Re: [PATCH] Define wc_wmb, a write barrier for PCI write combining

[Roland Dreier <rdreier@cisco.com>]

    Jesse> I don't think it addresses the flushing issue you seem to
    Jesse> be concerned about though.  I don't know the exact
    Jesse> semantics of sfence, but I think bcrl is likely right that
    Jesse> it won't absolutely guarantee that your writes have hit the
    Jesse> device before proceeding (though it may do that on some CPU
    Jesse> implementations).

Yeah, I think that Bryan just wrote something different than what he
meant: there is no desire for wc_wmb() to make sure that writes via a
write-combining mapping have gone all the way to the device, any more
than a normal wmb() makes sure normal writes have gone all the way to
the device.  All that wc_wmb() needs to do is make sure that writes
via a write-combining mapping don't get passed by later writes.

This does speak to the need for precise documentation though :)

 - R.