Race between Py_SetRecursionLimit and _Py_LeaveRecursiveCallPy under free-threading

[hawkinsp]

Bug report

Bug description:

Under Python 3.13 at commit 65da5db, we see the following TSAN race:

import concurrent.futures
import functools
import sys
import threading

num_threads = 8

def count(n):
  if n > 0:
    return count(n - 1) + 1
  return 0

def closure(b):
  b.wait()
  sys.setrecursionlimit(100)
  count(50)

with concurrent.futures.ThreadPoolExecutor(max_workers=num_threads) as executor:
  for _ in range(100):
    b = threading.Barrier(num_threads)
    for _ in range(num_threads):
      executor.submit(functools.partial(closure, b))

TSAN report:

WARNING: ThreadSanitizer: data race (pid=3579557)
  Write of size 4 at 0x72a40001e02c by thread T8:
    #0 Py_SetRecursionLimit /usr/local/google/home/phawkins/p/cpython/Python/ceval.c:272:35 (python3.13+0x3dd302) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #1 sys_setrecursionlimit_impl /usr/local/google/home/phawkins/p/cpython/./Python/sysmodule.c:1319:5 (python3.13+0x4bab67) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #2 sys_setrecursionlimit /usr/local/google/home/phawkins/p/cpython/./Python/clinic/sysmodule.c.h:509:20 (python3.13+0x4bab67)
    #3 cfunction_vectorcall_O /usr/local/google/home/phawkins/p/cpython/Objects/methodobject.c:512:24 (python3.13+0x28a3c5) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #4 _PyObject_VectorcallTstate /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_call.h:168:11 (python3.13+0x1eafea) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #5 PyObject_Vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/call.c:327:12 (python3.13+0x1eafea)
    #6 _PyEval_EvalFrameDefault /usr/local/google/home/phawkins/p/cpython/Python/generated_cases.c.h:813:23 (python3.13+0x3e290b) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #7 _PyEval_EvalFrame /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_ceval.h:119:16 (python3.13+0x3dea3a) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #8 _PyEval_Vector /usr/local/google/home/phawkins/p/cpython/Python/ceval.c:1807:12 (python3.13+0x3dea3a)
    #9 _PyFunction_Vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/call.c (python3.13+0x1eb65f) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #10 _PyObject_VectorcallTstate /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_call.h:168:11 (python3.13+0x572352) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #11 partial_vectorcall /usr/local/google/home/phawkins/p/cpython/./Modules/_functoolsmodule.c:252:16 (python3.13+0x572352)
    #12 _PyVectorcall_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:273:16 (python3.13+0x1eb2d3) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #13 _PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:348:16 (python3.13+0x1eb2d3)
    #14 PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:373:12 (python3.13+0x1eb355) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #15 _PyEval_EvalFrameDefault /usr/local/google/home/phawkins/p/cpython/Python/generated_cases.c.h:1355:26 (python3.13+0x3e4af2) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #16 _PyEval_EvalFrame /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_ceval.h:119:16 (python3.13+0x3dea3a) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #17 _PyEval_Vector /usr/local/google/home/phawkins/p/cpython/Python/ceval.c:1807:12 (python3.13+0x3dea3a)
    #18 _PyFunction_Vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/call.c (python3.13+0x1eb65f) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #19 _PyObject_VectorcallTstate /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_call.h:168:11 (python3.13+0x1ef62f) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #20 method_vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/classobject.c:70:20 (python3.13+0x1ef62f)
    #21 _PyVectorcall_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:273:16 (python3.13+0x1eb2d3) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #22 _PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:348:16 (python3.13+0x1eb2d3)
    #23 PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:373:12 (python3.13+0x1eb355) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #24 thread_run /usr/local/google/home/phawkins/p/cpython/./Modules/_threadmodule.c:337:21 (python3.13+0x564a32) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #25 pythread_wrapper /usr/local/google/home/phawkins/p/cpython/Python/thread_pthread.h:243:5 (python3.13+0x4bddb7) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)

  Previous write of size 4 at 0x72a40001e02c by thread T7:
    #0 _Py_LeaveRecursiveCallPy /usr/local/google/home/phawkins/p/cpython/Python/ceval_macros.h:389:35 (python3.13+0x3f3993) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #1 _PyEval_EvalFrameDefault /usr/local/google/home/phawkins/p/cpython/Python/generated_cases.c.h:5257:13 (python3.13+0x3f3993)
    #2 _PyEval_EvalFrame /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_ceval.h:119:16 (python3.13+0x3dea3a) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #3 _PyEval_Vector /usr/local/google/home/phawkins/p/cpython/Python/ceval.c:1807:12 (python3.13+0x3dea3a)
    #4 _PyFunction_Vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/call.c (python3.13+0x1eb65f) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #5 _PyObject_VectorcallTstate /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_call.h:168:11 (python3.13+0x572352) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #6 partial_vectorcall /usr/local/google/home/phawkins/p/cpython/./Modules/_functoolsmodule.c:252:16 (python3.13+0x572352)
    #7 _PyVectorcall_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:273:16 (python3.13+0x1eb2d3) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #8 _PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:348:16 (python3.13+0x1eb2d3)
    #9 PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:373:12 (python3.13+0x1eb355) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #10 _PyEval_EvalFrameDefault /usr/local/google/home/phawkins/p/cpython/Python/generated_cases.c.h:1355:26 (python3.13+0x3e4af2) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #11 _PyEval_EvalFrame /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_ceval.h:119:16 (python3.13+0x3dea3a) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #12 _PyEval_Vector /usr/local/google/home/phawkins/p/cpython/Python/ceval.c:1807:12 (python3.13+0x3dea3a)
    #13 _PyFunction_Vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/call.c (python3.13+0x1eb65f) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #14 _PyObject_VectorcallTstate /usr/local/google/home/phawkins/p/cpython/./Include/internal/pycore_call.h:168:11 (python3.13+0x1ef62f) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #15 method_vectorcall /usr/local/google/home/phawkins/p/cpython/Objects/classobject.c:70:20 (python3.13+0x1ef62f)
    #16 _PyVectorcall_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:273:16 (python3.13+0x1eb2d3) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #17 _PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:348:16 (python3.13+0x1eb2d3)
    #18 PyObject_Call /usr/local/google/home/phawkins/p/cpython/Objects/call.c:373:12 (python3.13+0x1eb355) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #19 thread_run /usr/local/google/home/phawkins/p/cpython/./Modules/_threadmodule.c:337:21 (python3.13+0x564a32) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)
    #20 pythread_wrapper /usr/local/google/home/phawkins/p/cpython/Python/thread_pthread.h:243:5 (python3.13+0x4bddb7) (BuildId: 19c569fa942016d8ac49d19fd40ccb1ddded939b)

Looks like the per-thread update of p->py_recursion_remaining is racy.

CPython versions tested on:

3.13

Operating systems tested on:

Linux

Linked PRs

• gh-128717: Stop-the-world when setting the recursion limit #128741
• [3.13] gh-128717: Stop-the-world when setting the recursion limit (GH-128741) #128757

[ZeroIntensity]

Yeah, modifying other thread states as done in Py_SetRecursionLimit isn't particularly thread-safe. I think we should just make a stop-the-world pause before setting the values.