bpo-39360: Ensure all workers exit when finalizing a multiprocessing Pool (GH-19009)

When the pull is not used via the context manager or terminate() is called, there is a system in multiprocessing.util that handles finalization of all pools via an atexit handler (the Finalize) class. This class registers the _terminate_pool handler in the registry of finalizers of the module, and that registry is called on interpreter exit via _exit_function. The problem is that the "happy" path with the context manager or manual call to finalize() does some extra steps that _terminate_pool does not. The step that is not executed when the atexit() handler calls _terminate_pool is pinging the _change_notifier queue to unblock the maintenance threads.

This commit moves the notification to the _terminate_pool function so is called from both code paths.

Co-authored-by: Pablo Galindo <[email protected]>

Author: isidentical

Lib/multiprocessing/pool.py

@@ -651,8 +651,6 @@ def close(self):
     def terminate(self):
         util.debug('terminating pool')
         self._state = TERMINATE
-        self._worker_handler._state = TERMINATE
-        self._change_notifier.put(None)
         self._terminate()
 
     def join(self):
@@ -682,7 +680,12 @@ def _terminate_pool(cls, taskqueue, inqueue, outqueue, pool, change_notifier,
         # this is guaranteed to only be called once
         util.debug('finalizing pool')
 
+        # Notify that the worker_handler state has been changed so the
+        # _handle_workers loop can be unblocked (and exited) in order to
+        # send the finalization sentinel all the workers.
         worker_handler._state = TERMINATE
+        change_notifier.put(None)
+
         task_handler._state = TERMINATE
 
         util.debug('helping task handler/workers to finish')

Lib/test/_test_multiprocessing.py

@@ -2780,6 +2780,24 @@ def test_pool_worker_lifetime_early_close(self):
         for (j, res) in enumerate(results):
             self.assertEqual(res.get(), sqr(j))
 
+    def test_worker_finalization_via_atexit_handler_of_multiprocessing(self):
+        # tests cases against bpo-38744 and bpo-39360
+        cmd = '''if 1:
+            from multiprocessing import Pool
+            problem = None
+            class A:
+                def __init__(self):
+                    self.pool = Pool(processes=1)
+            def test():
+                global problem
+                problem = A()
+                problem.pool.map(float, tuple(range(10)))
+            if __name__ == "__main__":
+                test()
+        '''
+        rc, out, err = test.support.script_helper.assert_python_ok('-c', cmd)
+        self.assertEqual(rc, 0)
+
 #
 # Test of creating a customized manager class
 #

Misc/NEWS.d/next/Library/2020-03-15-05-41-05.bpo-39360.cmcU5p.rst

@@ -0,0 +1,4 @@
+Ensure all workers exit when finalizing a :class:`multiprocessing.Pool` implicitly via the module finalization
+handlers of multiprocessing. This fixes a deadlock situation that can be experienced when the Pool is not
+properly finalized via the context manager or a call to ``multiprocessing.Pool.terminate``. Patch by Batuhan Taskaya
+and Pablo Galindo.