On Tue, 2018-03-27 at 09:36 +1100, Benjamin Herrenschmidt wrote:

I don't kow, it used to be the case, at least that's what drove us to
define things the way we did.

Maybe things changed, but if that's the case, nobody knows for sure,
and we probably want to get Linus POV on the matter.

I know I still write drivers that do not add a wmb in that case because
I expect things to work without it.

If that has changed, we probably can relax some of the barriers in our
implementations of writel on a number of architectures, but not before
auditing a bunch more drivers to make sure they have the write wmb()'s

Note also that this was the entire point behind the definition of
the _relaxed() accessors, to lift that specific ordering guarantee.

If you now says that memory + writel requires a wmb() in between
then you made writel be identical to writel_relaxed.

You might notice that Documentation/driver-api/device-io.rst makes no
mention of wmb() at all.

Cheers,
Ben.