msg/async: mark down local_connection before draining the stack.
`AsyncMessenger` has `local_connection` – an instance of
`AsyncConnection` used to send messages to itself (loop back).
Connections are handled by `EventCenter` in its dedicated
thread. When shutting down a messenger, it must be ensured
there is no task left in the `EventCenter`'s queue. Otherwise
a use-after-free can happen. That's the reason why `shutdown()`
of `AsyncMessenger` does `mark_down_all()` on connections
**before** draining the stack. The latter actually injects a task
into all `EventCenter` instances and waits for its completion
(synchronization barrier).
However, that's not the case for `local_connection`. Without
the patch it's marked down by the messenger's destructor
far **after** the synchronization barrier. This behavior is
dangerous when an implementation of `mark_down()` creates
a new task to be executed inside the boundaries of corresponding
`EventCenter` instance.
The fix unifies handling of `local_connection` with other
connections in the aspect of the shutdown phase.
Fixes: https://tracker.ceph.com/issues/43667
Signed-off-by: Radoslaw Zarzynski <rzarzyns@redhat.com>
(cherry picked from commit
c2f61b95f50f88bfbfe4f46fcc447762c119bae1)