Re: [PATCH] memcg: move LRU size accounting on reparenting instead of copying it

On Fri 21-08-26 19:47:07, Shakeel Butt wrote:
> When a memory cgroup is offlined its LRU folios are reparented to the
> parent.  lruvec_reparent_lru() splices the child's lists into the
> parent's and credits the parent with the child's per-zone
> lru_zone_size[], but never clears the child's copy, so the size is
> copied rather than moved.  lru_gen_reparent_memcg() does the same for
> MGLRU.
> 
> The parent is left correct, credited with exactly the folios it took
> over.  The stale value sits on the child and nothing will correct it:
> folio->memcg_data now resolves to the parent, so every later
> update_lru_size() for those folios goes there.
> 
> Dying cgroups are not freed immediately and mem_cgroup_iter() still
> walks them, so shrink_lruvec() keeps being called on them.
> get_scan_count() reads the phantom counter through lruvec_lru_size() and
> the scan loop then grinds through nr[] in SWAP_CLUSTER_MAX steps against
> an empty list, for as long as the dead cgroup lives.  Under MGLRU the
> MGLRU scanner runs instead, but count_shadow_nodes() sums all of
> NR_LRU_LISTS through lruvec_lru_size() and over-budgets the shadow node
> limit just the same.
> 
> On one 251 GiB host a sweep of every mz->lru_zone_size[] found 380
> counters describing folios on no list at all: 124777314 pages, 476 GiB,
> 1.89x the machine's RAM, across 57 cgroups.  All were on memcgs with
> CSS_DYING set and CSS_ONLINE clear, and parent/child pairs reported
> byte-identical sizes.
> 
> LRU_UNEVICTABLE needs its size moved too.  Its list is deliberately not
> spliced because lruvec_init() poisons the head - the unevictable LRU is
> imaginary and folios are never threaded on it - but the size is kept by
> lruvec_add_folio()/lruvec_del_folio() and those folios account to the
> parent from here on.
> 
> This depends on commit bf4ade7dbd76 ("memcg: keep folio's objcg same as
> its node") and must not be backported ahead of it.  Without that
> invariant a folio's objcg can belong to another node, so a folio already
> spliced onto the parent's list can still resolve to the child's lruvec
> until the objcg's node is reparented in a later iteration of
> memcg_reparent_objcgs(); clearing the child's counter early then lets
> lruvec_del_folio() underflow it and trip the WARN_ONCE()/VM_BUG_ON() in
> mem_cgroup_update_lru_size().
> 
> Fixes: 07a6e9a2c199 ("mm: vmscan: prepare for reparenting traditional LRU folios")
> Fixes: f304652609ea ("mm: vmscan: prepare for reparenting MGLRU folios")
> Cc: <[email protected]> # After: bf4ade7dbd76: memcg: keep folio's objcg same as its node
> Signed-off-by: Shakeel Butt <[email protected]>

Acked-by: Michal Hocko <[email protected]>
Thanks!

> ---
>  mm/folio.c  | 9 +++++++++
>  mm/vmscan.c | 5 +++++
>  2 files changed, 14 insertions(+)
> 
> diff --git a/mm/folio.c b/mm/folio.c
> index 59c477120b9a..c02dcea9c03c 100644
> --- a/mm/folio.c
> +++ b/mm/folio.c
> @@ -1130,7 +1130,16 @@ static void lruvec_reparent_lru(struct lruvec *child_lruvec,
>  	for_each_managed_zone_pgdat(zone, NODE_DATA(nid), zid, MAX_NR_ZONES - 1) {
>  		unsigned long size = mem_cgroup_get_zone_lru_size(child_lruvec, lru, zid);
>  
> +		if (!size)
> +			continue;
> +
> +		/*
> +		 * The folios are accounted to the parent from now on, so the
> +		 * size has to be moved, not just copied. Leaving it behind
> +		 * makes the dying child describe folios it no longer owns.
> +		 */
>  		mem_cgroup_update_lru_size(parent_lruvec, lru, zid, size);
> +		mem_cgroup_update_lru_size(child_lruvec, lru, zid, -(long)size);
>  	}
>  }
>  
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index fe7f0c52a18c..561eeec5628c 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -4635,7 +4635,12 @@ void lru_gen_reparent_memcg(struct mem_cgroup *memcg, struct mem_cgroup *parent,
>  		for_each_managed_zone_pgdat(zone, NODE_DATA(nid), zid, MAX_NR_ZONES - 1) {
>  			unsigned long size = mem_cgroup_get_zone_lru_size(child_lruvec, lru, zid);
>  
> +			if (!size)
> +				continue;
> +
> +			/* Move the accounting, do not duplicate it. */
>  			mem_cgroup_update_lru_size(parent_lruvec, lru, zid, size);
> +			mem_cgroup_update_lru_size(child_lruvec, lru, zid, -(long)size);
>  		}
>  	}
>  }
> -- 
> 2.53.0-Meta

-- 
Michal Hocko
SUSE Labs