work.
        """
        try:
            proc = next(w for w in self._pool if w.pid == pid)
        except StopIteration:
            return logger.warning('process with pid=%s already exited', pid)
        assert proc.inqW_fd not in self._fileno_to_inq
        assert proc.inqW_fd not in self._all_inqueues
        self._waiting_to_start.discard(proc)
        self._fileno_to_inq[proc.inqW_fd] = proc
        self._fileno_to_synq[proc.synqW_fd] = proc
        self._all_inqueues.add(proc.inqW_fd)

    def on_job_process_down(self, job, pid_gone):
        """Called for each job when the process assigned to it exits."""
        if job._write_to and not job._write_to._is_alive():
            # job was partially written
            self.on_partial_read(job, job._write_to)
        elif job._scheduled_for and not job._scheduled_for._is_alive():
            # job was only scheduled to be written to this process,
            # but no data was sent so put it back on the outbound_buffer.
            self._put_back(job)

    def on_job_process_lost(self, job, pid, exitcode):
        """Called when the process executing job' exits.

        This happens when the process job'
        was assigned to exited by mysterious means (error exitcodes and
        signals).
        """
        self.mark_as_worker_lost(job, exitcode)

    def human_write_stats(self):
        if self.write_stats is None:
            return 'N/A'
        vals = list(values(self.write_stats))
        total = sum(vals)

        def per(v, total):
            return '{0:.2%}'.format((float(v) / total) if v else 0)

        return {
            'total': total,
            'avg': per(total / len(self.write_stats) if total else 0, total),
            'all': ', '.join(per(v, total) for v in vals),
            'raw': ', '.join(map(str, vals)),
            'strategy': SCHED_STRATEGY_TO_NAME.get(
                self.sched_strategy, self.sched_strategy,
            ),
            'inqueues': {
                'total': len(self._all_inqueues),
                'active': len(self._active_writes),
            }
        }

    def _process_cleanup_queues(self, proc):
        """Called to clean up queues after process exit."""
        if not proc.dead:
            try:
                self._queues[self._find_worker_queues(proc)] = None
            except (KeyError, ValueError):
                pass

    @staticmethod
    def _stop_task_handler(task_handler):
        """Called at shutdown to tell processes that we're shutting down."""
        for proc in task_handler.pool:
            try:
                setblocking(proc.inq._writer, 1)
            except (OSError, IOError):
                pass
            else:
                try:
                    proc.inq.put(None)
                except OSError as exc:
                    if exc.errno != errno.EBADF:
                        raise

    def create_result_handler(self):
        return super(AsynPool, self).create_result_handler(
            fileno_to_outq=self._fileno_to_outq,
            on_process_alive=self.on_process_alive,
        )

    def _process_register_queues(self, proc, queues):
        """Mark new ownership for ``queues`` to update fileno indices."""
        assert queues in self._queues
        b = len(self._queues)
        self._queues[queues] = proc
        assert b == len(self._queues)

    def _find_worker_queues(self, proc):
        """Find the queues owned by ``proc``."""
        try:
            return next(q for q, owner in items(self._queues)
                        if owner == proc)
        except StopIteration:
            raise ValueError(proc)

    def _setup_queues(self):
        # this is only used by the original pool that used a shared
        # queue for all processes.
        self._quick_put = None

        # these attributes are unused by this class, but we'll still
        # have to initialize them for compatibility.
        self._inqueue = self._outqueue = \
            self._quick_get = self._poll_result = None

    def process_flush_queues(self, proc):
        """Flush all queues.

        Including the outbound buffer, so that
        all tasks that haven't been started will be discarded.

        In Celery this is called whenever the transport connection is lost
        (consumer restart), and when a process is terminated.
        """
        resq = proc.outq._reader
        on_state_change = self._result_handler.on_state_change
        fds = {resq}
        while fds and not resq.closed and self._state != TERMINATE:
            readable, _, _ = _select(fds, None, fds, timeout=0.01)
            if readable:
                try:
                    task = resq.recv()
                except (OSError, IOError, EOFError) as exc:
                    _errno = getattr(exc, 'errno', None)
                    if _errno == errno.EINTR:
                        continue
                    elif _errno == errno.EAGAIN:
                        break
                    elif _errno not in UNAVAIL:
                        debug('got %r while flushing process %r',
                              exc, proc, exc_info=1)
                    break
                else:
                    if task is None:
                        debug('got sentinel while flushing process %r', proc)
                        break
                    else:
                        on_state_change(task)
            else:
                break

    def on_partial_read(self, job, proc):
        """Called when a job was partially written to exited child."""
        # worker terminated by signal:
        # we cannot reuse the sockets again, because we don't know if
        # the process wrote/read anything frmo them, and if so we cannot
        # restore the message boundaries.
        if not job._accepted:
            # job was not acked, so find another worker to send it to.
            self._put_back(job)
        writer = _get_job_writer(job)
        if writer:
            self._active_writers.discard(writer)
            del writer

        if not proc.dead:
            proc.dead = True
            # Replace queues to avoid reuse
            before = len(self._queues)
            try:
                queues = self._find_worker_queues(proc)
                if self.destroy_queues(queues, proc):
                    self._queues[self.create_process_queues()] = None
            except ValueError:
                pass
            assert len(self._queues) == before

    def destroy_queues(self, queues, proc):
        """Destroy queues that can no longer be used.

        This way they can be replaced by new usable sockets.
        """
        assert not proc._is_alive()
        self._waiting_to_start.discard(proc)
        removed = 1
        try:
            self._queues.pop(queues)
        except KeyError:
            removed = 0
        try:
            self.on_inqueue_close(queues[0]._writer.fileno(), proc)
        except IOError:
            pass
        for queue in queues:
            if queue:
                for sock in (queue._reader, queue._writer):
                    if not sock.closed:
                        try:
                            sock.close()
                        except (IOError, OSError):
                            pass
        return removed

    def _create_payload(self, type_, args,
                        dumps=_pickle.dumps, pack=struct.pack,
                        protocol=HIGHEST_PROTOCOL):
        body = dumps((type_, args), protocol=protocol)
        size = len(body)
        header = pack(b'>I', size)
        return header, body, size

    @classmethod
    def _set_result_sentinel(cls, _outqueue, _pool):
        # unused
        pass

    def _help_stuff_finish_args(self):
        # Pool._help_stuff_finished is a classmethod so we have to use this
        # trick to modify the arguments passed to it.
        return (self._pool,)

    @classmethod
    def _help_stuff_finish(cls, pool):
        # pylint: disable=arguments-differ
        debug(
            'removing tasks from inqueue until task handler finished',
        )
        fileno_to_proc = {}
        inqR = set()
        for w in pool:
            try:
                fd = w.inq._reader.fileno()
                inqR.add(fd)
                fileno_to_proc[fd] = w
            except IOError:
                pass
        while inqR:
            readable, _, again = _select(inqR, timeout=0.5)
            if again:
                continue
            if not readable:
                break
            for fd in readable:
                fileno_to_proc[fd].inq._reader.recv()
            sleep(0)

    @property
    def timers(self):
        return {self.maintain_pool: 5.0}