gh-155363: Fix QSBR slot leak on thread state creation failure (gh-155365)

In the free-threaded build, new_threadstate() reserves a QSBR thread-state
slot before it can still fail for other reasons, but the failure paths only
called free_threadstate(), which does not know about the reservation.
Restructure code so failure path doesn't leak.

Author: nascheme

Include/internal/pycore_code.h

@@ -582,6 +582,10 @@ PyAPI_FUNC(_Py_CODEUNIT *) _PyCode_GetTLBC(PyCodeObject *co);
 // Returns the reserved index or -1 on error.
 extern int32_t _Py_ReserveTLBCIndex(PyInterpreterState *interp);
 
+// Release an index returned by _Py_ReserveTLBCIndex() that was never stored
+// in a PyThreadState.
+extern void _Py_UnreserveTLBCIndex(PyInterpreterState *interp, int32_t index);
+
 // Release the current thread's index into thread-local bytecode arrays
 extern void _Py_ClearTLBCIndex(_PyThreadStateImpl *tstate);
 

Misc/NEWS.d/next/Core_and_Builtins/2026-08-07-13-40-12.gh-issue-155363.Qk3Vt9.rst

@@ -0,0 +1,4 @@
+Fix a leak in the :term:`free-threaded build` when creating a thread state
+fails after an internal QSBR slot has been reserved for it.  The slot could
+never be reclaimed, so the QSBR array grew without bound across repeated
+failures.

Objects/codeobject.c

@@ -3314,14 +3314,20 @@ _Py_ReserveTLBCIndex(PyInterpreterState *interp)
 }
 
 void
-_Py_ClearTLBCIndex(_PyThreadStateImpl *tstate)
+_Py_UnreserveTLBCIndex(PyInterpreterState *interp, int32_t index)
 {
-    PyInterpreterState *interp = ((PyThreadState *)tstate)->interp;
     if (interp->config.tlbc_enabled) {
-        _PyIndexPool_FreeIndex(&interp->tlbc_indices, tstate->tlbc_index);
+        _PyIndexPool_FreeIndex(&interp->tlbc_indices, index);
     }
 }
 
+void
+_Py_ClearTLBCIndex(_PyThreadStateImpl *tstate)
+{
+    PyInterpreterState *interp = ((PyThreadState *)tstate)->interp;
+    _Py_UnreserveTLBCIndex(interp, tstate->tlbc_index);
+}
+
 static _PyCodeArray *
 _PyCodeArray_New(Py_ssize_t size)
 {

Python/pystate.c

@@ -1667,21 +1667,23 @@ new_threadstate(PyInterpreterState *interp, int whence)
         return NULL;
     }
 
-#ifdef Py_GIL_DISABLED
-    Py_ssize_t qsbr_idx = _Py_qsbr_reserve(interp);
-    if (qsbr_idx < 0) {
+#ifdef Py_STATS
+    // The PyStats structure is quite large and is allocated separated from
+    // tstate.
+    if (!_PyStats_ThreadInit(interp, tstate)) {
         free_threadstate(tstate);
         return NULL;
     }
+#endif
+#ifdef Py_GIL_DISABLED
     int32_t tlbc_idx = _Py_ReserveTLBCIndex(interp);
     if (tlbc_idx < 0) {
         free_threadstate(tstate);
         return NULL;
     }
-#endif
-#ifdef Py_STATS
-    // The PyStats structure is quite large and is allocated separated from tstate.
-    if (!_PyStats_ThreadInit(interp, tstate)) {
+    Py_ssize_t qsbr_idx = _Py_qsbr_reserve(interp);
+    if (qsbr_idx < 0) {
+        _Py_UnreserveTLBCIndex(interp, tlbc_idx);
         free_threadstate(tstate);
         return NULL;
     }