]> granicus.if.org Git - python/commitdiff
bpo-35378: Fix multiprocessing.Pool references (GH-11627)
authorPablo Galindo <Pablogsal@gmail.com>
Mon, 11 Feb 2019 17:29:00 +0000 (17:29 +0000)
committerGitHub <noreply@github.com>
Mon, 11 Feb 2019 17:29:00 +0000 (17:29 +0000)
Changes in this commit:

1. Use a _strong_ reference between the Pool and associated iterators
2. Rework PR #8450 to eliminate a cycle in the Pool.

There is no test in this commit because any test that automatically tests this behaviour needs to eliminate the pool before joining the pool to check that the pool object is garbaged collected/does not hang. But doing this will potentially leak threads and processes (see https://bugs.python.org/issue35413).

Lib/multiprocessing/pool.py
Lib/test/_test_multiprocessing.py
Misc/NEWS.d/next/Library/2019-01-21-02-15-20.bpo-35378.4oF03i.rst [new file with mode: 0644]

index bfb2769ba6ec1e95ac160a839941b81b9170e8e8..18a56f8524b4b6faeeb04f10fda9e467cb188d50 100644 (file)
@@ -151,8 +151,9 @@ class Pool(object):
     '''
     _wrap_exception = True
 
-    def Process(self, *args, **kwds):
-        return self._ctx.Process(*args, **kwds)
+    @staticmethod
+    def Process(ctx, *args, **kwds):
+        return ctx.Process(*args, **kwds)
 
     def __init__(self, processes=None, initializer=None, initargs=(),
                  maxtasksperchild=None, context=None):
@@ -190,7 +191,10 @@ class Pool(object):
 
         self._worker_handler = threading.Thread(
             target=Pool._handle_workers,
-            args=(self, )
+            args=(self._cache, self._taskqueue, self._ctx, self.Process,
+                  self._processes, self._pool, self._inqueue, self._outqueue,
+                  self._initializer, self._initargs, self._maxtasksperchild,
+                  self._wrap_exception)
             )
         self._worker_handler.daemon = True
         self._worker_handler._state = RUN
@@ -236,43 +240,61 @@ class Pool(object):
                 f'state={self._state} '
                 f'pool_size={len(self._pool)}>')
 
-    def _join_exited_workers(self):
+    @staticmethod
+    def _join_exited_workers(pool):
         """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]
+        for i in reversed(range(len(pool))):
+            worker = pool[i]
             if worker.exitcode is not None:
                 # worker exited
                 util.debug('cleaning up worker %d' % i)
                 worker.join()
                 cleaned = True
-                del self._pool[i]
+                del pool[i]
         return cleaned
 
     def _repopulate_pool(self):
+        return self._repopulate_pool_static(self._ctx, self.Process,
+                                            self._processes,
+                                            self._pool, self._inqueue,
+                                            self._outqueue, self._initializer,
+                                            self._initargs,
+                                            self._maxtasksperchild,
+                                            self._wrap_exception)
+
+    @staticmethod
+    def _repopulate_pool_static(ctx, Process, processes, pool, inqueue,
+                                outqueue, initializer, initargs,
+                                maxtasksperchild, wrap_exception):
         """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._wrap_exception)
-                            )
+        for i in range(processes - len(pool)):
+            w = Process(ctx, target=worker,
+                        args=(inqueue, outqueue,
+                              initializer,
+                              initargs, maxtasksperchild,
+                              wrap_exception))
             w.name = w.name.replace('Process', 'PoolWorker')
             w.daemon = True
             w.start()
-            self._pool.append(w)
+            pool.append(w)
             util.debug('added worker')
 
-    def _maintain_pool(self):
+    @staticmethod
+    def _maintain_pool(ctx, Process, processes, pool, inqueue, outqueue,
+                       initializer, initargs, maxtasksperchild,
+                       wrap_exception):
         """Clean up any exited workers and start replacements for them.
         """
-        if self._join_exited_workers():
-            self._repopulate_pool()
+        if Pool._join_exited_workers(pool):
+            Pool._repopulate_pool_static(ctx, Process, processes, pool,
+                                         inqueue, outqueue, initializer,
+                                         initargs, maxtasksperchild,
+                                         wrap_exception)
 
     def _setup_queues(self):
         self._inqueue = self._ctx.SimpleQueue()
@@ -331,7 +353,7 @@ class Pool(object):
         '''
         self._check_running()
         if chunksize == 1:
-            result = IMapIterator(self._cache)
+            result = IMapIterator(self)
             self._taskqueue.put(
                 (
                     self._guarded_task_generation(result._job, func, iterable),
@@ -344,7 +366,7 @@ class Pool(object):
                     "Chunksize must be 1+, not {0:n}".format(
                         chunksize))
             task_batches = Pool._get_tasks(func, iterable, chunksize)
-            result = IMapIterator(self._cache)
+            result = IMapIterator(self)
             self._taskqueue.put(
                 (
                     self._guarded_task_generation(result._job,
@@ -360,7 +382,7 @@ class Pool(object):
         '''
         self._check_running()
         if chunksize == 1:
-            result = IMapUnorderedIterator(self._cache)
+            result = IMapUnorderedIterator(self)
             self._taskqueue.put(
                 (
                     self._guarded_task_generation(result._job, func, iterable),
@@ -372,7 +394,7 @@ class Pool(object):
                 raise ValueError(
                     "Chunksize must be 1+, not {0!r}".format(chunksize))
             task_batches = Pool._get_tasks(func, iterable, chunksize)
-            result = IMapUnorderedIterator(self._cache)
+            result = IMapUnorderedIterator(self)
             self._taskqueue.put(
                 (
                     self._guarded_task_generation(result._job,
@@ -388,7 +410,7 @@ class Pool(object):
         Asynchronous version of `apply()` method.
         '''
         self._check_running()
-        result = ApplyResult(self._cache, callback, error_callback)
+        result = ApplyResult(self, callback, error_callback)
         self._taskqueue.put(([(result._job, 0, func, args, kwds)], None))
         return result
 
@@ -417,7 +439,7 @@ class Pool(object):
             chunksize = 0
 
         task_batches = Pool._get_tasks(func, iterable, chunksize)
-        result = MapResult(self._cache, chunksize, len(iterable), callback,
+        result = MapResult(self, chunksize, len(iterable), callback,
                            error_callback=error_callback)
         self._taskqueue.put(
             (
@@ -430,16 +452,20 @@ class Pool(object):
         return result
 
     @staticmethod
-    def _handle_workers(pool):
+    def _handle_workers(cache, taskqueue, ctx, Process, processes, pool,
+                        inqueue, outqueue, initializer, initargs,
+                        maxtasksperchild, wrap_exception):
         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()
+        while thread._state == RUN or (cache and thread._state != TERMINATE):
+            Pool._maintain_pool(ctx, Process, processes, pool, inqueue,
+                                outqueue, initializer, initargs,
+                                maxtasksperchild, wrap_exception)
             time.sleep(0.1)
         # send sentinel to stop workers
-        pool._taskqueue.put(None)
+        taskqueue.put(None)
         util.debug('worker handler exiting')
 
     @staticmethod
@@ -656,13 +682,14 @@ class Pool(object):
 
 class ApplyResult(object):
 
-    def __init__(self, cache, callback, error_callback):
+    def __init__(self, pool, callback, error_callback):
+        self._pool = pool
         self._event = threading.Event()
         self._job = next(job_counter)
-        self._cache = cache
+        self._cache = pool._cache
         self._callback = callback
         self._error_callback = error_callback
-        cache[self._job] = self
+        self._cache[self._job] = self
 
     def ready(self):
         return self._event.is_set()
@@ -692,6 +719,7 @@ class ApplyResult(object):
             self._error_callback(self._value)
         self._event.set()
         del self._cache[self._job]
+        self._pool = None
 
 AsyncResult = ApplyResult       # create alias -- see #17805
 
@@ -701,8 +729,8 @@ AsyncResult = ApplyResult       # create alias -- see #17805
 
 class MapResult(ApplyResult):
 
-    def __init__(self, cache, chunksize, length, callback, error_callback):
-        ApplyResult.__init__(self, cache, callback,
+    def __init__(self, pool, chunksize, length, callback, error_callback):
+        ApplyResult.__init__(self, pool, callback,
                              error_callback=error_callback)
         self._success = True
         self._value = [None] * length
@@ -710,7 +738,7 @@ class MapResult(ApplyResult):
         if chunksize <= 0:
             self._number_left = 0
             self._event.set()
-            del cache[self._job]
+            del self._cache[self._job]
         else:
             self._number_left = length//chunksize + bool(length % chunksize)
 
@@ -724,6 +752,7 @@ class MapResult(ApplyResult):
                     self._callback(self._value)
                 del self._cache[self._job]
                 self._event.set()
+                self._pool = None
         else:
             if not success and self._success:
                 # only store first exception
@@ -735,6 +764,7 @@ class MapResult(ApplyResult):
                     self._error_callback(self._value)
                 del self._cache[self._job]
                 self._event.set()
+                self._pool = None
 
 #
 # Class whose instances are returned by `Pool.imap()`
@@ -742,15 +772,16 @@ class MapResult(ApplyResult):
 
 class IMapIterator(object):
 
-    def __init__(self, cache):
+    def __init__(self, pool):
+        self._pool = pool
         self._cond = threading.Condition(threading.Lock())
         self._job = next(job_counter)
-        self._cache = cache
+        self._cache = pool._cache
         self._items = collections.deque()
         self._index = 0
         self._length = None
         self._unsorted = {}
-        cache[self._job] = self
+        self._cache[self._job] = self
 
     def __iter__(self):
         return self
@@ -761,12 +792,14 @@ class IMapIterator(object):
                 item = self._items.popleft()
             except IndexError:
                 if self._index == self._length:
+                    self._pool = None
                     raise StopIteration from None
                 self._cond.wait(timeout)
                 try:
                     item = self._items.popleft()
                 except IndexError:
                     if self._index == self._length:
+                        self._pool = None
                         raise StopIteration from None
                     raise TimeoutError from None
 
@@ -792,6 +825,7 @@ class IMapIterator(object):
 
             if self._index == self._length:
                 del self._cache[self._job]
+                self._pool = None
 
     def _set_length(self, length):
         with self._cond:
@@ -799,6 +833,7 @@ class IMapIterator(object):
             if self._index == self._length:
                 self._cond.notify()
                 del self._cache[self._job]
+                self._pool = None
 
 #
 # Class whose instances are returned by `Pool.imap_unordered()`
@@ -813,6 +848,7 @@ class IMapUnorderedIterator(IMapIterator):
             self._cond.notify()
             if self._index == self._length:
                 del self._cache[self._job]
+                self._pool = None
 
 #
 #
@@ -822,7 +858,7 @@ class ThreadPool(Pool):
     _wrap_exception = False
 
     @staticmethod
-    def Process(*args, **kwds):
+    def Process(ctx, *args, **kwds):
         from .dummy import Process
         return Process(*args, **kwds)
 
index bc1072d4b30dc497294de12f7c6f4b2b59055e50..d93303bb9ccacf26a95fb92e25ee217bb58d2de2 100644 (file)
@@ -2593,7 +2593,6 @@ class _TestPool(BaseTestCase):
             pool = None
             support.gc_collect()
 
-
 def raising():
     raise KeyError("key")
 
diff --git a/Misc/NEWS.d/next/Library/2019-01-21-02-15-20.bpo-35378.4oF03i.rst b/Misc/NEWS.d/next/Library/2019-01-21-02-15-20.bpo-35378.4oF03i.rst
new file mode 100644 (file)
index 0000000..bb57f71
--- /dev/null
@@ -0,0 +1,6 @@
+Fix a reference issue inside :class:`multiprocessing.Pool` that caused
+the pool to remain alive if it was deleted without being closed or
+terminated explicitly. A new strong reference is added to the pool
+iterators to link the lifetime of the pool to the lifetime of its
+iterators so the pool does not get destroyed if a pool iterator is
+still alive.