RFC on writel and writel_relaxed

[Benjamin Herrenschmidt]

On Mon, 2018-03-26 at 16:50 -0600, Jason Gunthorpe wrote:
> On Tue, Mar 27, 2018 at 09:36:11AM +1100, Benjamin Herrenschmidt wrote:
> > On Mon, 2018-03-26 at 16:27 -0600, Jason Gunthorpe wrote:
> > > > Otherwise almost all drivers out there are broken which I very much
> > > > doubt :-)
> > > 
> > > But.. Sinan is right, you look anywhere in the driver tree and you
> > > find stuff like this:
> > > 
> > > drivers/net/ethernet/intel/i40e/i40e_txrx.c
> > > 
> > >         /* Force memory writes to complete before letting h/w
> > >          * know there are new descriptors to fetch.
> > >          */
> > >         wmb();
> > > 
> > > 
> > > It is *systemic*
> > 
> > Yes, because they all copied e1000e :-) If you look at the comment in
> > there, it does say it's only for weakly ordered archs such as ia64, and
> > even then, probably predates Linus strong statement on the matter.
> 
> Hahah, sure I'll buy that..
> 
> But still, if this really is the case, a *strong* statement in
> barriers.txt to that effect (and not an example demanding the wmb()!)
> would be very helpful for those of us that have to review driver code!

I agree, and that Mellanox bug you pointed me to seems to indicate that
this may not even be true on x86 anymore ...

I think we might need to revisit this properly...

Cheers,
Ben.