[PATCH] arm64: allow ioremap_cache() to use existing RAM mappings

[Catalin Marinas]

On Wed, Oct 23, 2013 at 02:46:12PM +0100, Mark Salter wrote:
> On Wed, 2013-10-23 at 10:18 +0100, Catalin Marinas wrote:
> > On Mon, 2013-10-21 at 14:36 +0100, msalter at redhat.com wrote:
> > > diff --git a/arch/arm64/mm/ioremap.c b/arch/arm64/mm/ioremap.c
> > > index 1725cd6..fb44b3d 100644
> > > --- a/arch/arm64/mm/ioremap.c
> > > +++ b/arch/arm64/mm/ioremap.c
> > > @@ -79,6 +79,21 @@ void __iounmap(volatile void __iomem *io_addr)
> > >  {
> > >  	void *addr = (void *)(PAGE_MASK & (unsigned long)io_addr);
> > >  
> > > +	/* Nothing to do for normal memory. See ioremap_cache() */
> > > +	if (pfn_valid(__virt_to_phys(addr) >> PAGE_SHIFT))
> > > +		return;
> > 
> > addr here can be some I/O address mapped previously, so __virt_to_phys()
> > is not valid (you don't actually get the pfn by shifting).
> > 
> 
> Yeah, that's ugly. The thought was that only the kernel mapping of RAM
> would yield a valid address from __virt_to_phys(). Anything else, like
> a mapping of I/O space would lead to an invalid PFN. There's probably a
> clearer way of doing that that. Other than that, is the general concept
> of the patch reasonable?

I think the concept is fine. You could change the check on
VMALLOC_START/END or just always create a new mapping as long as it has
the same caching attributes (PROT_NORMAL).

-- 
Catalin