Bugfix: don’t remove SubstructureRedirect event mask temporarily

This fixes race conditions, for example when i3bar gets reconfigured
after the available outputs change. In that specific case, i3bar sends a
ConfigureWindow request (see
b5693d6fb3/i3bar/src/xcb.c (L376))
which normally is turned into a ConfigureRequest that i3 largely
ignores, only the dock client’s height is considered (see
b5693d6fb3/src/handlers.c (L390)).

Turning ConfigureWindow into ConfigureRequest is only done when the
SubstructureRedirect event mask is set, and because we temporarily
removed _all_ events from our mask, the ConfigureWindow request went
through unmodified.

This in turn lead to the i3bar client window (not its decoration frame)
being positioned at e.g. y=1304, whereas dock client windows should
always end up at x=0 y=0 within their decoration frame. The result of
the i3bar client window being out of the visible space was either a
black i3bar or graphics corruption.

This also fixes issue #1904, I think. I couldn’t reproduce issue #1904
specifically, but when i3bar is in the misconfigured state, it will
receive a VisibilityNotify event, telling i3bar that it is obscured.
This would explain why i3bar sent a SIGSTOP in issue #1904.

fixes #1904

src/x.c

@@ -984,7 +984,10 @@ void x_push_changes(Con *con) {
 
     DLOG("-- PUSHING WINDOW STACK --\n");
     //DLOG("Disabling EnterNotify\n");
-    uint32_t values[1] = {XCB_NONE};
+    /* We need to keep SubstructureRedirect around, otherwise clients can send
+     * ConfigureWindow requests and get them applied directly instead of having
+     * them become ConfigureRequests that i3 handles. */
+    uint32_t values[1] = {XCB_EVENT_MASK_SUBSTRUCTURE_REDIRECT};
     CIRCLEQ_FOREACH_REVERSE(state, &state_head, state) {
         if (state->mapped)
             xcb_change_window_attributes(conn, state->id, XCB_CW_EVENT_MASK, values);