Re: [PATCH] mm: Stop kswapd early when nothing's waiting for it to free pages

[Shakeel Butt <shakeelb@google.com>]

On Wed, Feb 26, 2020 at 4:15 AM Hillf Danton <hdanton@sina.com> wrote:
>
>
> On Tue, 25 Feb 2020 14:30:03 -0800 Shakeel Butt wrote:
> >
> > BTW we are seeing a similar situation in our production environment.
> > We have swappiness=0, no swap from kswapd (because we don't swapout on
> > pressure, only on cold age) and too few file pages, the kswapd goes
> > crazy on shrink_slab and spends 100% cpu on it.
>
> Dunno if swappiness is able to put peace on your kswapd.
>
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -2631,8 +2631,14 @@ static inline bool should_continue_recla
>          */
>         pages_for_compaction = compact_gap(sc->order);
>         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
> -       if (get_nr_swap_pages() > 0)
> -               inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
> +       do {
> +               struct lruvec *lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
> +               struct mem_cgroup *memcg = lruvec_memcg(lruvec);
> +               int swappiness = mem_cgroup_swappiness(memcg);
> +
> +               if (swappiness && get_nr_swap_pages() > 0)

Thanks for finding this. I think we also need to check sc->may_swap as
well. Can you please send a signed-off patch? It may or maynot help
kswapd but I think this is needed.

> +                       inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
> +       } while (0);
>
>         return inactive_lru_pages > pages_for_compaction;
>  }
>