bitbake: compat/utils: Add copy of python multiprocessing pool for pre 2.7.3 issues
authorRichard Purdie <richard.purdie@linuxfoundation.org>
Wed, 6 Feb 2013 00:28:08 +0000 (00:28 +0000)
committerRichard Purdie <richard.purdie@linuxfoundation.org>
Wed, 6 Feb 2013 13:13:01 +0000 (13:13 +0000)
python 2.7 shows hangs with issues in its pool implmenetation. Rather than
try and hack around these, add a copy of the working pool implementation
to the compat module from 2.7.3.

(Bitbake rev: c9eb742637131e8dbd526d2ad9b458abea0a2d87)

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
bitbake/lib/bb/compat.py
bitbake/lib/bb/utils.py

index 1466da2..ea4e23a 100644 (file)
@@ -239,3 +239,690 @@ class OrderedDict(dict):
     def viewitems(self):
         "od.viewitems() -> a set-like object providing a view on od's items"
         return ItemsView(self)
+
+# Multiprocessing pool code imported from python 2.7.3. Previous versions of 
+# python have issues in this code which hang pool usage
+
+#
+# Module providing the `Pool` class for managing a process pool
+#
+# multiprocessing/pool.py
+#
+# Copyright (c) 2006-2008, R Oudkerk
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#
+# 1. Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+# 2. Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+# 3. Neither the name of author nor the names of any contributors may be
+#    used to endorse or promote products derived from this software
+#    without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND
+# ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+# ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+# OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+# HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+# LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+# OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+# SUCH DAMAGE.
+#
+import threading
+import Queue
+import itertools
+import collections
+import time
+
+import multiprocessing
+from multiprocessing import Process, cpu_count, TimeoutError, pool
+from multiprocessing.util import Finalize, debug
+
+#
+# Constants representing the state of a pool
+#
+
+RUN = 0
+CLOSE = 1
+TERMINATE = 2
+
+#
+# Miscellaneous
+#
+
+def mapstar(args):
+    return map(*args)
+
+class MaybeEncodingError(Exception):
+    """Wraps possible unpickleable errors, so they can be
+    safely sent through the socket."""
+
+    def __init__(self, exc, value):
+        self.exc = repr(exc)
+        self.value = repr(value)
+        super(MaybeEncodingError, self).__init__(self.exc, self.value)
+
+    def __str__(self):
+        return "Error sending result: '%s'. Reason: '%s'" % (self.value,
+                                                             self.exc)
+
+    def __repr__(self):
+        return "<MaybeEncodingError: %s>" % str(self)
+
+def worker(inqueue, outqueue, initializer=None, initargs=(), maxtasks=None):
+    assert maxtasks is None or (type(maxtasks) == int and maxtasks > 0)
+    put = outqueue.put
+    get = inqueue.get
+    if hasattr(inqueue, '_writer'):
+        inqueue._writer.close()
+        outqueue._reader.close()
+
+    if initializer is not None:
+        initializer(*initargs)
+
+    completed = 0
+    while maxtasks is None or (maxtasks and completed < maxtasks):
+        try:
+            task = get()
+        except (EOFError, IOError):
+            debug('worker got EOFError or IOError -- exiting')
+            break
+
+        if task is None:
+            debug('worker got sentinel -- exiting')
+            break
+
+        job, i, func, args, kwds = task
+        try:
+            result = (True, func(*args, **kwds))
+        except Exception, e:
+            result = (False, e)
+        try:
+            put((job, i, result))
+        except Exception as e:
+            wrapped = MaybeEncodingError(e, result[1])
+            debug("Possible encoding error while sending result: %s" % (
+                wrapped))
+            put((job, i, (False, wrapped)))
+        completed += 1
+    debug('worker exiting after %d tasks' % completed)
+
+
+class Pool(object):
+    '''
+    Class which supports an async version of the `apply()` builtin
+    '''
+    Process = Process
+
+    def __init__(self, processes=None, initializer=None, initargs=(),
+                 maxtasksperchild=None):
+        self._setup_queues()
+        self._taskqueue = Queue.Queue()
+        self._cache = {}
+        self._state = RUN
+        self._maxtasksperchild = maxtasksperchild
+        self._initializer = initializer
+        self._initargs = initargs
+
+        if processes is None:
+            try:
+                processes = cpu_count()
+            except NotImplementedError:
+                processes = 1
+        if processes < 1:
+            raise ValueError("Number of processes must be at least 1")
+
+        if initializer is not None and not hasattr(initializer, '__call__'):
+            raise TypeError('initializer must be a callable')
+
+        self._processes = processes
+        self._pool = []
+        self._repopulate_pool()
+
+        self._worker_handler = threading.Thread(
+            target=Pool._handle_workers,
+            args=(self, )
+            )
+        self._worker_handler.daemon = True
+        self._worker_handler._state = RUN
+        self._worker_handler.start()
+
+
+        self._task_handler = threading.Thread(
+            target=Pool._handle_tasks,
+            args=(self._taskqueue, self._quick_put, self._outqueue, self._pool)
+            )
+        self._task_handler.daemon = True
+        self._task_handler._state = RUN
+        self._task_handler.start()
+
+        self._result_handler = threading.Thread(
+            target=Pool._handle_results,
+            args=(self._outqueue, self._quick_get, self._cache)
+            )
+        self._result_handler.daemon = True
+        self._result_handler._state = RUN
+        self._result_handler.start()
+
+        self._terminate = Finalize(
+            self, self._terminate_pool,
+            args=(self._taskqueue, self._inqueue, self._outqueue, self._pool,
+                  self._worker_handler, self._task_handler,
+                  self._result_handler, self._cache),
+            exitpriority=15
+            )
+
+    def _join_exited_workers(self):
+        """Cleanup after any worker processes which have exited due to reaching
+        their specified lifetime.  Returns True if any workers were cleaned up.
+        """
+        cleaned = False
+        for i in reversed(range(len(self._pool))):
+            worker = self._pool[i]
+            if worker.exitcode is not None:
+                # worker exited
+                debug('cleaning up worker %d' % i)
+                worker.join()
+                cleaned = True
+                del self._pool[i]
+        return cleaned
+
+    def _repopulate_pool(self):
+        """Bring the number of pool processes up to the specified number,
+        for use after reaping workers which have exited.
+        """
+        for i in range(self._processes - len(self._pool)):
+            w = self.Process(target=worker,
+                             args=(self._inqueue, self._outqueue,
+                                   self._initializer,
+                                   self._initargs, self._maxtasksperchild)
+                            )
+            self._pool.append(w)
+            w.name = w.name.replace('Process', 'PoolWorker')
+            w.daemon = True
+            w.start()
+            debug('added worker')
+
+    def _maintain_pool(self):
+        """Clean up any exited workers and start replacements for them.
+        """
+        if self._join_exited_workers():
+            self._repopulate_pool()
+
+    def _setup_queues(self):
+        from multiprocessing.queues import SimpleQueue
+        self._inqueue = SimpleQueue()
+        self._outqueue = SimpleQueue()
+        self._quick_put = self._inqueue._writer.send
+        self._quick_get = self._outqueue._reader.recv
+
+    def apply(self, func, args=(), kwds={}):
+        '''
+        Equivalent of `apply()` builtin
+        '''
+        assert self._state == RUN
+        return self.apply_async(func, args, kwds).get()
+
+    def map(self, func, iterable, chunksize=None):
+        '''
+        Equivalent of `map()` builtin
+        '''
+        assert self._state == RUN
+        return self.map_async(func, iterable, chunksize).get()
+
+    def imap(self, func, iterable, chunksize=1):
+        '''
+        Equivalent of `itertools.imap()` -- can be MUCH slower than `Pool.map()`
+        '''
+        assert self._state == RUN
+        if chunksize == 1:
+            result = IMapIterator(self._cache)
+            self._taskqueue.put((((result._job, i, func, (x,), {})
+                         for i, x in enumerate(iterable)), result._set_length))
+            return result
+        else:
+            assert chunksize > 1
+            task_batches = Pool._get_tasks(func, iterable, chunksize)
+            result = IMapIterator(self._cache)
+            self._taskqueue.put((((result._job, i, mapstar, (x,), {})
+                     for i, x in enumerate(task_batches)), result._set_length))
+            return (item for chunk in result for item in chunk)
+
+    def imap_unordered(self, func, iterable, chunksize=1):
+        '''
+        Like `imap()` method but ordering of results is arbitrary
+        '''
+        assert self._state == RUN
+        if chunksize == 1:
+            result = IMapUnorderedIterator(self._cache)
+            self._taskqueue.put((((result._job, i, func, (x,), {})
+                         for i, x in enumerate(iterable)), result._set_length))
+            return result
+        else:
+            assert chunksize > 1
+            task_batches = Pool._get_tasks(func, iterable, chunksize)
+            result = IMapUnorderedIterator(self._cache)
+            self._taskqueue.put((((result._job, i, mapstar, (x,), {})
+                     for i, x in enumerate(task_batches)), result._set_length))
+            return (item for chunk in result for item in chunk)
+
+    def apply_async(self, func, args=(), kwds={}, callback=None):
+        '''
+        Asynchronous equivalent of `apply()` builtin
+        '''
+        assert self._state == RUN
+        result = ApplyResult(self._cache, callback)
+        self._taskqueue.put(([(result._job, None, func, args, kwds)], None))
+        return result
+
+    def map_async(self, func, iterable, chunksize=None, callback=None):
+        '''
+        Asynchronous equivalent of `map()` builtin
+        '''
+        assert self._state == RUN
+        if not hasattr(iterable, '__len__'):
+            iterable = list(iterable)
+
+        if chunksize is None:
+            chunksize, extra = divmod(len(iterable), len(self._pool) * 4)
+            if extra:
+                chunksize += 1
+        if len(iterable) == 0:
+            chunksize = 0
+
+        task_batches = Pool._get_tasks(func, iterable, chunksize)
+        result = MapResult(self._cache, chunksize, len(iterable), callback)
+        self._taskqueue.put((((result._job, i, mapstar, (x,), {})
+                              for i, x in enumerate(task_batches)), None))
+        return result
+
+    @staticmethod
+    def _handle_workers(pool):
+        thread = threading.current_thread()
+
+        # Keep maintaining workers until the cache gets drained, unless the pool
+        # is terminated.
+        while thread._state == RUN or (pool._cache and thread._state != TERMINATE):
+            pool._maintain_pool()
+            time.sleep(0.1)
+        # send sentinel to stop workers
+        pool._taskqueue.put(None)
+        debug('worker handler exiting')
+
+    @staticmethod
+    def _handle_tasks(taskqueue, put, outqueue, pool):
+        thread = threading.current_thread()
+
+        for taskseq, set_length in iter(taskqueue.get, None):
+            i = -1
+            for i, task in enumerate(taskseq):
+                if thread._state:
+                    debug('task handler found thread._state != RUN')
+                    break
+                try:
+                    put(task)
+                except IOError:
+                    debug('could not put task on queue')
+                    break
+            else:
+                if set_length:
+                    debug('doing set_length()')
+                    set_length(i+1)
+                continue
+            break
+        else:
+            debug('task handler got sentinel')
+
+
+        try:
+            # tell result handler to finish when cache is empty
+            debug('task handler sending sentinel to result handler')
+            outqueue.put(None)
+
+            # tell workers there is no more work
+            debug('task handler sending sentinel to workers')
+            for p in pool:
+                put(None)
+        except IOError:
+            debug('task handler got IOError when sending sentinels')
+
+        debug('task handler exiting')
+
+    @staticmethod
+    def _handle_results(outqueue, get, cache):
+        thread = threading.current_thread()
+
+        while 1:
+            try:
+                task = get()
+            except (IOError, EOFError):
+                debug('result handler got EOFError/IOError -- exiting')
+                return
+
+            if thread._state:
+                assert thread._state == TERMINATE
+                debug('result handler found thread._state=TERMINATE')
+                break
+
+            if task is None:
+                debug('result handler got sentinel')
+                break
+
+            job, i, obj = task
+            try:
+                cache[job]._set(i, obj)
+            except KeyError:
+                pass
+
+        while cache and thread._state != TERMINATE:
+            try:
+                task = get()
+            except (IOError, EOFError):
+                debug('result handler got EOFError/IOError -- exiting')
+                return
+
+            if task is None:
+                debug('result handler ignoring extra sentinel')
+                continue
+            job, i, obj = task
+            try:
+                cache[job]._set(i, obj)
+            except KeyError:
+                pass
+
+        if hasattr(outqueue, '_reader'):
+            debug('ensuring that outqueue is not full')
+            # If we don't make room available in outqueue then
+            # attempts to add the sentinel (None) to outqueue may
+            # block.  There is guaranteed to be no more than 2 sentinels.
+            try:
+                for i in range(10):
+                    if not outqueue._reader.poll():
+                        break
+                    get()
+            except (IOError, EOFError):
+                pass
+
+        debug('result handler exiting: len(cache)=%s, thread._state=%s',
+              len(cache), thread._state)
+
+    @staticmethod
+    def _get_tasks(func, it, size):
+        it = iter(it)
+        while 1:
+            x = tuple(itertools.islice(it, size))
+            if not x:
+                return
+            yield (func, x)
+
+    def __reduce__(self):
+        raise NotImplementedError(
+              'pool objects cannot be passed between processes or pickled'
+              )
+
+    def close(self):
+        debug('closing pool')
+        if self._state == RUN:
+            self._state = CLOSE
+            self._worker_handler._state = CLOSE
+
+    def terminate(self):
+        debug('terminating pool')
+        self._state = TERMINATE
+        self._worker_handler._state = TERMINATE
+        self._terminate()
+
+    def join(self):
+        debug('joining pool')
+        assert self._state in (CLOSE, TERMINATE)
+        self._worker_handler.join()
+        self._task_handler.join()
+        self._result_handler.join()
+        for p in self._pool:
+            p.join()
+
+    @staticmethod
+    def _help_stuff_finish(inqueue, task_handler, size):
+        # task_handler may be blocked trying to put items on inqueue
+        debug('removing tasks from inqueue until task handler finished')
+        inqueue._rlock.acquire()
+        while task_handler.is_alive() and inqueue._reader.poll():
+            inqueue._reader.recv()
+            time.sleep(0)
+
+    @classmethod
+    def _terminate_pool(cls, taskqueue, inqueue, outqueue, pool,
+                        worker_handler, task_handler, result_handler, cache):
+        # this is guaranteed to only be called once
+        debug('finalizing pool')
+
+        worker_handler._state = TERMINATE
+        task_handler._state = TERMINATE
+
+        debug('helping task handler/workers to finish')
+        cls._help_stuff_finish(inqueue, task_handler, len(pool))
+
+        assert result_handler.is_alive() or len(cache) == 0
+
+        result_handler._state = TERMINATE
+        outqueue.put(None)                  # sentinel
+
+        # We must wait for the worker handler to exit before terminating
+        # workers because we don't want workers to be restarted behind our back.
+        debug('joining worker handler')
+        if threading.current_thread() is not worker_handler:
+            worker_handler.join(1e100)
+
+        # Terminate workers which haven't already finished.
+        if pool and hasattr(pool[0], 'terminate'):
+            debug('terminating workers')
+            for p in pool:
+                if p.exitcode is None:
+                    p.terminate()
+
+        debug('joining task handler')
+        if threading.current_thread() is not task_handler:
+            task_handler.join(1e100)
+
+        debug('joining result handler')
+        if threading.current_thread() is not result_handler:
+            result_handler.join(1e100)
+
+        if pool and hasattr(pool[0], 'terminate'):
+            debug('joining pool workers')
+            for p in pool:
+                if p.is_alive():
+                    # worker has not yet exited
+                    debug('cleaning up worker %d' % p.pid)
+                    p.join()
+
+class ApplyResult(object):
+
+    def __init__(self, cache, callback):
+        self._cond = threading.Condition(threading.Lock())
+        self._job = multiprocessing.pool.job_counter.next()
+        self._cache = cache
+        self._ready = False
+        self._callback = callback
+        cache[self._job] = self
+
+    def ready(self):
+        return self._ready
+
+    def successful(self):
+        assert self._ready
+        return self._success
+
+    def wait(self, timeout=None):
+        self._cond.acquire()
+        try:
+            if not self._ready:
+                self._cond.wait(timeout)
+        finally:
+            self._cond.release()
+
+    def get(self, timeout=None):
+        self.wait(timeout)
+        if not self._ready:
+            raise TimeoutError
+        if self._success:
+            return self._value
+        else:
+            raise self._value
+
+    def _set(self, i, obj):
+        self._success, self._value = obj
+        if self._callback and self._success:
+            self._callback(self._value)
+        self._cond.acquire()
+        try:
+            self._ready = True
+            self._cond.notify()
+        finally:
+            self._cond.release()
+        del self._cache[self._job]
+
+#
+# Class whose instances are returned by `Pool.map_async()`
+#
+
+class MapResult(ApplyResult):
+
+    def __init__(self, cache, chunksize, length, callback):
+        ApplyResult.__init__(self, cache, callback)
+        self._success = True
+        self._value = [None] * length
+        self._chunksize = chunksize
+        if chunksize <= 0:
+            self._number_left = 0
+            self._ready = True
+            del cache[self._job]
+        else:
+            self._number_left = length//chunksize + bool(length % chunksize)
+
+    def _set(self, i, success_result):
+        success, result = success_result
+        if success:
+            self._value[i*self._chunksize:(i+1)*self._chunksize] = result
+            self._number_left -= 1
+            if self._number_left == 0:
+                if self._callback:
+                    self._callback(self._value)
+                del self._cache[self._job]
+                self._cond.acquire()
+                try:
+                    self._ready = True
+                    self._cond.notify()
+                finally:
+                    self._cond.release()
+
+        else:
+            self._success = False
+            self._value = result
+            del self._cache[self._job]
+            self._cond.acquire()
+            try:
+                self._ready = True
+                self._cond.notify()
+            finally:
+                self._cond.release()
+
+#
+# Class whose instances are returned by `Pool.imap()`
+#
+
+class IMapIterator(object):
+
+    def __init__(self, cache):
+        self._cond = threading.Condition(threading.Lock())
+        self._job = multiprocessing.pool.job_counter.next()
+        self._cache = cache
+        self._items = collections.deque()
+        self._index = 0
+        self._length = None
+        self._unsorted = {}
+        cache[self._job] = self
+
+    def __iter__(self):
+        return self
+
+    def next(self, timeout=None):
+        self._cond.acquire()
+        try:
+            try:
+                item = self._items.popleft()
+            except IndexError:
+                if self._index == self._length:
+                    raise StopIteration
+                self._cond.wait(timeout)
+                try:
+                    item = self._items.popleft()
+                except IndexError:
+                    if self._index == self._length:
+                        raise StopIteration
+                    raise TimeoutError
+        finally:
+            self._cond.release()
+
+        success, value = item
+        if success:
+            return value
+        raise value
+
+    __next__ = next                    # XXX
+
+    def _set(self, i, obj):
+        self._cond.acquire()
+        try:
+            if self._index == i:
+                self._items.append(obj)
+                self._index += 1
+                while self._index in self._unsorted:
+                    obj = self._unsorted.pop(self._index)
+                    self._items.append(obj)
+                    self._index += 1
+                self._cond.notify()
+            else:
+                self._unsorted[i] = obj
+
+            if self._index == self._length:
+                del self._cache[self._job]
+        finally:
+            self._cond.release()
+
+    def _set_length(self, length):
+        self._cond.acquire()
+        try:
+            self._length = length
+            if self._index == self._length:
+                self._cond.notify()
+                del self._cache[self._job]
+        finally:
+            self._cond.release()
+
+#
+# Class whose instances are returned by `Pool.imap_unordered()`
+#
+
+class IMapUnorderedIterator(IMapIterator):
+
+    def _set(self, i, obj):
+        self._cond.acquire()
+        try:
+            self._items.append(obj)
+            self._index += 1
+            self._cond.notify()
+            if self._index == self._length:
+                del self._cache[self._job]
+        finally:
+            self._cond.release()
+
+
index 7e81df5..83159a6 100644 (file)
@@ -839,4 +839,12 @@ def process_profilelog(fn):
     pout.flush()
     pout.close()  
 
+#
+# Work around multiprocessing pool bugs in python < 2.7.3
+#
+def multiprocessingpool(*args, **kwargs):
+    if sys.version_info < (2, 7, 3):
+        return bb.compat.Pool(*args, **kwargs)
+    else:
+        return multiprocessing.pool.Pool(*args, **kwargs)