I'm not entirely sure when this regressed, but now we have a code path:

 gdk_window_begin_paint() 
   gdk_window_get_content()
     gdk_window_ref_impl_surface()
       gdk_x11_ref_cairo_surface
         hook_surface_changed()
     cairo_surface_destroy()
       [surface is destroyed]
         [mime data is removed]
           on_surface_changed,
             
So A) we have a surface being created and destroyed on every call to gdk_window_get_content() B) The hack to use cairo_surface_set_mime_data() to get notification when the window contents actually change is mostly ineffective and every call to gdk_window_begin_paint() will cause a frame to be triggered, even if no damage is actually painted.

I think I added the hack mostly to handle cases where begin_paint() wasn't called at all - maybe when GDK_EXPOSURE_MASK is not set on a window, so that would still work - but it's certainly concerning that we're triggering it off a getter.

gdk_window_create_similar_surface() has similar thrashing behavior where a cairo surface is created and destroyed.