gh-127266: avoid data races when updating type slots

In the free-threaded build, avoid data races caused by updating type
slots or type flags after the type was initially created.  For those
(typically rare) cases, use the stop-the-world mechanism.  Remove the
use of atomics when reading or writing type flags.  The use of atomics
is not sufficient to avoid races (since flags are sometimes read without
a lock and without atomics) and are no longer required.

Author: nascheme

Include/internal/pycore_interp_structs.h

@@ -667,8 +667,11 @@ struct _Py_interp_cached_objects {
 
     /* object.__reduce__ */
     PyObject *objreduce;
+#ifndef Py_GIL_DISABLED
+    /* resolve_slotdups() */
     PyObject *type_slots_pname;
     pytype_slotdef *type_slots_ptrs[MAX_EQUIV];
+#endif
 
     /* TypeVar and related types */
     PyTypeObject *generic_type;

Include/internal/pycore_object.h

@@ -313,7 +313,7 @@ extern int _PyDict_CheckConsistency(PyObject *mp, int check_content);
 // Fast inlined version of PyType_HasFeature()
 static inline int
 _PyType_HasFeature(PyTypeObject *type, unsigned long feature) {
-    return ((FT_ATOMIC_LOAD_ULONG_RELAXED(type->tp_flags) & feature) != 0);
+    return ((type->tp_flags) & feature) != 0;
 }
 
 extern void _PyType_InitCache(PyInterpreterState *interp);

Include/internal/pycore_typeobject.h

@@ -134,7 +134,6 @@ extern int _PyType_AddMethod(PyTypeObject *, PyMethodDef *);
 extern void _PyType_SetFlagsRecursive(PyTypeObject *self, unsigned long mask,
                                       unsigned long flags);
 
-extern unsigned int _PyType_GetVersionForCurrentState(PyTypeObject *tp);
 PyAPI_FUNC(void) _PyType_SetVersion(PyTypeObject *tp, unsigned int version);
 PyTypeObject *_PyType_LookupByVersion(unsigned int version);
 

Include/object.h

@@ -620,6 +620,12 @@ given type object has a specified feature.
 #define Py_TPFLAGS_HAVE_FINALIZE (1UL << 0)
 #define Py_TPFLAGS_HAVE_VERSION_TAG   (1UL << 18)
 
+// Flag values for ob_flags (16 bits available, if SIZEOF_VOID_P > 4).
+#define _Py_IMMORTAL_FLAGS (1 << 0)
+#define _Py_STATICALLY_ALLOCATED_FLAG (1 << 2)
+#if defined(Py_GIL_DISABLED) && defined(Py_DEBUG)
+#define _Py_TYPE_REVEALED_FLAG (1 << 3)
+#endif
 
 #define Py_CONSTANT_NONE 0
 #define Py_CONSTANT_FALSE 1
@@ -776,11 +782,7 @@ PyType_HasFeature(PyTypeObject *type, unsigned long feature)
     // PyTypeObject is opaque in the limited C API
     flags = PyType_GetFlags(type);
 #else
-#   ifdef Py_GIL_DISABLED
-        flags = _Py_atomic_load_ulong_relaxed(&type->tp_flags);
-#   else
-        flags = type->tp_flags;
-#   endif
+    flags = type->tp_flags;
 #endif
     return ((flags & feature) != 0);
 }

Include/refcount.h

@@ -19,9 +19,6 @@ immortal. The latter should be the only instances that require
 cleanup during runtime finalization.
 */
 
-#define _Py_STATICALLY_ALLOCATED_FLAG 4
-#define _Py_IMMORTAL_FLAGS 1
-
 #if SIZEOF_VOID_P > 4
 /*
 In 64+ bit systems, any object whose 32 bit reference count is >= 2**31

Misc/NEWS.d/next/Core_and_Builtins/2025-03-14-13-08-20.gh-issue-127266._tyfBp.rst

@@ -0,0 +1,6 @@
+In the free-threaded build, avoid data races caused by updating type slots
+or type flags after the type was initially created.  For those (typically
+rare) cases, use the stop-the-world mechanism.  Remove the use of atomics
+when reading or writing type flags.  The use of atomics is not sufficient to
+avoid races (since flags are sometimes read without a lock and without
+atomics) and are no longer required.