[PATCH v2 08/25] arch: introduce memremap()

[Toshi Kani]

On Wed, 2015-07-29 at 11:33 -0700, Dan Williams wrote:
> On Wed, Jul 29, 2015 at 11:27 AM, Luis R. Rodriguez <mcgrof at suse.com> 
> wrote:
> > On Wed, Jul 29, 2015 at 08:50:04AM +0200, Christoph Hellwig wrote:
> > > On Mon, Jul 27, 2015 at 04:26:03PM -0700, Dan Williams wrote:
> > > > Oh, because all we have at this point is ioremap_cache() which
> > > > silently falls back.  It's not until the introduction of
> > > > arch_memremp() where we update the arch code to break that 
> > > > behavior.
> > > 
> > > Ok, makes sense.  Might be worth to document in the changelog.
> > > 
> > > > That said, I think it may be beneficial to allow a fallback if the
> > > > user cares.  So maybe memremap() can call plain ioremap() if
> > > > MEMREMAP_STRICT is not set and none of the other mapping types are
> > > > satisfied.
> > > 
> > > Is there a real use case for it?  Fallback APIs always seem confusing
> > > and it might make more sense to do this in the caller(s) that 
> > > actually
> > > need it.
> > 
> > It seems semantics-wise we are trying to separate these two really, so 
> > I agree
> > with this. Having a fallback would onloy make things more complicated 
> > for any
> > sanitizer / checker / etc, and I don't think the practical gains of 
> > having a
> > fallback outweight the gains of having a clear semantic separation on 
> > intended
> > memory type and interactions with it.
> > 
> 
> Yup, consider it dropped.  Drivers that want fallback behavior can do
> it explicitly.

I agree in general.  However, for the drivers to be able to fall back
properly, they will need to know the cache type they can fall back with. 
 ioremap() falls back to the cache type of an existing mapping to avoid
aliasing.

Thanks,
-Toshi