Re: [PATCH v5] mm/rmap: do not add fully unmapped large folio to deferred split list

[David Hildenbrand]

On 01.05.24 15:38, Zi Yan wrote:
> On 1 May 2024, at 9:24, Alexander Gordeev wrote:
>
> > On Fri, Apr 26, 2024 at 03:02:53PM -0400, Zi Yan wrote:
> >
> > Hi Zi,
> >
> > It increasingly looks like this commit is crashing on s390 since
> > 2024-04-30 in linux-next. If I do not miss something - since it
> > was included in mm-everything.
> >
> > > @@ -1553,9 +1558,10 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
> > >   		 * page of the folio is unmapped and at least one page
> > >   		 * is still mapped.
> > >   		 */
> > > -		if (folio_test_large(folio) && folio_test_anon(folio))
> > > -			if (level == RMAP_LEVEL_PTE || nr < nr_pmdmapped)
> > > -				deferred_split_folio(folio);
> > > +		if (folio_test_anon(folio) &&
> > > +		    list_empty(&folio->_deferred_list) &&
> >
> > An attempt to reference folio->_deferred_list causes the crash below.
>
> So if you remove this line, the crash no longer happens? It looks strange to
> me that referencing a anonymous folio's _deferred_list would cause a crash.
> Hmm, unless the folio is order-0.
>
> Can you try the patch below and see if it fixes the crash? It moves partially_mapped
> ahead to exclude order-0 folios.
>
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 087a79f1f611..2d27c92bb6d5 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1557,9 +1557,8 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>                   * page of the folio is unmapped and at least one page
>                   * is still mapped.
>                   */
> -               if (folio_test_anon(folio) &&
> -                   list_empty(&folio->_deferred_list) &&
> -                   partially_mapped)
> +               if (folio_test_anon(folio) && partially_mapped &&
> +                   list_empty(&folio->_deferred_list))
>                          deferred_split_folio(folio);

Yes, that should fix it and is the right thing to do. For small folios, 
partially_mapped will always be false.

--
Cheers,

David / dhildenb