Create by Billow.Jen,2020.3.8
利用linux kernel 动态追踪技术,排查问题本身就可能会变成一个非常有趣的过程,让我们遇到线上的诡异问题就感到兴奋,就仿佛好不容易又逮着机会,可以去解一道迷人的谜题。
负载均衡模块独立主机部署,运行BIRD、quagga、nftbl、contiv、ovs等进程,实现网络NAT功能。 内核版本:centos7-4.19.8
在调用nftlb写入nft规则的过程中,发现部分主机性能低(因操作系统为centos且内核进行了升级,红帽服务不支持。) 过程如下: 1、执行命令:
time ip netns exec vpc067871207-lbalance curl -H "Key: spidernet" -X POST http://127.0.0.1:06787/farms --data '{"farms":[{"name":"b1","family":"ipv4","virtual-addr":"XX.XX.XX.XX","virtual-ports":"1001","mode":"snat","protocol":"tcp","scheduler":"weight","state":"up","backends":[{"name":"bck01","ip-addr":"192.168.0.1","port":"80","weight":"5","mark":"0x00000001","priority":"1","state":"up"}]}]}'
该命令正常应该在0.2ms内完成,但在部分主机上耗时到2、3s。 2、利用strace命令跟踪:
strace -tt -T -v -f -s 1024 -p 2727846
发现耗时在sendmsg(AF_NETLINK…)函数上,有长有短,短的在几十us,但有四个长在0.5s左右的sendmsg,如: [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-eAfNbx6U-1584208688223)(en-resource://database/1982:1)] 3、继续分析主机进程,发现一个进程kworker一直100% [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-QxeEl9NA-1584208688224)(en-resource://database/1984:1)] 4、观察主机性能,nmon如下:[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-ewVOFHNl-1584208688225)(en-resource://database/1990:1)] 此时主机内存、cpu、网络、存储都很空闲,但中断比正常主机高1~2倍。 5、用perf工具查看kworker耗时 [外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-CEyDfmjJ-1584208688226)(en-resource://database/1988:1)] 6、网卡中断,经确认(/proc/interrupts)网卡中断分布在每个cpu上,网络中断均衡 7、继续ftrace,分析事件:
cd /sys/kernel/debug/tracing
echo wakeup_rt > current_tracer
echo 0 > options/function-trace
echo 1 > events/enable
echo 1 > tracing_on
echo 0 > tracing_max_latency
chrt -f 5 sleep 1
echo 0 > tracing_on
cat trace
结果没有针对性,不具备进一步分析条件。初步问题分析完了,下步进行内核态深度分析。
内核该kworker进程的性能影响了work处理效率,导致内核较慢,响应sendmsg的work延迟,导致curl耗时长。
ftrace、kprobe、perf、operf/oprofile、systemtap都是跟踪内核的好工具,但有所区别。 Linux内核提供的基础设施:
kprobe作为轻量级内核调试工具,在诊断内核bug时有着先天独厚的优势,相关其他工具,kprobe有如下优点: 1、不用更新内核 2、可以以模块的形式加载进内核,用完后直接卸载即可,不会对内核造成污染 3、动态跟踪,自由构造,灵巧轻便。
经过大量实验探索,依赖少并可深入自定义的,kprobe胜出。
调试源码见附录,kprobe使用及原理自行学习,本文不做基础介绍,重点在思路、步骤,主要展示方法论的形成过程。 准备工作,说明下kprobe的参数示例:
struct file *file = (struct file *)regs->di;
//因为x86的参数传递规则是di,si,dx,cx,r8,r9,所以di就是vfs_write的第一个参数。arm默认是r0,r1,r2,r3,相应的取r0
看看kworker在做什么:
[root@04b280305 kprobe]# cat /proc/352560/stack
[<0>] insert_work+0x6e/0xa0
[<0>] __queue_work+0x131/0x3b0
[<0>] queue_work_on+0x28/0x40
[<0>] rht_deferred_worker+0x8b/0x3e0
[<0>] process_one_work+0x179/0x390
[<0>] worker_thread+0x4f/0x3e0
[<0>] kthread+0x105/0x140
[<0>] ret_from_fork+0x35/0x40
[<0>] 0xffffffffffffffff
可知,rht_deferred_worker是特定任务,需要重点分析。 写kprobe代码对rht_deferred_worker函数加探针,分析rht_deferred_worker事件(/var/log/messages,当时dmesg没记录被覆盖了,所以从messages中直接看):
[root@04b280305 kprobe]# grep rht_deferred_worker /var/log/messages |tail
Mar 9 21:35:14 04b280305 kernel: <rht_deferred_worker> pre_handler: 813f99c0
Mar 9 21:35:14 04b280305 kernel: <rht_deferred_worker> post_handler: 813f99c0
Mar 9 21:35:14 04b280305 kernel: <rht_deferred_worker> pre_handler: 813f99c0
Mar 9 21:35:14 04b280305 kernel: <rht_deferred_worker> post_handler: 813f99c0
Mar 9 21:35:14 04b280305 kernel: <rht_deferred_worker> pre_handler: 813f99c0
看看813f99c0是什么work:
[root@04b280305 kprobe]# cat /proc/kallsyms |grep 813f99c0
ffffffff813f99c0 t rht_deferred_worker
分析rht_deferred_worker源码,得知在重新分配或者收缩失败时,会触发插入work:rht_deferred_worker,此处是关键。
static void rht_deferred_worker(struct work_struct *work)
{
struct rhashtable *ht;
struct bucket_table *tbl;
int err = 0;
ht = container_of(work, struct rhashtable, run_work);
mutex_lock(&ht->mutex);
tbl = rht_dereference(ht->tbl, ht);
tbl = rhashtable_last_table(ht, tbl);
if (rht_grow_above_75(ht, tbl))
err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
err = rhashtable_shrink(ht);
else if (tbl->nest)
err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
if (!err)//出错时,返回非0,执行不到
err = rhashtable_rehash_table(ht);
mutex_unlock(&ht->mutex);
if (err)
schedule_work(&ht->run_work);
/*上行是关键,造成了死循环*/
}
确认下是alloc失败还是shrink失败: 写kprobe代码对 rhashtable_rehash_alloc函数加探针,查看结果:
Mar 9 22:50:03 04b280305 kernel: Planted kprobe at rhashtable_rehash_alloc+0x0/0x50
Mar 9 22:50:07 04b280305 kernel: kprobe at rhashtable_rehash_alloc+0x0/0x50 unregistered
只执行了一次,说明不是alloc的问题,继续分析shrink,rhashtable_shrink不是系统导出符号,不能kprobe,看下源码:
static int rhashtable_shrink(struct rhashtable *ht)
{
struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
unsigned int nelems = atomic_read(&ht->nelems);
unsigned int size = 0;
if (nelems)
size = roundup_pow_of_two(nelems * 3 / 2);
if (size < ht->p.min_size)
size = ht->p.min_size;
if (old_tbl->size <= size)
return 0;
if (rht_dereference(old_tbl->future_tbl, ht))
return -EEXIST;
return rhashtable_rehash_alloc(ht, old_tbl, size);
}
写了个test_shirnk的探测函数,一安装服务器崩了… 去看看高版本内核,从5.1版本解决了此问题:
static void rht_deferred_worker(struct work_struct *work)
{
struct rhashtable *ht;
struct bucket_table *tbl;
int err = 0;
ht = container_of(work, struct rhashtable, run_work);
mutex_lock(&ht->mutex);
tbl = rht_dereference(ht->tbl, ht);
tbl = rhashtable_last_table(ht, tbl);
if (rht_grow_above_75(ht, tbl))
err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
err = rhashtable_shrink(ht);
else if (tbl->nest)
err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
/*此块代码,修复了BUG*/
if (!err || err == -EEXIST) {
int nerr;
nerr = rhashtable_rehash_table(ht);
err = err ?: nerr;
}
mutex_unlock(&ht->mutex);
if (err)
schedule_work(&ht->run_work);
}
BUG说明: rhashtable: Still do rehash when we get EEXIST As it stands if a shrink is delayed because of an outstanding rehash, we will go into a rescheduling loop without ever doing the rehash.
This patch fixes this by still carrying out the rehash and then rescheduling so that we can shrink after the completion of the rehash should it still be necessary.
The return value of EEXIST captures this case and other cases (e.g., another thread expanded/rehashed the table at the same time) where we should still proceed with the rehash.
Fixes: da20420 (“rhashtable: Add nested tables”)
英文不知所云,我自己的理解是: 哈希表收缩时(rhashtable_shrink),会去判断该表是否被间接引用,如果存在则不缩容,返回-EEXIST。 v5.1之前的版本,该返回值会触发重新生成一个rht_deferred_worker的work,这样就形成了递归,如果间接引用一直存在则形成死循环,导致CPU 100%。 v5.1之后的版本,该返回值会触发重新哈希,若此期间间接引用还存在则返回0,不再会触发产生新rht_deferred_worker的work,以此来解决BUG。
探测完注意尽快卸载驱动,影响性能,会打印大量日志,并且可能会把服务器搞崩
从https://www.kernel.org下载patch文件,升级解决。
$ diff -up linux-4.19.8-1/lib/rhashtable.c linux-4.19.8-2/lib/rhashtable.c > rht_patch
$ patch -p1 < rht_patch
遇到操作系统性能或功能问题,可采用如下步骤定位:
内核调试博大精深,本文非常有局限性,具体问题还得具体分析
推荐几个链接: https://zhuanlan.zhihu.com/p/71437161 https://www.csdndoc.com/article/8015807 https://blog.csdn.net/yuntongsf/article/details/78707576
Talk is cheap,show you the code! kprobe_kworker.c
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/kprobes.h>
#define MAX_SYMBOL_LEN 64
static char symbol[MAX_SYMBOL_LEN] = "rht_deferred_worker";
//rht_deferred_worker可根据需要改成要探测的函数
module_param_string(symbol, symbol, sizeof(symbol), 0644);
/* For each probe you need to allocate a kprobe structure */
static struct kprobe kp = {
.symbol_name = symbol,
};
/* kprobe pre_handler: called just before the probed instruction is executed */
static int handler_pre(struct kprobe *p, struct pt_regs *regs)
{
#ifdef CONFIG_X86
pr_info("<%s>,pid = %ld\n",current->comm, current->pid);
struct work_struct *work = (struct work_struct *)regs->di;
pr_info("<%s> pre_handler: %x\n",p->symbol_name,work->func);
#endif
#ifdef CONFIG_ARM64
pr_info("<%s> pre_handler: p->addr = %pF, pc = 0x%lx,"
" pstate = 0x%lx\n",
p->symbol_name, p->addr, (long)regs->pc, (long)regs->pstate);
#endif
/* A dump_stack() here will give a stack backtrace */
return 0;
}
/* kprobe post_handler: called after the probed instruction is executed */
static void handler_post(struct kprobe *p, struct pt_regs *regs,
unsigned long flags)
{
#ifdef CONFIG_X86
struct work_struct *work = (struct work_struct *)regs->di;
pr_info("<%s> post_handler: %x\n",p->symbol_name,work->func);
// pr_info("<%s> post_handler: p->addr = %pF, flags = 0x%lx\n",
// p->symbol_name, p->addr, regs->flags);
#endif
#ifdef CONFIG_ARM64
pr_info("<%s> post_handler: p->addr = %pF, pstate = 0x%lx\n",
p->symbol_name, p->addr, (long)regs->pstate);
#endif
}
/*
* fault_handler: this is called if an exception is generated for any
* instruction within the pre- or post-handler, or when Kprobes
* single-steps the probed instruction.
*/
static int handler_fault(struct kprobe *p, struct pt_regs *regs, int trapnr)
{
pr_info("fault_handler: p->addr = %pF, trap #%dn", p->addr, trapnr);
/* Return 0 because we don't handle the fault. */
return 0;
}
static int __init kprobe_init(void)
{
int ret;
kp.pre_handler = handler_pre;
kp.post_handler = handler_post;
kp.fault_handler = handler_fault;
ret = register_kprobe(&kp);
if (ret < 0) {
pr_err("register_kprobe failed, returned %d\n", ret);
return ret;
}
pr_info("Planted kprobe at %pF\n", kp.addr);
return 0;
}
static void __exit kprobe_exit(void)
{
unregister_kprobe(&kp);
pr_info("kprobe at %pF unregistered\n", kp.addr);
}
module_init(kprobe_init)
module_exit(kprobe_exit)
MODULE_LICENSE("GPL");
Makefile
obj-m := kprobe_kworker.o
KBUILD_EXTRA_SYMBOLS:=/mod_a/Module.symvers
CROSS_COMPILE=''
KDIR := /lib/modules/4.19.8-1.el7.elrepo.x86_64/build
all:
make -C $(KDIR) M=$(PWD) modules
clean:
rm -f *.ko *.o *.mod.o *.mod.c .*.cmd *.symvers modul*
执行命令
make
insmod kprobe_kworker.ko
rmmod kprobe_kworker.ko
dmesg查看结果
dmesg
版权声明:本文内容由互联网用户自发贡献,该文观点仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 举报,一经查实,本站将立刻删除。