[PATCH] arm64/mm: Intercept pfn changes in set_pte_at()

[Mark Rutland]

On Tue, Nov 22, 2022 at 09:57:49AM +0000, Will Deacon wrote:
> On Tue, Nov 22, 2022 at 01:43:17PM +0530, Anshuman Khandual wrote:
> > 
> > 
> > On 11/18/22 19:43, Will Deacon wrote:
> > > On Wed, Nov 16, 2022 at 08:40:01AM +0530, Anshuman Khandual wrote:
> > >> Changing pfn on a user page table mapped entry, without first going through
> > >> break-before-make (BBM) procedure is unsafe. This just updates set_pte_at()
> > >> to intercept such changes, via an updated pgattr_change_is_safe(). This new
> > >> check happens via __check_racy_pte_update(), which has now been renamed as
> > >> __check_safe_pte_update().
> > >>
> > >> Cc: Catalin Marinas <catalin.marinas at arm.com>
> > >> Cc: Will Deacon <will at kernel.org>
> > >> Cc: Mark Rutland <mark.rutland at arm.com> 
> > >> Cc: Andrew Morton <akpm at linux-foundation.org> 
> > >> Cc: linux-arm-kernel at lists.infradead.org
> > >> Cc: linux-kernel at vger.kernel.org
> > >> Signed-off-by: Anshuman Khandual <anshuman.khandual at arm.com>
> > >> ---
> > >> This applies on v6.1-rc4
> > >>
> > >>  arch/arm64/include/asm/pgtable.h | 8 ++++++--
> > >>  arch/arm64/mm/mmu.c              | 8 +++++++-
> > >>  2 files changed, 13 insertions(+), 3 deletions(-)
> > > 
> > > I remember Mark saying that BBM is sometimes violated by the core code in
> > > cases where the pte isn't actually part of a live pgtable (e.g. if it's on
> > > the stack or part of a newly allocated table). Won't that cause false
> > > positives here?
> > 
> > Could you please elaborate ? If the pte is not on a live page table, then
> > pte_valid() will return negative on such entries. So any update there will
> > be safe. I am wondering, how this change will cause false positives which
> > would not have been possible earlier.
> 
> I don't think pte_valid() will always return false for these entries.
> Consider, for example, ptes which are valid but which live in a table that
> is not reachable by the MMU. I think this is what Mark had in mind, but it
> would be helpful if he could chime in with the specific example he ran into.

Yup -- that was the case I had in mind. IIRC I hit that in the past when trying
to do something similar, but I can't recall exactly where that was. I suspect
that was probably to do with page migration or huge page splitting/merging.

Looking around, at least __split_huge_zero_page_pmd() and
__split_huge_pmd_locked() do something like that, creating a temporary pmd
entry on the stack, populating a table of non-live but valid ptes, then
plumbing it into the real pmd.

We'd need to check that there aren't other cases like that.

Thanks,
Mark.