From: jiebin sun <jiebin.sun@intel.com>

Remove the redundant updating of stats_flush_threshold. If the
global var stats_flush_threshold has exceeded the trigger value
for __mem_cgroup_flush_stats, further increment is unnecessary.

Apply the patch and test the pts/hackbench-1.0.0 Count:4 (160 threads).

Score gain: 1.95x
Reduce CPU cycles in __mod_memcg_lruvec_state (44.88% -> 0.12%)

CPU: ICX 8380 x 2 sockets
Core number: 40 x 2 physical cores
Benchmark: pts/hackbench-1.0.0 Count:4 (160 threads)

Signed-off-by: Jiebin Sun <jiebin.sun@intel.com>
---
 mm/memcontrol.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index abec50f31fe6..9e8c6f24c694 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -626,7 +626,14 @@ static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
 
 	x = __this_cpu_add_return(stats_updates, abs(val));
 	if (x > MEMCG_CHARGE_BATCH) {
-		atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
+		/*
+		 * If stats_flush_threshold exceeds the threshold
+		 * (>num_online_cpus()), cgroup stats update will be triggered
+		 * in __mem_cgroup_flush_stats(). Increasing this var further
+		 * is redundant and simply adds overhead in atomic update.
+		 */
+		if (atomic_read(&stats_flush_threshold) <= num_online_cpus())
+			atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
 		__this_cpu_write(stats_updates, 0);
 	}
 }
-- 
2.31.1

On Sat, Jul 23, 2022 at 12:49:49AM +0800, Jiebin Sun wrote:
> From: jiebin sun <jiebin.sun@intel.com>
> 
> Remove the redundant updating of stats_flush_threshold. If the
> global var stats_flush_threshold has exceeded the trigger value
> for __mem_cgroup_flush_stats, further increment is unnecessary.
> 
> Apply the patch and test the pts/hackbench-1.0.0 Count:4 (160 threads).
> 
> Score gain: 1.95x
> Reduce CPU cycles in __mod_memcg_lruvec_state (44.88% -> 0.12%)
> 
> CPU: ICX 8380 x 2 sockets
> Core number: 40 x 2 physical cores
> Benchmark: pts/hackbench-1.0.0 Count:4 (160 threads)
> 
> Signed-off-by: Jiebin Sun <jiebin.sun@intel.com>

Acked-by: Muchun Song <songmuchun@bytedance.com>

Thanks.

On Sat, Jul 23, 2022 at 12:49:49AM +0800, Jiebin Sun wrote:
> From: jiebin sun <jiebin.sun@intel.com>
> 
> Remove the redundant updating of stats_flush_threshold. If the
> global var stats_flush_threshold has exceeded the trigger value
> for __mem_cgroup_flush_stats, further increment is unnecessary.
> 
> Apply the patch and test the pts/hackbench-1.0.0 Count:4 (160 threads).
> 
> Score gain: 1.95x
> Reduce CPU cycles in __mod_memcg_lruvec_state (44.88% -> 0.12%)
> 
> CPU: ICX 8380 x 2 sockets
> Core number: 40 x 2 physical cores
> Benchmark: pts/hackbench-1.0.0 Count:4 (160 threads)
> 
> Signed-off-by: Jiebin Sun <jiebin.sun@intel.com>

Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>

Good optimization, thanks!

>
>On Sat, Jul 23, 2022 at 12:49:49AM +0800, Jiebin Sun wrote:
>> From: jiebin sun <jiebin.sun@intel.com>
>>
>> Remove the redundant updating of stats_flush_threshold. If the global
>> var stats_flush_threshold has exceeded the trigger value for
>> __mem_cgroup_flush_stats, further increment is unnecessary.
>>
>> Apply the patch and test the pts/hackbench-1.0.0 Count:4 (160 threads).
>>
>> Score gain: 1.95x
>> Reduce CPU cycles in __mod_memcg_lruvec_state (44.88% -> 0.12%)
>>
>> CPU: ICX 8380 x 2 sockets
>> Core number: 40 x 2 physical cores
>> Benchmark: pts/hackbench-1.0.0 Count:4 (160 threads)
>>
>> Signed-off-by: Jiebin Sun <jiebin.sun@intel.com>
>
>Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
>
>Good optimization, thanks!

Looks good. Nice performance improvement.

Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>

On Sat, Jul 23, 2022 at 12:49:49AM +0800, Jiebin Sun wrote:
> From: jiebin sun <jiebin.sun@intel.com>
> 
> Remove the redundant updating of stats_flush_threshold. If the
> global var stats_flush_threshold has exceeded the trigger value
> for __mem_cgroup_flush_stats, further increment is unnecessary.
> 
> Apply the patch and test the pts/hackbench-1.0.0 Count:4 (160 threads).
> 
> Score gain: 1.95x
> Reduce CPU cycles in __mod_memcg_lruvec_state (44.88% -> 0.12%)
> 
> CPU: ICX 8380 x 2 sockets
> Core number: 40 x 2 physical cores
> Benchmark: pts/hackbench-1.0.0 Count:4 (160 threads)
> 
> Signed-off-by: Jiebin Sun <jiebin.sun@intel.com>

Yes, this makes sense. No need to dirty a cacheline if we are already
over the threshold.

Acked-by: Shakeel Butt <shakeelb@google.com>