epoll_create 创建的 文件描述符和其他文件描述符一样,是被fork出的子进程继承的,那也就是子进程可以使用这个epoll fd添加感兴趣的io(epoll_ctl),然后是可以影响到父进程的epoll_wait。比如,子进程中注册了一个io写事件后,因为某种原因挂起来了,导致父进程的epoll_wait频繁返回,CPU占用率飙升。看下下面的演示代码:
void DoWrite(int epollfd) {
5 int fd = socket(AF_LOCAL, SOCK_STREAM, 0);
6 if (fd < 0) {
7 perror("socket ");
8 exit(-1);
9 }
10 struct sockaddr_un peer_addr;
11 memset(&peer_addr, 0, sizeof(peer_addr));
12 peer_addr.sun_family = AF_LOCAL;
13 const char *ipc_path = "/home/longzhiri/my_code/nettest/localnettest.ipc";
14 strncpy(peer_addr.sun_path, ipc_path, sizeof(peer_addr.sun_path)-1);
15 if (connect(fd, (struct sockaddr *)&peer_addr, SUN_LEN(&peer_addr))<0){
16 perror("connect ");
17 exit(-1);
18 }
19 struct epoll_event ev, events[10];
20 ev.events = EPOLLOUT;
21 ev.data.fd = fd;
22 if (epoll_ctl(epollfd, EPOLL_CTL_ADD, fd, &ev) < 0) {
23 perror("epoll_ctl");
24 exit(-1);
25 }
26
27 sleep(10);
28 epoll_ctl(epollfd, EPOLL_CTL_DEL, fd, &ev);
29 printf("child process exit ");
30 }
31
32 int main(int argc, char *argv[]) {
33 int epollfd = epoll_create(10);
34 if (epollfd < 0) {
35 perror("epoll_create");
36 return -1;
37 }
38
39 struct epoll_event events[10];
40 int pid = fork();
41 if (pid < 0) {
42 perror("fork");
43 return -1;
44 } else if (pid > 0) {
45 for (;;) {
46 int nfds = epoll_wait(epollfd, events, 10, -1);
47 if (nfds < 0) {
48 perror("epoll_wait 1");
49 return -1;
50 }
51 printf("wake up +++ ");
52 }
53 } else {
54 DoWrite(epollfd);
55 }
56 return 0;
57 }
5 int fd = socket(AF_LOCAL, SOCK_STREAM, 0);
6 if (fd < 0) {
7 perror("socket ");
8 exit(-1);
9 }
10 struct sockaddr_un peer_addr;
11 memset(&peer_addr, 0, sizeof(peer_addr));
12 peer_addr.sun_family = AF_LOCAL;
13 const char *ipc_path = "/home/longzhiri/my_code/nettest/localnettest.ipc";
14 strncpy(peer_addr.sun_path, ipc_path, sizeof(peer_addr.sun_path)-1);
15 if (connect(fd, (struct sockaddr *)&peer_addr, SUN_LEN(&peer_addr))<0){
16 perror("connect ");
17 exit(-1);
18 }
19 struct epoll_event ev, events[10];
20 ev.events = EPOLLOUT;
21 ev.data.fd = fd;
22 if (epoll_ctl(epollfd, EPOLL_CTL_ADD, fd, &ev) < 0) {
23 perror("epoll_ctl");
24 exit(-1);
25 }
26
27 sleep(10);
28 epoll_ctl(epollfd, EPOLL_CTL_DEL, fd, &ev);
29 printf("child process exit ");
30 }
31
32 int main(int argc, char *argv[]) {
33 int epollfd = epoll_create(10);
34 if (epollfd < 0) {
35 perror("epoll_create");
36 return -1;
37 }
38
39 struct epoll_event events[10];
40 int pid = fork();
41 if (pid < 0) {
42 perror("fork");
43 return -1;
44 } else if (pid > 0) {
45 for (;;) {
46 int nfds = epoll_wait(epollfd, events, 10, -1);
47 if (nfds < 0) {
48 perror("epoll_wait 1");
49 return -1;
50 }
51 printf("wake up +++ ");
52 }
53 } else {
54 DoWrite(epollfd);
55 }
56 return 0;
57 }
真实环境看起来很难犯这种错误,但其实在使用第三方库的时候,因为隐藏不少逻辑,犯错的概率就高了。比如在使用libevent的时候,fork出的子进程想要复用父进程的event_base是必需调用event_reinit的,但是不了解这个问题的人就会直接拿父进程的event_base来使用。看下event_reinit的代码,底层使用了epoll的重新epoll_create出一个epollfd的,而不是重用父进程的,也就是没有调用event_reinit,注册的读写事件,其实都是在父进程中那个epoll_wait那里返回。