Re: [PATCH 2/5] mm: prepare to premature release of per-node lruvec_stat_cpu

From: Johannes Weiner
Date: Mon Mar 11 2019 - 13:17:53 EST


On Thu, Mar 07, 2019 at 03:00:30PM -0800, Roman Gushchin wrote:
> Similar to the memcg's vmstats_percpu, per-memcg per-node stats
> consists of percpu- and atomic counterparts, and we do expect
> that both coexist during the whole life-cycle of the memcg.
>
> To prepare for a premature release of percpu per-node data,
> let's pretend that lruvec_stat_cpu is a rcu-protected pointer,
> which can be NULL. This patch adds corresponding checks whenever
> required.
>
> Signed-off-by: Roman Gushchin <guro@xxxxxx>

Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>

> @@ -4430,7 +4436,8 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
> if (!pn)
> return 1;
>
> - pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
> + rcu_assign_pointer(pn->lruvec_stat_cpu,
> + alloc_percpu(struct lruvec_stat));
> if (!pn->lruvec_stat_cpu) {

Nitpick: wouldn't this have to use rcu_dereference()? Might be cleaner
to use an intermediate variable and only assign after the NULL check.