Re: [PATCH v8 2/6] mm/memory-failure: surface unhandlable kernel pages as -ENOTRECOVERABLE

[Breno Leitao <leitao@debian.org>]

On Fri, Jun 05, 2026 at 11:42:53AM +0200, David Hildenbrand (Arm) wrote:
> On 6/5/26 11:35, Breno Leitao wrote:
> > On Wed, Jun 03, 2026 at 10:33:04AM +0800, Miaohe Lin wrote:
> >> On 2026/6/2 17:41, David Hildenbrand (Arm) wrote:
> >>>
> >>> Races are fine. We might miss some pages, but that can happen on races either way.
> >>>
> >>>
> >>> I'd just do something like
> >>>
> >>> if (PageReserved(page))
> >>> 	return true;
> >>>
> >>> head = compound_head(page);
> >>
> >> If @head is split just after compound_head. And then @head is freed into buddy and re-allocated as slab
> >> page while @page is still in the buddy. We would panic on this scene as @head is PageSlab. But we were
> >> supposed to successfully handle @page. Or am I miss something?
> > 
> > You're right that it is racy, but I think it is an acceptable race here.
> > 
> 
> I mean, any such races can currently already happen one way or the other?
> 
> Really, the only way to not get races is to tryget the (compound)page,
> revalidate that the page is still part of the compound page.
> 
> I'm not sure if that's really a good idea.
> 
> But my memory is a bit vague in which scenarios we already hold a page reference
> here to prevent any concurrent freeing?

No, we don't hold one here in the case that matters.

HWPoisonKernelOwned() runs at the very top of get_any_page(), before
try_again: and before __get_hwpoison_page(). The first refcount taken in
the whole path is the folio_try_get() inside __get_hwpoison_page(), which
runs *after* the short-circuit.

So get_any_page() itself never holds a reference at the check -- the only way
one exists is if the caller passed MF_COUNT_INCREASED (count_increased ==
true).

So on the MCE/GHES path -- the one this panic option exists for -- no
reference is held when HWPoisonKernelOwned() does its compound_head() +
PageSlab()/PageTable()/PageLargeKmalloc() checks.

Given that, I'd rather keep it racy and take no refcount than add a
tryget + revalidate purely for this check. As I've said earleir, an operator
who enabled it has chosen to crash rather than run on corrupted memory;
mis-attributing one such rare, genuinely-poisoned page is within that contract.