gh-140257: fix data race on eval_breaker during finalization (#140265)

Author: ashm-dev

Misc/NEWS.d/next/Core_and_Builtins/2025-10-17-20-23-19.gh-issue-140257.8Txmem.rst

@@ -0,0 +1,2 @@
+Fix data race between interpreter_clear() and take_gil() on eval_breaker
+during finalization with daemon threads.

Python/pystate.c

@@ -763,10 +763,11 @@ interpreter_clear(PyInterpreterState *interp, PyThreadState *tstate)
 
     Py_CLEAR(interp->audit_hooks);
 
-    // At this time, all the threads should be cleared so we don't need atomic
-    // operations for instrumentation_version or eval_breaker.
+    // gh-140257: Threads have already been cleared, but daemon threads may
+    // still access eval_breaker atomically via take_gil() right before they
+    // hang. Use an atomic store to prevent data races during finalization.
     interp->ceval.instrumentation_version = 0;
-    tstate->eval_breaker = 0;
+    _Py_atomic_store_uintptr(&tstate->eval_breaker, 0);
 
     for (int i = 0; i < _PY_MONITORING_UNGROUPED_EVENTS; i++) {
         interp->monitors.tools[i] = 0;