On Wed, Apr 04, 2018 at 10:08:55AM -0400, Johannes Weiner wrote:
> On Sat, Mar 24, 2018 at 09:08:59AM -0700, Tejun Heo wrote:
> > @@ -91,6 +91,9 @@ struct mem_cgroup_stat_cpu {
> > unsigned long events[MEMCG_NR_EVENTS];
> > unsigned long nr_page_events;
> > unsigned long targets[MEM_CGROUP_NTARGETS];
> > +
> > + /* for cgroup rstat delta calculation */
> > + unsigned long last_events[MEMCG_NR_EVENTS];
> > };
> >
> > struct mem_cgroup_reclaim_iter {
> > @@ -233,7 +236,11 @@ struct mem_cgroup {
> >
> > struct mem_cgroup_stat_cpu __percpu *stat_cpu;
> > atomic_long_t stat[MEMCG_NR_STAT];
> > - atomic_long_t events[MEMCG_NR_EVENTS];
> > +
> > + /* events is managed by cgroup rstat */
> > + unsigned long long events[MEMCG_NR_EVENTS]; /* local */
> > + unsigned long long tree_events[MEMCG_NR_EVENTS]; /* subtree */
> > + unsigned long long pending_events[MEMCG_NR_EVENTS];/* propagation */
>
> The lazy updates are neat, but I'm a little concerned at the memory
> footprint. On a 64-cpu machine for example, this adds close to 9000
> words to struct mem_cgroup. And we really only need the accuracy for
> the 4 cgroup items in memory.events, not all VM events and stats.
>
> Why not restrict the patch to those? It would also get rid of the
> weird sharing between VM and cgroup enums.
In fact, I wonder if we need per-cpuness for MEMCG_LOW, MEMCG_HIGH
etc. in the first place. They describe super high-level reclaim and
OOM events, so they're not nearly as hot as other VM events and
stats. We could probably just have a per-memcg array of atomics.
