bug in identity map for 4KB pages?

[Mark Rutland]

On Wed, Jul 29, 2015 at 12:49:53PM +0100, Ard Biesheuvel wrote:
> On 29 July 2015 at 13:42, Mark Rutland <mark.rutland at arm.com> wrote:
> > On Wed, Jul 29, 2015 at 08:47:10AM +0100, Ard Biesheuvel wrote:
> >> On 29 July 2015 at 04:37, Stuart Yoder <stuart.yoder at freescale.com> wrote:
> >> > Our system has RAM at a high address, and previously required using 48-bit VA
> >> > in order to have an idmap that covered all of RAM.
> >> >
> >> > In testing on 4.2-rc4, which now contains support for the increased VA range
> >> > of the idmap without using 48-bit VA, I'm finding that things work for
> >> > 64KB pages, but do not for 4KB pages.
> >> >
> >> > Is there any known limitation here with 4KB pages?  Any ideas?
> >> >
> >>
> >> You probably have memory at 0x8000_0000 and at 0x80_8000_0000, right?
> >> So the physical arrangement still requires more than the 39 bits of
> >> virtual address space you get with 3 levels, even if the ID map can
> >> cope now. That is why you get the 0x40_0000_0000 virtual address:
> >> __phys_to_virt() just wraps to a positive number.
> >
> > Ah, I see. So it's the linear mapping rather than the idmap which is the
> > problem.
> >
> > We cut memory which would fall below the start of the linear map in
> > early_init_dt_add_memory_arch, by cutting memory below phys_offset.
> >
> > It looks like we already have the logic for cutting memory beyond the
> > end of the linear map, so we should just need to override the limit.
> >
> > Stuart, does the below patch prevent the panic you see?
> >
> 
> Wouldn't something like this make more sense?
> 
> diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> index 597831bdddf3..64480b65ef17 100644
> --- a/arch/arm64/mm/init.c
> +++ b/arch/arm64/mm/init.c
> @@ -158,6 +158,15 @@ early_param("mem", early_mem);
> 
>  void __init arm64_memblock_init(void)
>  {
> +       /*
> +        * Remove the memory that we will not be able to cover
> +        * with the linear mapping.
> +        */
> +       const s64 linear_region_size = -(s64)PAGE_OFFSET;
> +
> +       memblock_remove(0, memstart_addr);
> +       memblock_remove(memstart_addr + linear_region_size, ULLONG_MAX);

The maths is more correct, certainly.

I'd prefer to handle the truncation in early_init_dt_add_memory_arch
because it already has the requisite logic, and keeping the truncation
in one place should keep things less confusing.

To that end I'll respin my patch using PAGE_OFFSET to determine the
physical memory limit.

Unless there's something I've missed?

Thanks,
Mark.