gh-107080: Fix Py_TRACE_REFS Crashes Under Isolated Subinterpreters (gh-107567)

The linked list of objects was a global variable, which broke isolation between interpreters, causing crashes. To solve this, we've moved the linked list to each interpreter.

Author: ericsnowcurrently

Include/internal/pycore_object.h

@@ -292,8 +292,8 @@ extern void _PyDebug_PrintTotalRefs(void);
 
 #ifdef Py_TRACE_REFS
 extern void _Py_AddToAllObjects(PyObject *op, int force);
-extern void _Py_PrintReferences(FILE *);
-extern void _Py_PrintReferenceAddresses(FILE *);
+extern void _Py_PrintReferences(PyInterpreterState *, FILE *);
+extern void _Py_PrintReferenceAddresses(PyInterpreterState *, FILE *);
 #endif
 
 

Include/internal/pycore_object_state.h

@@ -11,17 +11,22 @@ extern "C" {
 struct _py_object_runtime_state {
 #ifdef Py_REF_DEBUG
     Py_ssize_t interpreter_leaks;
-#else
-    int _not_used;
 #endif
+    int _not_used;
 };
 
 struct _py_object_state {
 #ifdef Py_REF_DEBUG
     Py_ssize_t reftotal;
-#else
-    int _not_used;
 #endif
+#ifdef Py_TRACE_REFS
+    /* Head of circular doubly-linked list of all objects.  These are linked
+     * together via the _ob_prev and _ob_next members of a PyObject, which
+     * exist only in a Py_TRACE_REFS build.
+     */
+    PyObject refchain;
+#endif
+    int _not_used;
 };
 
 

Include/internal/pycore_runtime_init.h

@@ -157,6 +157,7 @@ extern PyTypeObject _PyExc_MemoryError;
                 { .threshold = 10, }, \
             }, \
         }, \
+        .object_state = _py_object_state_INIT(INTERP), \
         .dtoa = _dtoa_state_INIT(&(INTERP)), \
         .dict_state = _dict_state_INIT, \
         .func_state = { \
@@ -186,6 +187,16 @@ extern PyTypeObject _PyExc_MemoryError;
         .context_ver = 1, \
     }
 
+#ifdef Py_TRACE_REFS
+# define _py_object_state_INIT(INTERP) \
+    { \
+        .refchain = {&INTERP.object_state.refchain, &INTERP.object_state.refchain}, \
+    }
+#else
+# define _py_object_state_INIT(INTERP) \
+    { 0 }
+#endif
+
 
 // global objects
 

Misc/NEWS.d/next/Core and Builtins/2023-08-02-12-24-51.gh-issue-107080.PNolFU.rst

@@ -0,0 +1,4 @@
+Trace refs builds (``--with-trace-refs``) were crashing when used with
+isolated subinterpreters.  The problematic global state has been isolated to
+each interpreter.  Other fixing the crashes, this change does not affect
+users.

Objects/object.c

@@ -159,11 +159,8 @@ _PyDebug_PrintTotalRefs(void) {
    Do not call them otherwise, they do not initialize the object! */
 
 #ifdef Py_TRACE_REFS
-/* Head of circular doubly-linked list of all objects.  These are linked
- * together via the _ob_prev and _ob_next members of a PyObject, which
- * exist only in a Py_TRACE_REFS build.
- */
-static PyObject refchain = {&refchain, &refchain};
+
+#define REFCHAIN(interp) &interp->object_state.refchain
 
 /* Insert op at the front of the list of all objects.  If force is true,
  * op is added even if _ob_prev and _ob_next are non-NULL already.  If
@@ -188,10 +185,11 @@ _Py_AddToAllObjects(PyObject *op, int force)
     }
 #endif
     if (force || op->_ob_prev == NULL) {
-        op->_ob_next = refchain._ob_next;
-        op->_ob_prev = &refchain;
-        refchain._ob_next->_ob_prev = op;
-        refchain._ob_next = op;
+        PyObject *refchain = REFCHAIN(_PyInterpreterState_GET());
+        op->_ob_next = refchain->_ob_next;
+        op->_ob_prev = refchain;
+        refchain->_ob_next->_ob_prev = op;
+        refchain->_ob_next = op;
     }
 }
 #endif  /* Py_TRACE_REFS */
@@ -2229,20 +2227,21 @@ _Py_ForgetReference(PyObject *op)
         _PyObject_ASSERT_FAILED_MSG(op, "negative refcnt");
     }
 
-    if (op == &refchain ||
+    PyObject *refchain = REFCHAIN(_PyInterpreterState_GET());
+    if (op == refchain ||
         op->_ob_prev->_ob_next != op || op->_ob_next->_ob_prev != op)
     {
         _PyObject_ASSERT_FAILED_MSG(op, "invalid object chain");
     }
 
 #ifdef SLOW_UNREF_CHECK
     PyObject *p;
-    for (p = refchain._ob_next; p != &refchain; p = p->_ob_next) {
+    for (p = refchain->_ob_next; p != refchain; p = p->_ob_next) {
         if (p == op) {
             break;
         }
     }
-    if (p == &refchain) {
+    if (p == refchain) {
         /* Not found */
         _PyObject_ASSERT_FAILED_MSG(op,
                                     "object not found in the objects list");
@@ -2258,11 +2257,15 @@ _Py_ForgetReference(PyObject *op)
  * interpreter must be in a healthy state.
  */
 void
-_Py_PrintReferences(FILE *fp)
+_Py_PrintReferences(PyInterpreterState *interp, FILE *fp)
 {
     PyObject *op;
+    if (interp == NULL) {
+        interp = _PyInterpreterState_Main();
+    }
     fprintf(fp, "Remaining objects:\n");
-    for (op = refchain._ob_next; op != &refchain; op = op->_ob_next) {
+    PyObject *refchain = REFCHAIN(interp);
+    for (op = refchain->_ob_next; op != refchain; op = op->_ob_next) {
         fprintf(fp, "%p [%zd] ", (void *)op, Py_REFCNT(op));
         if (PyObject_Print(op, fp, 0) != 0) {
             PyErr_Clear();
@@ -2274,34 +2277,42 @@ _Py_PrintReferences(FILE *fp)
 /* Print the addresses of all live objects.  Unlike _Py_PrintReferences, this
  * doesn't make any calls to the Python C API, so is always safe to call.
  */
+// XXX This function is not safe to use if the interpreter has been
+// freed or is in an unhealthy state (e.g. late in finalization).
+// The call in Py_FinalizeEx() is okay since the main interpreter
+// is statically allocated.
 void
-_Py_PrintReferenceAddresses(FILE *fp)
+_Py_PrintReferenceAddresses(PyInterpreterState *interp, FILE *fp)
 {
     PyObject *op;
+    PyObject *refchain = REFCHAIN(interp);
     fprintf(fp, "Remaining object addresses:\n");
-    for (op = refchain._ob_next; op != &refchain; op = op->_ob_next)
+    for (op = refchain->_ob_next; op != refchain; op = op->_ob_next)
         fprintf(fp, "%p [%zd] %s\n", (void *)op,
             Py_REFCNT(op), Py_TYPE(op)->tp_name);
 }
 
+/* The implementation of sys.getobjects(). */
 PyObject *
 _Py_GetObjects(PyObject *self, PyObject *args)
 {
     int i, n;
     PyObject *t = NULL;
     PyObject *res, *op;
+    PyInterpreterState *interp = _PyInterpreterState_GET();
 
     if (!PyArg_ParseTuple(args, "i|O", &n, &t))
         return NULL;
-    op = refchain._ob_next;
+    PyObject *refchain = REFCHAIN(interp);
+    op = refchain->_ob_next;
     res = PyList_New(0);
     if (res == NULL)
         return NULL;
-    for (i = 0; (n == 0 || i < n) && op != &refchain; i++) {
+    for (i = 0; (n == 0 || i < n) && op != refchain; i++) {
         while (op == self || op == args || op == res || op == t ||
                (t != NULL && !Py_IS_TYPE(op, (PyTypeObject *) t))) {
             op = op->_ob_next;
-            if (op == &refchain)
+            if (op == refchain)
                 return res;
         }
         if (PyList_Append(res, op) < 0) {
@@ -2313,6 +2324,8 @@ _Py_GetObjects(PyObject *self, PyObject *args)
     return res;
 }
 
+#undef REFCHAIN
+
 #endif
 
 

Python/pylifecycle.c

@@ -1921,11 +1921,11 @@ Py_FinalizeEx(void)
     }
 
     if (dump_refs) {
-        _Py_PrintReferences(stderr);
+        _Py_PrintReferences(tstate->interp, stderr);
     }
 
     if (dump_refs_fp != NULL) {
-        _Py_PrintReferences(dump_refs_fp);
+        _Py_PrintReferences(tstate->interp, dump_refs_fp);
     }
 #endif /* Py_TRACE_REFS */
 
@@ -1961,11 +1961,11 @@ Py_FinalizeEx(void)
      */
 
     if (dump_refs) {
-        _Py_PrintReferenceAddresses(stderr);
+        _Py_PrintReferenceAddresses(tstate->interp, stderr);
     }
 
     if (dump_refs_fp != NULL) {
-        _Py_PrintReferenceAddresses(dump_refs_fp);
+        _Py_PrintReferenceAddresses(tstate->interp, dump_refs_fp);
         fclose(dump_refs_fp);
     }
 #endif /* Py_TRACE_REFS */