[issue38323] asyncio: MultiLoopWatcher has a race condition (test_asyncio: test_close_kill_running() hangs on AMD64 RHEL7 Refleaks 3.x)

[STINNER Victor]

STINNER Victor <vstinner at python.org> added the comment:

When I reproduce test_cancel_make_subprocess_transport_exec() hang, the problem is that the C signal handler is called with SIGCHLD when the child process completes, but the Python signal handler is not called.

Python is "blocked" in a selector (maybe select.select(), it doesn't matter). I guess that the selector is interrupted by a signal (even if asyncio calls signal.setinterrupt(SIGCHLD, False)), but since the signal handler doesn't raise an exception, the syscall is restarted: see the PEP 475.

I understood that the asyncio event loop only gets the opportunity to call the Python signal handler if there is a pending asyncio event (call_soon, call_timer, event on a tracked FD, whatever). If the signal arrives when the event loop is idle, the Python signal handler will never be called since the selector is called with timeout=0 (blocking mode).

MultiLoopChildWatcher must ensures that the event loop is awaken when it receives a signal by using signal.setwakeup(). This is done by _UnixSelectorEventLoop.add_signal_handler(). Maybe MultiLoopChildWatcher could reuse this function, rather than calling directly signal.signal().

----------

_______________________________________
Python tracker <report at bugs.python.org>
<https://bugs.python.org/issue38323>
_______________________________________