[3.13] gh-146313: Fix multiprocessing ResourceTracker deadlock after os.fork() (GH-146316) (#148426)

gh-146313: Fix multiprocessing ResourceTracker deadlock after os.fork() (GH-146316)

`ResourceTracker.__del__` (added in gh-88887 circa Python 3.12) calls
os.waitpid(pid, 0) which blocks indefinitely if a process created via os.fork()
still holds the tracker pipe's write end. The tracker never sees EOF, never
exits, and the parent hangs at interpreter shutdown.

Fix with two layers:

- **At-fork handler.** An os.register_at_fork(after_in_child=...)
  handler closes the inherited pipe fd in the child unless a preserve
  flag is set. popen_fork.Popen._launch() sets the flag before its
  fork so mp.Process(fork) children keep the fd and reuse the parent's
  tracker (preserving gh-80849). Raw os.fork() children close the fd,
  letting the parent reap promptly.

- **Timeout safety-net.** _stop_locked() gains a wait_timeout
  parameter. When called from `__del__`, it polls with WNOHANG using
  exponential backoff for up to 1 second instead of blocking
  indefinitely. The at-fork handler makes this unreachable in
  well-behaved paths; it remains for abnormal shutdowns.
(cherry picked from commit 3a7df632c9)

Co-authored-by: Gregory P. Smith <68491+gpshead@users.noreply.github.com>
Co-authored-by: Itamar Oren <itamarost@gmail.com>

Lib/multiprocessing/popen_fork.py

@@ -64,7 +64,17 @@ class Popen(object):
         code = 1
         parent_r, child_w = os.pipe()
         child_r, parent_w = os.pipe()
-        self.pid = os.fork()
+        # gh-146313: Tell the resource tracker's at-fork handler to keep
+        # the inherited pipe fd so this child reuses the parent's tracker
+        # (gh-80849) rather than closing it and launching its own.
+        from .resource_tracker import _fork_intent
+        _fork_intent.preserve_fd = True
+        try:
+            self.pid = os.fork()
+        finally:
+            # Reset in both parent and child so the flag does not leak
+            # into a subsequent raw os.fork() or nested Process launch.
+            _fork_intent.preserve_fd = False
         if self.pid == 0:
             try:
                 atexit._clear()

Lib/multiprocessing/resource_tracker.py

@@ -20,6 +20,7 @@ import os
 import signal
 import sys
 import threading
+import time
 import warnings
 from collections import deque
 
@@ -79,6 +80,10 @@ class ResourceTracker(object):
         # The reader should understand all formats.
         self._use_simple_format = True
 
+        # Set to True by _stop_locked() if the waitpid polling loop ran to
+        # its timeout without reaping the tracker.  Exposed for tests.
+        self._waitpid_timed_out = False
+
     def _reentrant_call_error(self):
         # gh-109629: this happens if an explicit call to the ResourceTracker
         # gets interrupted by a garbage collection, invoking a finalizer (*)
@@ -91,16 +96,51 @@ class ResourceTracker(object):
         # making sure child processess are cleaned before ResourceTracker
         # gets destructed.
         # see https://github.com/python/cpython/issues/88887
-        self._stop(use_blocking_lock=False)
+        # gh-146313: use a timeout to avoid deadlocking if a forked child
+        # still holds the pipe's write end open.
+        self._stop(use_blocking_lock=False, wait_timeout=1.0)
 
-    def _stop(self, use_blocking_lock=True):
+    def _after_fork_in_child(self):
+        # gh-146313: Called in the child right after os.fork().
+        #
+        # The tracker process is a child of the *parent*, not of us, so we
+        # could never waitpid() it anyway.  Clearing _pid means our __del__
+        # becomes a no-op (the early return for _pid is None).
+        #
+        # Whether we keep the inherited _fd depends on who forked us:
+        #
+        #   - multiprocessing.Process with the 'fork' start method sets
+        #     _fork_intent.preserve_fd before forking.  The child keeps the
+        #     fd and reuses the parent's tracker (gh-80849).  This is safe
+        #     because multiprocessing's atexit handler joins all children
+        #     before the parent's __del__ runs, so by then the fd copies
+        #     are gone and the parent can reap the tracker promptly.
+        #
+        #   - A raw os.fork() leaves the flag unset.  We close the fd in the child after forking so
+        #     the parent's __del__ can reap the tracker without waiting
+        #     for the child to exit.  If we later need a tracker, ensure_running()
+        #     will launch a fresh one.
+        self._lock._at_fork_reinit()
+        self._reentrant_messages.clear()
+        self._pid = None
+        self._exitcode = None
+        if (self._fd is not None and
+            not getattr(_fork_intent, 'preserve_fd', False)):
+            fd = self._fd
+            self._fd = None
+            try:
+                os.close(fd)
+            except OSError:
+                pass
+
+    def _stop(self, use_blocking_lock=True, wait_timeout=None):
         if use_blocking_lock:
             with self._lock:
-                self._stop_locked()
+                self._stop_locked(wait_timeout=wait_timeout)
         else:
             acquired = self._lock.acquire(blocking=False)
             try:
-                self._stop_locked()
+                self._stop_locked(wait_timeout=wait_timeout)
             finally:
                 if acquired:
                     self._lock.release()
@@ -110,6 +150,10 @@ class ResourceTracker(object):
         close=os.close,
         waitpid=os.waitpid,
         waitstatus_to_exitcode=os.waitstatus_to_exitcode,
+        monotonic=time.monotonic,
+        sleep=time.sleep,
+        WNOHANG=getattr(os, 'WNOHANG', None),
+        wait_timeout=None,
     ):
         # This shouldn't happen (it might when called by a finalizer)
         # so we check for it anyway.
@@ -126,7 +170,30 @@ class ResourceTracker(object):
         self._fd = None
 
         try:
-            _, status = waitpid(self._pid, 0)
+            if wait_timeout is None:
+                _, status = waitpid(self._pid, 0)
+            else:
+                # gh-146313: A forked child may still hold the pipe's write
+                # end open, preventing the tracker from seeing EOF and
+                # exiting.  Poll with WNOHANG to avoid blocking forever.
+                deadline = monotonic() + wait_timeout
+                delay = 0.001
+                while True:
+                    result_pid, status = waitpid(self._pid, WNOHANG)
+                    if result_pid != 0:
+                        break
+                    remaining = deadline - monotonic()
+                    if remaining <= 0:
+                        # The tracker is still running; it will be
+                        # reparented to PID 1 (or the nearest subreaper)
+                        # when we exit, and reaped there once all pipe
+                        # holders release their fd.
+                        self._pid = None
+                        self._exitcode = None
+                        self._waitpid_timed_out = True
+                        return
+                    delay = min(delay * 2, remaining, 0.1)
+                    sleep(delay)
         except ChildProcessError:
             self._pid = None
             self._exitcode = None
@@ -312,12 +379,24 @@ class ResourceTracker(object):
 
         self._ensure_running_and_write(msg)
 
+# gh-146313: Per-thread flag set by .popen_fork.Popen._launch() just before
+# os.fork(), telling _after_fork_in_child() to keep the inherited pipe fd so
+# the child can reuse this tracker (gh-80849).  Unset for raw os.fork() calls,
+# where the child instead closes the fd so the parent's __del__ can reap the
+# tracker.  Using threading.local() keeps multiple threads calling
+# popen_fork.Popen._launch() at once from clobbering eachothers intent.
+_fork_intent = threading.local()
+
 _resource_tracker = ResourceTracker()
 ensure_running = _resource_tracker.ensure_running
 register = _resource_tracker.register
 unregister = _resource_tracker.unregister
 getfd = _resource_tracker.getfd
 
+# gh-146313: See _after_fork_in_child docstring.
+if hasattr(os, 'register_at_fork'):
+    os.register_at_fork(after_in_child=_resource_tracker._after_fork_in_child)
+
 
 def _decode_message(line):
     if line.startswith(b'{'):

Lib/test/_test_multiprocessing.py

@@ -6007,8 +6007,9 @@ class TestResourceTracker(unittest.TestCase):
     def _is_resource_tracker_reused(conn, pid):
         from multiprocessing.resource_tracker import _resource_tracker
         _resource_tracker.ensure_running()
-        # The pid should be None in the child process, expect for the fork
-        # context. It should not be a new value.
+        # The pid should be None in the child (the at-fork handler clears
+        # it for fork; spawn/forkserver children never had it set).  It
+        # should not be a new value.
         reused = _resource_tracker._pid in (None, pid)
         reused &= _resource_tracker._check_alive()
         conn.send(reused)
@@ -6093,6 +6094,183 @@ class TestResourceTracker(unittest.TestCase):
             # restore sigmask to what it was before executing test
             signal.pthread_sigmask(signal.SIG_SETMASK, orig_sigmask)
 
+    @only_run_in_forkserver_testsuite("avoids redundant testing.")
+    def test_resource_tracker_fork_deadlock(self):
+        # gh-146313: ResourceTracker.__del__ used to deadlock if a forked
+        # child still held the pipe's write end open when the parent
+        # exited, because the parent would block in waitpid() waiting for
+        # the tracker to exit, but the tracker would never see EOF.
+        cmd = '''if 1:
+            import os, signal
+            from multiprocessing.resource_tracker import ensure_running
+            ensure_running()
+            if os.fork() == 0:
+                signal.pause()
+                os._exit(0)
+            # parent falls through and exits, triggering __del__
+        '''
+        proc = subprocess.Popen([sys.executable, '-c', cmd],
+                                start_new_session=True)
+        try:
+            try:
+                proc.wait(timeout=support.SHORT_TIMEOUT)
+            except subprocess.TimeoutExpired:
+                self.fail(
+                    "Parent process deadlocked in ResourceTracker.__del__"
+                )
+            self.assertEqual(proc.returncode, 0)
+        finally:
+            try:
+                os.killpg(proc.pid, signal.SIGKILL)
+            except ProcessLookupError:
+                pass
+            proc.wait()
+
+    @only_run_in_forkserver_testsuite("avoids redundant testing.")
+    def test_resource_tracker_mp_fork_reuse_and_prompt_reap(self):
+        # gh-146313 / gh-80849: A child started via multiprocessing.Process
+        # with the 'fork' start method should reuse the parent's resource
+        # tracker (the at-fork handler preserves the inherited pipe fd),
+        # *and* the parent should be able to reap the tracker promptly
+        # after joining the child, without hitting the waitpid timeout.
+        cmd = textwrap.dedent('''
+            import multiprocessing as mp
+            from multiprocessing.resource_tracker import _resource_tracker
+
+            def child(conn):
+                # Prove we can talk to the parent's tracker by registering
+                # and unregistering a dummy resource over the inherited fd.
+                # If the fd were closed, ensure_running would launch a new
+                # tracker and _pid would be non-None.
+                _resource_tracker.register("x", "dummy")
+                _resource_tracker.unregister("x", "dummy")
+                conn.send((_resource_tracker._fd is not None,
+                           _resource_tracker._pid is None,
+                           _resource_tracker._check_alive()))
+
+            if __name__ == "__main__":
+                mp.set_start_method("fork")
+                _resource_tracker.ensure_running()
+                r, w = mp.Pipe(duplex=False)
+                p = mp.Process(target=child, args=(w,))
+                p.start()
+                child_has_fd, child_pid_none, child_alive = r.recv()
+                p.join()
+                w.close(); r.close()
+
+                # Now simulate __del__: the child has exited and released
+                # its fd copy, so the tracker should see EOF and exit
+                # promptly -- no timeout.
+                _resource_tracker._stop(wait_timeout=5.0)
+                print(child_has_fd, child_pid_none, child_alive,
+                      _resource_tracker._waitpid_timed_out,
+                      _resource_tracker._exitcode)
+        ''')
+        rc, out, err = script_helper.assert_python_ok('-c', cmd)
+        parts = out.decode().split()
+        self.assertEqual(parts, ['True', 'True', 'True', 'False', '0'],
+            f"unexpected: {parts!r} stderr={err!r}")
+
+    @only_run_in_forkserver_testsuite("avoids redundant testing.")
+    def test_resource_tracker_raw_fork_prompt_reap(self):
+        # gh-146313: After a raw os.fork() the at-fork handler closes the
+        # child's inherited fd, so the parent can reap the tracker
+        # immediately -- even while the child is still alive -- rather
+        # than waiting out the 1s timeout.
+        cmd = textwrap.dedent('''
+            import os, signal
+            from multiprocessing.resource_tracker import _resource_tracker
+
+            _resource_tracker.ensure_running()
+            r, w = os.pipe()
+            pid = os.fork()
+            if pid == 0:
+                os.close(r)
+                # Report whether our fd was closed by the at-fork handler.
+                os.write(w, b"1" if _resource_tracker._fd is None else b"0")
+                os.close(w)
+                signal.pause()  # stay alive so parent's reap is meaningful
+                os._exit(0)
+            os.close(w)
+            child_fd_closed = os.read(r, 1) == b"1"
+            os.close(r)
+
+            # Child is still alive and paused.  Because it closed its fd
+            # copy, our close below is the last one and the tracker exits.
+            _resource_tracker._stop(wait_timeout=5.0)
+
+            os.kill(pid, signal.SIGKILL)
+            os.waitpid(pid, 0)
+            print(child_fd_closed,
+                  _resource_tracker._waitpid_timed_out,
+                  _resource_tracker._exitcode)
+        ''')
+        rc, out, err = script_helper.assert_python_ok('-c', cmd)
+        parts = out.decode().split()
+        self.assertEqual(parts, ['True', 'False', '0'],
+            f"unexpected: {parts!r} stderr={err!r}")
+
+    @only_run_in_forkserver_testsuite("avoids redundant testing.")
+    def test_resource_tracker_lock_reinit_after_fork(self):
+        # gh-146313: If a parent thread held the tracker's lock at fork
+        # time, the child would inherit the held lock and deadlock on
+        # its next ensure_running().  The at-fork handler reinits it.
+        cmd = textwrap.dedent('''
+            import os, threading
+            from multiprocessing.resource_tracker import _resource_tracker
+
+            held = threading.Event()
+            release = threading.Event()
+            def hold():
+                with _resource_tracker._lock:
+                    held.set()
+                    release.wait()
+            t = threading.Thread(target=hold)
+            t.start()
+            held.wait()
+
+            pid = os.fork()
+            if pid == 0:
+                ok = _resource_tracker._lock.acquire(timeout=5.0)
+                os._exit(0 if ok else 1)
+
+            release.set()
+            t.join()
+            _, status = os.waitpid(pid, 0)
+            print(os.waitstatus_to_exitcode(status))
+        ''')
+        rc, out, err = script_helper.assert_python_ok(
+            '-W', 'ignore::DeprecationWarning', '-c', cmd)
+        self.assertEqual(out.strip(), b'0',
+            f"child failed to acquire lock: stderr={err!r}")
+
+    @only_run_in_forkserver_testsuite("avoids redundant testing.")
+    def test_resource_tracker_safety_net_timeout(self):
+        # gh-146313: When an mp.Process(fork) child holds the preserved
+        # fd and the parent calls _stop() without joining (simulating
+        # abnormal shutdown), the safety-net timeout should fire rather
+        # than deadlocking.
+        cmd = textwrap.dedent('''
+            import multiprocessing as mp
+            import signal
+            from multiprocessing.resource_tracker import _resource_tracker
+
+            if __name__ == "__main__":
+                mp.set_start_method("fork")
+                _resource_tracker.ensure_running()
+                p = mp.Process(target=signal.pause)
+                p.start()
+                # Stop WITHOUT joining -- child still holds preserved fd
+                _resource_tracker._stop(wait_timeout=0.5)
+                print(_resource_tracker._waitpid_timed_out)
+                p.terminate()
+                p.join()
+        ''')
+        rc, out, err = script_helper.assert_python_ok('-c', cmd)
+        self.assertEqual(out.strip(), b'True',
+            f"safety-net timeout did not fire: stderr={err!r}")
+
+
 class TestSimpleQueue(unittest.TestCase):
 
     @classmethod

Misc/NEWS.d/next/Library/2026-03-22-23-42-22.gh-issue-146313.RtDeAd.rst

@@ -0,0 +1,4 @@
+Fix a deadlock in :mod:`multiprocessing`'s resource tracker
+where the parent process could hang indefinitely in :func:`os.waitpid`
+during interpreter shutdown if a child created via :func:`os.fork` still
+held the resource tracker's pipe open.