On Wed, Jul 28, 2021 at 12:43 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Tue, 27 Jul 2021 18:22:43 -0700 Shakeel Butt <shakeelb@xxxxxxxxxx> wrote: > > > We used to have per-cpu memcg and lruvec stats and the readers have to > > traverse and sum the stats from each cpu. This summing was racy and may > > expose transient negative values. So, an explicit check was added to > > avoid such scenarios. Now these stats are moved to rstat infrastructure > > and are no more per-cpu, so we can remove the fixup for transient > > negative values. > > We can't do anything about the same code in lruvec_page_state_local()? lruvec_page_state_local() is used by cgroup v1's memory.numa_stat for cgroup local stats (not hierarchical) and are still per-cpu. To make it non-per-cpu, we have to add 'long state_local[NR_VM_NODE_STAT_ITEMS]' in 'struct lruvec_stats' and do aggregation in rstat flushing. So, paying the cpu traversal cost with more memory usage. I am not sure if it is worth it.
