I think this works.

Author: ZeroIntensity

Include/internal/pycore_pystate.h

@@ -296,6 +296,9 @@ _Py_AssertHoldsTstateFunc(const char *func)
 #define _Py_AssertHoldsTstate()
 #endif
 
+void _PyThreadState_Decref(PyThreadState *tstate);
+void _PyThreadState_Incref(PyThreadState *tstate);
+
 #ifdef __cplusplus
 }
 #endif

Python/ceval.c

@@ -2467,14 +2467,20 @@ PyEval_SetTraceAllThreads(Py_tracefunc func, PyObject *arg)
     _PyRuntimeState *runtime = &_PyRuntime;
     HEAD_LOCK(runtime);
     PyThreadState* ts = PyInterpreterState_ThreadHead(interp);
+    /* gh-132296: We need to prevent the thread state
+       from getting concurrently deallocated.
+       We can't stop-the-world because _PyEval_SetTrace() is re-entrant. */
+    _PyThreadState_Incref(ts);
     HEAD_UNLOCK(runtime);
 
     while (ts) {
         if (_PyEval_SetTrace(ts, func, arg) < 0) {
             PyErr_FormatUnraisable("Exception ignored in PyEval_SetTraceAllThreads");
         }
         HEAD_LOCK(runtime);
+        PyThreadState *old = ts;
         ts = PyThreadState_Next(ts);
+        _PyThreadState_Decref(old);
         HEAD_UNLOCK(runtime);
     }
 }

Python/pystate.c

@@ -1485,6 +1485,22 @@ decref_threadstate(_PyThreadStateImpl *tstate)
     }
 }
 
+static void
+_PyThreadState_Decref(PyThreadState *tstate)
+{
+    assert(tstate != NULL);
+    _PyThreadStateImpl *impl = (_PyThreadStateImpl *)tstate;
+    decref_threadstate(tstate);
+}
+
+static void
+_PyThreadState_Incref(PyThreadState *tstate)
+{
+    assert(tstate != NULL);
+    _PyThreadStateImpl *impl = (_PyThreadStateImpl *)tstate;
+    _Py_atomic_add_ssize(&impl->refcount, 1);
+}
+
 /* Get the thread state to a minimal consistent state.
    Further init happens in pylifecycle.c before it can be used.
    All fields not initialized here are expected to be zeroed out,