diff mbox series

[v2] mm: make pcp_decay_high working better with NOHZ full

Message ID 1729238277-26683-1-git-send-email-mengensun@tencent.com (mailing list archive)
State New
Headers show
Series [v2] mm: make pcp_decay_high working better with NOHZ full | expand

Commit Message

MengEn Sun Oct. 18, 2024, 7:57 a.m. UTC 
From: MengEn Sun <mengensun@tencent.com>

When a cpu entring NOHZ full, quiet_vmstat may flush percpu
zonestats and nodestats.

The vmstat_shepherd only check percpu zonestats and nodestats
to determine whether it is necessary to fire vmstat_update on
the target cpu for now.

If a process on a certain CPU allocates a large amount of memory,
then frees that memory, and subsequently the CPU enters NOHZ, and
the process not freeing and allocating memory anymore,the
vmstat_update not being executed on the cpu. Because
vmstat_shepherd may not see zonestats and nodestats of the cpu
changed, so may resulting in vmstat_update on the cpu not fired
for a long time.

While, This seems to be fine:
- if freeing and allocating memory occur later, it may the
  high_max may be adjust automatically
- If memory is tight, the memory reclamation process will
  release the pcp

Whatever, we make vmstat_shepherd to checking whether we need
decay pcp high_max, and fire pcp_decay_high early if we need.

Fixes: 51a755c56dc0 ("mm: tune PCP high automatically")
Reviewed-by: Jinliang Zheng <alexjlzheng@tencent.com>
Signed-off-by: MengEn Sun <mengensun@tencent.com>
---
changelog:
v1: https://lore.kernel.org/lkml/20241012154328.015f57635566485ad60712f3@linux-foundation.org/T/#t
v2: Make the commit message clearer by adding some comments.
---
 mm/vmstat.c | 9 +++++++++
 1 file changed, 9 insertions(+)
diff mbox series

Patch

diff --git a/mm/vmstat.c b/mm/vmstat.c
index 1917c034c045..07b494b06872 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -2024,8 +2024,17 @@  static bool need_update(int cpu)
 
 	for_each_populated_zone(zone) {
 		struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
+		struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
 		struct per_cpu_nodestat *n;
 
+		/* per_cpu_nodestats and per_cpu_zonestats maybe flush when cpu
+		 * entering NOHZ full, see quiet_vmstat. so, we check pcp
+		 * high_{min,max} to determine whether it is necessary to run
+		 * decay_pcp_high on the corresponding CPU
+		 */
+		if (pcp->high_max > pcp->high_min)
+			return true;
+
 		/*
 		 * The fast way of checking if there are any vmstat diffs.
 		 */