Data race creating Tcl interpreters concurrently via `_tkinter.create` under free-threading

Bug report

Bug description:

_tkinter declares Py_MOD_GIL_NOT_USED, so under --disable-gil the GIL no longer serializes calls into it. _tkinter.create() creates a new Tcl interpreter with Tcl_CreateInterp():

https://github.com/python/cpython/blob/22a6c51c94a4fde986b8964f1d36d5ec3ac20dcc/Modules/_tkinter.c#L631-L643

reached from _tkinter_create_impl:

https://github.com/python/cpython/blob/22a6c51c94a4fde986b8964f1d36d5ec3ac20dcc/Modules/_tkinter.c#L3456-L3483

_tkinter does not serialize interpreter creation, so two threads calling _tkinter.create() run Tcl_CreateInterp() concurrently. That triggers Tcl's first-time global initialization, where Tcl_MutexLock lazily initializes a static mutex. That init is not concurrency-safe, so one thread's pthread_mutex_init (write) races with another thread's pthread_mutex_lock (atomic read) on the same Tcl global.

Reproducer:

import _tkinter
from threading import Thread

def worker():
    for _ in range(20000):
        try:
            _tkinter.create(None, '', 'Tk', False, 1, False, False, None)
        except Exception:
            pass

ts = [Thread(target=worker) for _ in range(12)]
for t in ts: t.start()
for t in ts: t.join()

TSAN Report (Tested on Linux, Tcl 8.6.14, with useTk=False):

==================
WARNING: ThreadSanitizer: data race (pid=650542)
  Atomic read of size 1 at 0x72a0000145d0 by thread T2:
    #0 pthread_mutex_lock <null> 
    #1 TclCreateExecEnv /usr/src/tcl8.6-8.6.14+dfsg-1build1/generic/tclExecute.c:936:5 
    #2 cfunction_vectorcall_FASTCALL /cpython/Objects/methodobject.c:449:24 
    #3 _PyObject_VectorcallTstate /cpython/./Include/internal/pycore_call.h:144:11 
    #4 PyObject_Vectorcall /cpython/Objects/call.c:327:12 
    #5 _Py_VectorCallInstrumentation_StackRefSteal /cpython/Python/ceval.c:768:11 
    #6 _PyEval_EvalFrameDefault /cpython/Python/generated_cases.c.h:1906:35  

  Previous write of size 1 at 0x72a0000145d0 by thread T1 (mutexes: write M0):
    #0 pthread_mutex_init <null> 
    #1 Tcl_MutexLock /usr/src/tcl8.6-8.6.14+dfsg-1build1/unix/tclUnixThrd.c:430:6 
    #2 cfunction_vectorcall_FASTCALL /cpython/Objects/methodobject.c:449:24 
    #3 _PyObject_VectorcallTstate /cpython/./Include/internal/pycore_call.h:144:11 
    #4 PyObject_Vectorcall /cpython/Objects/call.c:327:12 
    #5 _Py_VectorCallInstrumentation_StackRefSteal /cpython/Python/ceval.c:768:11 
    #6 _PyEval_EvalFrameDefault /cpython/Python/generated_cases.c.h:1906:35  

  Location is heap block of size 16384 at 0x72a000014000 allocated by thread T1:
    #0 malloc <null> 
    #1 GetBlocks /usr/src/tcl8.6-8.6.14+dfsg-1build1/generic/tclThreadAlloc.c:1044:17 
    #2 cfunction_vectorcall_FASTCALL /cpython/Objects/methodobject.c:449:24 
    #3 _PyObject_VectorcallTstate /cpython/./Include/internal/pycore_call.h:144:11 
    #4 PyObject_Vectorcall /cpython/Objects/call.c:327:12 
    #5 _Py_VectorCallInstrumentation_StackRefSteal /cpython/Python/ceval.c:768:11 
    #6 _PyEval_EvalFrameDefault /cpython/Python/generated_cases.c.h:1906:35 

  Mutex M0 (0x7fffb1dff7a0) created at:
    #0 pthread_mutex_lock <null>
    #1 Tcl_MutexLock /usr/src/tcl8.6-8.6.14+dfsg-1build1/unix/tclUnixThrd.c:423:2 
    #2 _PyImport_RunModInitFunc /cpython/./Python/importdl.c:436:19 
    #3 import_run_extension /cpython/Python/import.c:2167:14
    #4 _imp_create_dynamic_impl /cpython/Python/import.c:5565:11 
    #5 _imp_create_dynamic /cpython/Python/clinic/import.c.h:489:20 
    #6 cfunction_vectorcall_FASTCALL /cpython/Objects/methodobject.c:449:24 
    #7 _PyVectorcall_Call /cpython/Objects/call.c:273:16 
    #8 _PyObject_Call /cpython/Objects/call.c:348:16
    #9 PyObject_Call /cpython/Objects/call.c:373:12
    #10 _PyEval_EvalFrameDefault /cpython/Python/generated_cases.c.h:2831:38 

SUMMARY: ThreadSanitizer: data race (/cpython/cpython-tsan/bin/python3.16t+0xfad6e)  in pthread_mutex_lock
==================

On macOS with Tcl/Tk 9.0 the same reproducer surfaces the _tkinter module-global data races (PyOS_InputHook, tcl_lock). With useTk=True it crashes inside Tk's own display initialization.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-156342

Commencez dans Modules/_tkinter.c, au niveau de Tcl_CreateInterp et de _tkinter_create_impl, puis reproduisez les appels concurrents avec une build de CPython free-threaded utilisant ThreadSanitizer. Comparez la race Tcl sous Linux avec les races au niveau global du module sous macOS décrites dans le rapport. C’est terminé lorsque la création de l’interpréteur ne signale plus les races identifiées et ne plante plus avec le reproducer indiqué ; gh-156342 est déjà lié pour le contexte.