Just noticed this while looking at the code to asyncio.Task.__step

asyncio Futures have two different error states: they can have an arbitrary exception set, or they can be marked as cancelled.

asyncio Tasks handle this by detecting what exception was raised by the task code: if it's a CancelledError, then the mark the Future as cancelled, and if it's anything else, they set that as the Future's exception.

There is also a special case handled inside the 'except StopIteration' clause in Task.__step. If we request that a Task be cancelled, but then the task exits normally before we have a chance to throw in a CancelledError, then we also want mark the Future as cancelled. BUT, this special case is handled incorrectly: it does Future.set_exception(CancelledError()), instead of Future.cancel(). Normally it's impossible for a task to end up with its exception set to CancelledError, but it does happen in this one weird edge case, basically as a race condition.

Here's some sample code to illustrate the problem (tested on 3.7):

------

import asyncio

# This gets cancelled normally
async def cancel_early():
    asyncio.current_task().cancel()
    await asyncio.sleep(1)

async def cancel_late():
    asyncio.current_task().cancel()
    # No sleep, so CancelledError doesn't get delivered until after the
    # task exits

async def main():
    t_early = asyncio.create_task(cancel_early())
    await asyncio.sleep(0.1)
    print(f"t_early.cancelled(): {t_early.cancelled()!r}")

    t_late = asyncio.create_task(cancel_late())
    await asyncio.sleep(0.1)
    print(f"t_late.cancelled(): {t_late.cancelled()!r}")
    print(f"t_late.exception(): {t_late.exception()!r}")

asyncio.run(main())

Output:

t_early.cancelled(): True
t_late.cancelled(): False
t_late.exception(): CancelledError()

The obvious fix would be to modify the 'except StopIteration' branch to handle this case by calling super().cancel() instead of super().set_exception(...).

Alternatively, I could see an argument that asyncio.Task should always preserve the CancelledError, so that e.g. you can get a proper traceback. In that case we'd need to change the 'except CancelledError' branch to call super().set_exception(...) instead of super().cancel(). This would also need some more changes, like overriding .cancelled() to check for a stored exception and then return isinstance(stored_exc, CancelledError), and maybe others... I'm not sure of the full consequences.

But handling these two cases differently is definitely wrong, that part I'm sure of :-)

The obvious fix would be to modify the 'except StopIteration' branch to handle this case by calling super().cancel() instead of super().set_exception(...).

Yeah, I think this is the solution we should do in 3.8.

New changeset edad4d8 by Carol Willing (Yury Selivanov) in branch 'master':
bpo-38248: Fix inconsistent immediate asyncio.Task cancellation (GH-16330)
edad4d8

New changeset 16cec13 by Carol Willing (Miss Islington (bot)) in branch '3.8':
bpo-38248: Fix inconsistent immediate asyncio.Task cancellation (GH-16330) (GH-16383)
16cec13

New changeset efe74b6 by Victor Stinner in branch 'master':
bpo-38321: Fix _asynciomodule.c compiler warning (GH-16493)
efe74b6

New changeset bfe1f74 by Victor Stinner in branch '3.8':
[3.8] bpo-3832: Fix compiler warnings (GH-16518)
bfe1f74

Can this issue be closed now?