[PATCH v6 07/26] mm: Preserve the PG_arch_* flags in __split_huge_page_tail()

[David Hildenbrand]

On 06.07.20 18:30, Catalin Marinas wrote:
> On Mon, Jul 06, 2020 at 04:16:13PM +0200, David Hildenbrand wrote:
>> On 03.07.20 17:36, Catalin Marinas wrote:
>>> When a huge page is split into normal pages, part of the head page flags
>>> are transferred to the tail pages. However, the PG_arch_* flags are not
>>> part of the preserved set.
>>>
>>> PG_arch_1 is currently used by the arch code to handle cache maintenance
>>> for user space (either for I-D cache coherency or for D-cache aliases
>>> consistent with the kernel mapping). Since splitting a huge page does
>>> not change the physical or virtual address of a mapping, additional
>>> cache maintenance for the tail pages is unnecessary. Preserving the
>>> PG_arch_1 flag from the head page in the tail pages would not break the
>>> current use-cases.
>>
>> ^ is fairly arm64 specific, no? (I remember that the semantics are
>> different e.g., on s390x).
> 
> Not entirely arm64 specific. Apart from s390 and x86, I think all the
> other architectures use this flag for cache maintenance (I guess they
> followed the cachetlb.rst suggestion). My understanding of the s390 and
> x86 is that transferring this flag from the head of a compound page to
> the tail pages should not cause any issue. We don't even document
> anywhere that this flag is meant to disappear on huge page splitting. I
> guess no-one noticed because clearing it is relatively benign.

On s390x, PG_arch_1 indicates (s390/kernel/uv.c:arch_make_page_accessible())
- kernel page tables
- for hugetlbfs pages, that storage keys are initialized for that page
  (IIRC KVM only)
- a user space page might be encrypted/secure (KVM only)

The latter does not support hugetlbfs/THP. KVM does not support THP. So
on s390x the bit should never be set in that context and, therefore,
also won't be affected by this change.

> 
> But if there are concerns, I'm happy to guard it with something like
> __ARCH_WANT_PG_ARCH_HEAD_TAIL (I need to think of a more suggestive
> name).

I guess we can avoid that if we properly check+document all users.
(ignoring x86 and s390x behavior here might be dangerous, although my
gut feeling is that it's ok for both)

-- 
Thanks,

David / dhildenb