Re: [PATCH] mm/memory_hotplug: don't offline a hwpoisoned page still in the page cache

[Mani Tofigh <mt4022@columbia.edu>]

On 11/2/26 1:56 AM, David Hildenbrand (Arm) wrote:
> 
> What if we still have the page pinned somehow? (vmscplice etc).
> 
> Wouldn't it be better to make sure that poisoning code holds exactly one
> reference (whoever set hwpoison) and check here that the refcount is 1 ?
> 
> Pagecache references would turn it > 1.


Yes right, a pin isn't covered (neither before nor with this patch).

I don't think refcount == 1 can replace the mapping check though. In the
race from the changelog memory_failure() has set PG_hwpoison but is 
still waiting in get_hwpoison_page(), so it doesn't hold its reference 
yet, and a clean page-cache folio then has just the one reference from 
the page cache. memory_failure() also doesn't take a reference on the 
kernel pages it ignores, and a poisoned page that gets freed later sits 
at 0.

So I'd keep the mapping check and add "refcount <= 1" on top as a
follow-up, since that also needs do_migrate_range() to drop the 
reference it takes when it takes a poisoned folio off the LRU. Does that 
sound good?

I'll send a v2 soon with a small do_migrate_range() change from the 
bot's review.

Mani