Re: [PATCH] vmscan,migrate: fix double-decrement on node stats when demoting pages

[Gregory Price]

On Sun, Oct 27, 2024 at 10:24:10PM -0700, Shakeel Butt wrote:
> On Fri, Oct 25, 2024 at 10:17:24AM GMT, Gregory Price wrote:
> > When numa balancing is enabled with demotion, vmscan will call
> > migrate_pages when shrinking LRUs.  Successful demotions will
> > cause node vmstat numbers to double-decrement, leading to an
> > imbalanced page count.  The result is dmesg output like such:
> > 
> > $ cat /proc/sys/vm/stat_refresh
> > 
> > [77383.088417] vmstat_refresh: nr_isolated_anon -103212
> > [77383.088417] vmstat_refresh: nr_isolated_file -899642
> > 
> > This negative value may impact compaction and reclaim throttling.
> > 
> > The double-decrement occurs in the migrate_pages path:
> > 
> > caller to shrink_folio_list decrements the count
> >   shrink_folio_list
> >     demote_folio_list
> >       migrate_pages
> >         migrate_pages_batch
> >           migrate_folio_move
> >             migrate_folio_done
> >               mod_node_page_state(-ve) <- second decrement
> > 
> > This path happens for SUCCESSFUL migrations, not failures. Typically
> > callers to migrate_pages are required to handle putback/accounting for
> > failures, but this is already handled in the shrink code.
> > 
> > When accounting for migrations, instead do not decrement the count
> > when the migration reason is MR_DEMOTION. As of v6.11, this demotion
> > logic is the only source of MR_DEMOTION.
> > 
> > Signed-off-by: Gregory Price <gourry@xxxxxxxxxx>
> > Fixes: 26aa2d199d6f2 ("mm/migrate: demote pages during reclaim")
> > Cc: stable@xxxxxxxxxxxxxxx
> 
> Reviewed-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> 
> This patch looks good for stable backports. For future I wonder if
> instead of migrate_pages(), the caller providing the isolated folios,
> manages the isolated stats (increments and decrements) similar to how
> reclaim does it.
>

Note that even if you provided the folios, you'd likely still end up in
migrate_pages_batch/migrate_folio_move and subsequently the same accounting
path.  Probably there's some refactoring we can do to make the accounting
more obvious - it is very subtle here.
 
> > ---
> >  mm/migrate.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/mm/migrate.c b/mm/migrate.c
> > index 923ea80ba744..e3aac274cf16 100644
> > --- a/mm/migrate.c
> > +++ b/mm/migrate.c
> > @@ -1099,7 +1099,7 @@ static void migrate_folio_done(struct folio *src,
> >  	 * not accounted to NR_ISOLATED_*. They can be recognized
> >  	 * as __folio_test_movable
> >  	 */
> > -	if (likely(!__folio_test_movable(src)))
> > +	if (likely(!__folio_test_movable(src)) && reason != MR_DEMOTION)
> >  		mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
> >  				    folio_is_file_lru(src), -folio_nr_pages(src));
> >  
> > -- 
> > 2.43.0
> >