Mirrors_BOOKS/LDD-LinuxDeviceDrivers
1
0
Fork 0
mirror of https://github.com/gatieme/LDD-LinuxDeviceDrivers.git synced 2026-09-22 12:43:39 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

description: EAS sched-domain overutilized

Browse Source
This commit is contained in:
Cheng Jian
2022-11-25 18:11:33 +08:00
parent b017085aa0
commit ca034e3b82
5 changed files with 79 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
study/kernel/00-DESCRIPTION/DEBUGGING.md
+16 -2
View File
@@ -357,16 +357,30 @@ $reclaim = current\_mem \times reclaim\_ratio \times max(0,1 – \frac{psi_some}
## 11.3 TOPDOWN
-------
[Support standalone metrics and metric groups for perf](https://lore.kernel.org/all/20170831194036.30146-1-andi@firstfloor.org)
### 11.3.1 stat topdown
-------
| 时间 | 作者 | 特性 | 描述 | 是否合入主线 | 链接 |
|:----:|:----:|:---:|:----:|:---------:|:----:|
| 2016/05/24 | Andi Kleen <andi@firstfloor.org> | [perf stat: Basic support for TopDown in perf stat](https://lore.kernel.org/all/1464119559-17203-1-git-send-email-andi@firstfloor.org) | perf stat 支持 topdown 分析. | v1 ☑ 4.8-rc1 | [PatchWork 1/4](https://lore.kernel.org/all/1464119559-17203-1-git-send-email-andi@firstfloor.org) |
| 2021/02/02 | Kan Liang <kan.liang@linux.intel.com> | [perf stat: Support L2 Topdown events](https://lore.kernel.org/lkml/1612296553-21962-9-git-send-email-kan.liang@linux.intel.com) | perf stat 支持 Level 2 级别 topdown 分析. [perf core PMU support for Sapphire Rapids (User tools)](https://lore.kernel.org/lkml/1612296553-21962-1-git-send-email-kan.liang@linux.intel.com) 系列中的其中一个补丁. | v1 ☑ 5.12-rc1 | [PatchWork 1/4](https://lore.kernel.org/lkml/1612296553-21962-9-git-send-email-kan.liang@linux.intel.com) |
### 11.3.2 metricsgroup topdown
-------
| 时间 | 作者 | 特性 | 描述 | 是否合入主线 | 链接 |
|:----:|:----:|:---:|:----:|:---------:|:----:|
| 2017/08/31 | Andi Kleen <andi@firstfloor.org> | [Support standalone metrics and metric groups for perf](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?id=b90f1333ef08d2a497ae239798868b046f4e3a97) | 向 perf-stat 添加对 JSON 文件中指定的独立度量 metrics 的支持. 度量 metrics 使用多个事件来计算更高级别结果 (例如 IPC, TOPDOWN 等) 的公式. 对于更复杂的度量(比如 pipeline, TOPDOWN), 需要具有特定于微架构的配置, 因此将度量 metrics 与 JSON 事件列表联系起来是有意义的. 以前的度量始终与事件关联, 并随该事件自动启用. 但现在改变它, 可以有独立的指标. 它们与事件处于相同的 JSON 数据结构中, 但没有事件名称, 只有度量名称. 同时允许在度量组中组织度量, 这允许一次选择几个相关度量的快捷方式. | v3 ☑✓ 4.15-rc1 | [LORE v3,0/11](https://lore.kernel.org/all/20170831194036.30146-2-andi@firstfloor.org) |
| 2017/08/31 | Andi Kleen <andi@firstfloor.org> | [perf arm64 metricgroup support](https://lore.kernel.org/all/20170831194036.30146-1-andi@firstfloor.org) | 为 perf stat 添加[对 JSON 文件中指定的独立指标](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=b18f3e365019de1a5b26a851e123f0aedcce881f) 的通用支持. 这些指标是一个公式, 它使用多个事件来计算更高级别的结果(例如 IPC, TOPDOWN 分析等). 添加一个新的 `-M /--metrics` 选项来添加指定的度量或度量组. 并增加了对 Intel X86 平台的支持. 通过这些 JSON 文件定义的事件组合度量, 可以很好的支持 TOPDOWN 分析. | v3 ☑ 4.15-rc1 | [PatchWork v3,00/11](https://lore.kernel.org/all/20170831194036.30146-1-andi@firstfloor.org) |
| 2020/09/11 | Kan Liang <kan.liang@linux.intel.com> | [TopDown metrics support for Ice Lake (perf tool)](https://lore.kernel.org/lkml/20200911144808.27603-1-kan.liang@linux.intel.com) | 为 perf metrics 分析增加对 Ice Lake 的支持. 将原本 group 重命名为 topdown. | v3 ☑ 5.10-rc1 | [PatchWork v3,0/4](https://lore.kernel.org/lkml/20200911144808.27603-1-kan.liang@linux.intel.com) |
| 2020/09/11 | Kan Liang <kan.liang@linux.intel.com> | [TopDown metrics support for Ice Lake (perf tool)](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?id=328781df86fa8f9299219c74ffae876bd625c6d1) | 为 perf metrics 分析增加对 Ice Lake 的支持. 将原本 group 重命名为 topdown. | v3 ☑ 5.10-rc1 | [PatchWork v3,0/4](https://lore.kernel.org/lkml/20200911144808.27603-1-kan.liang@linux.intel.com) |
| 2021/04/07 | John Garry <john.garry@huawei.com> | [perf arm64 metricgroup support](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?id=0cc177cfc95d565e1a458136a592b0bd6d487db0) | perf 支持 HiSilicon hip08 平台的 topdown metric. 支持到 Level 3. 自此鲲鹏 920 的 ARM64 服务器上, 可以使用:<br>`sudo perf stat -M TopDownL1 sleeep 1`<br>来进行 TopDown 分析了. | v1 ☑ 5.13-rc1 | [PatchWork 0/5](https://patchwork.kernel.org/project/linux-arm-kernel/cover/1614784938-27080-1-git-send-email-john.garry@huawei.com)<br>*-*-*-*-*-*-*-* <br>[PatchWork v2,0/6](https://patchwork.kernel.org/project/linux-arm-kernel/cover/1616668398-144648-1-git-send-email-john.garry@huawei.com)<br>*-*-*-*-*-*-*-* <br>[PatchWork v3,0/6](https://patchwork.kernel.org/project/linux-arm-kernel/cover/1617791570-165223-1-git-send-email-john.garry@huawei.com) |
| 2022/05/28 | zhengjun <zhengjun.xing@linux.intel.com> | [perf vendor events intel: Add metrics for Sapphirerapids](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?id=1bcca2b1bd67f3c0e5c3a88ed16c6389f01a5b31) | TODO | v2 ☑✓ 5.19-rc1 | [LORE v2,0/2](https://lore.kernel.org/all/20220528095933.1784141-1-zhengjun.xing@linux.intel.com) |
| 2022/08/25 | zhengjun.xing@linux.intel.com <zhengjun.xing@linux.intel.com> | [perf stat: Capitalize topdown metricsnel.org/all/20220825015458.3252239-1-zhengjun.xing@linux.intel.com) | TODO | v1 ☐☑✓ | [LORE](https://lore.kernel.org/9-1-zhengjun.xing@linux.intel.com) |
| 2022/10/21 | Shang XiaoJing <shangxiaojing@huawei.com> | [perf vendor events arm64: Fix incorrect metrics and improve readability](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?id=e9229d5b6254a75291536f582652c599957344d2) | TODO | v2 ☑✓ 6.1-rc3 | [LORE v2,0/3](https://lore.kernel.org/all/20221021105035.10000-1-shangxiaojing@huawei.com) |
| 2022/10/21 | Shang XiaoJing <shangxiaojing@huawei.com> | [perf vendor events arm64: Fix incorrect Hisi hip08 L3 metrics](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?id=e9229d5b6254a75291536f582652c599957344d2) | TODO | v2 ☑✓ 6.1-rc3 | [LORE v2,0/3](https://lore.kernel.org/all/20221021105035.10000-1-shangxiaojing@huawei.com) |
## 11.3 Userspace counter access
study/kernel/00-DESCRIPTION/MEMORY_MANAGER.md
+2
View File
@@ -641,6 +641,8 @@ MTE 实现了锁和密钥访问内存. 这样在内存访问期间, 可以在内
[Intel Revs Its Linear Address Masking Patches For Linux](https://www.phoronix.com/scan.php?page=news_item&px=Intel-LAM-Linux-v5)
[Intel Linear Address Masking "LAM" Ready For Linux 6.2](https://www.phoronix.com/news/Intel-LAM-Linux-6.2)
| 时间 | 作者 | 特性 | 描述 | 是否合入主线 | 链接 |
|:----:|:----:|:---:|:----:|:---------:|:----:|
| 2021/02/05 | "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> | [Linear Address Masking enabling](https://patchwork.kernel.org/project/linux-mm/cover/20210205151631.43511-1-kirill.shutemov@linux.intel.com) | [线性地址屏蔽(LAM)](https://software.intel.com/content/dam/develop/external/us/en/documents-tps/architecture-instruction-set-extensions-programming-reference.pdf) 修改应用于 64 位线性地址的检查, 允许软件将未翻译的地址位用于元数据. 手册参见 [ISE, Chapter 14](https://patchwork.kernel.org/project/linux-mm/cover/20210205151631.43511-1-kirill.shutemov@linux.intel.com). 代码参见 [kas/linux.git](https://git.kernel.org/pub/scm/linux/kernel/git/kas/linux.git/log/?h=lam). | RFC ☐ | [PatchWork RFC,0/9](https://patchwork.kernel.org/project/linux-mm/cover/20210205151631.43511-1-kirill.shutemov@linux.intel.com)<br>*-*-*-*-*-*-*-* <br>[LORE v1,0/8](https://lore.kernel.org/r/20220610143527.22974-1-kirill.shutemov@linux.intel.com)<br>*-*-*-*-*-*-*-* <br>[LORE v1,0/11](https://lore.kernel.org/r/20220815041803.17954-1-kirill.shutemov@linux.intel.com)<br>*-*-*-*-*-*-*-* <br>[2022/08/30 LORE v1,0/11](https://lore.kernel.org/r/20220830010104.1282-1-kirill.shutemov@linux.intel.com)<br>*-*-*-*-*-*-*-* <br>[2022/09/30 LORE v1,0/14](https://lore.kernel.org/r/20220930144758.30232-1-kirill.shutemov@linux.intel.com) |
study/kernel/00-DESCRIPTION/OPEN_SOURCE.md
+1
View File
@@ -100,6 +100,7 @@
| 5.18 | NA | NA | NA |
| 5.19 | NA | NA | NA |
| 6.0 | NA | NA | [Linux 6.0 Supporting New Intel/AMD Hardware, Performance Improvements & Much More](https://www.phoronix.com/review/linux-60-features), [6.0-rc1](https://www.phoronix.com/news/Linux-6.0-rc1-Released) |
| 6.1 | NA | NA | [Linux 6.1 Features Include Initial Rust Code, MGLRU, New AMD CPU Features, More Security](https://www.phoronix.com/review/linux-61-features) |
# 6 业界会议
-------
study/kernel/00-DESCRIPTION/README.md
-2
View File
@@ -112,10 +112,8 @@ Linux 内核特性演进史
* [深入理解Linux内存管理-之-目录导航_OSKernelLAB-CSDN博客](https://kernel.blog.csdn.net/article/details/52384965)
​
* [Linux进程管理与调度-之-目录导航_OSKernelLAB-CSDN博客](https://kernel.blog.csdn.net/article/details/51456569)
​
### 1.2.3 根基(本章的重点内容)
-------
study/kernel/00-DESCRIPTION/SCHEDULER.md
+60
View File
@@ -4190,6 +4190,66 @@ Donnefort 称: 边距删除使内核能够充分利用能量模型, 任务更有
| 2021/12/20 | Vincent Donnefort <vincent.donnefort@arm.com> | [Fix stuck overutilized](https://lkml.kernel.org/lkml/20211220114323.22811-1-vincent.donnefort@arm.com) | NA | v1 ☐ | [LORE 0/3](https://lkml.kernel.org/lkml/20211220114323.22811-1-vincent.donnefort@arm.com) |
| 2022/10/06 | Pierre Gondois <pierre.gondois@arm.com> | [sched/fair: feec() improvement](https://lore.kernel.org/all/20221006081052.3862167-1-pierre.gondois@arm.com) | TODO | v2 ☐☑✓ | [LORE v2,0/1](https://lore.kernel.org/all/20221006081052.3862167-1-pierre.gondois@arm.com) |
#### 7.2.3.5 sched-domain overutilized
-------
EAS 按照能效进行选核等操作也是有一定开销的, 因此调度器 更倾向于在系统负载不高仍有余力时使用 EAS 按照能效进行调度, 而当系统负载已经很高时, 不再使用 EAS, 而是回退到原生 SMP NICE 的情况. 这就需要一种标记系统是否过载的方法.
[commit 2802bf3cd936 ("sched/fair: Add over-utilization/tipping point indicator")](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=2802bf3cd936fe2c8033a696d375a4d9d3974de4) 引入了 root-domain 的 overutilized, 当一个 CPU 的利用率(util_avg) 超过一定阈值(默认为 80%) 的时, 就认为当前 CPU 过载了, 同时也会认为整个系统也是过载的.
1. 当系统没有过载时, 通过 EAS 根据任务的实际利用率(util_avg)来进行任务的 SELECT TASK RQ, 在获得能效更高的任务分配的同时, 又不至于剥夺任何任务的执行机会. 此时将跳过 CFS 原生基于 load_avg 的 SMP NICE 的负载均衡器.
2. 当系统过载时时, 则禁用 EAS, 使能 CFS Load Balancing, 根据附加了任务权重信息的 load_avg 来进行负载均衡, 将任务分布到尽可能多的 CPU 上, 以保持 SMP NICE.
EAS 原生的 overutilized 机制非常保守, 一旦发现某个 CPU 出现了 cpu_overutilized(), 则会直接反映到 root_domain 上, 即任务整个系统都是过载的, 从而禁用 EAS, 并使能 CFS Load Balancing. 这样本来一些 CPU 或者 sched_domain 本身能从能效感知策略中获益的, 现在无法再获得任何收益.
因此不少厂商都会基于主线的策略进行优化. Linaro 为 ANDROID/AOSP 贡献了 [ANDROID: sched: Per-Sched-domain over utilization](https://git.codelinaro.org/clo/la/kernel/msm-4.14/-/commit/0dca2fc973a98de742d6df894135f04cacabb6b5) 机制, 将原来保守的 root_domain->overutilized 引入到 SD_SHARE_PKG_RESOURCES 级别. 从而一个 CPU 过载时不再会直接影响整个系统(root_domain), 而是只影响单个 sched_domain.
1. 当 CPU 过载时, CPU 所在调度域将被标记为过载.
2. 当异构系统 CPU 上存在 misfit 的任务时, 父调度域级别的负载平衡就显的很有意义, 则标记所有 SD_ASYM_CPUCAPACITY 的层级为过载的, 已确保在异构层级上选到 fits_capacity() 的 CPU.
3. 注意标记同构层级的调度域过载是没有意义的, 因为 misfit 的情况下, 同构的调度域内 CPU capacity 相同. 当某个调度域整体负载偏高, 此时可能需要 Load Balancing 有效地对任务进行负载均衡, 则标记其父调度域过载.
| 时间 | 作者 | 特性 | 描述 | 是否合入主线 | 链接 |
|:---:|:----:|:---:|:----:|:---------:|:----:|
| 2021/05/04 | Thara Gopinath <thara.gopinath@linaro.org> | [ANDROID: sched: Per-Sched-domain over utilization](https://github.com/aosp-mirror/kernel_common/commit/addef37808728c719d8c095a75bcf81befdacdaf) | per sched-domain 级别的 utilization. | v3 ☐☑✓ | [LORE](https://github.com/aosp-mirror/kernel_common/commit/addef37808728c719d8c095a75bcf81befdacdaf) |
#### 7.2.3.6 latency sensitive
-------
AOSP 4.14 中 prefer_idle 的配置.
```cpp
70a86fb70704 sched/fair: fix prefer_idle behaviour // QCOM MSM-4.14 自研补丁
e709f59a8cd2 ANDROID: restrict store of prefer_idle as boolean
dc6f7513623f ANDROID: sched/fair: return idle CPU immediately for prefer_idle
5f574ff1c8c7 ANDROID: sched/fair: add idle state filter to prefer_idle case
19c03afff299 ANDROID: sched/fair: remove order from CPU selection
03429a533a2d ANDROID:sched/fair: prefer energy efficient CPUs for !prefer_idle tasks
f240e4440655 ANDROID: Add find_best_target to minimise energy calculation overhead
c0ff131c88f6 ANDROID: sched: fair: Bypass energy-aware wakeup for prefer-idle tasks
159c14f03977 ANDROID: sched: fair/tune: Add schedtune with cgroups interface
```
AOSP 4.19 中
```cpp
f609a2239f83 ANDROID: sched/core: Move SchedTune task API into UtilClamp wrappers
d0eb1f35140e ANDROID: sched/fair: Make the EAS wake-up prefer-idle aware
c06013206c84 ANDROID: sched/fair: Bypass energy computation for prefer_idle tasks
c27c56105dca ANDROID: Add find_best_target to minimise energy calculation overhead
68dbff9ce92a ANDROID: sched: fair/tune: Add schedtune with cgroups interface
```
5.0 之后, EAS 合入主线, AOSP 5.4 及其之后的版本引入 commit 760b82c9b88d ("ANDROID: sched/fair: Bias EAS placement for latency") 为 EAS 增加 latency 感知的能力. 1232/5000
在 find_energy_efficient_cpu() 中添加一个延迟敏感 latency sensitive 的 case, 模仿 [android-4.19 以及之前版本 prefer-idle 所做的事情()](https://android.googlesource.com/kernel/common.git/+/c27c56105dcaaae54ecc39ef33fbfac87a1486fc).
#### 7.2.3.x EAS timeline
-------