asyncio.subprocess blocked event loop unexpectedly on MacOS

[starstreammm]

Bug report

Bug description:

After upgrading to 3.14.7 & macOS 27, this bug has happened.

When pausing the subprocess created by asyncio.create_subprocess_exec, it will blocked the whole event loop on MacOS.

Here is a minimal test code:

import asyncio
import psutil

async def main():
    proc = await asyncio.create_subprocess_exec(
        "sleep",
        "1000",
    )
    psutil.Process(proc.pid).suspend()

    print("child pid:", proc.pid, flush=True)

    count = 0
    while True:
        await asyncio.sleep(1)
        print(f"alive {count}", flush=True)
        count += 1

asyncio.run(main())

Expectly, it should print:

child pid: xxxxx
alive 0
alive 1
...

However, it just prints child pid: xxxxx, without any alive.

This issue isn't caused by psutil because the same result was got with kill -STOP xxxxx.

This runs well on Windows, but not on MacOS. This didn't happen on the previous version of 3.14.

CPython versions tested on:

3.14

Operating systems tested on:

macOS, Windows

[picnixz]

cc @kumaraditya303

[picnixz]

@starstreammm Do you mean that on Python 3.14.6 it works well? or is it on 3.13 that it works well? (this is to confirm whether it's a release blocker or not)

[starstreammm]

@starstreammm Do you mean that on Python 3.14.6 it works well? or is it on 3.13 that it works well? (this is to confirm whether it's a release blocker or not)

I updated my Python version yesterday. It had been quite a while since my last update, but I can confirm that I was already using 3.14 after the update. This issue didn't happen on that version. But I don't remember which version I used before exactly.

[picnixz]

I'm going to let Kumar have a look at this issue and decide whether this is a release-blocker. At least, it's a regression AFAICT.

[deadlovelll]

hey, i just took a quick look with my own tooling (on macos too)

The trace:

[EventLoopBlock] 'main' blocked the loop for 0.735s (threshold: 0.1s)
Stack at the moment of blocking:
  File "/Users/timofeiivankov/cpython/repro.py", line 199, in <module>
    asyncio.run(main())
  File "/Users/timofeiivankov/cpython/Lib/asyncio/runners.py", line 205, in run
    return runner.run(main)
  File "/Users/timofeiivankov/cpython/Lib/asyncio/runners.py", line 128, in run
    return self._loop.run_until_complete(task)
  File "/Users/timofeiivankov/cpython/Lib/asyncio/base_events.py", line 712, in run_until_complete
    self.run_forever()
  File "/Users/timofeiivankov/cpython/Lib/asyncio/base_events.py", line 683, in run_forever
    self._run_once()
  File "/Users/timofeiivankov/cpython/Lib/asyncio/base_events.py", line 2090, in _run_once
    handle._run()
  File "/Users/timofeiivankov/cpython/Lib/asyncio/events.py", line 135, in _run
    return super()._run()
  File "/Users/timofeiivankov/cpython/Lib/asyncio/events.py", line 91, in _run
    self._context.run(self._callback, *self._args)
  File "/Users/timofeiivankov/cpython/Lib/asyncio/unix_events.py", line 1003, in _reap_and_notify
    pid, returncode = self._reap(loop, expected_pid)
  File "/Users/timofeiivankov/cpython/Lib/asyncio/unix_events.py", line 1008, in _reap
    pid, status = os.waitpid(expected_pid, 0)

-> Culprit: _reap() at /Users/timofeiivankov/cpython/Lib/asyncio/unix_events.py:1008

So _reap is probably the cause, or the code near it. If no one objects, I can work on the PR

[picnixz]

Can you bisect the faulty commit first? @deadlovelll

[deadlovelll]

Can you bisect the faulty commit first? @deadlovelll

Faulty commit - 2875d1d (#153810). Fix is needed for main, 3.15 and 3.14