前几天在现网部署软件时,发现一个进程占用CPU非常非常高,仔细探查原因,发现是处理消息时pthread_cond_timedwait阻塞失败,或者说,没有到达预定的时间就已经超时返回。
代码示例如下:
#include
#include
#include time.h>
using namespace std;
class Ebupt
{
public:
Ebupt();
virtual ~Ebupt();
void dealMsg(long wait_ns);
private:
pthread_mutex_t mutex;
pthread_cond_t cond;
};
Ebupt::Ebupt()
{
pthread_mutex_init(&mutex, NULL);
pthread_cond_init(&cond, NULL);
}
Ebupt::~Ebupt()
{
pthread_mutex_destroy(&mutex);
pthread_cond_destroy(&cond);
}
void Ebupt::dealMsg(long wait_ns)
{
pthread_mutex_lock(&mutex);
struct timeval now;
gettimeofday(&now, NULL);
struct timespec abstime;
if (now.tv_usec*1000 + (wait_ns%1000000000) >= 1000000000)
{
abstime.tv_sec = now.tv_sec + wait_ns/1000000000 + 1;
abstime.tv_nsec = (now.tv_usec*1000 + wait_ns%1000000000)%1000000000;
}
else
{
abstime.tv_sec = now.tv_sec + wait_ns/1000000000;
abstime.tv_nsec = now.tv_usec*1000 + wait_ns%1000000000;
}
pthread_cond_timedwait(&cond, &mutex, &abstime);
pthread_mutex_unlock(&mutex);
}
int main()
{
Ebupt e;
struct timeval now;
while (true)
{
gettimeofday(&now, NULL);
cout<<"++"<<now.tv_sec<<":"<<now.tv_usec<200000000);
gettimeofday(&now, NULL);
cout<<"--"<<now.tv_sec<<":"<<now.tv_usec<0;
}
编译及输出如下:
[ismp@cn3 20171026]$ g++ -o main main.C -lpthread
[ismp@cn3 20171026]$ ./main
++1509023506:721641
--1509023506:721706
++1509023506:721710
--1509023506:721716
++1509023506:721718
--1509023506:721724
++1509023506:721726
--1509023506:721731
++1509023506:721733
--1509023506:721739
++1509023506:721741
--1509023506:721750
++1509023506:721753
--1509023506:721761
++1509023506:721763
--1509023506:721769
……
(CTRL+C)
理论上,我没有signal,那么应该阻塞200ms,再从阻塞中超时返回,但实际上,并没有阻塞,而是如同脱缰的野马,直接超时返回,由于dealMsg还是在一个while循环中,就如同死循环一般,CPU高当然很正常。
top看下嘞:
top - 21:15:52 up 419 days, 7:30, 2 users, load average: 9.57, 8.94, 8.32
Tasks: 241 total, 3 running, 238 sleeping, 0 stopped, 0 zombie
Cpu(s): 10.6%us, 63.1%sy, 0.0%ni, 24.6%id, 0.0%wa, 0.0%hi, 1.6%si, 0.0%st
Mem: 32879016k total, 32578784k used, 300232k free, 217448k buffers
Swap: 2097144k total, 749020k used, 1348124k free, 28921976k cached
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
22096 ismp 20 0 13904 1116 956 S 3.7 0.0 0:01.84 main
20409 ismp 20 0 109m 1956 1556 S 0.0 0.0 0:00.02 bash
就是个无阻塞死循环…
这个简单的示例还好,CPU飙到了4%不到,但是我那个进程直接飙到了70%多…
后来找了诸多问题,曾经想过,是不是gettimeofday使用的时钟和pthread_cond_timedwait实际使用的时钟不是同一个?
那我改改试试,如下:
#include
#include
#include
#include
using namespace std;
class Ebupt
{
public:
Ebupt();
virtual ~Ebupt();
void dealMsg(long wait_ns);
private:
pthread_mutex_t mutex;
pthread_cond_t cond;
};
Ebupt::Ebupt()
{
pthread_mutex_init(&mutex, NULL);
pthread_cond_init(&cond, NULL);
}
Ebupt::~Ebupt()
{
pthread_mutex_destroy(&mutex);
pthread_cond_destroy(&cond);
}
void Ebupt::dealMsg(long wait_ns)
{
pthread_mutex_lock(&mutex);
struct timespec now;
clock_gettime(CLOCK_REALTIME, &now);
struct timespec abstime;
if (now.tv_nsec + (wait_ns%1000000000) >= 1000000000)
{
abstime.tv_sec = now.tv_sec + wait_ns/1000000000 + 1;
abstime.tv_nsec = (now.tv_nsec + wait_ns%1000000000)%1000000000;
}
else
{
abstime.tv_sec = now.tv_sec + wait_ns/1000000000;
abstime.tv_nsec = now.tv_nsec + wait_ns%1000000000;
}
pthread_cond_timedwait(&cond, &mutex, &abstime);
pthread_mutex_unlock(&mutex);
}
int main()
{
Ebupt e;
struct timeval now;
while (true)
{
gettimeofday(&now, NULL);
cout<<"++"<":"<200000000);
gettimeofday(&now, NULL);
cout<<"--"<":"<return 0;
}
[ismp@cn3 20171026]$ g++ -o main main.C -lpthread -lrt
[ismp@cn3 20171026]$ ./main
++1509024234:822675
--1509024234:822733
++1509024234:822737
--1509024234:822748
++1509024234:822751
--1509024234:822761
……
(CTRL+C)
还是没有阻塞,看来并不是那个(gettimeofday和pthread_cond_timedwait使用的时钟不是同一个)原因。
如果我给条件变量加上属性试试?如下:
#include
#include
#include
#include
using namespace std;
class Ebupt
{
……
Ebupt::Ebupt()
{
pthread_mutex_init(&mutex, NULL);
pthread_condattr_t condattr;
pthread_condattr_init(&condattr);
pthread_condattr_setclock(&condattr, CLOCK_REALTIME);
pthread_cond_init(&cond, &condattr);
pthread_condattr_destroy(&condattr);
}
……(同上)
[ismp@cn3 20171026]$ g++ -o main main.C -lpthread -lrt
[ismp@cn3 20171026]$ ./main
++1509024510:358162
--1509024510:358221
++1509024510:358225
--1509024510:358236
++1509024510:358239
--1509024510:358249
……
(CTRL+C)
后来无意中发现,解决这个问题可以换个时钟,使用MONOTONIC这个时钟:
#include
#include
#include
#include
using namespace std;
class Ebupt
{
……
Ebupt::Ebupt()
{
pthread_mutex_init(&mutex, NULL);
pthread_condattr_t condattr;
pthread_condattr_init(&condattr);
pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
pthread_cond_init(&cond, &condattr);
pthread_condattr_destroy(&condattr);
}
……
void Ebupt::dealMsg(long wait_ns)
{
pthread_mutex_lock(&mutex);
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
struct timespec abstime;
if (now.tv_nsec + (wait_ns%1000000000) >= 1000000000)
{
abstime.tv_sec = now.tv_sec + wait_ns/1000000000 + 1;
abstime.tv_nsec = (now.tv_nsec + wait_ns%1000000000)%1000000000;
}
else
{
abstime.tv_sec = now.tv_sec + wait_ns/1000000000;
abstime.tv_nsec = now.tv_nsec + wait_ns%1000000000;
}
pthread_cond_timedwait(&cond, &mutex, &abstime);
pthread_mutex_unlock(&mutex);
}
……
[ismp@cn3 20171026]$ g++ -o main main.C -lpthread -lrt
[ismp@cn3 20171026]$ ./main
++1509024798:440277
--1509024798:640389
++1509024798:640400
--1509024798:840413
++1509024798:840424
--1509024799:40507
++1509024799:40517
--1509024799:240565
++1509024799:240581
--1509024799:440595
(CTRL+C)
也就是说,最后解决办法是:
给条件变量设置时钟,使用MONOTONIC,而不使用REALTIME。
MONOTONIC使用的是jiffies变量来计算时间,是一个单调递增的时间,代表boot当前机器的时间,在boot后jiffies初始化为0;
REALTIME使用的是xtime,而这个xtime是在boot后从主板上的硬件时钟(RTC)读取的,运行时刻也会受到特权用户(例如root)使用类似date的命令影响;例如你设定在1h后超时,但是如果在这个阻塞的时间窗口中,你使用date命令将系统时间(或者叫做wall time)调整到1h之后,那么阻塞的语句会立刻超时返回,一如我们的pthread_cond_timedwait。
其实到最后也没有找出到底是什么原因导致的pthread_cond_timedwait阻塞失败,只是偶然间得出的临时的解决办法,后续有时间再研究为何pthread_cond_timedwait阻塞失败吧…
后记:
发现现网的进程的CPU占比都有点不太正常:
top - 21:45:30 up 419 days, 8:00, 1 user, load average: 8.85, 8.37, 8.38
Tasks: 238 total, 4 running, 234 sleeping, 0 stopped, 0 zombie
Cpu(s): 10.4%us, 65.4%sy, 0.0%ni, 23.5%id, 0.0%wa, 0.0%hi, 0.7%si, 0.0%st
Mem: 32879016k total, 32650184k used, 228832k free, 218716k buffers
Swap: 2097144k total, 749020k used, 1348124k free, 28992212k cached
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
12303 sdc 20 0 3517m 864m 7176 S 344.1 2.7 5125261h java
9 root 20 0 0 0 0 S 50.5 0.0 144940:15 ksoftirqd/1
13 root 20 0 0 0 0 R 48.9 0.0 157182:21 ksoftirqd/2
4 root 20 0 0 0 0 R 46.9 0.0 153791:36 ksoftirqd/0
33 root 20 0 0 0 0 S 46.5 0.0 148379:24 ksoftirqd/7
21 root 20 0 0 0 0 R 44.2 0.0 156277:16 ksoftirqd/4
29 root 20 0 0 0 0 S 43.2 0.0 154775:19 ksoftirqd/6
17 root 20 0 0 0 0 S 30.9 0.0 174973:53 ksoftirqd/3
25 root 20 0 0 0 0 S 10.0 0.0 156328:27 ksoftirqd/5
27888 www 20 0 177m 121m 1900 S 1.3 0.4 1167:11 nginx
41 root 20 0 0 0 0 S 0.3 0.0 17:36.77 events/6
21937 sdc 20 0 134m 7564 1136 S 0.3 0.0 57:06.09 redis-server
24218 ismp 20 0 15164 1344 944 R 0.3 0.0 0:00.01 top
27890 www 20 0 180m 124m 1900 S 0.3 0.4 1163:55 nginx
27891 www 20 0 170m 114m 1912 S 0.3 0.4 1069:01 nginx
1 root 20 0 19348 852 544 S 0.0 0.0 0:01.41 init
2 root 20 0 0 0 0 S 0.0 0.0 0:00.00 kthreadd
3 root RT 0 0 0 0 S 0.0 0.0 1:02.86 migration/0
5 root RT 0 0 0 0 S 0.0 0.0 0:00.00 migration/0
尤其是java后台进程和ksoftirqd。
我猜测java是不是也是底层使用了条件变量结果没有阻塞住?
后来的后来…重启了一下现网的机器,各个进程占用的CPU就降下来了,然后也不会再出现上面阻塞失败的问题了……
如果有小伙伴曾经有见过这个问题,欢迎指教哈,嘿嘿~