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

[Luis R. Rodriguez]

On Wed, Jul 29, 2015 at 06:00:04PM -0600, Toshi Kani wrote:
> On Wed, 2015-07-29 at 23:43 +0200, Luis R. Rodriguez wrote:
> > On Wed, Jul 29, 2015 at 03:00:38PM -0600, Toshi Kani wrote:
> > > 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. 
> > > 
> > 
> > That would depend on the purpose of the region and the driver developer 
> > should in theory know best. One issue with this of course is that, as 
> > we've discovered, the semantics of on the ioremap*() variant front 
> > regarding cache types is not clearly well defined, or at least it may be 
> > only well defined implicitly on x86 only, so the driver developer can only 
> > *hope* for the best across architectures. Our ambiguity on our semantics 
> > on ioremap*() variants therefore means driver developers can resonably be 
> > puzzled by what fallbacks to use. That also means architectures 
> > maintainers should not whip driver developers for any misuse. Such 
> > considerations are why although we're now revisiting semantics for 
> > ioremap*() variants I was in hopes we could be at least somewhat
> > pedantic about memremap() semantics.
> 
> I agree.  However, there are a few exceptions like /dev/mem, which can map a
> target range without knowledge.

Still, the expectation to require support for overlapping ioremap() calls would
seem to be more of an exception than the norm, so I'd argue that requiring
callers who know they do need overlapping support to be explicit about it may
help us simplify expecations on semantics in other areas of the kernel.

> > For instance since memremap() only has 2 types right now can a respective
> > fallback be documented as an alternative to help with this ? Or can we not
> > generalize this ?  One for MEMREMAP_WB and one for MEMREMAP_WT ?
> 
> Yes, if a target range can be only mapped by memremap().  However, there is
> no restriction that a range can be mapped with a single interface alone. 
>  For example, a range can be mapped with remap_pfn_range() to user space
> with any cache type.  So, in theory, memremap() can overlap with any other
> types.

Shouldn't that be an issue or area of concern ? It seems the flakiness on
ioremap() and its wide array flexibility would spill over the any semantics
which folks would be trying to set out with memremap(). That should not stop
the evolution of memremap() though, just pointing out that perhaps we should be
a bit more restrictive over how things can criss-cross and who areas can do it.

> > > ioremap() falls back to the cache type of an existing mapping to avoid
> > > aliasing.
> > 
> > Does that assume an existing ioremap*() call was used on a bigger range?
> > Do you know if that happens to only be the case for x86 (I'd think so)
> > or if its the same for other architectures ?
> 
> In the /dev/mem example, it is remap_pfn_range().  I think other archs have
> the same issue, but I do not know if they fall back in case of overlapping
> call. 

What should happen if remap_pfn_range() was used with pgprot_writecombine()
and later memremap() is used for instance with a smaller overlappin section,
or perhaps bigger?

  Luis