Bug#851754: gdm autologin ignores wayland session

[David Härdeman]

Package: gdm3
Version: 3.22.3-1
Followup-For: Bug #851754

Hi,

I've dug through the gdm3 sources and I can confirm this.

The problem is that autologin is racy.

In daemon/gdm-session.c, gdm_session_setup() and
gdm_session_setup_for_user() both do:

	send_setup[_for_user](self, service_name);
	gdm_session_defaults_changed(self);

send_setup[_for_user]() does asynchronous dbus calls and return
immediately, meaning that gdm_session_defaults_changed() is called
*before* the dbus call is finished and before the user's saved session
has been set.

gdm_session_defaults_changed() calls:
  --> update_session_type()
    --> gdm_session_is_wayland_session()
      --> get_session_filename()
        --> get_session_name() /* no session selected */
          --> get_default_session_name() /* no saved_session */
            --> get_fallback_session_name()
	
The default session returned by get_fallback_session_name() is "default"
on Debian systems (that's a Debian specific patch, upstream uses "gnome"
as the default) which gets converted to "default.desktop" before
returning to gdm_session_is_wayland_session().

gdm_session_is_wayland_session() then calls load_key_file_for_file(self,
"default.desktop"...) which looks for "default.desktop" in the
directories returned by get_system_session_dirs() in the following
order:

                "/etc/X11/sessions/",
                DMCONFDIR "/Sessions/",
                DATADIR "/gdm/BuiltInSessions/",
                DATADIR "/xsessions/",
		DATADIR "/wayland-sessions/"

"default.desktop" is found in DATADIR "/gdm/BuiltInSessions/", i.e.
/usr/share/gdm/BuiltInSessions/default.desktop and the full path is
returned to gdm_session_is_wayland_session() which checks the path to
determine if the session is a wayland session: 

        if (full_path != NULL && strstr (full_path, "/wayland-sessions/") != NULL) {
                is_wayland_session = TRUE;
        }

This means that gdm_session_is_wayland_session() will always return
false for the default session.

update_session_type() uses the return value (is_wayland_session) as
follows:

        if (is_wayland_session) {
                set_session_type (self, "wayland");
        } else {
                set_session_type (self, NULL);
        }

gdm_session_start_session() is later called, after the real user session
has been set but update_session_type() is *not* called when that happens.

gdm_session_start_session() does:

        gboolean                is_x11 = TRUE;
	...
	#ifdef ENABLE_WAYLAND_SUPPORT
        is_x11 = g_strcmp0 (self->priv->session_type, "wayland") != 0;
	#endif

        if (display_mode == GDM_SESSION_DISPLAY_MODE_LOGIND_MANAGED ||
            display_mode == GDM_SESSION_DISPLAY_MODE_NEW_VT) {
                run_launcher = TRUE;
        }
	...
		command = get_session_command (self);
		...
		if (run_launcher) {
                        if (is_x11) {
                                program = g_strdup_printf (LIBEXECDIR "/gdm-x-session %s %s\"%s\"",
                                                           run_xsession_script? "--run-script " : "",
                                                           allow_remote_connections? "--allow-remote-connections " : "",
                                                           command);
                        } else {
                                program = g_strdup_printf (LIBEXECDIR "/gdm-wayland-session \"%s\"",
                                                           command);
                        }
		}

Therefore, is_x11 will always be true here.

I couldn't find the right point in time to call update_session_type()
because of all the async dbus calls and callbacks, so I stopped here.

TL;DR; - seems like an upstream bug.

-- 
David Härdeman