python / python/cpython

Free-threading: `signal.pthread_kill` deadlocks child threads

未关闭
#150,873 6 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

extension-modules topic-free-threading type-bug
主要语言
Python
星标
77.2k
派生
35.9k
PR 合并指标
PR 指标待抓取

描述

Context

The Dask threaded scheduler (e.g. dask.array or dask.dataframe without dask.distributed) blocks the main thread when the user calls compute() and internally spawns a concurrent.futures.ThreadPoolExecutor to schedule tasks. If a user hits Ctrl+C, they expect the program to terminate soon (at most waiting for the tasks that are halfway through execution to terminate).

Bug report

This sequence frequently deadlocks on Linux x64 free-threaded:

  1. main thread calls threading.Thread.start
  2. another thread calls signal.pthread_kill(main_thread, signal.SIGINT), which lands after start() has returned but before the thread's target function has started executing
  3. main thread handles the resulting KeyboardInterrupt
  4. main thread calls .join() on the child thread started at point 1.
  5. frequently, .join() deadlocks. A second SIGINT (e.g. from Ctrl+C) unblocks it.

Reproducer

import signal
import threading

USE_BARRIER = False
NTHREADS = 20

def demo():
    main_thread = threading.get_ident()
    barrier = threading.Barrier(NTHREADS) if USE_BARRIER else None
    thread_can_die = threading.Event()

    def f() -> None:
        if barrier:
            barrier.wait()
        thread_can_die.wait()

    def interrupt() -> None:
        if barrier:
            barrier.wait()
        signal.pthread_kill(main_thread, signal.SIGINT)

    threads = []
    for _ in range(NTHREADS - 1):
        t = threading.Thread(target=f)
        t.start()
        threads.append(t)
    try:
        t = threading.Thread(target=interrupt)
        t.start()
        threads.append(t)
        for t in threads:
            t.join()
    except KeyboardInterrupt:
        pass
    else:
        assert False, "Expected KeyboardInterrupt"

    thread_can_die.set()
    for t in threads:
        t.join()


if __name__ == "__main__":
    i = 0
    while True:
        print(i, flush=True)
        demo()
        i += 1

On these setups, the above program hangs within a handful of cycles:

  • Linux x64 3.14.5t
  • Linux x64 git tip #7a468a101268d2b13105f94ae027df8b502d0c87 (built with Tools/pixi-packages/freethreading)

Cannot reproduce on

  • Linux x64 3.14.5

Setting USE_BARRIER=True, which ensures that all child threads are already fully started when the SIGINT lands on the main thread, makes the issue disappear.

CPython versions tested on:

CPython main branch, 3.14

Operating systems tested on:

Linux

贡献指南

打开贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

调研方向

首先在 Linux x64 free-threaded CPython 上运行提供的 reproducer,然后跟踪 threading.Thread.start()、threading.Thread.join() 和 signal.pthread_kill() 之间的交互。当 SIGINT 在子线程的 target 开始之前到达时 reproducer 不再死锁,同时预期的 KeyboardInterrupt 行为保持不变,即表示完成。

由索引模型根据 Issue 内容生成。

评估

技术栈
python
领域
operating-systems
Issue 类型
缺陷
难度
5/5
预计耗时
一周以上
活跃度
冷清
描述清晰度
基本清楚
新手友好度
45/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。