Re: [PATCH v2 0/2] mm: don't migrate folios containing hwpoisoned pages

[Yuanhe Shu]

On 10/28/26 09:02, David Hildenbrand wrote:
> Why is that a problem?
>
> Why do you care?

The flagged state is not debug-only.  A real error on an isolated
folio leaves PG_hwpoison set.  The data is still readable.  A THP
that cannot split is left on the LRU flagged, by design.  A copy
then propagates the poison silently: no marker, no hwpoison PTE,
no SIGBUS.

Thinking about it again, the copy layer is indeed the better
place.  So how about this: drop patch 2, and put the flag check in
the copy implementations instead.  The fallback patch already checks
is_page_hwpoison() before and after the copy on the !copy_mc archs.
The same checks in the copy_mc implementations would cover x86_64
and PPC64.  folio_mc_copy() then propagates the refusal to the
migration core.  Patch 1, the compaction skip, stands alone.  Does
that work for you?

-- 
Yuanhe