Investigate: Reinitialize EventChannel<_> for states with custom dispatchers

[azriel91]

In GitLab by @azriel91 on May 2, 2019, 23:03

For states with custom dispatchers (including every state that impl AppState), Systems which register ReaderIds against EventChannel<_>s will be dropped when the State ends. This means those event channels will persist in the World, and cause a memory leak -- the reader never reads, so the event channel holds onto the events that have not been read by every reader.

Attempting to re-initialize a channel in State#on_stop causes an index out of bounds error:

thread 'main' panicked at 'index out of bounds: the len is 0 but the index is 0', /rustc/e305df1846a6d985315917ae0c81b74af8b4e641\src\libcore\slice\mod.rs:2693:14

fn on_stop(&mut self, data: StateData<'_, GameData<'a, 'b>>) {
    // Replace the event channel, so that reader IDs from previous instantiations of this
    // `State` will not cause the buffer to grow indefinitely.
    let mut character_selection_ec = data
        .world
        .write_resource::<EventChannel<CharacterSelectionEvent>>();
    *character_selection_ec = EventChannel::<CharacterSelectionEvent>::new();
}

This is further complicated by event channels that are long lived, such as EventChannel<ControlInputEvent>.

---

Options:

• Try to implement remove(ReaderId) for EventChannel in shrev.

  This is part of shrev 1.0.

• Figure out what's still running the Systems that try to read the event channel.

  InputToControlInputSystem was doing a delta from the existing ControllerInput value against the InputEvent value, and sending a ControlInputEvent if it didn't match. This meant it would always send an event for new entities with a ControllerInput component.

  !123 changes the behaviour to send ControlInputEvents when InputEvents are sent.

• Try to drop / initialize the EventChannel somewhere else (note, AppState sets up the dispatcher before calling on_start on the delegate).

  Bad solution -- it's hard to discover who is replacing the EventChannel.

• Register all systems in the main dispatcher, use state specific status types to decide whether to:

  • Drop events.
  • Proceed with regular system logic.

• Try to drop the ReaderId in State#on_stop

  Needs more thought. See Improve ReaderId rustgd/shrev-rs#28 (comment)

[azriel91]

In GitLab by @azriel91 on Jul 19, 2019, 21:13

This shouldn't be an issue -- dropped ReaderIds unregister themselves on drop.