gh-134634: Use a pipe instead of a queue for Pool's change notifier

Lib/multiprocessing/pool.py

@@ -147,6 +147,30 @@ def _helper_reraises_exception(ex):
 # Class representing a process pool
 #
 
+class _ChangeNotifier:
+    """Pipe-based notifier that does not depend on POSIX named semaphores.
+
+    Replaces the use of multiprocessing.SimpleQueue for the pool's change
+    notification mechanism.  SimpleQueue's internal locks require sem_open(),
+    which fails on platforms without /dev/shm (AWS Lambda, Android, iOS).
+    The change notifier never crosses a process boundary, so named semaphores
+    were never needed.
+    """
+
+    def __init__(self):
+        from .connection import Pipe
+        self._reader, self._writer = Pipe(duplex=False)
+
+    def put(self, obj):
+        self._writer.send_bytes(b'\0')
+
+    def get(self):
+        self._reader.recv_bytes()
+
+    def empty(self):
+        return not self._reader.poll(0)
+
+
 class _PoolCache(dict):
     """
     Class that implements a cache for the Pool class that will notify
@@ -190,10 +214,7 @@ def __init__(self, processes=None, initializer=None, initargs=(),
         self._ctx = context or get_context()
         self._setup_queues()
         self._taskqueue = queue.SimpleQueue()
-        # The _change_notifier queue exist to wake up self._handle_workers()
-        # when the cache (self._cache) is empty or when there is a change in
-        # the _state variable of the thread that runs _handle_workers.
-        self._change_notifier = self._ctx.SimpleQueue()
+        self._change_notifier = _ChangeNotifier()
         self._cache = _PoolCache(notifier=self._change_notifier)
         self._maxtasksperchild = maxtasksperchild
         self._initializer = initializer

Lib/test/_test_multiprocessing.py

@@ -2828,6 +2828,9 @@ def raise_large_valuerror(wait):
 def identity(x):
     return x
 
+def _kill_self():
+    os.kill(os.getpid(), signal.SIGKILL)
+
 class CountedObject(object):
     n_instances = 0
 
@@ -3068,6 +3071,43 @@ def test_make_pool(self):
                 p.close()
                 p.join()
 
+    @warnings_helper.ignore_fork_in_thread_deprecation_warnings()
+    def test_change_notifier_no_semaphore(self):
+        # gh-134634: The pool's change notifier uses a pipe instead of
+        # a multiprocessing.SimpleQueue to avoid depending on sem_open(),
+        # which requires /dev/shm.
+        if self.TYPE == 'manager':
+            return
+        from multiprocessing.pool import _ChangeNotifier
+        p = self.Pool(2)
+        try:
+            self.assertIsInstance(p._change_notifier, _ChangeNotifier)
+            result = p.map(sqr, list(range(10)))
+            self.assertEqual(result, list(map(sqr, range(10))))
+        finally:
+            p.close()
+            p.join()
+
+    @warnings_helper.ignore_fork_in_thread_deprecation_warnings()
+    def test_change_notifier_drain(self):
+        # gh-134634: Verify that multiple rapid notifications (from many
+        # tasks completing and cache emptying) are properly drained and
+        # the pool shuts down without hanging.
+        if self.TYPE == 'manager':
+            return
+        p = self.Pool(4)
+        try:
+            # Submit many short tasks so notifications pile up faster
+            # than the worker handler can drain them.
+            results = [p.apply_async(sqr, (i,)) for i in range(200)]
+            for i, r in enumerate(results):
+                self.assertEqual(r.get(timeout=support.SHORT_TIMEOUT), sqr(i))
+        finally:
+            # close() adds another notification on top of the cache-empty
+            # ones.  If draining is broken, join() hangs.
+            p.close()
+            p.join()
+
     @warnings_helper.ignore_fork_in_thread_deprecation_warnings()
     def test_terminate(self):
         # Simulate slow tasks which take "forever" to complete
@@ -3331,6 +3371,34 @@ def test_pool_worker_lifetime_early_close(self):
         for (j, res) in enumerate(results):
             self.assertEqual(res.get(), sqr(j))
 
+    @warnings_helper.ignore_fork_in_thread_deprecation_warnings()
+    def test_pool_worker_killed_mid_task(self):
+        # gh-134634: Verify the worker handler detects a killed worker
+        # via its sentinel fd and replaces it, keeping the pool functional.
+        p = multiprocessing.Pool(3)
+        orig_count = len(p._pool)
+        self.assertEqual(orig_count, 3)
+
+        # Kill one worker.  The task will never return a result.
+        p.apply_async(_kill_self)
+
+        # Give the worker handler time to detect the death and replace.
+        deadline = time.monotonic() + support.SHORT_TIMEOUT
+        while time.monotonic() < deadline:
+            alive = [w for w in p._pool if w.is_alive()]
+            if len(alive) >= orig_count:
+                break
+            time.sleep(0.1)
+        self.assertEqual(len(p._pool), orig_count,
+                         "worker handler did not replace the killed worker")
+
+        # Pool should still be functional with the replacement worker.
+        result = p.map(sqr, list(range(5)))
+        self.assertEqual(result, list(map(sqr, range(5))))
+
+        p.terminate()
+        p.join()
+
     def test_pool_maxtasksperchild_invalid(self):
         for value in [0, -1, 0.5, "12"]:
             with self.assertRaises(ValueError):

Misc/NEWS.d/next/Library/2026-04-03-20-00-00.gh-issue-134634.Xk9fRe.rst

@@ -0,0 +1,5 @@
+:class:`multiprocessing.pool.Pool` and :class:`multiprocessing.pool.ThreadPool`
+no longer require POSIX named semaphores for the internal change notification
+mechanism.  This fixes ``Pool`` and ``ThreadPool`` on platforms where
+``sem_open()`` is unavailable, such as AWS Lambda (no ``/dev/shm``), Android,
+and iOS.