This causes a small performance hit when multiple GL contexts are used. As an
optimization we could use ARB_sync to only wait for the last draw call instead
of all GL commands.