Re: [PATCH] arm64: mm: mark tramp_pg_dir read-only

[Ard Biesheuvel]

On 19 June 2018 at 17:37, Will Deacon <will.deacon@xxxxxxx> wrote:
> On Tue, Jun 19, 2018 at 05:29:03PM +0200, Ard Biesheuvel wrote:
>> On 19 June 2018 at 17:28, Will Deacon <will.deacon@xxxxxxx> wrote:
>> > On Tue, Jun 19, 2018 at 05:23:41PM +0200, Ard Biesheuvel wrote:
>> >> On 19 June 2018 at 17:20, Will Deacon <will.deacon@xxxxxxx> wrote:
>> >> > Hi Ard,
>> >> >
>> >> > Sorry, I forgot to reply to this.
>> >> >
>> >> > On Wed, May 30, 2018 at 11:53:20AM +0200, Ard Biesheuvel wrote:
>> >> >> On 30 May 2018 at 11:14, Will Deacon <will.deacon@xxxxxxx> wrote:
>> >> >> > On Wed, May 30, 2018 at 12:48:06PM +0800, YaoJun wrote:
>> >> >> >> To protect against KSMA(Kernel Space Mirroring Attack), make
>> >> >> >> tramp_pg_dir read-only. The principle of KSMA is to insert a
>> >> >> >> carefully constructed PGD entry into the translation table.
>> >> >> >> The type of this entry is block, which maps the kernel text
>> >> >> >> and its access permissions bits are 01. The user process can
>> >> >> >> then modify kernel text directly through this mapping. In this
>> >> >> >> way, an arbitrary write can be converted to multiple arbitrary
>> >> >> >> writes.
>> >> >> >>
>> >> >> >> Signed-off-by: YaoJun <yaojun8558363@xxxxxxxxx>
>> >> >> >> ---
>> >> >> >>  arch/arm64/mm/mmu.c | 4 ++++
>> >> >> >>  1 file changed, 4 insertions(+)
>> >> >> >>
>> >> >> >> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>> >> >> >> index 2dbb2c9f1ec1..ac4b22c7e435 100644
>> >> >> >> --- a/arch/arm64/mm/mmu.c
>> >> >> >> +++ b/arch/arm64/mm/mmu.c
>> >> >> >> @@ -551,6 +551,10 @@ static int __init map_entry_trampoline(void)
>> >> >> >>       __create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
>> >> >> >>                            prot, pgd_pgtable_alloc, 0);
>> >> >> >>
>> >> >> >> +     update_mapping_prot(__pa_symbol(tramp_pg_dir),
>> >> >> >> +                             (unsigned long)tramp_pg_dir,
>> >> >> >> +                             PGD_SIZE, PAGE_KERNEL_RO);
>> >> >> >
>> >> >> > Hmm, I like the idea but is there a risk that the page table has been mapped
>> >> >> > as part of a block entry, which we can't safely split at this point (i.e.
>> >> >> > we'll run into one of the BUG_ONs in the mapping code)?
>> >> >> >
>> >> >>
>> >> >> We'd need to create a separate segment for it initially so the mapping
>> >> >> is already at the right granularity.
>> >> >
>> >> > Why do you think that's the case? I can't see anything that guarantees this
>> >> > for the page table itself.
>> >> >
>> >>
>> >> We'd need to pass NO_BLOCK_MAPPINGS to map_kernel_segment(),
>> >> obviously, but that shouldn't hurt since that segment is relatively
>> >> tiny anyway.
>> >
>> > Ah right, with NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS, I agree that we're good.
>> > Ideally, we'd move {idmap,swapper,tramp}_pg_dir into .rodata...
>> >
>>
>> idmap and tramp yes, but swapper needs to be modifiable at runtime, no?
>
> Right, but couldn't we swizzle the permissions in e.g. set_pmd? We could
> even predicate that on a sanity check of the prot.
>

Swizzle the permissions of the entire .rodata segment? That sounds
doable, but there is a whole class of data that belongs in this
category, and I think PaX/grsecurity had an API for that (but I don't
think anyone is upstreaming that yet). So let's not reinvent that
wheel for swapper_pg_dir only.