Re: [PATCH v2] mm: do not update memcg stats for NR_{FILE/SHMEM}_PMDMAPPED

[David Hildenbrand]

On 06.05.24 22:18, Yosry Ahmed wrote:
> On Mon, May 06, 2024 at 09:50:10PM +0200, David Hildenbrand wrote:
> > On 06.05.24 21:29, Yosry Ahmed wrote:
> > > Previously, all NR_VM_EVENT_ITEMS stats were maintained per-memcg,
> > > although some of those fields are not exposed anywhere. Commit
> > > 14e0f6c957e39 ("memcg: reduce memory for the lruvec and memcg stats")
> > > changed this such that we only maintain the stats we actually expose
> > > per-memcg via a translation table.
> > >
> > > Additionally, commit 514462bbe927b ("memcg: warn for unexpected events
> > > and stats") added a warning if a per-memcg stat update is attempted for
> > > a stat that is not in the translation table. The warning started firing
> > > for the NR_{FILE/SHMEM}_PMDMAPPED stat updates in the rmap code. These
> > > stats are not maintained per-memcg, and hence are not in the translation
> > > table.
> > >
> > > Do not use __lruvec_stat_mod_folio() when updating NR_FILE_PMDMAPPED and
> > > NR_SHMEM_PMDMAPPED. Use __mod_node_page_state() instead, which updates
> > > the global per-node stats only.
> > >
> > > Reported-by: syzbot+9319a4268a640e26b72b@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Closes: https://lore.kernel.org/lkml/0000000000001b9d500617c8b23c@xxxxxxxxxx
> > > Fixes: 514462bbe927 ("memcg: warn for unexpected events and stats")
> > > Acked-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> > > Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > > ---
> > >    mm/rmap.c | 15 +++++++++------
> > >    1 file changed, 9 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/mm/rmap.c b/mm/rmap.c
> > > index 12be4241474ab..ed7f820369864 100644
> > > --- a/mm/rmap.c
> > > +++ b/mm/rmap.c
> > > @@ -1435,13 +1435,14 @@ static __always_inline void __folio_add_file_rmap(struct folio *folio,
> > >    		struct page *page, int nr_pages, struct vm_area_struct *vma,
> > >    		enum rmap_level level)
> > >    {
> > > +	pg_data_t *pgdat = folio_pgdat(folio);
> > >    	int nr, nr_pmdmapped = 0;
> > >    	VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
> > >    	nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
> > >    	if (nr_pmdmapped)
> > > -		__lruvec_stat_mod_folio(folio, folio_test_swapbacked(folio) ?
> > > +		__mod_node_page_state(pgdat, folio_test_swapbacked(folio) ?
> > >    			NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
> > >    	if (nr)
> > >    		__lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr);
> > > @@ -1493,6 +1494,7 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
> > >    		enum rmap_level level)
> > >    {
> > >    	atomic_t *mapped = &folio->_nr_pages_mapped;
> > > +	pg_data_t *pgdat = folio_pgdat(folio);
> > >    	int last, nr = 0, nr_pmdmapped = 0;
> > >    	bool partially_mapped = false;
> > >    	enum node_stat_item idx;
> > > @@ -1540,13 +1542,14 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
> > >    	}
> > >    	if (nr_pmdmapped) {
> > > +		/* NR_{FILE/SHMEM}_PMDMAPPED are not maintained per-memcg */
> > >    		if (folio_test_anon(folio))
> > > -			idx = NR_ANON_THPS;
> > > -		else if (folio_test_swapbacked(folio))
> > > -			idx = NR_SHMEM_PMDMAPPED;
> > > +			__lruvec_stat_mod_folio(folio, NR_ANON_THPS, -nr_pmdmapped);
> > >    		else
> > > -			idx = NR_FILE_PMDMAPPED;
> > > -		__lruvec_stat_mod_folio(folio, idx, -nr_pmdmapped);
> > > +			__mod_node_page_state(pgdat,
> >
> > folio_pgdat(folio) should fit here easily. :)
> >
> > But I would actually suggest something like the following in mm/rmap.c
>
> I am not a big fan of this. Not because I don't like the abstraction,
> but because I think it doesn't go all the way. It abstracts a very
> certain case: updating nr_pmdmapped for file folios.

Right. It only removes some of the ugliness ;)

> I think if we opt for abstracting the stats updates in mm/rmap.c, we
> should go all the way with something like the following (probably split
> as two patches: refactoring then bug fix). WDYT about the below?
>
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 12be4241474ab..70d6f6309da01 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1269,6 +1269,28 @@ static void __page_check_anon_rmap(struct folio *folio, struct page *page,
>   		       page);
>   }
>   
> +static void __foio_mod_stat(struct folio *folio, int nr, int nr_pmdmapped)
> +{
> +	int idx;
> +
> +	if (nr) {
> +		idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
> +		__lruvec_stat_mod_folio(folio, idx, nr);
> +	}
> +	if (nr_pmdmapped) {
> +		if (folio_test_anon(folio)) {
> +			idx = NR_ANON_THPS;
> +			__lruvec_stat_mod_folio(folio, idx, nr_pmdmapped);
> +		} else {
> +			/* NR_*_PMDMAPPED are not maintained per-memcg */
> +			idx = folio_test_swapbacked(folio) ?
> +				NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED;
> +			__mod_node_page_state(folio_pgdat(folio), idx,
> +					      nr_pmdmapped);
> +		}
> +	}
> +}
> +

I didn't suggest that, because in the _anon and _file functions we'll 
end up introducing unnecessary folio_test_anon() checks that the 
compiler cannot optimize out.

But at least in the removal path it's a clear win.

--
Cheers,

David / dhildenb