Re: [PATCH] mm: avoid poison consumption when splitting THP

[David Hildenbrand]

On 11.09.25 04:14, Andrew Zaborowski wrote:
> Handling a memory failure pointing inside a huge page requires splitting
> the page.  The splitting logic uses a mechanism, implemented in
> migrate.c:try_to_map_unused_to_zeropage(), that inspects contents of
> individual pages to find zero-filled pages.  The read access to the
> contents may cause a new, synchronous exception like an x86 Machine
> Check, delivered before the initial memory_failure() finishes, ending
> in a crash.
>
> Luckily memory_failure() already sets the has_hwpoisoned flag on the
> folio right before try_to_split_thp_page().  Don't enable the shared
> zeropage mechanism (RMP_USE_SHARED_ZEROPAGE flag) down in
> __split_unmapped_folio() when the original folio has has_hwpoisoned.
>
> Note: we're disabling a potentially useful feature, some of the
> individual pages that aren't poisoned might be zero-filled.  One
> argument for not trying to add a mechanism to maybe re-scan them later,
> apart from code cost, is that the owning process is likely being
> killed and the memory released.
>
> Signed-off-by: Andrew Zaborowski <balrogg+code@xxxxxxxxx>
> ---

I would suggest just checking whether the page (PageHWPoison()) is 
poisoned before doing the check for zero. If set, just treat it as non-zero.

No need to stop the split.

You'll have to do that in two locations.

No need to mess with RMP_USE_SHARED_ZEROPAGE

--
Cheers

David / dhildenb