MMIO and gcc re-ordering issue

[Alan Cox]

> It's just another complicated thing for driver authors to get wrong.
> The other side of the coin is of course the cost.
> 
> The only thing I am absolutely sure of is that we should make a
> decision fast, document it, and just stick to it.

It's unfortunate that the __read/___write versions have other different
semantics. The default should definitely be ordered but having un-ordered
ones as an option would be good for zapping the hot paths that matter.

Most I/O paths aren't even worth the thinking time for non-ordered I/O.

Alan