[PATCH v3] arm64: mm: move zero page from .bss to right before swapper_pg_dir

[Ard Biesheuvel]

On 10 October 2016 at 00:10, Mark Rutland <mark.rutland at arm.com> wrote:
> On Fri, Oct 07, 2016 at 10:31:14AM +0100, Ard Biesheuvel wrote:
>> On 12 September 2016 at 17:15, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
>> > Move the statically allocated zero page from the .bss section to right
>> > before swapper_pg_dir. This allows us to refer to its physical address
>> > by simply reading TTBR1_EL1 (which always points to swapper_pg_dir and
>> > always has its ASID field cleared), and subtracting PAGE_SIZE.
>> >
>> > To protect the zero page from inadvertent modification, carve out a
>> > segment that covers it as well as idmap_pg_dir[], and mark it read-only
>> > in both the primary and the linear mappings of the kernel.
>
> [...]
>
>> > -       map_kernel_segment(pgd, _data, _end, PAGE_KERNEL, &vmlinux_data);
>> > +       map_kernel_segment(pgd, _data, __robss_start, PAGE_KERNEL,
>> > +                          &vmlinux_data);
>> > +       map_kernel_segment(pgd, __robss_start, __robss_end, PAGE_KERNEL_RO,
>> > +                          &vmlinux_robss);
>>
>> I realised it is actually unnecessary to map the idmap and the zero
>> page into the kernel mapping, so we could drop this line.
>
> Given that drivers use the zero page, I wouldn't be entirely surprised to see
> phys_to_virt(virt_to_phys(zero_page)) happen indirectly, and the end result
> read. Are we sure that doesn't happen anywhere?
>

That conversion would actually still work, it would be the direct
reference that is left unmapped. But given that it is mapped R/O
anyway (which is the whole point of the patch), it makes more sense to
follow the principle of least surprise, and make the direct symbol
dereference work as expected.

> For the idmap, I think we might walk that were we to take a fault (though
> perhaps we don't). Otherwise, unless we add a sysfs walker for it I guess we
> don't strictly need it in the linear map.
>

Likewise, this is the kernel mapping not the linear mapping. But given
how little this matters, please forget I said anything :-)