nginx cpu高排查

首先查看nginx的error日志，无异常打印。

cpu占用如下图所示：

top - 10:05:40 up 233 days, 16:28, 4 users, load average: 25.53, 25.39, 21.62
Tasks: 836 total, 20 running, 816 sleeping, 0 stopped, 0 zombie
Cpu(s): 51.8%us, 4.1%sy, 0.0%ni, 31.0%id, 6.1%wa, 0.0%hi, 6.9%si, 0.0%st
Mem: 257940M total, 254533M used, 3407M free, 6481M buffers
Swap: 0M total, 0M used, 0M free, 190273M cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
16645 root 20 0 15.6g 7.2g 7.1g R 99 2.9 19888:06 nginx
16643 root 20 0 15.6g 7.2g 7.1g R 98 2.9 19917:31 nginx
16641 root 20 0 15.6g 7.2g 7.1g R 98 2.9 19547:57 nginx
16661 root 20 0 15.6g 7.2g 7.1g R 97 2.9 19957:15 nginx
16659 root 20 0 15.6g 7.2g 7.1g R 97 2.9 20032:33 nginx
16637 root 20 0 15.6g 7.2g 7.1g R 96 2.9 21085:44 nginx
16640 root 20 0 15.6g 7.2g 7.1g R 96 2.9 16897:32 nginx
16644 root 20 0 15.6g 7.2g 7.1g R 96 2.9 19860:42 nginx
16638 root 20 0 15.6g 7.2g 7.1g R 96 2.9 16362:11 nginx
16648 root 20 0 15.6g 7.2g 7.1g R 96 2.9 16627:30 nginx
16669 root 20 0 15.6g 7.2g 7.1g R 96 2.9 17290:48 nginx
16653 root 20 0 15.6g 7.2g 7.1g R 96 2.9 17093:40 nginx
16656 root 20 0 15.6g 7.2g 7.1g R 96 2.9 19544:29 nginx

nginx的cpu占用很高，且处于用户态占用。单独查看16645进程的cpu占用如下：

top -H -p 16645

top - 10:06:09 up 233 days, 16:28, 4 users, load average: 26.95, 25.73, 21.83
Tasks: 68 total, 1 running, 67 sleeping, 0 stopped, 0 zombie
Cpu(s): 8.0%us, 5.5%sy, 0.0%ni, 63.8%id, 13.0%wa, 0.0%hi, 9.7%si, 0.0%st
Mem: 257940M total, 254387M used, 3552M free, 6481M buffers
Swap: 0M total, 0M used, 0M free, 190106M cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
16645 root 20 0 15.6g 7.2g 7.1g R 94 2.9 15689:54 nginx
18749 root 20 0 15.6g 7.2g 7.1g S 0 2.9 42:31.63 nginx
18751 root 20 0 15.6g 7.2g 7.1g S 0 2.9 63:29.10 nginx
18753 root 20 0 15.6g 7.2g 7.1g S 0 2.9 64:42.03 nginx
18755 root 20 0 15.6g 7.2g 7.1g S 0 2.9 65:18.44 nginx
18757 root 20 0 15.6g 7.2g 7.1g S 0 2.9 65:02.60 nginx

还好，是单独一个线程cpu高，以前遇到复杂的，是cpu各个线程之间跳着高。

想用perf看一下热点函数：

perf top -p 16645
If 'perf' is not a typo you can run the following command to lookup the package that contains the binary:
command-not-found perf
-bash: perf: command not found

发现命令没装，由于是线上环境，出现一次也不容易，所以只能attch上去手工采样，看这个线程在忙什么。

由于是多线程的进程，所以要确定线程号，然后看堆栈，info thr 确定线程号之后，切换到对应的堆栈，然后bt查看堆栈。

部分堆栈如下：

(gdb) info thr
Id Target Id Frame
68 Thread 0x7f4ef02e2700 (LWP 17129) "nginx" 0x00007f5154c1666c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
67 Thread 0x7f4eefae1700 (LWP 17132) "nginx" 0x00007f5154c1666c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0

......................//省略中间部分

2 Thread 0x7f4d71635700 (LWP 19135) "nginx" 0x00007f5154015b1d in nanosleep () from /lib64/libc.so.6
* 1 Thread 0x7f5155419720 (LWP 16645) "nginx" 0x00007f51547b4ac0 in ?? () from /usr/local/lib/libluajit-5.1.so.2

(gdb) thr 1
[Switching to thread 1 (Thread 0x7f5155419720 (LWP 16645))]

对应的堆栈如下：

(gdb) bt
#0 0x00007f51547b47e7 in ?? () from /usr/local/lib/libluajit-5.1.so.2
#1 0x00007f51547b4865 in ?? () from /usr/local/lib/libluajit-5.1.so.2
#2 0x00007f51547b4af4 in ?? () from /usr/local/lib/libluajit-5.1.so.2
#3 0x00007f51547b4c17 in ?? () from /usr/local/lib/libluajit-5.1.so.2
#4 0x00007f51547b489f in ?? () from /usr/local/lib/libluajit-5.1.so.2
#5 0x00007f51547b54aa in ?? () from /usr/local/lib/libluajit-5.1.so.2
#6 0x00007f515476c4eb in ?? () from /usr/local/lib/libluajit-5.1.so.2
#7 0x00000000005400ab in ngx_http_lua_run_thread (L=0x417ac378, r=0xc26fd0, ctx=0xed1800, nrets=1) at ./required/lua-nginx-module-0.9.19/src/ngx_http_lua_util.c:1015
#8 0x000000000052636a in ngx_http_lua_subrequest_resume (r=0xc26fd0) at ./required/lua-nginx-module-0.9.19/src/ngx_http_lua_subrequest.c:1595
#9 0x0000000000545b0f in ngx_http_lua_content_wev_handler (r=0xc26fd0) at ./required/lua-nginx-module-0.9.19/src/ngx_http_lua_contentby.c:131

由于用户态采样毕竟频率过低，最好要进行多次采样确定地址，连续多次采样，确认堆栈不变。

ngx_http_lua_run_thread 是nginx调用的lua第三方模块，根据堆栈r=0xc26fd0，我们可以获取到如下信息：

(gdb) p ((ngx_http_request_t*)0xc1cab0).request_line.data
$22 = (
u_char *) 0x1a4b5e0 "GET /001/2/ch00000090990000001073/2560000.m3u8?"...

很可惜，url打印不全，这个是由于默认情况下，gdb会根据性能要求，打印一定长度的字符串，修改下gdb的字符串打印配置如下：

(gdb) set print element 0           -----------这个0就是全部打印，如果是200，就默认打印200个字符，前提是没有遇到\0

比如设置为1，则打印是如下：

(gdb) set print element 1
(gdb) p ((ngx_http_request_t*)0xc1cab0).request_line.data
$28 = (u_char *) 0x1a4b5e0 "G"...

重新设置为0，打印的数据涉及到保密，就不展示了，反正获取到了url，接下来就轻松了，curl请求该url，cpu立刻飙升至100，完美复现。

后面的分析就不重要了，gdb跟一把ok。收工。