Re: [PATCH v2 3/4] arch: introduce memremap()

[Arnd Bergmann <arnd@arndb.de>]

On Saturday 30 May 2015 14:39:48 Dan Williams wrote:
> On Sat, May 30, 2015 at 2:00 PM, Arnd Bergmann <arnd@arndb.de> wrote:
> > On Saturday 30 May 2015, Dan Williams wrote:
> >>
> >> +/*
> >> + * memremap() is "ioremap" for cases where it is known that the resource
> >> + * being mapped does not have i/o side effects and the __iomem
> >> + * annotation is not applicable.
> >> + */
> >> +
> >> +static inline void *memremap(resource_size_t offset, size_t size)
> >> +{
> >> +       return (void __force *) ioremap(offset, size);
> >> +}
> >> +
> >> +static inline void *memremap_nocache(resource_size_t offset, size_t size)
> >> +{
> >> +       return (void __force *) ioremap_nocache(offset, size);
> >> +}
> >> +
> >> +static inline void *memremap_cache(resource_size_t offset, size_t size)
> >> +{
> >> +       return (void __force *) ioremap_cache(offset, size);
> >> +}
> >> +
> >
> > There are architectures on which the result of ioremap is not necessarily
> > a pointer, but instead indicates that the access is to be done through
> > some other indirect access, or require special instructions. I think implementing
> > the memremap() interfaces is generally helpful, but don't rely on the
> > ioremap implementation.
> 
> Is it enough to detect the archs where ioremap() does return an
> otherwise usable pointer and set ARCH_HAS_MEMREMAP, in the first take
> of this introduction?  Regardless, it seems that drivers should have
> Kconfig dependency checks for archs where ioremap can not be used in
> this manner.

Yes, that should work.

> > Adding both cached an uncached versions is also dangerous, because you
> > typically get either undefined behavior or a system checkstop when a
> > single page is mapped both cached and uncached at the same time. This
> > means that doing memremap() or memremap_nocache() on something that
> > may be part of the linear kernel mapping is a bug, and we should probably
> > check for that here.
> 
> Part of the reason for relying on ioremap() was to borrow its internal
> checks to fail attempts that try to remap ranges that are already in
> the kernel linear map.  Hmm, that's a guarantee x86 ioremap gives, but
> maybe that's not universal?

I haven't seen that check elsewhere. IIRC what ioremap() guarantees on ARM
is that if there is an existing boot-time mapping (similar to x86 fixmap,
but more commonly used), we use the same flags in the new ioremap and
override the ones that are provided by the caller.

	Arnd

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>