Re: [PATCH] mm: vmscan: memcontrol: remove mem_cgroup_select_victim_node()

[Johannes Weiner <hannes@cmpxchg.org>]

On Tue, Oct 29, 2019 at 04:47:53PM -0700, Shakeel Butt wrote:
> Since commit 1ba6fc9af35b ("mm: vmscan: do not share cgroup iteration
> between reclaimers"), the memcg reclaim does not bail out earlier based
> on sc->nr_reclaimed and will traverse all the nodes. All the reclaimable
> pages of the memcg on all the nodes will be scanned relative to the
> reclaim priority. So, there is no need to maintain state regarding which
> node to start the memcg reclaim from. Also KCSAN complains data races in
> the code maintaining the state.
> 
> This patch effectively reverts the commit 889976dbcb12 ("memcg: reclaim
> memory from nodes in round-robin order") and the commit 453a9bf347f1
> ("memcg: fix numa scan information update to be triggered by memory
> event").
> 
> Signed-off-by: Shakeel Butt <shakeelb@google.com>
> Reported-by: <syzbot+13f93c99c06988391efe@syzkaller.appspotmail.com>

Excellent, thanks Shakeel!
Acked-by: Johannes Weiner <hannes@cmpxchg.org>

Just a request on this bit:

> @@ -3360,16 +3358,9 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
>  		.may_unmap = 1,
>  		.may_swap = may_swap,
>  	};
> +	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
>  
>  	set_task_reclaim_state(current, &sc.reclaim_state);
> -	/*
> -	 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
> -	 * take care of from where we get pages. So the node where we start the
> -	 * scan does not need to be the current node.
> -	 */
> -	nid = mem_cgroup_select_victim_node(memcg);
> -
> -	zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];

This works, but it *is* somewhat fragile if we decide to add bail-out
conditions to reclaim again. And some numa nodes receiving slightly
less pressure than others could be quite tricky to debug.

Can we add a comment here that points out the assumption that the
zonelist walk is comprehensive, and that all nodes receive equal
reclaim pressure?

Also, I think we should use sc.gfp_mask & ~__GFP_THISNODE, so that
allocations with a physical node preference still do node-agnostic
reclaim for the purpose of cgroup accounting.