mmap question on ppc440

[Benjamin Herrenschmidt]

On Tue, 2007-11-06 at 07:50 +0100, Stefan Roese wrote:
> On Monday 05 November 2007, Josh Boyer wrote:
> > > I am attempting to access the CPLD on the AMCC Sequoia board from
> > > user-land.  I open /dev/mem, and mmap it, then try to access the
> > > resulting pointer.  That works fine when accessing physical addresses
> > > that correspond to RAM, but as soon as I try to access the CPLD at
> > > physical address 0xc0000000, I get an infinite machine check.
> >
> > That's because the CPLD is actually at physical address 0x1C0000000.
> > Yay for 36-bit physical addresses.
> 
> Right. Are you using arch/ppc or arch/powerpc? If it's arch/ppc you could
> give the following patch a try:
> 
> @@ -275,6 +275,14 @@
>  {
>         size_t size = vma->vm_end - vma->vm_start;
> 
> +#if defined(CONFIG_44x) && !defined(CONFIG_PPC_MERGE)
> +       /*
> +        * 2006-08-07: sr
> +        * Needed on 44x-er systems for 36bit addresses (like pci on 440gx)
> +        */
> +       vma->vm_pgoff = (fixup_bigphys_addr(vma->vm_pgoff << PAGE_SHIFT, size) >> PAGE_SHIFT);
> +#endif
> +
>         if (!valid_mmap_phys_addr_range(vma->vm_pgoff, size))
>                 return -EINVAL;

I think we need to ditch the bigphys fixup stuff and come up with a way
to make /dev/mem work with the actual 36 bits offsets (after all, it's
all pgoff, it should work).

The other problem is X of course... 32 bits X server currently cannot
cope with physical addresses > 32 bits at all. They will just blow up or
randomly scribble over /dev/mem.

The solution is libpciaccess and the new pci-rework branch of X which
uses it, but I haven't had a chance to test that properly yet on 4xx.

Ben.