gh-127266: avoid data races when updating type slots (gh-131174)

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.

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 = type->tp_flags;
-        flags = _Py_atomic_load_ulong_relaxed(&type->tp_flags);
-#   else
-        flags = type->tp_flags;
-#   endif
 #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

Lib/test/test_opcache.py

@@ -576,6 +576,7 @@ class TestRacesDoNotCrash(TestBase):
     # Careful with these. Bigger numbers have a higher chance of catching bugs,
     # but you can also burn through a *ton* of type/dict/function versions:
     ITEMS = 1000
+    SMALL_ITEMS = 100
     LOOPS = 4
     WRITERS = 2
 
@@ -619,7 +620,7 @@ class C:
                 __getitem__ = lambda self, item: None
 
             items = []
-            for _ in range(self.ITEMS):
+            for _ in range(self.SMALL_ITEMS):
                 item = C()
                 items.append(item)
             return items
@@ -790,7 +791,7 @@ class C:
                 __getattribute__ = lambda self, name: None
 
             items = []
-            for _ in range(self.ITEMS):
+            for _ in range(self.SMALL_ITEMS):
                 item = C()
                 items.append(item)
             return items

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.

Objects/typeobject.c

@@ -48,7 +48,7 @@ class object "PyObject *" "&PyBaseObject_Type"
          & ((1 << MCACHE_SIZE_EXP) - 1))
 
 #define MCACHE_HASH_METHOD(type, name)                                  \
-    MCACHE_HASH(FT_ATOMIC_LOAD_UINT32_RELAXED((type)->tp_version_tag),   \
+    MCACHE_HASH(FT_ATOMIC_LOAD_UINT_RELAXED((type)->tp_version_tag),   \
                 ((Py_ssize_t)(name)) >> 3)
 #define MCACHE_CACHEABLE_NAME(name)                             \
         PyUnicode_CheckExact(name) &&                           \
@@ -60,11 +60,19 @@ class object "PyObject *" "&PyBaseObject_Type"
 
 #ifdef Py_GIL_DISABLED
 
-// There's a global lock for mutation of types.  This avoids having to take
+// There's a global lock for types that ensures that tp_version_tag and
-// additional locks while doing various subclass processing which may result
+// _spec_cache are correctly updated if the type is modified.  It also protects
-// in odd behaviors w.r.t. running with the GIL as the outer type lock could
+// tp_mro, tp_bases, and tp_base.  This avoids having to take additional locks
-// be released and reacquired during a subclass update if there's contention
+// while doing various subclass processing which may result in odd behaviors
-// on the subclass lock.
+// w.r.t. running with the GIL as the outer type lock could be released and
+// reacquired during a subclass update if there's contention on the subclass
+// lock.
+//
+// Note that this lock does not protect updates of other type slots or the
+// tp_flags member.  Instead, we either ensure those updates are done before
+// the type has been revealed to other threads or we only do those updates
+// while the stop-the-world mechanism is active.  The slots and flags are read
+// in many places without holding a lock and without atomics.
 #define TYPE_LOCK &PyInterpreterState_Get()->types.mutex
 #define BEGIN_TYPE_LOCK() Py_BEGIN_CRITICAL_SECTION_MUT(TYPE_LOCK)
 #define END_TYPE_LOCK() Py_END_CRITICAL_SECTION()
@@ -74,8 +82,59 @@ class object "PyObject *" "&PyBaseObject_Type"
 
 #define END_TYPE_DICT_LOCK() Py_END_CRITICAL_SECTION2()
 
+#ifdef Py_DEBUG
+// Return true if the world is currently stopped.
+static bool
+types_world_is_stopped(void)
+{
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    return interp->stoptheworld.world_stopped;
+}
+#endif
+
+// Checks that the type has not yet been revealed (exposed) to other
+// threads.  The _Py_TYPE_REVEALED_FLAG flag is set by type_new() and
+// PyType_FromMetaclass() to indicate that a newly initialized type might be
+// revealed.  We only have ob_flags on 64-bit platforms.
+#if SIZEOF_VOID_P > 4
+#define TYPE_IS_REVEALED(tp) ((((PyObject *)(tp))->ob_flags & _Py_TYPE_REVEALED_FLAG) != 0)
+#else
+#define TYPE_IS_REVEALED(tp) 0
+#endif
+
+#ifdef Py_DEBUG
 #define ASSERT_TYPE_LOCK_HELD() \
-    _Py_CRITICAL_SECTION_ASSERT_MUTEX_LOCKED(TYPE_LOCK)
+    if (!types_world_is_stopped()) { _Py_CRITICAL_SECTION_ASSERT_MUTEX_LOCKED(TYPE_LOCK); }
+
+// Checks if we can safely update type slots or tp_flags.
+#define ASSERT_WORLD_STOPPED_OR_NEW_TYPE(tp) \
+    assert(!TYPE_IS_REVEALED(tp) || types_world_is_stopped())
+
+#define ASSERT_NEW_TYPE_OR_LOCKED(tp) \
+    if (TYPE_IS_REVEALED(tp)) { ASSERT_TYPE_LOCK_HELD(); }
+#else
+#define ASSERT_TYPE_LOCK_HELD()
+#define ASSERT_WORLD_STOPPED_OR_NEW_TYPE(tp)
+#define ASSERT_NEW_TYPE_OR_LOCKED(tp)
+#endif
+
+static void
+types_stop_world(void)
+{
+    assert(!types_world_is_stopped());
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    _PyEval_StopTheWorld(interp);
+    assert(types_world_is_stopped());
+}
+
+static void
+types_start_world(void)
+{
+    assert(types_world_is_stopped());
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    _PyEval_StartTheWorld(interp);
+    assert(!types_world_is_stopped());
+}
 
 #else
 
@@ -84,6 +143,12 @@ class object "PyObject *" "&PyBaseObject_Type"
 #define BEGIN_TYPE_DICT_LOCK(d)
 #define END_TYPE_DICT_LOCK()
 #define ASSERT_TYPE_LOCK_HELD()
+#define TYPE_IS_REVEALED(tp) 0
+#define ASSERT_WORLD_STOPPED_OR_NEW_TYPE(tp)
+#define ASSERT_NEW_TYPE_OR_LOCKED(tp)
+#define types_world_is_stopped() 1
+#define types_stop_world()
+#define types_start_world()
 
 #endif
 
@@ -346,21 +411,14 @@ _PyStaticType_GetBuiltins(void)
 static void
 type_set_flags(PyTypeObject *tp, unsigned long flags)
 {
-    if (tp->tp_flags & Py_TPFLAGS_READY) {
+    ASSERT_WORLD_STOPPED_OR_NEW_TYPE(tp);
-        // It's possible the type object has been exposed to other threads
+    tp->tp_flags = flags;
-        // if it's been marked ready.  In that case, the type lock should be
-        // held when flags are modified.
-        ASSERT_TYPE_LOCK_HELD();
-    }
-    // Since PyType_HasFeature() reads the flags without holding the type
-    // lock, we need an atomic store here.
-    FT_ATOMIC_STORE_ULONG_RELAXED(tp->tp_flags, flags);
 }
 
 static void
 type_set_flags_with_mask(PyTypeObject *tp, unsigned long mask, unsigned long flags)
 {
-    ASSERT_TYPE_LOCK_HELD();
+    ASSERT_WORLD_STOPPED_OR_NEW_TYPE(tp);
     unsigned long new_flags = (tp->tp_flags & ~mask) | flags;
     type_set_flags(tp, new_flags);
 }
@@ -498,6 +556,7 @@ static inline void
 set_tp_bases(PyTypeObject *self, PyObject *bases, int initial)
 {
     assert(PyTuple_Check(bases));
+    ASSERT_NEW_TYPE_OR_LOCKED(self);
     if (self->tp_flags & _Py_TPFLAGS_STATIC_BUILTIN) {
         // XXX tp_bases can probably be statically allocated for each
         // static builtin type.
@@ -542,7 +601,7 @@ clear_tp_bases(PyTypeObject *self, int final)
 static inline PyObject *
 lookup_tp_mro(PyTypeObject *self)
 {
-    ASSERT_TYPE_LOCK_HELD();
+    ASSERT_NEW_TYPE_OR_LOCKED(self);
     return self->tp_mro;
 }
 
@@ -1027,7 +1086,6 @@ PyType_Unwatch(int watcher_id, PyObject* obj)
 static void
 set_version_unlocked(PyTypeObject *tp, unsigned int version)
 {
-    ASSERT_TYPE_LOCK_HELD();
     assert(version == 0 || (tp->tp_versions_used != _Py_ATTR_CACHE_UNUSED));
 #ifndef Py_GIL_DISABLED
     PyInterpreterState *interp = _PyInterpreterState_GET();
@@ -1075,7 +1133,7 @@ type_modified_unlocked(PyTypeObject *type)
        We don't assign new version tags eagerly, but only as
        needed.
      */
-    ASSERT_TYPE_LOCK_HELD();
+    ASSERT_NEW_TYPE_OR_LOCKED(type);
     if (type->tp_version_tag == 0) {
         return;
     }
@@ -1106,6 +1164,8 @@ type_modified_unlocked(PyTypeObject *type)
         while (bits) {
             assert(i < TYPE_MAX_WATCHERS);
             if (bits & 1) {
+                // Note that PyErr_FormatUnraisable is potentially re-entrant
+                // and the watcher callback might be too.
                 PyType_WatchCallback cb = interp->type_watchers[i];
                 if (cb && (cb(type) < 0)) {
                     PyErr_FormatUnraisable(
@@ -1245,14 +1305,6 @@ _PyType_LookupByVersion(unsigned int version)
 #endif
 }
 
-unsigned int
-_PyType_GetVersionForCurrentState(PyTypeObject *tp)
-{
-    return tp->tp_version_tag;
-}
-
-
-
 #define MAX_VERSIONS_PER_CLASS 1000
 #if _Py_ATTR_CACHE_UNUSED < MAX_VERSIONS_PER_CLASS
 #error "_Py_ATTR_CACHE_UNUSED must be bigger than max"
@@ -1586,10 +1638,13 @@ type_set_abstractmethods(PyObject *tp, PyObject *value, void *Py_UNUSED(closure)
 
     BEGIN_TYPE_LOCK();
     type_modified_unlocked(type);
+    types_stop_world();
     if (abstract)
         type_add_flags(type, Py_TPFLAGS_IS_ABSTRACT);
     else
         type_clear_flags(type, Py_TPFLAGS_IS_ABSTRACT);
+    types_start_world();
+    ASSERT_TYPE_LOCK_HELD();
     END_TYPE_LOCK();
 
     return 0;
@@ -1624,8 +1679,8 @@ type_get_mro(PyObject *tp, void *Py_UNUSED(closure))
     return mro;
 }
 
-static PyTypeObject *best_base(PyObject *);
+static PyTypeObject *find_best_base(PyObject *);
-static int mro_internal(PyTypeObject *, PyObject **);
+static int mro_internal(PyTypeObject *, int, PyObject **);
 static int type_is_subtype_base_chain(PyTypeObject *, PyTypeObject *);
 static int compatible_for_assignment(PyTypeObject *, PyTypeObject *, const char *);
 static int add_subclass(PyTypeObject*, PyTypeObject*);
@@ -1640,13 +1695,15 @@ static int update_subclasses(PyTypeObject *type, PyObject *attr_name,
 static int recurse_down_subclasses(PyTypeObject *type, PyObject *name,
                                    update_callback callback, void *data);
 
+// Compute tp_mro for this type and all of its subclasses.  This
+// is called after __bases__ is assigned to an existing type.
 static int
 mro_hierarchy(PyTypeObject *type, PyObject *temp)
 {
     ASSERT_TYPE_LOCK_HELD();
 
     PyObject *old_mro;
-    int res = mro_internal(type, &old_mro);
+    int res = mro_internal(type, 0, &old_mro);
     if (res <= 0) {
         /* error / reentrance */
         return res;
@@ -1708,9 +1765,9 @@ mro_hierarchy(PyTypeObject *type, PyObject *temp)
 }
 
 static int
-type_set_bases_unlocked(PyTypeObject *type, PyObject *new_bases)
+type_check_new_bases(PyTypeObject *type, PyObject *new_bases, PyTypeObject **best_base)
 {
-    // Check arguments
+    // Check arguments, this is re-entrant due to the PySys_Audit() call
     if (!check_set_special_type_attr(type, new_bases, "__bases__")) {
         return -1;
     }
@@ -1759,20 +1816,29 @@ type_set_bases_unlocked(PyTypeObject *type, PyObject *new_bases)
     }
 
     // Compute the new MRO and the new base class
-    PyTypeObject *new_base = best_base(new_bases);
+    *best_base = find_best_base(new_bases);
-    if (new_base == NULL)
+    if (*best_base == NULL)
         return -1;
 
-    if (!compatible_for_assignment(type->tp_base, new_base, "__bases__")) {
+    if (!compatible_for_assignment(type->tp_base, *best_base, "__bases__")) {
         return -1;
     }
 
+    return 0;
+}
+
+static int
+type_set_bases_unlocked(PyTypeObject *type, PyObject *new_bases, PyTypeObject *best_base)
+{
+    ASSERT_TYPE_LOCK_HELD();
+
+    Py_ssize_t n;
     PyObject *old_bases = lookup_tp_bases(type);
     assert(old_bases != NULL);
     PyTypeObject *old_base = type->tp_base;
 
     set_tp_bases(type, Py_NewRef(new_bases), 0);
-    type->tp_base = (PyTypeObject *)Py_NewRef(new_base);
+    type->tp_base = (PyTypeObject *)Py_NewRef(best_base);
 
     PyObject *temp = PyList_New(0);
     if (temp == NULL) {
@@ -1796,7 +1862,10 @@ type_set_bases_unlocked(PyTypeObject *type, PyObject *new_bases)
            add to all new_bases */
         remove_all_subclasses(type, old_bases);
         res = add_all_subclasses(type, new_bases);
+        types_stop_world();
         update_all_slots(type);
+        types_start_world();
+        ASSERT_TYPE_LOCK_HELD();
     }
     else {
         res = 0;
@@ -1827,13 +1896,13 @@ type_set_bases_unlocked(PyTypeObject *type, PyObject *new_bases)
 
   bail:
     if (lookup_tp_bases(type) == new_bases) {
-        assert(type->tp_base == new_base);
+        assert(type->tp_base == best_base);
 
         set_tp_bases(type, old_bases, 0);
         type->tp_base = old_base;
 
         Py_DECREF(new_bases);
-        Py_DECREF(new_base);
+        Py_DECREF(best_base);
     }
     else {
         Py_DECREF(old_bases);
@@ -1848,9 +1917,13 @@ static int
 type_set_bases(PyObject *tp, PyObject *new_bases, void *Py_UNUSED(closure))
 {
     PyTypeObject *type = PyTypeObject_CAST(tp);
+    PyTypeObject *best_base;
     int res;
     BEGIN_TYPE_LOCK();
-    res = type_set_bases_unlocked(type, new_bases);
+    res = type_check_new_bases(type, new_bases, &best_base);
+    if (res == 0) {
+        res = type_set_bases_unlocked(type, new_bases, best_base);
+    }
     END_TYPE_LOCK();
     return res;
 }
@@ -3092,6 +3165,7 @@ static PyObject *
 class_name(PyObject *cls)
 {
     PyObject *name;
+    // Note that this is potentially re-entrant.
     if (PyObject_GetOptionalAttr(cls, &_Py_ID(__name__), &name) == 0) {
         name = PyObject_Repr(cls);
     }
@@ -3428,9 +3502,13 @@ mro_invoke(PyTypeObject *type)
     const int custom = !Py_IS_TYPE(type, &PyType_Type);
 
     if (custom) {
+        // Custom mro() method on metaclass.  This is potentially re-entrant.
+        // We are called either from type_ready() or from type_set_bases().
         mro_result = call_method_noarg((PyObject *)type, &_Py_ID(mro));
     }
     else {
+        // In this case, the mro() method on the type object is being used and
+        // we know that these calls are not re-entrant.
         mro_result = mro_implementation_unlocked(type);
     }
     if (mro_result == NULL)
@@ -3478,7 +3556,7 @@ mro_invoke(PyTypeObject *type)
      - Returns -1 in case of an error.
 */
 static int
-mro_internal_unlocked(PyTypeObject *type, int initial, PyObject **p_old_mro)
+mro_internal(PyTypeObject *type, int initial, PyObject **p_old_mro)
 {
     ASSERT_TYPE_LOCK_HELD();
 
@@ -3526,21 +3604,11 @@ mro_internal_unlocked(PyTypeObject *type, int initial, PyObject **p_old_mro)
     return 1;
 }
 
-static int
-mro_internal(PyTypeObject *type, PyObject **p_old_mro)
-{
-    int res;
-    BEGIN_TYPE_LOCK();
-    res = mro_internal_unlocked(type, 0, p_old_mro);
-    END_TYPE_LOCK();
-    return res;
-}
-
 /* Calculate the best base amongst multiple base classes.
    This is the first one that's on the path to the "solid base". */
 
 static PyTypeObject *
-best_base(PyObject *bases)
+find_best_base(PyObject *bases)
 {
     Py_ssize_t i, n;
     PyTypeObject *base, *winner, *candidate;
@@ -3629,6 +3697,7 @@ static int update_slot(PyTypeObject *, PyObject *);
 static void fixup_slot_dispatchers(PyTypeObject *);
 static int type_new_set_names(PyTypeObject *);
 static int type_new_init_subclass(PyTypeObject *, PyObject *);
+static bool has_slotdef(PyObject *);
 
 /*
  * Helpers for  __dict__ descriptor.  We don't want to expose the dicts
@@ -3826,7 +3895,7 @@ type_init(PyObject *cls, PyObject *args, PyObject *kwds)
 unsigned long
 PyType_GetFlags(PyTypeObject *type)
 {
-    return FT_ATOMIC_LOAD_ULONG_RELAXED(type->tp_flags);
+    return type->tp_flags;
 }
 
 
@@ -4604,6 +4673,10 @@ type_new_impl(type_new_ctx *ctx)
     }
 
     assert(_PyType_CheckConsistency(type));
+#if defined(Py_GIL_DISABLED) && defined(Py_DEBUG) && SIZEOF_VOID_P > 4
+    // After this point, other threads can potentally use this type.
+    ((PyObject*)type)->ob_flags |= _Py_TYPE_REVEALED_FLAG;
+#endif
 
     return (PyObject *)type;
 
@@ -4666,7 +4739,7 @@ type_new_get_bases(type_new_ctx *ctx, PyObject **type)
     }
 
     /* Calculate best base, and check that all bases are type objects */
-    PyTypeObject *base = best_base(ctx->bases);
+    PyTypeObject *base = find_best_base(ctx->bases);
     if (base == NULL) {
         return -1;
     }
@@ -5081,12 +5154,12 @@ PyType_FromMetaclass(
     }
 
     /* Calculate best base, and check that all bases are type objects */
-    PyTypeObject *base = best_base(bases);  // borrowed ref
+    PyTypeObject *base = find_best_base(bases);  // borrowed ref
     if (base == NULL) {
         goto finally;
     }
-    // best_base should check Py_TPFLAGS_BASETYPE & raise a proper exception,
+    // find_best_base() should check Py_TPFLAGS_BASETYPE & raise a proper
-    // here we just check its work
+    // exception, here we just check its work
     assert(_PyType_HasFeature(base, Py_TPFLAGS_BASETYPE));
 
     /* Calculate sizes */
@@ -5317,6 +5390,10 @@ PyType_FromMetaclass(
     }
 
     assert(_PyType_CheckConsistency(type));
+#if defined(Py_GIL_DISABLED) && defined(Py_DEBUG) && SIZEOF_VOID_P > 4
+    // After this point, other threads can potentally use this type.
+    ((PyObject*)type)->ob_flags |= _Py_TYPE_REVEALED_FLAG;
+#endif
 
  finally:
     if (PyErr_Occurred()) {
@@ -5610,8 +5687,6 @@ PyObject_GetItemData(PyObject *obj)
 static PyObject *
 find_name_in_mro(PyTypeObject *type, PyObject *name, int *error)
 {
-    ASSERT_TYPE_LOCK_HELD();
-
     Py_hash_t hash = _PyObject_HashFast(name);
     if (hash == -1) {
         *error = -1;
@@ -5920,9 +5995,13 @@ _PyType_CacheGetItemForSpecialization(PyHeapTypeObject *ht, PyObject *descriptor
 void
 _PyType_SetFlags(PyTypeObject *self, unsigned long mask, unsigned long flags)
 {
-    BEGIN_TYPE_LOCK();
+    unsigned long new_flags = (self->tp_flags & ~mask) | flags;
-    type_set_flags_with_mask(self, mask, flags);
+    if (new_flags != self->tp_flags) {
-    END_TYPE_LOCK();
+        types_stop_world();
+        // can't use new_flags here since they could be out-of-date
+        self->tp_flags = (self->tp_flags & ~mask) | flags;
+        types_start_world();
+    }
 }
 
 int
@@ -5969,9 +6048,9 @@ set_flags_recursive(PyTypeObject *self, unsigned long mask, unsigned long flags)
 void
 _PyType_SetFlagsRecursive(PyTypeObject *self, unsigned long mask, unsigned long flags)
 {
-    BEGIN_TYPE_LOCK();
+    types_stop_world();
     set_flags_recursive(self, mask, flags);
-    END_TYPE_LOCK();
+    types_start_world();
 }
 
 /* This is similar to PyObject_GenericGetAttr(),
@@ -6085,6 +6164,8 @@ _Py_type_getattro(PyObject *tp, PyObject *name)
     return _Py_type_getattro_impl(type, name, NULL);
 }
 
+// Called by type_setattro().  Updates both the type dict and
+// the type versions.
 static int
 type_update_dict(PyTypeObject *type, PyDictObject *dict, PyObject *name,
                  PyObject *value, PyObject **old_value)
@@ -6114,10 +6195,6 @@ type_update_dict(PyTypeObject *type, PyDictObject *dict, PyObject *name,
         return -1;
     }
 
-    if (is_dunder_name(name)) {
-        return update_slot(type, name);
-    }
-
     return 0;
 }
 
@@ -6175,7 +6252,9 @@ type_setattro(PyObject *self, PyObject *name, PyObject *value)
 
     PyObject *dict = type->tp_dict;
     if (dict == NULL) {
-        // We don't just do PyType_Ready because we could already be readying
+        // This is an unlikely case.  PyType_Ready has not yet been done and
+        // we need to initialize tp_dict.  We don't just do PyType_Ready
+        // because we could already be readying.
         BEGIN_TYPE_LOCK();
         dict = type->tp_dict;
         if (dict == NULL) {
@@ -6191,6 +6270,15 @@ type_setattro(PyObject *self, PyObject *name, PyObject *value)
     BEGIN_TYPE_DICT_LOCK(dict);
     res = type_update_dict(type, (PyDictObject *)dict, name, value, &old_value);
     assert(_PyType_CheckConsistency(type));
+    if (res == 0) {
+        if (is_dunder_name(name) && has_slotdef(name)) {
+            // The name corresponds to a type slot.
+            types_stop_world();
+            res = update_slot(type, name);
+            types_start_world();
+            ASSERT_TYPE_LOCK_HELD();
+        }
+    }
     END_TYPE_DICT_LOCK();
 
 done:
@@ -7120,15 +7208,10 @@ object_set_class(PyObject *self, PyObject *value, void *closure)
         return -1;
     }
 
-#ifdef Py_GIL_DISABLED
+    types_stop_world();
-    PyInterpreterState *interp = _PyInterpreterState_GET();
-    _PyEval_StopTheWorld(interp);
-#endif
     PyTypeObject *oldto = Py_TYPE(self);
     int res = object_set_class_world_stopped(self, newto);
-#ifdef Py_GIL_DISABLED
+    types_start_world();
-    _PyEval_StartTheWorld(interp);
-#endif
     if (res == 0) {
         if (oldto->tp_flags & Py_TPFLAGS_HEAPTYPE) {
             Py_DECREF(oldto);
@@ -8536,7 +8619,7 @@ type_ready_mro(PyTypeObject *type, int initial)
     }
 
     /* Calculate method resolution order */
-    if (mro_internal_unlocked(type, initial, NULL) < 0) {
+    if (mro_internal(type, initial, NULL) < 0) {
         return -1;
     }
     PyObject *mro = lookup_tp_mro(type);
@@ -11059,12 +11142,21 @@ resolve_slotdups(PyTypeObject *type, PyObject *name)
 {
     /* XXX Maybe this could be optimized more -- but is it worth it? */
 
+#ifdef Py_GIL_DISABLED
+    pytype_slotdef *ptrs[MAX_EQUIV];
+    pytype_slotdef **pp = ptrs;
+    /* Collect all slotdefs that match name into ptrs. */
+    for (pytype_slotdef *p = slotdefs; p->name_strobj; p++) {
+        if (p->name_strobj == name)
+            *pp++ = p;
+    }
+    *pp = NULL;
+#else
     /* pname and ptrs act as a little cache */
     PyInterpreterState *interp = _PyInterpreterState_GET();
 #define pname _Py_INTERP_CACHED_OBJECT(interp, type_slots_pname)
 #define ptrs _Py_INTERP_CACHED_OBJECT(interp, type_slots_ptrs)
     pytype_slotdef *p, **pp;
-    void **res, **ptr;
 
     if (pname != name) {
         /* Collect all slotdefs that match name into ptrs. */
@@ -11076,10 +11168,12 @@ resolve_slotdups(PyTypeObject *type, PyObject *name)
         }
         *pp = NULL;
     }
+#endif
 
     /* Look in all slots of the type matching the name. If exactly one of these
        has a filled-in slot, return a pointer to that slot.
        Otherwise, return NULL. */
+    void **res, **ptr;
     res = NULL;
     for (pp = ptrs; *pp; pp++) {
         ptr = slotptr(type, (*pp)->offset);
@@ -11089,11 +11183,25 @@ resolve_slotdups(PyTypeObject *type, PyObject *name)
             return NULL;
         res = ptr;
     }
-    return res;
+#ifndef Py_GIL_DISABLED
 #undef pname
 #undef ptrs
+#endif
+    return res;
 }
 
+// Return true if "name" corresponds to at least one slot definition.  This is
+// a more accurate but more expensive test compared to is_dunder_name().
+static bool
+has_slotdef(PyObject *name)
+{
+    for (pytype_slotdef *p = slotdefs; p->name_strobj; p++) {
+        if (p->name_strobj == name) {
+            return true;
+        }
+    }
+    return false;
+}
 
 /* Common code for update_slots_callback() and fixup_slot_dispatchers().
  *
@@ -11152,7 +11260,7 @@ resolve_slotdups(PyTypeObject *type, PyObject *name)
 static pytype_slotdef *
 update_one_slot(PyTypeObject *type, pytype_slotdef *p)
 {
-    ASSERT_TYPE_LOCK_HELD();
+    ASSERT_WORLD_STOPPED_OR_NEW_TYPE(type);
 
     PyObject *descr;
     PyWrapperDescrObject *d;
@@ -11275,7 +11383,7 @@ update_one_slot(PyTypeObject *type, pytype_slotdef *p)
 static int
 update_slots_callback(PyTypeObject *type, void *data)
 {
-    ASSERT_TYPE_LOCK_HELD();
+    ASSERT_WORLD_STOPPED_OR_NEW_TYPE(type);
 
     pytype_slotdef **pp = (pytype_slotdef **)data;
     for (; *pp; pp++) {
@@ -11293,7 +11401,7 @@ update_slot(PyTypeObject *type, PyObject *name)
     pytype_slotdef **pp;
     int offset;
 
-    ASSERT_TYPE_LOCK_HELD();
+    assert(types_world_is_stopped());
     assert(PyUnicode_CheckExact(name));
     assert(PyUnicode_CHECK_INTERNED(name));
 
@@ -11327,33 +11435,27 @@ update_slot(PyTypeObject *type, PyObject *name)
 static void
 fixup_slot_dispatchers(PyTypeObject *type)
 {
-    // This lock isn't strictly necessary because the type has not been
-    // exposed to anyone else yet, but update_ont_slot calls find_name_in_mro
-    // where we'd like to assert that the type is locked.
-    BEGIN_TYPE_LOCK();
-
     assert(!PyErr_Occurred());
     for (pytype_slotdef *p = slotdefs; p->name; ) {
         p = update_one_slot(type, p);
     }
-
-    END_TYPE_LOCK();
 }
 
+// Called when __bases__ is re-assigned.
 static void
 update_all_slots(PyTypeObject* type)
 {
     pytype_slotdef *p;
 
-    ASSERT_TYPE_LOCK_HELD();
+    assert(types_world_is_stopped());
-
-    /* Clear the VALID_VERSION flag of 'type' and all its subclasses. */
-    type_modified_unlocked(type);
 
     for (p = slotdefs; p->name; p++) {
         /* update_slot returns int but can't actually fail */
         update_slot(type, p->name_strobj);
     }
+
+    /* Clear the VALID_VERSION flag of 'type' and all its subclasses. */
+    type_modified_unlocked(type);
 }
 
 
@@ -11625,7 +11727,10 @@ PyType_Freeze(PyTypeObject *type)
     }
 
     BEGIN_TYPE_LOCK();
+    types_stop_world();
     type_add_flags(type, Py_TPFLAGS_IMMUTABLETYPE);
+    types_start_world();
+    ASSERT_TYPE_LOCK_HELD();
     type_modified_unlocked(type);
     END_TYPE_LOCK();
 

Python/ceval.c

@@ -138,6 +138,19 @@
 #endif
 
 
+static void
+check_invalid_reentrancy(void)
+{
+#if defined(Py_DEBUG) && defined(Py_GIL_DISABLED)
+    // In the free-threaded build, the interpreter must not be re-entered if
+    // the world-is-stopped.  If so, that's a bug somewhere (quite likely in
+    // the painfully complex typeobject code).
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    assert(!interp->stoptheworld.world_stopped);
+#endif
+}
+
+
 #ifdef Py_DEBUG
 static void
 dump_item(_PyStackRef item)
@@ -995,6 +1008,7 @@ PyObject* _Py_HOT_FUNCTION DONT_SLP_VECTORIZE
 _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int throwflag)
 {
     _Py_EnsureTstateNotNULL(tstate);
+    check_invalid_reentrancy();
     CALL_STAT_INC(pyeval_calls);
 
 #if USE_COMPUTED_GOTOS && !Py_TAIL_CALL_INTERP

Tools/tsan/suppressions_free_threading.txt

@@ -12,15 +12,12 @@
 
 # These warnings trigger directly in a CPython function.
 
-race_top:assign_version_tag
-race_top:_Py_slot_tp_getattr_hook
 race_top:dump_traceback
 race_top:fatal_error
 race_top:_PyFrame_GetCode
 race_top:_PyFrame_Initialize
 race_top:_PyObject_TryGetInstanceAttribute
 race_top:PyUnstable_InterpreterFrame_GetLine
-race_top:type_modified_unlocked
 race_top:write_thread_id
 
 # gh-129068: race on shared range iterators (test_free_threading.test_zip.ZipThreading.test_threading)
@@ -29,9 +26,6 @@ race_top:rangeiter_next
 # gh-129748: test.test_free_threading.test_slots.TestSlots.test_object
 race_top:mi_block_set_nextx
 
-# gh-127266: type slot updates are not thread-safe (test_opcache.test_load_attr_method_lazy_dict)
-race_top:update_one_slot
-
 # https://gist.github.com/mpage/6962e8870606cfc960e159b407a0cb40
 thread:pthread_create