`Subprocess.wait_for_exit` never resolves if process terminated before it is called

[meffmadd]

The wait_for_exit method eventually calls os.waitpid, which throws a ChildProcessError if no process with the specified pid exists. This exception is caught, the function just returns, and the Future never resolves.

if __name__ == '__main__':
    import asyncio
    from tornado.process import Subprocess

    async def f():
        p = Subprocess("ls")
        p.proc.wait()
        return await p.wait_for_exit()

    loop = asyncio.get_event_loop()
    loop.run_until_complete(f())

tornado/tornado/process.py

Line 348 in 65a9e48

return

Instead, the process could be retrieved from the _waiting dict and the return code could be accessed from the object directly.

[bdarnell]

As long as Tornado is the only thing that touches the SIGCHLD handler, the child process is guaranteed to exist (in a zombie state) until os.waitpid (or another wait function) is called on it once. Could something else be installing a SIGCHLD handler, ignoring SIGCHLD, or calling os.wait directly? (In this example it seems wrong to use both p.proc.wait() and p.wait_for_exit() but if you've seen it in the wild I doubt it's that simple)

[meffmadd]

Ok, interesting! As far as I understand, nothing else is happening. We were creating the subprocess (calling git init) and immediately called wait_for_exit on the following line.

However, we used the Subprocess class outside the tornado process (in a Celery worker instance). Could this have anything to do with it? We have now switched to Popen and wait, which works as intended. The docs don't mention anything about this.

[bdarnell]

However, we used the Subprocess class outside the tornado process (in a Celery worker instance).

I'm confused (maybe because I've never used celery) - there's still a tornado IOLoop (or asyncio event loop) running in the worker process, right? If not you wouldn't be able to await p.wait_for_exit().

Celery does touch the SIGCLD handler although I'm not sure if it's possible for this to be run in a way that would clobber tornado's SIGCLD handler. It doesn't appear to call asyncio's set_child_watcher function but if something in the stack does, it could also cause this problem.