Fix crashes on exit caused by wlroots listener checks

[Ferdi265]

This PR fixes sway crashing on exit due to event listeners not being removed from wlroots objects on exit.

Listeners fixed:

• server: all objects created during server_init()
• input/input-manager: wlr_backend.new_input, wlr_virtual_keyboard, wlr_virtual_pointer, wlr_keyboard_shortcuts_inhibit, wlr_transient_seat_manager
• desktop/idle-inhibit: wlr_idle_inhibit_manager
• input/text-input: wlr_text_input_manager and wlr_input_method_manager
• tree/container: wlr_scene_buffer.output_enter and wlr_scene_buffer.output_leave

This doesn't fix all wlroots objects with listener checks, just those that lead to immediate crashes on exit. I'll wait for a review of my general method of fixing this before continuing.

[Ferdi265]

On second thought, it might be more useful to remove the bulk of the listeners in server_fini() before calling destroy, which avoids adding dozens of destroy listeners.

Destroy listeners are still needed for a bunch of other components that aren't directly created in server.c, but that reduces the clutter by a lot.

[kennylevinsen]

On second thought, it might be more useful to remove the bulk of the listeners in server_fini() before calling destroy, which avoids adding dozens of destroy listeners.

If we could have a 1:1 mapping between what is set up in server_init and what is torn down in server_fini, that might be fine - but it is probably still easier to understand if all structs with listeners must have an appropriate destroy handler to remove them.

One thing is what is easiest to implement, another is what is easiest to maintain and catch errors within in the future.

[Ferdi265]

On second thought, it might be more useful to remove the bulk of the listeners in server_fini() before calling destroy, which avoids adding dozens of destroy listeners.

If we could have a 1:1 mapping between what is set up in server_init and what is torn down in server_fini, that might be fine - but it is probably still easier to understand if all structs with listeners must have an appropriate destroy handler to remove them.

One thing is what is easiest to implement, another is what is easiest to maintain and catch errors within in the future.

IMO components that have _init() and _finish() functions that create protocol objects that can't go away on their own should probably avoid extra destroy handlers and just remove the listeners in finish. Components where objects are created in reference to an outside lifetime (e.g. in response to a seat being created) should probably create destroy listeners to remove those. input-manager has a different problem where it can't remove its own listeners in destroy handlers due to it not having access to the main server struct, needing a new _finish() method that is called from server_fini().

This is sometimes hard to distinguish though; text_input.c for example uses global objects initialized in server_init(), but is created in response to seat objects which are destroyed after the wlroots ones. This lifetime mismatch makes this hard to deal with, since it needs to be cleaned up both in response to a seat going away and in response to the wlroots objects being destroyed, with both happening at exit (which would cause a crash due to double wl_list_remove()).

I'm currently handling these by making the _finish() function robust against being called twice, but this doesn't feel optimal.

I think I'll try to get this PR far enough to exit sway without crashes with this method, keeping as much as possible in separate commits for reviewability, and we'll sort out what the best path forward is during review.

[Ferdi265]

The current state makes sway exit correctly from a desktop containing a few konsole windows and a firefox window.

I'm not completely sure about the correct handling of all of these cases, and I'm sure there are a lot more objects with listeners that need to be removed before destroy, but those don't trigger on a regular exit in my setup. I've already started writing a script to search for all objects in wlroots that have listener checks and their corresponding usages in sway, but that might take a while.

In the meantime, I think it's better to mark this as reviewable now and see where we go afterwards.

[Ferdi265]

If we could have a 1:1 mapping between what is set up in server_init and what is torn down in server_fini, that might be fine

Addendum to this: this works fine for server_init, (and saves a ton of destroy handlers) since all those objects are created/destroyed only at init and fini, but this doesn't work for some of the more dynamic objects or listeners created elsewhere in sway.

The current implementation has this 1:1 mapping for objects created in server_init() in the first commit.

[Ferdi265]

Diff to state before last force-push:

diff --git a/sway/server.c b/sway/server.c
index 683ce169..79c8f542 100644
--- a/sway/server.c
+++ b/sway/server.c
@@ -458,6 +458,7 @@ bool server_init(struct sway_server *server) {
 
 void server_fini(struct sway_server *server) {
        // remove listeners
+       wl_list_remove(&server->renderer_lost.link);
        wl_list_remove(&server->new_output.link);
        wl_list_remove(&server->layer_shell_surface.link);
        wl_list_remove(&server->xdg_shell_toplevel.link);
@@ -474,14 +475,12 @@ void server_fini(struct sway_server *server) {
        wl_list_remove(&server->xdg_activation_v1_request_activate.link);
        wl_list_remove(&server->xdg_activation_v1_new_token.link);
        wl_list_remove(&server->request_set_cursor_shape.link);
-#if WLR_HAS_XWAYLAND
-       wl_list_remove(&server->xwayland_surface.link);
-       wl_list_remove(&server->xwayland_ready.link);
-#endif
        input_manager_finish(server);
 
        // TODO: free sway-specific resources
 #if WLR_HAS_XWAYLAND
+       wl_list_remove(&server->xwayland_surface.link);
+       wl_list_remove(&server->xwayland_ready.link);
        wlr_xwayland_destroy(server->xwayland.wlr_xwayland);
 #endif
        wl_display_destroy_clients(server->wl_display);