Use mutex rather than critical sections.

For TYPE_LOCK, replace critical sections with a simple mutex.  A number
of changes were needed to avoid deadlocking on reentrancy (trying to
re-acquire the mutex). Split _PyType_LookupRefAndVersion() into a locked
and unlocked version.  Remove atomic operations where they are not
required. Remove some cases of TYPE_LOCK being held that is not
required.

Author: nascheme

Include/internal/pycore_typeobject.h

@@ -144,6 +144,8 @@ struct types_state {
         managed_static_type_state initialized[_Py_MAX_MANAGED_STATIC_EXT_TYPES];
     } for_extensions;
     PyMutex mutex;
+    // used to check correct usage of the above mutex
+    unsigned long long mutex_tid;
 
     // Borrowed references to type objects whose
     // tp_version_tag % TYPE_VERSION_CACHE_SIZE

Include/object.h

@@ -576,8 +576,8 @@ given type object has a specified feature.
 /* Objects behave like an unbound method */
 #define Py_TPFLAGS_METHOD_DESCRIPTOR (1UL << 17)
 
-/* Unused. Legacy flag */
-#define Py_TPFLAGS_VALID_VERSION_TAG  (1UL << 19)
+/* Type structure is potentially exposed (revealed) to other threads */
+#define Py_TPFLAGS_EXPOSED  (1UL << 19)
 
 /* Type is abstract and cannot be instantiated */
 #define Py_TPFLAGS_IS_ABSTRACT (1UL << 20)