On Jun 26, 2015 3:05 PM, "Benjamin Herrenschmidt" <benh@kernel.crashing.org>
wrote:
>
> On Fri, 2015-06-26 at 21:31 +0200, Luis R. Rodriguez wrote:
> > > Yeah either, we need to fix our relaxed implementation (patch
> > > welcome :-)
> >
> > Yikes, so although powerpc has useful heuristics to automatically
> > enable WC the default write ops have been nullifying its effects?
>
> The heuristic is for userspace mapping which don't use the kernel
> accessors. To cut a long story short, this was all done back in the day
> to speed up X.org which doesn't use fancy accessors with barriers to
> access the framebuffer (neither does the kernel fbdev layer).

Great.

>
> > Shouldn't this have been blatantly obvious in performance benchmarks
> > and expectations?
>
> > If not a change should give considerable performance
> > gains... If the WC was nullified on powerpc then the experience and
> > value of the heuristics have less value and even if they are going
> > to be only considered for userspace mmap() it should be taken with
> > a bit grain of salt it seems.
>
> It wasn't nullified for the main user at the time, the fb. And I
> mentioned an IB adapter or two for which the code had been hand tuned.

This still means there could be some affected drivers when used on powerpc,
no?

  Luis