gh-111926: Avoid locking in PyType_IsSubtype

Include/internal/pycore_pyatomic_ft_wrappers.h

@@ -20,18 +20,23 @@ extern "C" {
 #endif
 
 #ifdef Py_GIL_DISABLED
+#define FT_ATOMIC_LOAD_PTR(value) _Py_atomic_load_ptr(&value)
 #define FT_ATOMIC_LOAD_SSIZE(value) _Py_atomic_load_ssize(&value)
 #define FT_ATOMIC_LOAD_SSIZE_RELAXED(value) \
     _Py_atomic_load_ssize_relaxed(&value)
+#define FT_ATOMIC_STORE_PTR(value, new_value) \
+    _Py_atomic_store_ptr(&value, new_value)
 #define FT_ATOMIC_STORE_PTR_RELAXED(value, new_value) \
     _Py_atomic_store_ptr_relaxed(&value, new_value)
 #define FT_ATOMIC_STORE_PTR_RELEASE(value, new_value) \
     _Py_atomic_store_ptr_release(&value, new_value)
 #define FT_ATOMIC_STORE_SSIZE_RELAXED(value, new_value) \
     _Py_atomic_store_ssize_relaxed(&value, new_value)
 #else
+#define FT_ATOMIC_LOAD_PTR(value) value
 #define FT_ATOMIC_LOAD_SSIZE(value) value
 #define FT_ATOMIC_LOAD_SSIZE_RELAXED(value) value
+#define FT_ATOMIC_STORE_PTR(value, new_value) value = new_value
 #define FT_ATOMIC_STORE_PTR_RELAXED(value, new_value) value = new_value
 #define FT_ATOMIC_STORE_PTR_RELEASE(value, new_value) value = new_value
 #define FT_ATOMIC_STORE_SSIZE_RELAXED(value, new_value) value = new_value

Objects/tupleobject.c

@@ -1161,6 +1161,13 @@ static inline int
 maybe_freelist_push(PyTupleObject *op)
 {
 #ifdef WITH_FREELISTS
+#ifdef Py_GIL_DISABLED
+    if (_PyObject_GC_IS_SHARED_INLINE((PyObject *) op)) {
+        // There may still be threads that are concurrently reading from the
+        // tuple.
+        return 0;
+    }
+#endif
     struct _Py_object_freelists *freelists = _Py_object_freelists_GET();
     if (Py_SIZE(op) == 0) {
         return 0;

Objects/typeobject.c

@@ -2,6 +2,7 @@
 
 #include "Python.h"
 #include "pycore_abstract.h"      // _PySequence_IterSearch()
+#include "pycore_pyatomic_ft_wrappers.h"
 #include "pycore_call.h"          // _PyObject_VectorcallTstate()
 #include "pycore_code.h"          // CO_FAST_FREE
 #include "pycore_dict.h"          // _PyDict_KeysSize()
@@ -366,15 +367,14 @@ clear_tp_bases(PyTypeObject *self)
 static inline PyObject *
 lookup_tp_mro(PyTypeObject *self)
 {
-    ASSERT_TYPE_LOCK_HELD();
-    return self->tp_mro;
+    return FT_ATOMIC_LOAD_PTR(self->tp_mro);
 }
 
 PyObject *
 _PyType_GetMRO(PyTypeObject *self)
 {
+    PyObject *mro = lookup_tp_mro(self);
 #ifdef Py_GIL_DISABLED
-    PyObject *mro = _Py_atomic_load_ptr_relaxed(&self->tp_mro);
     if (mro == NULL) {
         return NULL;
     }
@@ -388,7 +388,7 @@ _PyType_GetMRO(PyTypeObject *self)
     END_TYPE_LOCK()
     return mro;
 #else
-    return Py_XNewRef(lookup_tp_mro(self));
+    return Py_XNewRef(mro);
 #endif
 }
 
@@ -404,7 +404,13 @@ set_tp_mro(PyTypeObject *self, PyObject *mro)
         /* Other checks are done via set_tp_bases. */
         _Py_SetImmortal(mro);
     }
-    self->tp_mro = mro;
+#ifdef Py_GIL_DISABLED
+    if (self->tp_mro != NULL) {
+        // Allow concurrent reads from PyType_IsSubtype
+        _PyObject_GC_SET_SHARED_INLINE(self->tp_mro);
+    }
+#endif
+    FT_ATOMIC_STORE_PTR(self->tp_mro, mro);
 }
 
 static inline void
@@ -2341,14 +2347,7 @@ is_subtype_with_mro(PyObject *a_mro, PyTypeObject *a, PyTypeObject *b)
 int
 PyType_IsSubtype(PyTypeObject *a, PyTypeObject *b)
 {
-#ifdef Py_GIL_DISABLED
-    PyObject *mro = _PyType_GetMRO(a);
-    int res = is_subtype_with_mro(mro, a, b);
-    Py_XDECREF(mro);
-    return res;
-#else
     return is_subtype_with_mro(lookup_tp_mro(a), a, b);
-#endif
 }
 
 /* Routines to do a method lookup in the type without looking in the