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

["Luis R. Rodriguez" <mcgrof@suse.com>]

On Mon, Jul 27, 2015 at 04:31:09PM -0700, Dan Williams wrote:
> On Mon, Jul 27, 2015 at 4:17 PM, Luis R. Rodriguez <mcgrof@suse.com> wrote:
> > On Fri, Jul 24, 2015 at 10:38:42PM -0400, Dan Williams wrote:
> >> Existing users of ioremap_cache() are mapping memory that is known in
> >> advance to not have i/o side effects.  These users are forced to cast
> >> away the __iomem annotation, or otherwise neglect to fix the sparse
> >> errors thrown when dereferencing pointers to this memory.  Provide
> >> memremap() as a non __iomem annotated ioremap_*() in the case when
> >> ioremap is otherwise a pointer to memory.
> >
> > Ok so a special use case.
> >
> >> Outside of ioremap() and
> >> ioremap_nocache(), the expectation is that most calls to
> >> ioremap_<type>() are seeking memory-like semantics (e.g.  speculative
> >> reads, and prefetching permitted).  These callsites can be moved to
> >> memremap() over time.
> >
> > Such memory-like smantics are not well defined yet and this has caused
> > issues over expectations over a slew of APIs. As you note above
> > your own defined 'semantics' so far for memremap are just that there are
> > no i/o side effects, when the mapped memory is just a pointer to memory,
> > as such I do not think its fair to set the excpectations that we'll
> > switch all other ioremap_*() callers to the same memremap() API. Now,
> > it may be a good idea to use something similar, ie, to pass flags,
> > but setting the expectations outright to move to memremap() without having
> > any agreement having been made over semantics seems uncalled for at this
> > point in time, specially when you are noting that the expectations for
> > both sets of calls are different.
> >
> > So perhaps:
> >
> > "
> > Outside of ioremap() and ioremap_nocache(), all other ioremap_<type>()
> > variant calls are seeking memory-like semantics (e.g.  speculative
> > reads, and prefetching permitted) and all calls expecations currently
> > differ depending on architecture. Once and if we get agreement on such
> > semantics we may be able to move such ioremap_*() variant calls to
> > a similar API where the semantics required are clearly spelled out
> > and well defined and passed as arguments.
> 
> I still think ioremap_wc(), and now ioremap_uc(), are special and are
> not obvious candidates for conversion to memremap().

OK great, then we're in strong agreement, so removing the "Outside of
ioremap() and... over time" might be best then? Otherwise what I posted
seems to reflect the state of affairs better?

> >> +void *memremap(resource_size_t offset, size_t size, unsigned long flags)
> >> +{
> >> +     int is_ram = region_is_ram(offset, size);
> >> +     void *addr = NULL;
> >> +
> >> +     if (is_ram < 0) {
> >> +             WARN_ONCE(1, "memremap attempted on mixed range %pa size: %zu\n",
> >> +                             &offset, size);
> >> +             return NULL;
> >> +     }
> >> +
> >> +     /* Try all mapping types requested until one returns non-NULL */
> >> +     if (flags & MEMREMAP_CACHE) {
> >> +             flags &= ~MEMREMAP_CACHE;
> >> +             /*
> >> +              * MEMREMAP_CACHE is special in that it can be satisifed
> >> +              * from the direct map.  Some archs depend on the
> >> +              * capability of memremap() to autodetect cases where
> >> +              * the requested range is potentially in "System RAM"
> >> +              */
> >> +             if (is_ram)
> >> +                     addr = __va(offset);
> >
> > The no MMU case seems to match this, can that be switch to this?
> 
> Hmm, it may be possible to consolidate all the NOMMU cases here.
> That's a good incremental cleanup once these base patches are merged.

Sure fine by me.

  Luis