使用trace-event解决系统不能深度睡眠的问题
最近遇到一个问题，系统不能睡眠到c7s, 只能睡眠到c3. (c-state不能到c7s, cpu的c-state, c0是运行态，其它状态都是idle态，睡眠的越深，c-state的值越大)
这时候第一感觉是不是系统很忙导致, 使用pert top看一下耗cpu的进程和热点函数：
1perf top -E 100 --stdio > perf-top.txt
2 19.85% perf [.] __symbols__insert
3 7.68% perf [.] rb_next
4 4.60% libc-2.26.so [.] __strcmp_sse2_unaligned
5 4.20% libelf-0.168.so [.] gelf_getsym
6 3.92% perf [.] dso__load_sym
7 3.86% libc-2.26.so [.] _int_malloc
8 3.60% libc-2.26.so [.] __libc_calloc
9 3.30% libc-2.26.so [.] vfprintf
10 2.95% perf [.] rb_insert_color
11 2.61% [kernel] [k] prepare_exit_to_usermode
12 2.51% perf [.] machine__map_x86_64_entry_trampolines
13 2.31% perf [.] symbol__new
14 2.22% [kernel] [k] do_syscall_64
15 2.11% libc-2.26.so [.] __strlen_avx2
发现系统中只有perf工具本身比较耗cpu :(
然后就想到是不是系统中某个进程搞的鬼，不让cpu睡眠到c7s. 这时候使用trace event监控一下系统中sched_switch事件. 使用trace-cmd工具监控所有cpu上的sched_switch(进程切换)事件30秒:
#trace-cmd record -e sched:sched_switch -M -1 sleep 30
2CPU0 data recorded at offset=0x63e000
3 102400 bytes in size
4CPU1 data recorded at offset=0x657000
5 8192 bytes in size
6CPU2 data recorded at offset=0x659000
7 20480 bytes in size
8CPU3 data recorded at offset=0x65e000
9 20480 bytes in size
使用trace-cmd report 查看一下监控结果，但是查看这样的原始数据不够直观，没有某个进程被切换到的统计信息:
1#trace-cmd report
2cpus=4
3 trace-cmd-19794 [001] 225127.464466: sched_switch: trace-cmd:19794 [120] S ==> swapper/1:0 [120]
4 trace-cmd-19795 [003] 225127.464601: sched_switch: trace-cmd:19795 [120] S ==> swapper/3:0 [120]
5 sleep-19796 [002] 225127.464792: sched_switch: sleep:19796 [120] S ==> swapper/2:0 [120]
6 -0 [003] 225127.471948: sched_switch: swapper/3:0 [120] R ==> rcu_sched:11 [120]
7 rcu_sched-11 [003] 225127.471950: sched_switch: rcu_sched:11 [120] W ==> swapper/3:0 [120]
8 -0 [003] 225127.479959: sched_switch: swapper/3:0 [120] R ==> rcu_sched:11 [120]
9 rcu_sched-11 [003] 225127.479960: sched_switch: rcu_sched:11 [120] W ==> swapper/3:0 [120]
10 -0 [003] 225127.487959: sched_switch: swapper/3:0 [120] R ==> rcu_sched:11 [120]
11 rcu_sched-11 [003] 225127.487961: sched_switch: rcu_sched:11 [120] W ==> swapper/3:0 [120]
12 -0 [002] 225127.491959: sched_switch: swapper/2:0 [120] R ==> kworker/2:2:19735 [120]
13 kworker/2:2-19735 [002] 225127.491972: sched_switch: kworker/2:2:19735 [120] W ==> swapper/2:0 [120]...
trace-cmd report 的结果使用正则表达式过滤一下，然后排序统计：
1trace-cmd report | grep -o '==> [^ ]\+:\?' | sort | uniq -c
2 3 ==> irqbalance:1034
3 3 ==> khugepaged:43
4 20 ==> ksoftirqd/0:10
5 1 ==> ksoftirqd/1:18
6 18 ==> ksoftirqd/3:30
7 1 ==> kthreadd:19798
8 1 ==> kthreadd:2
9 4 ==> kworker/0:0:19785
10 1 ==> kworker/0:1:19736
11 5 ==> kworker/0:1:19798
12 5 ==> kworker/0:1H:364
13 53 ==> kworker/0:2:19614
14 19 ==> kworker/1:1:7665
15 30 ==> tuned:19498
19 ...
发现可疑线程tuned，30秒内被切换到运行了30次，其它线程都是常规线程。
此时查看一下系统中是否开启了tuned服务:
果真是系统开启了tuned服务，然后拉起了名字为tuned的线程.
查看一下tuned服务的配置文件：
localhost:/home/jeff # tuned-adm active
Current active profile: sap-hana
localhost:/home/jeff # cat /usr/lib/tuned/sap-hana/tuned.conf
[main]
summary=Optimize for SAP NetWeaver, SAP HANA and HANA based products
[cpu]
force_latency = 70
发现关于cpu这一项，设置强制延迟时间为70秒 force_latency = 70 ,这个是为了优化HANA数据库。
到底force_latency怎样起作用，经过一顿搜索，发现这个值是被设置进了/dev/cpu_dma_latency
使用lsof /dev/cpu_dma_latency, 发现tuned线程确实是在操作这个文件
#lsof /dev/cpu_dma_latency
COMMAND PID USER FD TYPE DEVICE SIZE/OFF NODE NAME
tuned 18734 root 9w CHR 10,60 0t0 11400 /dev/cpu_dma_latency
而且Linux内核文档也说明了/dev/cpu_dma_latency文件,如果要对它进行写操作，要open之后写数据之后不close,如果释放掉了文件描述符它就又会恢复到默认值，这也印证了上面lsof /dev/cpu_dma_latency是有输出结果的.
https://github.com/torvalds/linux/blob/v5.8/Documentation/trace/coresight/coresight-cpu-debug.rst
As specified in the PM QoS documentation the requested parameter
will stay in effect until the file descriptor is released. For example:
# exec 3<> /dev/cpu_dma_latency; echo 0 >&3
...
Do some work...
...
# exec 3<>-
查看一下/dev/cpu_dma_latency文件的内容，确实是70，也就是(force_latency = 70)
localhost:/home/jeff # cat /dev/cpu_dma_latency | hexdump -Cv
00000000 46 00 00 00 |F...|
localhost:/home/jeff # echo $((0x46))
70
此时查看一下系统中cpu各个睡眠态的描述和延迟时间值:
# cd /sys/devices/system/cpu/cpu0/cpuidle/
# for state in * ; do
echo -e \
"STATE: $state\t\
DESC: $(cat $state/desc)\t\
NAME: $(cat $state/name)\t\
LATENCY: $(cat $state/latency)\t\
RESIDENCY: $(cat $state/residency)"
done
发现C3态的延迟时间是33微秒，C4的延时时间是133微秒，所以(force_latency = 70) ，
系统就只能睡眠到C3了 .(延迟时间就是从此睡眠态唤醒到运行态的时间)
STATE: state0 DESC: CPUIDLE CORE POLL IDLE NAME: POLL LATENCY: 0 RESIDENCY: 0
STATE: state1 DESC: MWAIT 0x00 NAME: C1 LATENCY: 2 RESIDENCY: 2
STATE: state2 DESC: MWAIT 0x01 NAME: C1E LATENCY: 10 RESIDENCY: 20
STATE: state3 DESC: MWAIT 0x10 NAME: C3 LATENCY: 33 RESIDENCY: 100
STATE: state4 DESC: MWAIT 0x20 NAME: C6 LATENCY: 133 RESIDENCY: 400
STATE: state5 DESC: MWAIT 0x32 NAME: C7s LATENCY: 166 RESIDENCY: 500
此时关闭tuned 服务, 再查看一下 /dev/cpu_dma_latency的值，变成了默认的2000秒
localhost:/home/jeff # tuned-adm off
localhost:/home/jeff # cat /dev/cpu_dma_latency | hexdump -Cv
00000000 00 94 35 77 |..5w|
localhost:/home/jeff # echo $((0x77359400))
2000000000
然后验证一下，此时系统可以睡眠到C7s了，此问题得到解决 :)
解决此问题，主要用到了Linux内核本身提供的trace-event.
所以任何一个功能都不能小看，内核就是这样，一般看上去很无聊的功能，被一些工程师用很认真的态度打磨出来之后，潜力还是非常大的:)
本公众号持续分享实际工作和学习中关于linux内核的知识总结.
偶尔也会出一些视频分享，最近根据Linux实际工程中的底层需求，设计了一个视频《Linux常见锁和lockup检测机制》发布在了阅码场平台。点击左下角阅读原文可以一键报名和试看。