gh-102304: Move _Py_RefTotal to _PyRuntimeState (gh-102543)

The essentially eliminates the global variable, with the associated benefits. This is also a precursor to isolating this bit of state to PyInterpreterState.

Folks that currently read _Py_RefTotal directly would have to start using _Py_GetGlobalRefTotal() instead.

#102304

Author: ericsnowcurrently

Include/cpython/object.h

@@ -11,7 +11,10 @@ PyAPI_FUNC(void) _Py_ForgetReference(PyObject *);
 #endif
 
 #ifdef Py_REF_DEBUG
-PyAPI_FUNC(Py_ssize_t) _Py_GetRefTotal(void);
+/* These are useful as debugging aids when chasing down refleaks. */
+PyAPI_FUNC(Py_ssize_t) _Py_GetGlobalRefTotal(void);
+#  define _Py_GetRefTotal() _Py_GetGlobalRefTotal()
+PyAPI_FUNC(Py_ssize_t) _Py_GetLegacyRefTotal(void);
 #endif
 
 

Include/internal/pycore_object.h

@@ -46,7 +46,8 @@ PyAPI_DATA(Py_ssize_t) _Py_RefTotal;
 extern void _Py_AddRefTotal(Py_ssize_t);
 extern void _Py_IncRefTotal(void);
 extern void _Py_DecRefTotal(void);
-#  define _Py_DEC_REFTOTAL() _Py_RefTotal--
+
+#  define _Py_DEC_REFTOTAL() _PyRuntime.object_state.reftotal--
 #endif
 
 // Increment reference count by n
@@ -225,6 +226,7 @@ static inline void _PyObject_GC_UNTRACK(
 #endif
 
 #ifdef Py_REF_DEBUG
+extern void _Py_FinalizeRefTotal(_PyRuntimeState *);
 extern void _PyDebug_PrintTotalRefs(void);
 #endif
 

Include/internal/pycore_object_state.h

@@ -0,0 +1,23 @@
+#ifndef Py_INTERNAL_OBJECT_STATE_H
+#define Py_INTERNAL_OBJECT_STATE_H
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef Py_BUILD_CORE
+#  error "this header requires Py_BUILD_CORE define"
+#endif
+
+struct _py_object_runtime_state {
+#ifdef Py_REF_DEBUG
+    Py_ssize_t reftotal;
+#else
+    int _not_used;
+#endif
+};
+
+
+#ifdef __cplusplus
+}
+#endif
+#endif /* !Py_INTERNAL_OBJECT_STATE_H */

Include/internal/pycore_runtime.h

@@ -15,6 +15,7 @@ extern "C" {
 #include "pycore_global_objects.h"  // struct _Py_global_objects
 #include "pycore_import.h"          // struct _import_runtime_state
 #include "pycore_interp.h"          // PyInterpreterState
+#include "pycore_object_state.h"    // struct _py_object_runtime_state
 #include "pycore_parser.h"          // struct _parser_runtime_state
 #include "pycore_pymem.h"           // struct _pymem_allocators
 #include "pycore_pyhash.h"          // struct pyhash_runtime_state
@@ -150,6 +151,7 @@ typedef struct pyruntimestate {
     void *open_code_userdata;
     _Py_AuditHookEntry *audit_hook_head;
 
+    struct _py_object_runtime_state object_state;
     struct _Py_float_runtime_state float_state;
     struct _Py_unicode_runtime_state unicode_state;
 

Include/object.h

@@ -494,14 +494,9 @@ you can count such references to the type object.)
 extern Py_ssize_t _Py_RefTotal;
 #    define _Py_INC_REFTOTAL() _Py_RefTotal++
 #    define _Py_DEC_REFTOTAL() _Py_RefTotal--
-#  elif defined(Py_BUILD_CORE) && !defined(Py_BUILD_CORE_MODULE)
-extern void _Py_IncRefTotal(void);
-extern void _Py_DecRefTotal(void);
-#    define _Py_INC_REFTOTAL() _Py_IncRefTotal()
-#    define _Py_DEC_REFTOTAL() _Py_DecRefTotal()
 #  elif !defined(Py_LIMITED_API) || Py_LIMITED_API+0 > 0x030C0000
-extern void _Py_IncRefTotal_DO_NOT_USE_THIS(void);
-extern void _Py_DecRefTotal_DO_NOT_USE_THIS(void);
+PyAPI_FUNC(void) _Py_IncRefTotal_DO_NOT_USE_THIS(void);
+PyAPI_FUNC(void) _Py_DecRefTotal_DO_NOT_USE_THIS(void);
 #    define _Py_INC_REFTOTAL() _Py_IncRefTotal_DO_NOT_USE_THIS()
 #    define _Py_DEC_REFTOTAL() _Py_DecRefTotal_DO_NOT_USE_THIS()
 #  endif

Makefile.pre.in

@@ -1699,6 +1699,7 @@ PYTHON_HEADERS= \
 		$(srcdir)/Include/internal/pycore_moduleobject.h \
 		$(srcdir)/Include/internal/pycore_namespace.h \
 		$(srcdir)/Include/internal/pycore_object.h \
+		$(srcdir)/Include/internal/pycore_object_state.h \
 		$(srcdir)/Include/internal/pycore_obmalloc.h \
 		$(srcdir)/Include/internal/pycore_obmalloc_init.h \
 		$(srcdir)/Include/internal/pycore_pathconfig.h \

Objects/object.c

@@ -54,37 +54,71 @@ _PyObject_CheckConsistency(PyObject *op, int check_content)
 
 
 #ifdef Py_REF_DEBUG
+/* We keep the legacy symbol around for backward compatibility. */
 Py_ssize_t _Py_RefTotal;
 
+static inline Py_ssize_t
+get_legacy_reftotal(void)
+{
+    return _Py_RefTotal;
+}
+#endif
+
+#ifdef Py_REF_DEBUG
+
+#  define REFTOTAL(runtime) \
+    (runtime)->object_state.reftotal
+
+static inline void
+reftotal_increment(_PyRuntimeState *runtime)
+{
+    REFTOTAL(runtime)++;
+}
+
 static inline void
-reftotal_increment(void)
+reftotal_decrement(_PyRuntimeState *runtime)
 {
-    _Py_RefTotal++;
+    REFTOTAL(runtime)--;
 }
 
 static inline void
-reftotal_decrement(void)
+reftotal_add(_PyRuntimeState *runtime, Py_ssize_t n)
 {
-    _Py_RefTotal--;
+    REFTOTAL(runtime) += n;
 }
 
+static inline Py_ssize_t get_global_reftotal(_PyRuntimeState *);
+
+/* We preserve the number of refs leaked during runtime finalization,
+   so they can be reported if the runtime is initialized again. */
+// XXX We don't lose any information by dropping this,
+// so we should consider doing so.
+static Py_ssize_t last_final_reftotal = 0;
+
 void
-_Py_AddRefTotal(Py_ssize_t n)
+_Py_FinalizeRefTotal(_PyRuntimeState *runtime)
 {
-    _Py_RefTotal += n;
+    last_final_reftotal = get_global_reftotal(runtime);
+    REFTOTAL(runtime) = 0;
 }
 
-Py_ssize_t
-_Py_GetRefTotal(void)
+static inline Py_ssize_t
+get_global_reftotal(_PyRuntimeState *runtime)
 {
-    return _Py_RefTotal;
+    /* For an update from _Py_RefTotal first. */
+    Py_ssize_t legacy = get_legacy_reftotal();
+    return REFTOTAL(runtime) + legacy + last_final_reftotal;
 }
 
+#undef REFTOTAL
+
 void
 _PyDebug_PrintTotalRefs(void) {
+    _PyRuntimeState *runtime = &_PyRuntime;
     fprintf(stderr,
             "[%zd refs, %zd blocks]\n",
-            _Py_GetRefTotal(), _Py_GetAllocatedBlocks());
+            get_global_reftotal(runtime), _Py_GetAllocatedBlocks());
+    /* It may be helpful to also print the "legacy" reftotal separately. */
 }
 #endif /* Py_REF_DEBUG */
 
@@ -139,30 +173,50 @@ _Py_NegativeRefcount(const char *filename, int lineno, PyObject *op)
                            filename, lineno, __func__);
 }
 
-/* This is exposed strictly for use in Py_INCREF(). */
-PyAPI_FUNC(void)
+/* This is used strictly by Py_INCREF(). */
+void
 _Py_IncRefTotal_DO_NOT_USE_THIS(void)
 {
-    reftotal_increment();
+    reftotal_increment(&_PyRuntime);
 }
 
-/* This is exposed strictly for use in Py_DECREF(). */
-PyAPI_FUNC(void)
+/* This is used strictly by Py_DECREF(). */
+void
 _Py_DecRefTotal_DO_NOT_USE_THIS(void)
 {
-    reftotal_decrement();
+    reftotal_decrement(&_PyRuntime);
 }
 
 void
 _Py_IncRefTotal(void)
 {
-    reftotal_increment();
+    reftotal_increment(&_PyRuntime);
 }
 
 void
 _Py_DecRefTotal(void)
 {
-    reftotal_decrement();
+    reftotal_decrement(&_PyRuntime);
+}
+
+void
+_Py_AddRefTotal(Py_ssize_t n)
+{
+    reftotal_add(&_PyRuntime, n);
+}
+
+/* This includes the legacy total
+   and any carried over from the last runtime init/fini cycle. */
+Py_ssize_t
+_Py_GetGlobalRefTotal(void)
+{
+    return get_global_reftotal(&_PyRuntime);
+}
+
+Py_ssize_t
+_Py_GetLegacyRefTotal(void)
+{
+    return get_legacy_reftotal();
 }
 
 #endif /* Py_REF_DEBUG */
@@ -182,21 +236,18 @@ Py_DecRef(PyObject *o)
 void
 _Py_IncRef(PyObject *o)
 {
-#ifdef Py_REF_DEBUG
-    reftotal_increment();
-#endif
     Py_INCREF(o);
 }
 
 void
 _Py_DecRef(PyObject *o)
 {
-#ifdef Py_REF_DEBUG
-    reftotal_decrement();
-#endif
     Py_DECREF(o);
 }
 
+
+/**************************************/
+
 PyObject *
 PyObject_Init(PyObject *op, PyTypeObject *tp)
 {
@@ -2077,7 +2128,7 @@ void
 _Py_NewReference(PyObject *op)
 {
 #ifdef Py_REF_DEBUG
-    reftotal_increment();
+    reftotal_increment(&_PyRuntime);
 #endif
     new_reference(op);
 }

PCbuild/pythoncore.vcxproj

@@ -239,6 +239,7 @@
     <ClInclude Include="..\Include\internal\pycore_moduleobject.h" />
     <ClInclude Include="..\Include\internal\pycore_namespace.h" />
     <ClInclude Include="..\Include\internal\pycore_object.h" />
+    <ClInclude Include="..\Include\internal\pycore_object_state.h" />
     <ClInclude Include="..\Include\internal\pycore_obmalloc.h" />
     <ClInclude Include="..\Include\internal\pycore_obmalloc_init.h" />
     <ClInclude Include="..\Include\internal\pycore_pathconfig.h" />

PCbuild/pythoncore.vcxproj.filters

@@ -621,6 +621,9 @@
     <ClInclude Include="..\Include\internal\pycore_object.h">
       <Filter>Include\internal</Filter>
     </ClInclude>
+    <ClInclude Include="..\Include\internal\pycore_object_state.h">
+      <Filter>Include\internal</Filter>
+    </ClInclude>
     <ClInclude Include="..\Include\internal\pycore_obmalloc.h">
       <Filter>Include\internal</Filter>
     </ClInclude>

Python/pylifecycle.c

@@ -1930,6 +1930,7 @@ Py_FinalizeEx(void)
     if (show_ref_count) {
         _PyDebug_PrintTotalRefs();
     }
+    _Py_FinalizeRefTotal(runtime);
 #endif
 
 #ifdef Py_TRACE_REFS