Re: [PATCH 1/2] mm/zsmalloc: fix class per-fullness zspage counts

From: Sergey Senozhatsky
Date: Thu Jun 27 2024 - 20:52:00 EST


On (24/06/27 13:33), Andrew Morton wrote:
> On Thu, 27 Jun 2024 15:59:58 +0800 Chengming Zhou <chengming.zhou@xxxxxxxxx> wrote:
> > We always use insert_zspage() and remove_zspage() to update zspage's
> > fullness location, which will account correctly.
> >
> > But this special async free path use "splice" instead of remove_zspage(),
> > so the per-fullness zspage count for ZS_INUSE_RATIO_0 won't decrease.
> >
> > Fix it by decreasing when iterate over the zspage free list.
> >
> > ...
> >
> > Signed-off-by: Chengming Zhou <chengming.zhou@xxxxxxxxx>
> > +++ b/mm/zsmalloc.c
> > @@ -1883,6 +1883,7 @@ static void async_free_zspage(struct work_struct *work)
> >
> > class = zspage_class(pool, zspage);
> > spin_lock(&class->lock);
> > + class_stat_dec(class, ZS_INUSE_RATIO_0, 1);
> > __free_zspage(pool, class, zspage);
> > spin_unlock(&class->lock);
> > }
>
> What are the runtime effects of this bug? Should we backport the fix
> into earlier kernels? And are we able to identify the appropriate
> Fixes: target?

I don't think this has any run-time visible effects. Class stats
(ZS_OBJS_ALLOCATED and ZS_OBJS_INUSE) play their role during compaction
(defragmentation), but ZS_INUSE_RATIO_0 is for zspage fullness type,
moreover for empty zspage, which we don't look at during compaction.

With CONFIG_ZSMALLOC_STAT enabled we show pool/class stats to the users
via zs_stats_size_show() but ZS_INUSE_RATIO_0 is ignored. So no one
(external) should know what value is there and ZS_INUSE_RATIO_0 should
never be of any importance to zsmalloc (internally).

Code in question (async_free_zspage()) was introduced by 48b4800a1c6af
in 2016-07-26, so it's been a long time.