Re: [memcg] a8c49af3be: hackbench.throughput -13.7% regression

From: Yin Fengwei
Date: Wed Apr 27 2022 - 00:59:37 EST

Hi Yosry,

On 4/27/2022 12:34 PM, Yosry Ahmed wrote:
> I am not sure why this specific callsite causes regression, there are
> many callsites that modify stats similarly (whether through
> mod_memcg_state() directly or through other variants like
> mod_lruvec_state()). Maybe the kmem call path is exercised more
> heavily in this benchmark than other call paths that update stats.
perf data didn't show the function commented out either. No idea how
it impact the hackbench.

> The only seemingly expensive operation in the mod_memcg_state() path
> is the call to cgroup_rstat_updated() (through memcg_rstat_updated()).
> One idea off the top of my head is to batch calls to
> cgroup_rstat_updated(), similar to what 11192d9c124d ("memcg: flush
> stats only if updated") did on the flush side. I am interested to see
> what memcg maintainers think about this problem (and the proposed
> solution).
OK. Let's wait.

Yin, Fengwei