Re: [PATCH -next 3/5] mm/mglru: extend shrink_one for both lrugen and non-lrugen

[Johannes Weiner <hannes@cmpxchg.org>]

On Tue, Dec 09, 2025 at 01:25:55AM +0000, Chen Ridong wrote:
> From: Chen Ridong <chenridong@huawei.com>
> 
> Currently, flush_reclaim_state is placed differently between
> shrink_node_memcgs and shrink_many. shrink_many (only used for gen-LRU)
> calls it after each lruvec is shrunk, while shrink_node_memcgs calls it
> only after all lruvecs have been shrunk.
> 
> This patch moves flush_reclaim_state into shrink_node_memcgs and calls it
> after each lruvec. This unifies the behavior and is reasonable because:
> 
> 1. flush_reclaim_state adds current->reclaim_state->reclaimed to
>    sc->nr_reclaimed.
> 2. For non-MGLRU root reclaim, this can help stop the iteration earlier
>    when nr_to_reclaim is reached.
> 3. For non-root reclaim, the effect is negligible since flush_reclaim_state
>    does nothing in that case.
> 
> After moving flush_reclaim_state into shrink_node_memcgs, shrink_one can be
> extended to support both lrugen and non-lrugen paths. It will call
> try_to_shrink_lruvec for lrugen root reclaim and shrink_lruvec otherwise.
> 
> Signed-off-by: Chen Ridong <chenridong@huawei.com>
> ---
>  mm/vmscan.c | 57 +++++++++++++++++++++--------------------------------
>  1 file changed, 23 insertions(+), 34 deletions(-)
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 584f41eb4c14..795f5ebd9341 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -4758,23 +4758,7 @@ static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
>  	return nr_to_scan < 0;
>  }
>  
> -static void shrink_one(struct lruvec *lruvec, struct scan_control *sc)
> -{
> -	unsigned long scanned = sc->nr_scanned;
> -	unsigned long reclaimed = sc->nr_reclaimed;
> -	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
> -	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
> -
> -	try_to_shrink_lruvec(lruvec, sc);
> -
> -	shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
> -
> -	if (!sc->proactive)
> -		vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
> -			   sc->nr_reclaimed - reclaimed);
> -
> -	flush_reclaim_state(sc);
> -}
> +static void shrink_one(struct lruvec *lruvec, struct scan_control *sc);
>  
>  static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
>  {
> @@ -5760,6 +5744,27 @@ static inline bool should_continue_reclaim(struct pglist_data *pgdat,
>  	return inactive_lru_pages > pages_for_compaction;
>  }
>  
> +static void shrink_one(struct lruvec *lruvec, struct scan_control *sc)
> +{
> +	unsigned long scanned = sc->nr_scanned;
> +	unsigned long reclaimed = sc->nr_reclaimed;
> +	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
> +	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
> +
> +	if (lru_gen_enabled() && root_reclaim(sc))
> +		try_to_shrink_lruvec(lruvec, sc);
> +	else
> +		shrink_lruvec(lruvec, sc);

Yikes. So we end up with:

shrink_node_memcgs()
  shrink_one()
    if lru_gen_enabled && root_reclaim(sc)
      try_to_shrink_lruvec(lruvec, sc)
    else
      shrink_lruvec()
        if lru_gen_enabled && !root_reclaim(sc)
          lru_gen_shrink_lruvec(lruvec, sc)
            try_to_shrink_lruvec()

I think it's doing too much at once. Can you get it into the following
shape:

shrink_node_memcgs()
  for each memcg:
    if lru_gen_enabled:
      lru_gen_shrink_lruvec()
    else
      shrink_lruvec()

and handle the differences in those two functions? Then look for
overlap one level down, and so forth.