sched/numa：避免 remote private fault 让扫描周期被错误地翻倍

系列概况

标题：[PATCH v2] sched/numa: avoid doubling scan period for remote private faults
作者：Hongling Zeng zenghongling@kylinos.cn；Suggested-by: Zhan Xusheng zhanxusheng@xiaomi.com
版本：v2，单 patch（无后续回复）
规模：1 file changed, 10 insertions(+), 7 deletions(-)
修改文件：kernel/sched/fair.c
代码统计：update_task_scan_period() 内 3 处实质改动（第 3505、3529、3538 行附近）+ 两段注释修订
Message-ID：20260804061515.640924-1-zenghongling@kylinos.cn
完整性：commit message、Suggested-by、Signed-off-by、Changes from v1 齐全，签名链完整

补丁目的

NUMA 自动平衡为每个 task 维护 numa_scan_period，按最近一段时间的 fault 分布动态调整扫描节奏。旧实现用 local + shared == 0 判定 "无 relevant fault"，对只产生 remote private fault 的进程（local=0, shared=0, remote>0）会错误判为完全空闲，从而触发 numa_scan_period << 1，扫描被强制放慢一倍。本 patch 把 no-fault 判定改成 local + remote == 0，让 remote private fault 继续走到 locality-ratio 计算，同时纠正 ps_ratio、lr_ratio 两个分支注释与代码语义相反的问题。

旧流程的问题

判定口径错：把 "local=0 且 shared=0" 当成 "无 fault"，忽略 remote 字段，把大量 "远程私有访问为主" 的负载打入空闲档。
注释/代码语义错位：ps_ratio 分支注释写的是 "shared 多 → 没必要继续 fast scan"，但 ps_ratio 对应的正是 "private 多" 的场景；lr_ratio 分支同样语义错置。
两处错误同时出现在同一函数里，维护者很容易误读 NUMA 调整策略。

新流程

no-fault 条件：local + remote == 0，或 p->numa_faults_locality[2]（迁移失败过）→ 触发 numa_scan_period << 1。
ps_ratio 路径：专门处理 "大多访问是 private"，注释同步改正，扫描频率按 NUMA_PERIOD_SLOTS 节奏递增。
lr_ratio 路径：专门处理 "大多访问是 local"，注释去掉过时的 "shared memory" 表述。
行为差异：纯 remote private 负载不再被无条件翻倍；最终周期交给 locality-ratio 决定。

关键实现

/* kernel/sched/fair.c, update_task_scan_period() */
@@ update_task_scan_period(struct task_struct *p, ...)
-    if (local + shared == 0 || p->numa_faults_locality[2]) {
+    if (local + remote == 0  || p->numa_faults_locality[2]) {
         /* Completely idle or all activity is in areas that
          * are not of interest to automatic numa balancing.
          * Related to that, if there were failed migration then
          * it implies we are migrating too quickly or the local
          * node is overloaded. In either case, scan slower. */
         p->numa_scan_period = min(p->numa_scan_period_max,
                                   p->numa_scan_period << 1);
     }

     if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
-        /* Most memory accesses are shared with other tasks.
-         * There is no point in continuing fast NUMA scanning,
-         * since other tasks may just move the memory elsewhere. */
+        /* Most memory accesses are private. Slow down NUMA scanning
+         * since there is little shared memory to rebalance. */
         ...
     } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
-        /* Most memory accesses are local. There is no need to
-         * do fast NUMA scanning, since memory is already local. */
+        /* Most memory accesses are local. There is no need to
+         * do fast NUMA scanning, since memory is already local. */
         ...
     }

Patch 概览

本系列只有一个 patch，聚焦 update_task_scan_period() 内的判定逻辑与两段错置注释，没有触及 locality-ratio 计算本身、阈值（NUMA_PERIOD_SLOTS、NUMA_PERIOD_THRESHOLD）或 numa_faults_locality[2] 的语义。

类比

把 NUMA 扫描节奏想成 "体检频率"，把 fault 类型想成 "运动场景"：

local fault = 在家门口跑步；
shared fault = 在公司会议室开会；
remote fault = 出差去外地跑业务。

旧规则只看 "家门口 + 公司" 的步数，于是出差日被误判为 "整天没动"，下次体检直接拉长一倍。新规则改成 "家门口 + 外地"，出差日也算动过，按正常节奏安排体检；当你几乎只出差、不怎么开共享会议时，仍会因 "私事为主" 适当放慢扫描——但绝不会突然从 "出差日" 直接跳到 "完全不查"。

Highlight：风险与注意点

CPU 开销变化面：no-fault 判定改动会让 "纯 remote private" 负载的扫描更密，SCAN 路径 CPU 开销可能略增；建议跑 numactl/perf 基准实测。
作者已声明保留语义："this does not necessarily make the scan period shorter"——纯 private 负载最终仍会因 ps_ratio 涨到 NUMA_PERIOD_SLOTS 而变慢，不再被强制 << 1，但仍属调整。
失败迁移仍会触发翻倍：numa_faults_locality[2] 分支未改动，若一次失败 migration 后又被新逻辑识别为有 fault，仍可能走后续调整路径，需要继续观察真实工作负载。
注释错位可能影响后续 reviewer：ps_ratio / lr_ratio 注释颠倒的问题仅在表达层修复；若有维护者按 v1 注释做过行为推断，需要回看前面是否有要被追溯修正的判断。
未触及 numa_group/shared 权衡：随着 "remote shared" 概念未来可能引入，这里以 local + remote == 0 作为新基线，仍需要在 v3 或后续补丁里重新审视。
单 patch、零回复：本邮件列表只有 1 封 v2 提交，没有 reviewer 反馈；上游对 "remote shared" 是否存在的进一步澄清，要在落地后看 reply 走向。

版本变化（v1 → v2）

修订 commit message，去掉错误的 "speed up" 措辞。
patch 标题改为 "avoid doubling scan period for remote private faults"，更准确描述行为。
移除 lr_ratio 分支中关于 shared memory 的过时注释。
增加 "scan period comparison for clarity"，方便复核到底该不该进入翻倍分支。

一句话总结

把 update_task_scan_period() 的 no-fault 判定从 local + shared 改成 local + remote，避免只产生远程私有 fault 的 NUMA 负载被无条件翻倍降速；同时对齐 ps_ratio / lr_ratio 两个分支的注释语义。

+-----------------------------------------------------------+
|       old: local + shared == 0   vs   new: local + remote == 0 |
+-----------------------------------------------------------+

stats: local=0, shared=0, remote=N
   |
   v
+-------+----------------------------+      +----------------------------+
|  old  | treat as fully idle        |      | new: treat as has remote   |
|       | / no relevant fault        |      |     private fault          |
+-------+----------------------------+      +----------------------------+
   |                                                |
   v                                                v
period = min(period_max, period << 1)         continue into ps_ratio
(unconditional 2x slower)                      / lr_ratio calculation
   |                                          (locality-driven adjust)
   v                                                |
   |                                                v
must wait full scan window                  NUMA_PERIOD_SLOTS converges
before re-evaluate                          naturally toward steady state


+-----------------------------------------------------+
|          ps_ratio branch: comment fix               |
+-----------------------------------------------------+
- old (wrong):  "Most accesses are shared with other tasks"
+ new (correct): "Most accesses are private. Slow down
+                NUMA scanning ... little shared memory
+                to rebalance."

+-----------------------------------------------------+
|          lr_ratio branch: comment fix               |
+-----------------------------------------------------+
- old: stale "shared memory" wording
+ new: keep correct intent, remove stale shared wording.

[PATCH v2] sched/numa: avoid doubling scan period for remote private faults

LLM 分析