Bug report

Bug description:

On the proactor event loop (the default on Windows), BaseProactorEventLoop._start_serving()'s loop() callback runs protocol_factory() and transport creation in the same callback that re-arms accept(), guarded only by except OSError / except CancelledError. When protocol_factory() or _make_ssl_transport() / _make_socket_transport() raises for an accepted connection:

• a non-OSError (e.g. a failing protocol_factory) propagates out of loop(): the accepted socket is never closed (it leaks until garbage collection) and accept is not re-armed, so the server stops accepting new connections;
• an OSError is caught as 'Accept failed on a socket', which closes the listening socket, and still leaks the accepted socket.

This is the proactor counterpart of gh-155934 (selector accept) and gh-153133 (create_connection): an accepted socket must be closed when the transport is never created, and a per-connection failure must not tear down the listening socket or the accept loop.

Verified with mock-based tests in test_proactor_events (the same approach used for the selector fix in gh-155934), which run on all platforms.

CPython versions tested on:

CPython main

Operating systems tested on:

Windows (ProactorEventLoop)

Linked PRs

• gh-156365: Fix socket leak and stalled accept in the asyncio proactor server #156366