call_soon() call be called before the event loop starts and after run_forever(), the call will be executed by the next call to run_forever() (or run_until_complete).

But calling call_soon() after close() does not make sense and the user may be surprised to not see its function called. It can be obvious in a simple program, but not in a large application.

If anybody cares, here is a small patch to implement this. I ran the test suite and nothing else fails because of this change.

However I wonder if this is a slippery slope to go: If call_soon raises after the event loop is closed than everything else that registers an action with the loop should raise as well.

So for consistency this patch should grow quite a bit (unless create_connection, add_reader etc. already raise in this case).

If the decision is to go this path I would also suggest to add a new exception type for "Event loop is closed" so that it can be caught in client code.

YMMV

> If the decision is to go this path I would also suggest to add a new exception type for "Event loop is closed" so that it can be caught in client code.

I don't see the purpose of handling such exception. It's an obvious
bug, you must not handle bugs :)

> > If the decision is to go this path I would also suggest to add a new exception type for "Event loop is closed" so that it can be caught in client code.

> I don't see the purpose of handling such exception. It's an obvious
bug, you must not handle bugs :)

Unless you can't directly fix them. They might be inside a library and triggered during application shutdown.

> They might be inside a library and triggered during application shutdown.

In this case, fix the library.

call_soon_after_close.diff looks good but also incomplete: call_later, call_at and run_in_executor should also raise an exception if the event loop is closed.

You should also add call_soon_threadsafe to the test.

> So for consistency this patch should grow quite a bit (unless create_connection, add_reader etc. already raise in this case).

Yes, add_reader raises an exception and there is an unit test to ensure that.

create_task() is a little bit special, it does not schedule immediatly a coroutine. IMO it makes sense to fail if the loop was closed, I don't see how the task can be executed if the loop was executed, so calling create_task() on a closed loop looks like a bug.

I suggest to modify the following methods:
- call_soon, call_soon_threadsafe
- call_at, call_later
- run_in_executor
- create_task
- add_signal_handler
- subprocess_exec, subprocess_shell

Did I miss something?

I'm not sure that *all* asyncio methods should must call _check_closed(), it may make the code harder to read. If the basic functions like call_soon and add_reader already handle the closed status, I guess that all other methods will fail loudly, so they don't need to be modified.

For example, stop() calls call_soon() and so will also raises an exception if the loop is closed.

Maybe we should add almost all methods to the test checking that calling these methods on a closed loop fail.

Adding minimal code checks to make multiple tests pass sounds good to me.

New changeset 226f870b387d by Victor Stinner in branch '3.4':
Closes #22922: More EventLoop methods fail if the loop is closed. Initial patch
https://hg.python.org/cpython/rev/226f870b387d

I didn't change subprocess_exec() nor subprocess_shell() because they are coroutine. A coroutine object can only raise an exception when they are executed, whereas a closed loop cannot execute a coroutine object. I modified create_task() to fail if the event loop is closed, it should be enough.

Thanks Torsten Landschoff for your fix.

Windows buildbots are not happy:

http://buildbot.python.org/all/builders/AMD64%20Windows7%20SP1%203.x/builds/5385/steps/test/logs/stdio


======================================================================
ERROR: test_close (test.test_asyncio.test_windows_events.ProactorTests)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\proactor_events.py", line 394, in close
    self._close_self_pipe()
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\proactor_events.py", line 423, in _close_self_pipe
    self._self_reading_future.cancel()
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\windows_events.py", line 70, in cancel
    return super().cancel()
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\futures.py", line 230, in cancel
    self._schedule_callbacks()
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\futures.py", line 245, in _schedule_callbacks
    self._loop.call_soon(callback, self)
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\base_events.py", line 384, in call_soon
    handle = self._call_soon(callback, args, check_loop=True)
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\base_events.py", line 395, in _call_soon
    self._check_closed()
  File "C:\buildbot.python.org\3.x.kloth-win64\build\lib\asyncio\base_events.py", line 234, in _check_closed
    raise RuntimeError('Event loop is closed')
RuntimeError: Event loop is closed

New changeset 6a955db1f78f by Victor Stinner in branch 'default':
Issue #22922: Fix ProactorEventLoop.close()
https://hg.python.org/cpython/rev/6a955db1f78f

New changeset 0a7d956a2f2d by Victor Stinner in branch '3.4':
Issue #22922: Fix ProactorEventLoop.close()
https://hg.python.org/cpython/rev/0a7d956a2f2d

> Windows buildbots are not happy:

test_asyncio pass again on Windows buildbots, I close the issue.

New changeset 6c473f82309d by Victor Stinner in branch '3.4':
Issue #22922: Fix ProactorEventLoop.close()
https://hg.python.org/cpython/rev/6c473f82309d