gh-151518: Avoid STW starvation of attaching threads

Include/cpython/pystate.h

@@ -118,8 +118,7 @@ struct _ts {
 
     int _whence;
 
-    /* Thread state (_Py_THREAD_ATTACHED, _Py_THREAD_DETACHED, _Py_THREAD_SUSPENDED).
-       See Include/internal/pycore_pystate.h for more details. */
+    /* Thread state. See Include/internal/pycore_pystate.h for details. */
     int state;
 
     int py_recursion_remaining;

Include/internal/pycore_pystate.h

@@ -21,32 +21,32 @@ extern "C" {
 // interpreter at the same time. Only the "bound" thread may perform the
 // transitions between "attached" and "detached" on its own PyThreadState.
 //
-// The "suspended" state is used to implement stop-the-world pauses, such as
-// for cyclic garbage collection. It is only used in `--disable-gil` builds.
-// The "suspended" state is similar to the "detached" state in that in both
-// states the thread is not allowed to call most Python APIs. However, unlike
-// the "detached" state, a thread may not transition itself out from the
-// "suspended" state. Only the thread performing a stop-the-world pause may
-// transition a thread from the "suspended" state back to the "detached" state.
+// The "suspended" states are used to implement stop-the-world pauses and to
+// merge biased reference counts on behalf of detached threads. They are only
+// used in `--disable-gil` builds.
+// They are similar to the "detached" state in that the thread is not allowed
+// to call most Python APIs. A suspended thread trying to attach marks itself
+// as "suspended-waiting". Only the thread responsible for suspending it may
+// resume it, moving it to "detached" or "detached-waiting".
+// A "detached-waiting" thread must attach before it can be suspended again.
 //
 // The "shutting down" state is used when the interpreter is being finalized.
 // Threads in this state can't do anything other than block the OS thread.
 // (See _PyThreadState_HangThread).
 //
-// State transition diagram:
-//
-//            (bound thread)        (stop-the-world thread)
-// [attached]       <->       [detached]       <->       [suspended]
-//   |                                                        ^
-//   +---------------------------->---------------------------+
-//                          (bound thread)
-//
-// The (bound thread) and (stop-the-world thread) labels indicate which thread
-// is allowed to perform the transition.
-#define _Py_THREAD_DETACHED         0
-#define _Py_THREAD_ATTACHED         1
-#define _Py_THREAD_SUSPENDED        2
-#define _Py_THREAD_SHUTTING_DOWN    3
+// State transitions:
+//   Bound thread:          attached <-> detached
+//                         attached  -> suspended
+//                         suspended -> suspended-waiting
+//                         detached-waiting -> attached
+//   Suspending thread:     detached <-> suspended
+//                         suspended-waiting -> detached-waiting
+#define _Py_THREAD_DETACHED             0
+#define _Py_THREAD_ATTACHED             1
+#define _Py_THREAD_SUSPENDED            2
+#define _Py_THREAD_SHUTTING_DOWN        3
+#define _Py_THREAD_SUSPENDED_WAITING    4
+#define _Py_THREAD_DETACHED_WAITING     5
 
 
 /* Check if the current thread is the main thread.
@@ -162,8 +162,9 @@ extern void _PyThreadState_Suspend(PyThreadState *tstate);
 // Returns 1 on success, 0 if the thread was not in the "detached" state.
 extern int _PyThreadState_TrySuspendDetached(PyThreadState *tstate);
 
-// Undo a successful _PyThreadState_TrySuspendDetached(): switch the thread
-// back to "detached" and wake it if it is waiting to attach.
+// Resume a thread suspended by _PyThreadState_TrySuspendDetached() or a
+// stop-the-world pause: switch it back to "detached" or "detached-waiting"
+// and wake it if it is waiting to attach.
 extern void _PyThreadState_ResumeDetached(PyThreadState *tstate);
 #endif
 

Lib/test/test_free_threading/test_threading.py

@@ -1,5 +1,8 @@
 import unittest
-from test.support import threading_helper
+import textwrap
+
+from test import support
+from test.support import script_helper, threading_helper
 
 threading_helper.requires_working_threading(module=True)
 
@@ -22,5 +25,39 @@ def mutate_thread():
         threading_helper.run_concurrently([repr_thread, mutate_thread])
 
 
+class TestThreadState(unittest.TestCase):
+    @support.requires_subprocess()
+    def test_tight_stw_loop_does_not_starve_attach(self):
+        script = textwrap.dedent(f"""
+            import faulthandler
+
+            faulthandler.dump_traceback_later({support.SHORT_TIMEOUT}, exit=True)
+
+            import _testinternalcapi
+            import threading
+            import time
+
+            started = threading.Event()
+            stop = threading.Event()
+
+            def stop_the_world():
+                _testinternalcapi.test_stop_the_world()
+                started.set()
+                while not stop.is_set():
+                    _testinternalcapi.test_stop_the_world()
+
+            thread = threading.Thread(target=stop_the_world)
+            thread.start()
+            started.wait()
+            # Each reattachment must make progress between consecutive pauses.
+            for _ in range(50):
+                time.sleep(0.02)
+            stop.set()
+            thread.join()
+            faulthandler.cancel_dump_traceback_later()
+        """)
+        script_helper.assert_python_ok("-X", "gil=0", "-c", script)
+
+
 if __name__ == "__main__":
     unittest.main()

Misc/NEWS.d/next/Core_and_Builtins/2026-06-16-19-20-00.gh-issue-151518.e6v0Js.rst

@@ -0,0 +1,2 @@
+Fix a free-threaded stop-the-world fairness issue that could starve a thread
+reattaching after being suspended while detached.

Modules/_testinternalcapi.c

@@ -30,6 +30,7 @@
 #include "pycore_instruction_sequence.h"  // _PyInstructionSequence_New()
 #include "pycore_interpframe.h"   // _PyFrame_GetFunction()
 #include "pycore_jit.h"           // _PyJIT_AddressInJitCode()
+#include "pycore_lock.h"          // PyEvent_WaitTimed()
 #include "pycore_object.h"        // _PyObject_IsFreed()
 #include "pycore_optimizer.h"     // _Py_Executor_DependsOn
 #include "pycore_pathconfig.h"    // _PyPathConfig_ClearGlobal()
@@ -208,6 +209,23 @@ get_stack_margin(PyObject *self, PyObject *Py_UNUSED(args))
     return PyLong_FromSize_t(_PyOS_STACK_MARGIN_BYTES);
 }
 
+static PyObject *
+test_stop_the_world(PyObject *self, PyObject *Py_UNUSED(args))
+{
+#ifdef Py_GIL_DISABLED
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    // Request consecutive pauses without running Python code between them.
+    for (int i = 0; i < 100; i++) {
+        _PyEval_StopTheWorld(interp);
+        // Give detached threads time to try to reattach during the pause.
+        PyEvent event = {0};
+        PyEvent_WaitTimed(&event, 10 * 1000 * 1000, /*detach=*/0);
+        _PyEval_StartTheWorld(interp);
+    }
+#endif
+    Py_RETURN_NONE;
+}
+
 #ifdef MS_WINDOWS
 static const char *
 classify_address(uintptr_t addr, int jit_enabled, PyInterpreterState *interp)
@@ -3298,6 +3316,7 @@ static PyMethodDef module_functions[] = {
     {"get_c_recursion_remaining", get_c_recursion_remaining, METH_NOARGS},
     {"get_stack_pointer", get_stack_pointer, METH_NOARGS},
     {"get_stack_margin", get_stack_margin, METH_NOARGS},
+    {"test_stop_the_world", test_stop_the_world, METH_NOARGS},
     {"classify_stack_addresses", classify_stack_addresses, METH_VARARGS},
     {"get_jit_code_ranges", get_jit_code_ranges, METH_NOARGS},
     {"get_jit_backend", get_jit_backend, METH_NOARGS},

Python/pystate.c

@@ -1939,7 +1939,10 @@ tstate_delete_common(PyThreadState *tstate, int release_gil)
     if (tstate->next) {
         tstate->next->prev = tstate->prev;
     }
-    if (tstate->state != _Py_THREAD_SUSPENDED) {
+    int state = _Py_atomic_load_int_relaxed(&tstate->state);
+    if (state != _Py_THREAD_SUSPENDED &&
+        state != _Py_THREAD_SUSPENDED_WAITING)
+    {
         // Any ongoing stop-the-world request should not wait for us because
         // our thread is getting deleted.
         if (interp->stoptheworld.requested) {
@@ -2223,6 +2226,22 @@ tstate_try_attach(PyThreadState *tstate)
 #endif
 }
 
+static int
+tstate_try_attach_detached(PyThreadState *tstate, int *state)
+{
+#ifdef Py_GIL_DISABLED
+    assert(*state == _Py_THREAD_DETACHED ||
+           *state == _Py_THREAD_DETACHED_WAITING);
+    return _Py_atomic_compare_exchange_int(&tstate->state,
+                                           state,
+                                           _Py_THREAD_ATTACHED);
+#else
+    assert(tstate->state == _Py_THREAD_DETACHED);
+    tstate->state = _Py_THREAD_ATTACHED;
+    return 1;
+#endif
+}
+
 static void
 tstate_set_detached(PyThreadState *tstate, int detached_state)
 {
@@ -2237,10 +2256,20 @@ tstate_set_detached(PyThreadState *tstate, int detached_state)
 static void
 tstate_wait_attach(PyThreadState *tstate)
 {
-    do {
+    for (;;) {
         int state = _Py_atomic_load_int_relaxed(&tstate->state);
         if (state == _Py_THREAD_SUSPENDED) {
-            // Wait until we're switched out of SUSPENDED to DETACHED.
+            // Register an active attach waiter. The next stop-the-world
+            // request must let this thread attach before suspending it again.
+            if (!_Py_atomic_compare_exchange_int(
+                    &tstate->state, &state, _Py_THREAD_SUSPENDED_WAITING))
+            {
+                continue;
+            }
+            state = _Py_THREAD_SUSPENDED_WAITING;
+        }
+        if (state == _Py_THREAD_SUSPENDED_WAITING) {
+            // Park rechecks the state before sleeping, in case we were resumed.
             _PyParkingLot_Park(&tstate->state, &state, sizeof(tstate->state),
                                /*timeout=*/-1, NULL, /*detach=*/0);
         }
@@ -2249,10 +2278,13 @@ tstate_wait_attach(PyThreadState *tstate)
             _PyThreadState_HangThread(tstate);
         }
         else {
-            assert(state == _Py_THREAD_DETACHED);
+            assert(state == _Py_THREAD_DETACHED ||
+                   state == _Py_THREAD_DETACHED_WAITING);
+            if (tstate_try_attach_detached(tstate, &state)) {
+                return;
+            }
         }
-        // Once we're back in DETACHED we can re-attach
-    } while (!tstate_try_attach(tstate));
+    }
 }
 
 void
@@ -2394,8 +2426,20 @@ void
 _PyThreadState_ResumeDetached(PyThreadState *tstate)
 {
     assert(tstate != _PyThreadState_GET());
-    assert(_Py_atomic_load_int_relaxed(&tstate->state) == _Py_THREAD_SUSPENDED);
-    _Py_atomic_store_int(&tstate->state, _Py_THREAD_DETACHED);
+    int state = _Py_atomic_load_int_relaxed(&tstate->state);
+    int next_state;
+    do {
+        assert(state == _Py_THREAD_SUSPENDED ||
+               state == _Py_THREAD_SUSPENDED_WAITING);
+        if (state == _Py_THREAD_SUSPENDED_WAITING) {
+            next_state = _Py_THREAD_DETACHED_WAITING;
+        }
+        else {
+            next_state = _Py_THREAD_DETACHED;
+        }
+        // Retry if an attach waiter registered concurrently.
+    } while (!_Py_atomic_compare_exchange_int(
+                &tstate->state, &state, next_state));
     // Wake the thread if it is parked in tstate_wait_attach().
     _PyParkingLot_UnparkAll(&tstate->state);
 }
@@ -2442,6 +2486,8 @@ park_detached_threads(struct _stoptheworld_state *stw)
     _Py_FOR_EACH_STW_INTERP(stw, i) {
         _Py_FOR_EACH_TSTATE_UNLOCKED(i, t) {
             int state = _Py_atomic_load_int_relaxed(&t->state);
+            // DETACHED_WAITING threads remain counted until they attach and
+            // stop, so repeated pauses cannot prevent them from attaching.
             if (state == _Py_THREAD_DETACHED) {
                 // Atomically transition to "suspended" if in "detached" state.
                 if (_Py_atomic_compare_exchange_int(
@@ -2530,10 +2576,7 @@ start_the_world(struct _stoptheworld_state *stw)
     _Py_FOR_EACH_STW_INTERP(stw, i) {
         _Py_FOR_EACH_TSTATE_UNLOCKED(i, t) {
             if (t != stw->requester) {
-                assert(_Py_atomic_load_int_relaxed(&t->state) ==
-                       _Py_THREAD_SUSPENDED);
-                _Py_atomic_store_int(&t->state, _Py_THREAD_DETACHED);
-                _PyParkingLot_UnparkAll(&t->state);
+                _PyThreadState_ResumeDetached(t);
             }
         }
     }