Bug report

I don't have an actual failure to report: it's a potential issue I noticed while inspecting the code.

An asyncio.Runner on the main thread installs a SIGINT handler using signal.signal (and NOT loop.add_signal_handler for whatever reason). The signal handler could thus run at more-or-less any time and cannot rely on the asyncio machinery being in a consistent state. However, it directly interacts with the main task, calling main_task.done() and main_task.cancel():

    def _on_sigint(self, signum, frame, main_task):
        self._interrupt_count += 1
        if self._interrupt_count == 1 and not main_task.done():
            main_task.cancel()
            # wakeup loop if it is blocked by select() with long timeout
            self._loop.call_soon_threadsafe(lambda: None)
            return
        raise KeyboardInterrupt()

Task.cancel in particular seems like a complex piece of code that will depend on invariants, and that's before considering that one might have a user-provided event loop factory which in turn uses a custom task class.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux