Re: [RFC][PATCH] arm64: Add support for CONFIG_DEBUG_VIRTUAL

[Laura Abbott]

On 10/29/2016 02:39 AM, Ard Biesheuvel wrote:
> On 28 October 2016 at 23:07, Laura Abbott <labbott@xxxxxxxxxx> wrote:
> > > > > diff --git a/arch/arm64/mm/physaddr.c b/arch/arm64/mm/physaddr.c
> > > > > new file mode 100644
> > > > > index 0000000..6c271e2
> > > > > --- /dev/null
> > > > > +++ b/arch/arm64/mm/physaddr.c
> > > > > @@ -0,0 +1,17 @@
> > > > > +#include <linux/mm.h>
> > > > > +
> > > > > +#include <asm/memory.h>
> > > > > +
> > > > > +unsigned long __virt_to_phys(unsigned long x)
> > > > > +{
> > > > > +    phys_addr_t __x = (phys_addr_t)x;
> > > > > +
> > > > > +    if (__x & BIT(VA_BITS - 1)) {
> > > > > +        /* The bit check ensures this is the right range */
> > > > > +        return (__x & ~PAGE_OFFSET) + PHYS_OFFSET;
> > > > > +    } else {
> > > > > +        VIRTUAL_BUG_ON(x < kimage_vaddr || x > (unsigned long)_end);
> > > >
> > > >
> > > > IIUC, in (3) you were asking if the last check should be '>' or '>='?
> > > >
> > > > To match high_memory, I suspect the latter, as _end doesn't fall within
> > > > the mapped virtual address space.
> > >
> > > I was actually concerned about if _end would be correct with KASLR.
> > > Ard confirmed that it gets fixed up to be correct. I'll change the
> > > check to check for >=.
> >
> > While testing this, I found two places with __pa(_end) to get bounds,
> > one in arm64 code and one in memblock code. x86 gets away with this
> > because memblock is actually __pa_symbol and x86 does image placement
> > different and can check against the maximum image size. I think
> > including _end in __pa_symbol but excluding it from the generic
> > __virt_to_phys makes sense. It's a bit nicer than doing _end - 1 +
> > 1 everywhere.
>
> Could we redefine __pa_symbol() under CONFIG_DEBUG_VIRTUAL to
> something that checks (x >= kimage_vaddr + TEXT_OFFSET || x <=
> (unsigned long)_end), i.e., reject linear virtual addresses? (Assuming
> my understanding of the meaning of __pa_symbol() is correct)

Yes, that's going to be my approach. I originally prototyped this
with just x >= kimage_vaddr but kimage_vaddr + TEXT_OFFSET is
a tighter bound.

Thanks,
Laura