如何使用 eBPF 监控 Linux 内存 OOM killer:Linux 内存调优之 eBPF 监控内存 OOM killer 事件

知不可乎骤得,托遗响于悲风 —《赤壁赋》

写在前面


  • 博文内容涉及 使用 eBPF 监控内存 OOM killer 事件,并且采集当前系统的部分相关指标数据
  • 介绍了传统的监控方式以及使用 BPF/eBPF 的方式
  • 关于 OOM killer 是什么,以及对应的内核调优参数,博客没有涉及
  • 理解不足小伙伴帮忙指正 :),生活加油

知不可乎骤得,托遗响于悲风 —《赤壁赋》

持续分享技术干货,感兴趣小伙伴可以关注下 ^_^


下面实验用的 Linux 环境

1
2
3
4
5
6
7
8
9
10
11
12
13
[root@developer ~]# hostnamectl 
Static hostname: developer
Icon name: computer-vm
Chassis: vm
Machine ID: 7ad73f2b5f7046a2a389ca780f472467
Boot ID: cef15819a5c34efa92443b6eff608cc9
Virtualization: kvm
Operating System: openEuler 22.03 (LTS-SP4)
Kernel: Linux 5.10.0-250.0.0.154.oe2203sp4.aarch64
Architecture: arm64
Hardware Vendor: OpenStack Foundation
Hardware Model: OpenStack Nova
[root@developer ~]#

下面我们谈到的 BPF 或者 eBPF 代指整个 BPF/eBPF 技术

OOM Killer 事件: OOM Killer(Out-Of-Memory Killer)​是内核在系统内存严重不足时触发的紧急机制,通过终止进程释放内存以维持系统稳定,每个进程有一个 OOM 相关的分数,终止进程的时候基于这个分数进行处理,有一些内核参数可以控制 OOM Killer 的行为,生产中考虑QOS可以进行相关的配置,当然更合理的方式是使用Cgroup对不同进程的内存资源进行限制,这里不多讲,包括 OOM killer 打分机制等等感兴趣的小伙伴可以了解下。

传统的 OOM Killer 内存事件监控

传统的 OOM killer 历史数据查看一般通过内核日志,或者是Cgroup 内存子系统的事件计数器。

Cgroup 内存子系统有 OOM 相关的事件统计, memory.events 指标,是一个内存事件计数器:

1
2
3
4
5
6
7
8
┌──[root@liruilongs.github.io]-[/usr/lib/systemd/system] 
└─$cat /sys/fs/cgroup/memory/system.slice/tuned.service/memory.events
low 0
high 0
limit_in_bytes 0
oom 0
┌──[root@liruilongs.github.io]-[/usr/lib/systemd/system]
└─$

具体的参数指标说明:

  • low: 低内存压力事件次数
  • high: 高内存压力事件次数
  • limit_in_bytes: 达到内存限制的次数
  • oom: OOM(内存耗尽)触发次数。全为 0 表示无相关事件发生。

内核日志dmesg 可以显示详细的 OOM killer 进程相关数据

下面的日志:系统因内存耗尽触发了 ​​OOM Killer​​,终止了 stress-ng 进程(PID 39693)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
[root@liruilongs.github.io ~]# dmesg -T | grep -A 30  -i "Killed process 39693"
[日 5月 11 15:41:12 2025] Out of memory: Killed process 39693 (stress-ng) total-vm:2410396kB, anon-rss:1896300kB, file-rss:4kB, shmem-rss:60kB, UID:0 pgtables:3772kB oom_score_adj:1000
[日 5月 11 15:41:13 2025] stress-ng invoked oom-killer: gfp_mask=0x100dca(GFP_HIGHUSER_MOVABLE|__GFP_ZERO), order=0, oom_score_adj=1000
[日 5月 11 15:41:13 2025] CPU: 0 PID: 39692 Comm: stress-ng Kdump: loaded Not tainted 5.10.0-250.0.0.154.oe2203sp4.aarch64 #1
[日 5月 11 15:41:13 2025] Hardware name: OpenStack Foundation OpenStack Nova, BIOS 0.0.0 02/06/2015
[日 5月 11 15:41:13 2025] Call trace:
[日 5月 11 15:41:13 2025] dump_backtrace+0x0/0x214
[日 5月 11 15:41:13 2025] show_stack+0x20/0x2c
[日 5月 11 15:41:13 2025] dump_stack+0xf0/0x138
[日 5月 11 15:41:13 2025] dump_header+0x50/0x1b0
[日 5月 11 15:41:13 2025] oom_kill_process+0x258/0x270
[日 5月 11 15:41:13 2025] out_of_memory+0xf4/0x3b0
[日 5月 11 15:41:13 2025] __alloc_pages+0x1024/0x1214
[日 5月 11 15:41:13 2025] alloc_pages_vma+0xb4/0x3e0
[日 5月 11 15:41:13 2025] do_anonymous_page+0x1d4/0x784
[日 5月 11 15:41:13 2025] handle_pte_fault+0x19c/0x240
[日 5月 11 15:41:13 2025] __handle_mm_fault+0x1bc/0x3ac
[日 5月 11 15:41:13 2025] handle_mm_fault+0xf4/0x260
[日 5月 11 15:41:13 2025] do_page_fault+0x184/0x464
[日 5月 11 15:41:13 2025] do_translation_fault+0xb8/0xe4
[日 5月 11 15:41:13 2025] do_mem_abort+0x48/0xc0
[日 5月 11 15:41:13 2025] el0_da+0x44/0x80
[日 5月 11 15:41:13 2025] el0_sync_handler+0x68/0xc0
[日 5月 11 15:41:13 2025] el0_sync+0x160/0x180
[日 5月 11 15:41:13 2025] Mem-Info:
[日 5月 11 15:41:13 2025] active_anon:4575 inactive_anon:1652614 isolated_anon:0
active_file:21 inactive_file:34 isolated_file:0
unevictable:1551 dirty:0 writeback:0
slab_reclaimable:5171 slab_unreclaimable:10557
mapped:2877 shmem:10220 pagetables:6130 bounce:0
free:14963 free_pcp:25 free_cma:0
[日 5月 11 15:41:13 2025] Node 0 active_anon:18300kB inactive_anon:6610456kB active_file:76kB inactive_file:132kB unevictable:6204kB isolated(anon):0kB isolated(file):0kB mapped:11508kB dirty:0kB writeback:0kB shmem:40880kB shmem_thp: 0kB shmem_pmdmapped: 0kB anon_thp: 3829760kB writeback_tmp:0kB kernel_stack:6384kB all_unreclaimable? yes
[root@liruilongs.github.io ~]#

dmesg 日志来看,系统因内存耗尽触发了 OOM Killer,终止了 stress-ng 进程(PID 39693),内核日志可以看到详细的数据信息,

下面是日志中内存触发条件,也就是相关的第一条日志

1
2
3
[日 5月 11 15:41:12 2025] Out of memory: Killed process 39693 (stress-ng) 
total-vm:2410396kB, anon-rss:1896300kB, file-rss:4kB, shmem-rss:60kB,
UID:0 pgtables:3772kB oom_score_adj:1000

这条日志对应字段的含义:

  • total-vm:2410396kB:进程申请的总虚拟内存为 2.4GB(含物理内存和交换空间)
  • anon-rss:1896300kB:实际占用的匿名内存(堆/栈)为 1.8GB, rss 说明是驻留的物理内存
  • file-rss:4kB: 该进程当前通过文件映射占用了 4KB 物理内存(进程通过 mmap() 映射文件)
  • shmem-rss:60kB : 该进程当前通过共享内存占用了 60KB 物理内存(共享内存通常通过 shmget()tmpfs`(如 /dev/shm)实现)
  • UID:0 : 0 表示 ​​root 用户​​
  • pgtables:3772kB: 该进程的页表占用了约 3.7MB 内存
  • oom_score_adj:1000:进程的 OOM 评分为最高(1000),因此被内核选为牺牲者。

下面的是 对应的函数调用栈:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
[日 5月 11 15:41:13 2025] Call trace:
[日 5月 11 15:41:13 2025] dump_backtrace+0x0/0x214
[日 5月 11 15:41:13 2025] show_stack+0x20/0x2c
[日 5月 11 15:41:13 2025] dump_stack+0xf0/0x138
[日 5月 11 15:41:13 2025] dump_header+0x50/0x1b0
[日 5月 11 15:41:13 2025] oom_kill_process+0x258/0x270
[日 5月 11 15:41:13 2025] out_of_memory+0xf4/0x3b0
[日 5月 11 15:41:13 2025] __alloc_pages+0x1024/0x1214 # 尝试分配物理页失败
[日 5月 11 15:41:13 2025] alloc_pages_vma+0xb4/0x3e0
[日 5月 11 15:41:13 2025] do_anonymous_page+0x1d4/0x784 # 匿名页分配失败
[日 5月 11 15:41:13 2025] handle_pte_fault+0x19c/0x240 # 页表项错误处理
[日 5月 11 15:41:13 2025] __handle_mm_fault+0x1bc/0x3ac
[日 5月 11 15:41:13 2025] handle_mm_fault+0xf4/0x260
[日 5月 11 15:41:13 2025] do_page_fault+0x184/0x464
[日 5月 11 15:41:13 2025] do_translation_fault+0xb8/0xe4
[日 5月 11 15:41:13 2025] do_mem_abort+0x48/0xc0
[日 5月 11 15:41:13 2025] el0_da+0x44/0x80
[日 5月 11 15:41:13 2025] el0_sync_handler+0x68/0xc0
[日 5月 11 15:41:13 2025] el0_sync+0x160/0x180
...

函数调用栈下部分为内存分配失败 触发 进程 OOM 以及 oom_kill_process 的调用触发 内存杀手,后面部分为一些日志转储的操作。

下面为发生 OOM Kill 时候系统全局内存状态

1
2
3
4
5
6
7
[日 5月 11 15:41:13 2025] Mem-Info:
[日 5月 11 15:41:13 2025] active_anon:4575 inactive_anon:1652614 isolated_anon:0
active_file:21 inactive_file:34 isolated_file:0
unevictable:1551 dirty:0 writeback:0
slab_reclaimable:5171 slab_unreclaimable:10557
mapped:2877 shmem:10220 pagetables:6130 bounce:0
free:14963 free_pcp:25 free_cma:0

在内核日志中看到发生异常是的内存具体指标数据

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
active_anon:4575kB    # 活跃的匿名内存(进程堆/栈等动态分配的内存)
inactive_anon:1652614kB # 不活跃的匿名内存(长期未使用的堆/栈)
isolated_anon:0kB # 被隔离的匿名内存(通常为内存故障隔离
active_file:21kB # 活跃的文件映射内存(如页缓存、打开的文件)
inactive_file:34kB # 不活跃的文件映射内存
isolated_file:0kB # 被隔离的文件映射内存
unevictable:1551kB # 无法被交换或回收的内存(如 mlock 锁定的内存)
slab_reclaimable:5171kB # 可回收的 Slab 缓存(如内核对象池)
slab_unreclaimable:10557kB # 不可回收的 Slab 缓存(如内核数据结构)
shmem:10220kB # 共享内存(如 tmpfs、IPC 通信)
mapped:2877kB # 文件映射内存(如共享库、内存映射文件)
pagetables:6130kB # 进程虚拟地址到物理地址的映射表
free:14963kB # 完全空闲的物理内存
free_pcp:25kB # 每 CPU 空闲内存(用于本地分配)
free_cma:0kB # CMA(连续内存分配器)空闲区

同时会输出对应的 NUMA 节点 0 的内存指标,可以看到几乎完全耗尽(尤其是匿名内存),且透明大页占用显著(anon_thp: 3829760kB),加剧了内存碎片化问题

1
[日 5月 11 15:41:13 2025] Node 0 active_anon:18300kB inactive_anon:6610456kB active_file:76kB inactive_file:132kB unevictable:6204kB isolated(anon):0kB isolated(file):0kB mapped:11508kB dirty:0kB writeback:0kB shmem:40880kB shmem_thp: 0kB shmem_pmdmapped: 0kB anon_thp: 3829760kB writeback_tmp:0kB kernel_stack:6384kB all_unreclaimable? yes

使用 BPF 的方式

对于 BPF 的监控,主要通过 BPF 和 bpftrace 的 oomkill 工具,我们可以在触发 OOM killer 事件之后,观察到系统平均负载等一些其他的信息,原理是通过动态插桩内核函数 oom_kill_process(),捕获 OOM Killer 触发事件

比如平均负载信息可以在 OOM 发生时提供整个系统状态的一些 上下文信息,展示出系统整体是正在变忙还是处于稳定状态,以及那个进程触发了 OOM Killer 和,被 OOM Killer 杀掉的进程是那个等数据。

用内存测试工具简单复现一下 OOM killer,我们看看如何监控,这里需要把交换分区禁用掉,要不换页进程(kswapd)疯狂的输出,不太容易触发 OOM Killer

1
[root@liruilongs.github.io ~]# swapoff -a  # 临时禁用

stress-ng 对 Linux 系统内存施加高压负载

1
2
3
4
5
[root@liruilongs.github.io ~]# stress-ng --vm 4 --vm-bytes 9.5G  --timeout 60s
stress-ng: info: [37336] setting to a 60 second run per stressor
stress-ng: info: [37336] dispatching hogs: 4 vm
^[c^Cstress-ng: info: [37336] successful run completed in 40.87s
[root@liruilongs.github.io ~]#

通过 free 命令观察内存使用情况,中间的那一次输出可以直观的看到内存使用情况

1
2
3
4
5
6
7
8
9
10
11
12
13
[root@liruilongs.github.io ~]# free -h -s 10  -c 3
total used free shared buff/cache available
Mem: 6.5Gi 815Mi 5.6Gi 37Mi 293Mi 5.7Gi
Swap: 0B 0B 0B

total used free shared buff/cache available
Mem: 6.5Gi 6.4Gi 183Mi 39Mi 113Mi 139Mi
Swap: 0B 0B 0B

total used free shared buff/cache available
Mem: 6.5Gi 4.5Gi 2.0Gi 39Mi 84Mi 2.0Gi
Swap: 0B 0B 0B
[root@liruilongs.github.io ~]#

通过 oomkill 工具观察 OOM Killer 情况

内存分配失败调用栈,上面的 BPF 工具实际上是在 oom_kill_process 内核函数处埋点实现的​

可以看到触发的进程主要是 stress-ng(内存压力测试工具)持续申请内存,导致系统物理内存耗尽。部分系统进程(如 oeaware、Xvnc)也触发 OOM,说明内存竞争激烈,系统整体处于高压状态。通过负载指标​​:loadavg 值较高(如 4.59),表明 CPU 资源负载在升高 。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
[root@liruilongs.github.io ~]# /usr/share/bcc/tools/oomkill 
Tracing OOM kills... Ctrl-C to stop.
15:41:14 Triggered by PID 1039 ("oeaware"), OOM kill of PID 39693 ("stress-ng"), 1704429 pages, loadavg: 4.34 2.87 1.77 6/396 39695
15:41:15 Triggered by PID 39692 ("stress-ng"), OOM kill of PID 39692 ("stress-ng"), 1704429 pages, loadavg: 4.34 2.87 1.77 5/396 39696
15:41:16 Triggered by PID 39696 ("stress-ng"), OOM kill of PID 39694 ("stress-ng"), 1704429 pages, loadavg: 4.31 2.89 1.78 5/396 39697
15:41:17 Triggered by PID 39698 ("stress-ng"), OOM kill of PID 39695 ("stress-ng"), 1704429 pages, loadavg: 4.31 2.89 1.78 5/396 39699
15:41:19 Triggered by PID 1039 ("oeaware"), OOM kill of PID 39696 ("stress-ng"), 1704429 pages, loadavg: 4.31 2.89 1.78 5/396 39700
15:41:20 Triggered by PID 2121 ("ibus-ui-gtk3"), OOM kill of PID 39697 ("stress-ng"), 1704429 pages, loadavg: 4.31 2.89 1.78 6/396 39701
15:41:22 Triggered by PID 39699 ("stress-ng"), OOM kill of PID 39698 ("stress-ng"), 1704429 pages, loadavg: 4.29 2.91 1.80 5/396 39701
15:41:23 Triggered by PID 39700 ("stress-ng"), OOM kill of PID 39700 ("stress-ng"), 1704429 pages, loadavg: 4.29 2.91 1.80 6/396 39702
15:41:24 Triggered by PID 39701 ("stress-ng"), OOM kill of PID 39699 ("stress-ng"), 1704429 pages, loadavg: 4.29 2.91 1.80 5/396 39704
15:41:25 Triggered by PID 39702 ("stress-ng"), OOM kill of PID 39701 ("stress-ng"), 1704429 pages, loadavg: 4.29 2.91 1.80 5/396 39704
15:41:26 Triggered by PID 39703 ("stress-ng"), OOM kill of PID 39702 ("stress-ng"), 1704429 pages, loadavg: 4.59 3.00 1.83 5/396 39705
15:41:27 Triggered by PID 39705 ("stress-ng"), OOM kill of PID 39703 ("stress-ng"), 1704429 pages, loadavg: 4.59 3.00 1.83 5/396 39706
15:41:29 Triggered by PID 1304 ("Xvnc"), OOM kill of PID 39704 ("stress-ng"), 1704429 pages, loadavg: 4.59 3.00 1.83 6/396 39708
15:41:30 Triggered by PID 1492 ("lightdm-gtk-gre"), OOM kill of PID 39705 ("stress-ng"), 1704429 pages, loadavg: 4.59 3.00 1.83 5/395 39708
15:41:31 Triggered by PID 39709 ("stress-ng"), OOM kill of PID 39706 ("stress-ng"), 1704429 pages, loadavg: 4.94 3.10 1.87 8/395 39710

看下一下输出的指标信息,已第一条日志为例

1
15:41:14 Triggered by PID 1039 ("oeaware"), OOM kill of PID 39693 ("stress-ng"), 1704429 pages, loadavg: 4.34 2.87 1.77 6/396 39695。
字段 含义
Triggered by PID 触发 OOM 的进程 PID(如内存申请者)
OOM kill of PID 被 OOM Killer 终止的进程 PID
1704429 pages 被终止进程占用的物理内存页数(1页=4KB,换算为 6.8GB)
loadavg 系统负载(1分钟/5分钟/15分钟平均负载)
6/396 当前可运行进程数/总进程数
39695 最后被创建的进程 PID

当然上面的输出的功能有些简单,如果我们希望获取更多的数据信息,我们可以通过修改原来脚本的方式实现

bpftrace 对应的脚本

https://github.com/brendangregg/bpf-perf-tools-book/blob/master/originals/Ch07_Memory/oomkill.bt

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
[root@liruilongs.github.io ~]# cat  /usr/share/bpftrace/tools/oomkill.bt 
#!/usr/bin/bpftrace
/*
* oomkill Trace OOM killer.
* For Linux, uses bpftrace and eBPF.
*
* This traces the kernel out-of-memory killer, and prints basic details,
* including the system load averages. This can provide more context on the
* system state at the time of OOM: was it getting busier or steady, based
* on the load averages? This tool may also be useful to customize for
* investigations; for example, by adding other task_struct details at the
* time of the OOM, or other commands in the system() call.
*
* This currently works by using kernel dynamic tracing of oom_kill_process().
*
* USAGE: oomkill.bt
*
* Copyright 2018 Netflix, Inc.
* Licensed under the Apache License, Version 2.0 (the "License")
*
* 07-Sep-2018 Brendan Gregg Created this.
*/

#include <linux/oom.h>

BEGIN
{
printf("Tracing oom_kill_process()... Hit Ctrl-C to end.\n");
}

kprobe:oom_kill_process
{
$oc = (struct oom_control *)arg0;
time("%H:%M:%S ");
printf("Triggered by PID %d (\"%s\"), ", pid, comm);
printf("OOM kill of PID %d (\"%s\"), %d pages, loadavg: ",
$oc->chosen->pid, $oc->chosen->comm, $oc->totalpages);
cat("/proc/loadavg");
}
[root@liruilongs.github.io ~]#

通过动态插桩内核函数 oom_kill_process(),捕获 OOM Killer 触发事件,同时输出了一些其他的指标信息

自定义 OOM Killer 发生时的性能指标采集

下面是最上面脚本的基础上添加的一些他的指标数据采集,从而实现在 OOM Killer 发生时快速的定位问题

添加 /proc/meminfo 的全局内存指标信息,meminfo提供了系统范围内内存统计数据的超集,包括了vmstat、top、free和procinfo的信息

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
#include <linux/oom.h>

BEGIN
{
printf("Tracing oom_kill_process()... Hit Ctrl-C to end.\n");
}

kprobe:oom_kill_process
{
$oc = (struct oom_control *)arg0;
$task = $oc->chosen;
time("%H:%M:%S ");
printf("Triggered by PID %d (\"%s\"), ", pid, comm);
printf("OOM kill of PID %d (\"%s\"), %d pages, loadavg: ",
$oc->chosen->pid, $oc->chosen->comm, $oc->totalpages);
cat("/proc/loadavg");
print("当前系统内存性能统计信息:");
cat("/proc/meminfo");
}

修改脚本后的再次监控指标采集

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
[root@developer tools]# vim oomkill.bt
[root@developer tools]# ./oomkill.bt
Attaching 2 probes...
Tracing oom_kill_process()... Hit Ctrl-C to end.
19:06:46 Triggered by PID 1039 ("oeaware"), OOM kill of PID 1528049 ("stress-ng"), 1704429 pages, loadavg: 4.14 2.26 1.35 5/405 1528051
当前系统内存性能统计信息:
MemTotal: 6817716 kB
MemFree: 1494936 kB
MemAvailable: 1442388 kB
Buffers: 372 kB
Cached: 155720 kB
SwapCached: 0 kB
Active: 31476 kB
Inactive: 5156300 kB
Active(anon): 31388 kB
Inactive(anon): 5117004 kB
Active(file): 88 kB
Inactive(file): 39296 kB
Unevictable: 6236 kB
Mlocked: 76 kB
SwapTotal: 0 kB
SwapFree: 0 kB
Dirty: 0 kB
Writeback: 0 kB
AnonPages: 5038288 kB
Mapped: 21148 kB
Shmem: 116328 kB
KReclaimable: 21692 kB
Slab: 66816 kB
SReclaimable: 21692 kB
SUnreclaim: 45124 kB
KernelStack: 6528 kB
PageTables: 22160 kB
NFS_Unstable: 0 kB
Bounce: 0 kB
WritebackTmp: 0 kB
CommitLimit: 3408856 kB
Committed_AS: 12252200 kB
VmallocTotal: 135290159040 kB
VmallocUsed: 16152 kB
VmallocChunk: 0 kB
Percpu: 2800 kB
HardwareCorrupted: 0 kB
AnonHugePages: 4009984 kB
ShmemHugePages: 0 kB
ShmemPmdMapped: 0 kB
FileHugePages: 0 kB
FilePmdMapped: 0 kB
CmaTotal: 0 kB
CmaFree: 0 kB
HugePages_Total: 0
HugePages_Free: 0
HugePages_Rsvd: 0
HugePages_Surp: 0
Hugepagesize: 2048 kB
Hugetlb: 0 kB

对于内核态的内存分配,可以添加 /proc/slabinfo 相关指标的采集

上面是全局的内存信息的指标采集,当然也可以采集对被 kill 进程以及 创建的进程的相关的指标信息

下面的为提供的 BCC 版本的 oomkill 工具

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
[root@liruilongs.github.io ~]# cat /usr/share/bcc/tools/oomkill 
#!/usr/bin/python3
#
# oomkill Trace oom_kill_process(). For Linux, uses BCC, eBPF.
#
# This traces the kernel out-of-memory killer, and prints basic details,
# including the system load averages. This can provide more context on the
# system state at the time of OOM: was it getting busier or steady, based
# on the load averages? This tool may also be useful to customize for
# investigations; for example, by adding other task_struct details at the time
# of OOM.
#
# Copyright 2016 Netflix, Inc.
# Licensed under the Apache License, Version 2.0 (the "License")
#
# 09-Feb-2016 Brendan Gregg Created this.

from bpfcc import BPF
from time import strftime

# linux stats
loadavg = "/proc/loadavg"

# define BPF program
bpf_text = """
#include <uapi/linux/ptrace.h>
#include <linux/oom.h>

struct data_t {
u32 fpid;
u32 tpid;
u64 pages;
char fcomm[TASK_COMM_LEN];
char tcomm[TASK_COMM_LEN];
};

BPF_PERF_OUTPUT(events);

void kprobe__oom_kill_process(struct pt_regs *ctx, struct oom_control *oc, const char *message)
{
unsigned long totalpages;
struct task_struct *p = oc->chosen;
struct data_t data = {};
u32 pid = bpf_get_current_pid_tgid() >> 32;
data.fpid = pid;
data.tpid = p->pid;
data.pages = oc->totalpages;
bpf_get_current_comm(&data.fcomm, sizeof(data.fcomm));
bpf_probe_read_kernel(&data.tcomm, sizeof(data.tcomm), p->comm);
events.perf_submit(ctx, &data, sizeof(data));
}
"""

# process event
def print_event(cpu, data, size):
event = b["events"].event(data)
with open(loadavg) as stats:
avgline = stats.read().rstrip()
print(("%s Triggered by PID %d (\"%s\"), OOM kill of PID %d (\"%s\")"
", %d pages, loadavg: %s") % (strftime("%H:%M:%S"), event.fpid,
event.fcomm.decode('utf-8', 'replace'), event.tpid,
event.tcomm.decode('utf-8', 'replace'), event.pages, avgline))

# initialize BPF
b = BPF(text=bpf_text)
print("Tracing OOM kills... Ctrl-C to stop.")
b["events"].open_perf_buffer(print_event)
while 1:
try:
b.perf_buffer_poll()
except KeyboardInterrupt:
exit()
[root@liruilongs.github.io ~]#

我们对这个工具做一些简单的修改, /proc/pid/status 用于展示当前进程的一些基本指标

1
2
3
4
5
6
7
8
9
10
11
12
# process event
def print_event(cpu, data, size):
event = b["events"].event(data)
with open(loadavg) as stats:
avgline = stats.read().rstrip()
with open("/proc/"+ str(event.fpid) +"/status" ) as statm:
statmtable = statm.read().rstrip()
print(("%s Triggered by PID %d (\"%s\"), OOM kill of PID %d (\"%s\")"
", %d pages, loadavg: %s") % (strftime("%H:%M:%S"), event.fpid,
event.fcomm.decode('utf-8', 'replace'), event.tpid,
event.tcomm.decode('utf-8', 'replace'), event.pages, avgline))
print("新进程指标信息: &s",statmtable)

下面为输出结果

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
[root@developer tools]# ./oomkill
Tracing OOM kills... Ctrl-C to stop.
19:42:24 Triggered by PID 1539774 ("stress-ng"), OOM kill of PID 1539775 ("stress-ng"), 1704429 pages, loadavg: 2.40 1.72 1.07 5/407 1539778
新进程指标信息: &s Name: stress-ng
Umask: 0077
State: R (running)
Tgid: 1539774
Ngid: 0
Pid: 1539774
PPid: 1539770
TracerPid: 0
Uid: 0 0 0 0
Gid: 0 0 0 0
FDSize: 64
Groups: 0
NStgid: 1539774
NSpid: 1539774
NSpgid: 1539769
NSsid: 32074
VmPeak: 2410404 kB
VmSize: 2410404 kB
VmLck: 0 kB
VmPin: 0 kB
VmHWM: 1415824 kB
VmRSS: 1415824 kB
RssAnon: 1415760 kB
RssFile: 4 kB
RssShmem: 60 kB
VmData: 2369388 kB
VmStk: 132 kB
VmExe: 1416 kB
VmLib: 2888 kB
VmPTE: 2828 kB
VmSwap: 0 kB
HugetlbPages: 0 kB
CoreDumping: 0
THP_enabled: 1
Threads: 1
SigQ: 0/26524
SigPnd: 0000000000000000
ShdPnd: 0000000000000000
SigBlk: 0000000000000000
SigIgn: 0000000008300a00
SigCgt: 000000002380e0af
CapInh: 0000000000000000
CapPrm: 000001ffffffffff
CapEff: 000001ffffffffff
CapBnd: 000001ffffffffff
CapAmb: 0000000000000000
NoNewPrivs: 0
Seccomp: 0
Seccomp_filters: 0
Speculation_Store_Bypass: vulnerable
Cpus_allowed: f
Cpus_allowed_list: 0-3
Mems_allowed: 00000000,00000000,00000000,00000001
Mems_allowed_list: 0
voluntary_ctxt_switches: 32
nonvoluntary_ctxt_switches: 837
Cpus_preferred: 0
Cpus_preferred_list:

我们可以同时采集到 VmRSS 等有用的指标信息。

BPF 跟踪工具可以给各种内存行为提供更多的信息,可以用 BPF 跟踪软件事件系统调用缺页错误相关的跟踪点来分析;还可以使用 kprobes 跟踪内核中内存分配的函数;或使用 uprobes 来跟踪库函数、应用程序运行时,以及应用程序自带的内存分配器;或使用 USDT 探针来跟踪 libc 内存分配器事件;以及使用 PMC对内存访问进行溢出采样。感兴趣的小伙伴可以深入了解。

博文部分内容参考

© 文中涉及参考链接内容版权归原作者所有,如有侵权请告知 :)


《BPF Performance Tools》


© 2018-至今 liruilonger@gmail.com, 保持署名-非商用-相同方式共享(CC BY-NC-SA 4.0)

如何使用 eBPF 监控 Linux 内存 OOM killer:Linux 内存调优之 eBPF 监控内存 OOM killer 事件

https://liruilongs.github.io/2025/04/10/待发布/如何使用 eBPF 监控 Linux 内存 OOM killer:Linux 内存调优之 eBPF 监控内存 OOM killer 事件/

发布于

2025-04-10

更新于

2025-05-19

许可协议

评论
Your browser is out-of-date!

Update your browser to view this website correctly.&npsb;Update my browser now

×