Hm. Why wouldn't it work? The user wrote to the physical page backing both the non-doubled and doubled extent.

Re: magic buffers. Try this on a machine with a VIVT data cache and have fun debugging. The kernel will let you make the mappings, and then hilarity will ensue ;)