1. libev简介
libev是个高性能跨平台的事件驱动框架,支持io事件,超时事件,子进程状态改变通知,信号通知,文件状态改变通知,还能用来实现wait/notify机制。libev对每种监听事件都用一个ev_type类型的数据结构表示,如ev_io, ev_timer, ev_child, ev_async分别用来表示文件监听器, timeout监听器, 子进程状态监听器, 同步事件监听器.
libev支持优先级, libev一次loop收集的事件按优先级先排序, 优先级高的事件回调先执行, 优先级低的后执行, 相同优先级则按事件到达顺序执行. libev优先级从[-2, 2], 默认优先级为0,libev注册watcher的流程如下:
static void type_cb(EV_P_ ev_type *watcher, int revents)
{
// callback
}
static void ev_test()
{
#ifdef EV_MULTIPLICITY
struct ev_loop *loop;
#else
int loop;
#endif
ev_type *watcher;
loop = ev_default_loop(0);
watcher = (ev_type *)calloc(1, sizeof(*watcher));
assert(loop && watcher);
ev_type_init(watcher, type_cb, ...);
ev_start(EV_A_ watcher);
ev_run(EV_A_ 0);
/* 资源回收 */
ev_loop_destroy(EV_A);
free(watcher);
}
libev注册watcher可以分为四个步骤:
- 创建一个loop和watcher
- 初始化watcher,主要设置callback函数和定义watcher的参数
- 激活watcher
- 启动libev,开始loop收集事件
2. ev_io
libev内部使用后端select, poll, epoll(linux专有), kqueue(drawin), port(solaris10)实现io事件监听, 用户可以指定操作系统支持的后端或者由libev自动选择使用哪个后端,如linux平台上用户可以强制指定libev使用select作为后端。libev支持单例模式和多例模式, 假设我们链接的是多例模式的libev库, 且watcher使用默认优先级0.
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <unistd.h>
#include <ev.h>
static void io_cb(struct ev_loop *loop, ev_io *watcher, int revents)
{
char buf[1024] = {0};
/* 参数watcher即注册时的watcher */
read(watcher->fd, buf, sizeof(buf) - 1);
fprintf(stdout, "%s\n", buf);
ev_break(loop, EVBREAK_ALL);
}
static void io_test()
{
struct ev_loop *loop;
ev_io *io_watcher;
/* 指定libev使用epoll机制,关闭环境变量对libev影响 */
loop = ev_default_loop(EVFLAG_NOENV | EVBACKEND_EPOLL);
io_watcher = (ev_io *)calloc(1, sizeof(*io_watcher));
assert(("can not alloc memory", loop && io_watcher));
/* 设置监听标准输入的可读事件和回调函数 */
ev_io_init(io_watcher, io_cb, STDIN_FILENO, EV_READ);
ev_io_start(loop, io_watcher);
/* libev开启loop */
ev_run(loop, 0);
/* 资源回收 */
ev_loop_destroy(loop);
free(io_watcher);
}
int main(void)
{
io_test();
return 0;
}
ev_io_init(ev, cb, fd, event)
- ev:ev_io
- cb: 回调函数
- fd:socket,pipe等句柄
- event:事件类型(EV_READ/EV_WRITE)
Makefile
target := main
CC := clang
CFLAGS += -g -Wall -fPIC
LDFLAGS += -lev
src += \
main.c
obj := $(patsubst %.c, %.o, $(src))
$(target):$(obj)
$(CC) -o $@ $^ $(LDFLAGS)
%.o:%.c
$(CC) -o $@ -c $< $(CFLAGS)
.PHONY:clean
clean:
@rm -rf $(target) *.o
libev内部使用一个大的循环来收集各种watcher注册的事件,如果没有注册ev_timer和ev_periodic,则libev内部使用的后端采用59.743s作为超时事件,如果select作为后端,则select的超时设置为59.743s,这样可以降低cpu占用率,对一个fd可以注册的watcher数量不受限(或者说只受内存限制),比如可以对标准输入的可读事件注册100个watcher,当有用户输入时所有100个watcher的回调都能执行(当然回调中还是只有一个read操作成功)。
3. ev_timer
ev_timer可以用来实现定时器, ev_timer不受墙上时间影响,如设置一个1小时定时器,把当前系统时间调快1小时不能让定时器立刻超时,超时依旧发生在1小时后,如下是一个简单的例子:
static void timer_cb(struct ev_loop *loop, ev_timer *w, int revents)
{
fprintf(stdout, "%fs timeout\n", w->repeat);
ev_break(loop, EVBREAK_ALL);
}
static void timer_test()
{
struct ev_loop *loop;
ev_timer *timer_watcher;
loop = ev_default_loop(EVFLAG_NOENV);
timer_watcher = calloc(1, sizeof(*timer_watcher));
assert(("can not alloc memory", loop && timer_watcher));
ev_timer_init(timer_watcher, timer_cb, 0., 3600.);
ev_timer_start(loop, timer_watcher);
ev_run(loop, 0);
ev_loop_destroy(loop);
free(timer_watcher);
}
ev_timer_init(ev, cb, ofs, iva)
- ev: ev_timer
- cb: 回调函数
- ofs,iva: 超时事件为(now + ofs + ival * N), now为当前时间,N为正整数
如果ival参数为0,则timer是一次性的定时器,超时后libev自动stop timer
可以在回调中重新设置timer超时,并重新启动timer。
static void timer_cb(struct ev_loop *loop, ev_timer *watcher, int revents)
{
fprintf(stdout, "%fs timeout\n", watcher->repeat);
watcher->repeat += 5.;
ev_timer_again(loop, watcher);
}
上面介绍了libev是在一个大的循环中监听所有watcher的事件,只有ev_io类型的watcher时,libev后端以59.743s作为超时(如select超时),这时用户注册一个3s的timer,那么libev会不会因为后端超时太长导致定时器检测非常不准呢?答案是不会,libev保证后端超时时间不大于定时器超时时间,注册一个3s timer,则libev自动调整到3s以内loop一次,这样保证timer超时能及时被检测到,同时也带来更高的cpu占用率。
4. ev_periodic
ev_timer做为定时器很方便,但是对应指定到某时刻发生超时就比较困难,比如每天00:00:00触发超时,每天08:00开灯,18:00关灯等等,ev_periodic可以很好的应付这种场景,ev_periodic基于墙上时间,所以受墙上时间影响,如注册1小时后超时的ev_peroidic,同时系统时间调快1小时,ev_periodic立马能超时。如下例子指定每天凌晨发生超时:
static void periodic_cb(struct ev_loop *loop, ev_periodic *watcher, int revents)
{
fprintf(stdout, "00:00:00 now, time to sleep");
}
ev_tstamp my_schedule(ev_periodic *watcher, ev_tstamp now)
{
time_t cur;
struct tm tm;
time(&cur);
localtime_r(&cur, &tm);
tm.tm_wday += 1;
tm.tm_hour = 0;
tm.tm_sec = 0;
tm.tm_min = 0;
tm.tm_mon -= 1;
tm.tm_year -= 1900;
return mktime(&tm);
}
static void periodic_test()
{
struct ev_loop *loop;
ev_periodic *periodic_watcher;
loop = ev_default_loop(0);
periodic_watcher = (ev_periodic *)calloc(1, sizeof(*periodic_watcher));
assert(("can not alloc memory", loop && periodic_watcher));
ev_periodic_init(periodic_watcher, periodic_cb, 0, 0, my_schedule);
ev_periodic_start(loop, periodic_watcher);
ev_run(loop, 0);
ev_loop_destroy(loop);
free(periodic_watcher);
}
ev_periodic_init(ev, cb, ofs, ival,schedule)
- ev:ev_periodic
- cb:回调函数
- ofs,ival:ofs + ceil((now - ofs) / ival) * ival // now表示当前时间戳
- schedule:用户自定义函数,该函数返回下一次ev_periodic超时时间
- 如果schedule不为空,则ev_periodic超时时间为:ofs + ceil((now - ofs) / ival) * ival,表示从当前时间开始,经过ofs时间后所有能被ival整数的点,注册ev_periodic,ifs = 1, ival = 10, 当前时间为1604649458(2020-10-615:57:38),则ev_periodic第一次经过2s就发生超时了。
- 如果schedule存在,则ofs,ival参数被忽略,ev_periodic超时时间由schedule()返回值指定
5. ev_child
libev支持监听子进程状态变化, 如子进程退出, 内部用waitpid去实现, libev限制只能用default loop去监听子进程状态变化, 如果以ev_loop_new()创建的loop则不行, 通过ev_default_loop()创建default loop时libev内部自动注册了SIGCHILD信号处理函数, 需要在自己代码处理SIGCHILD的话, 可以在ev_default_loop()之后注册SIGCHIL处理以覆盖libev中的默认处理, 如下是一个简单的例子:
static void child_cb(struct ev_loop *loop, ev_child *watcher, int revents)
{
fprintf(stdout, "pid:%d exit, status:%d\n", watcher->rpid, watcher->rstatus);
}
static void child_test()
{
pid_t pid;
struct ev_loop *loop;
ev_child *child_watcher;
switch (pid = fork()) {
case 0:
sleep(5);
fprintf(stdout, "child_pid:%d\n", getpid());
exit(EXIT_SUCCESS);
default:
{
loop = ev_default_loop(0);
child_watcher = (ev_child*)calloc(1, sizeof(*child_watcher));
assert(("can not alloc memory", loop && child_watcher));
ev_child_init(child_watcher, child_cb, 0, 1);
ev_child_start(loop, child_watcher);
ev_run(loop, 0);
/* 资源回收 */
ev_loop_destroy(loop);
free(child_watcher);
}
}
}
ev_child_init(ev, cb, pid, trace)
- ev:ev_child
- cb:回调函数
- pid:子进程pid
- trace:设置为1
个人觉得libev的default loop默认注册SIGCHILD处理并不好,最好还是在自己代码中做处理。
6. ev_async
可以通过libev的async来实现wait/notify机制, 用户注册多个ev_async监听器, 在其他地方调用ev_async_send()即可触发ev_async注册的回调, libev内部用eventfd(linux平台)和pipe(win32)实现, 个人觉得linux平台上直接用eventfd更完美, 如下是简单例子.
static void *routine(void *args)
{
static size_t count = 0;
ev_async *watcher = (ev_async *)args;
struct ev_loop *loop = (struct ev_loop *)watcher->data;
while (count++ < 10) {
ev_async_send(loop, watcher);
sleep(1);
}
return NULL;
}
static void async_cb(struct ev_loop *loop, ev_async *watcher, int revents)
{
fprintf(stdout, "get the order, start move...\n");
}
static void async_test()
{
pthread_t pid;
struct ev_loop *loop;
ev_async *async_watcher;
loop = ev_default_loop(0);
async_watcher = (ev_async *)calloc(1, sizeof(*async_watcher));
assert(("can not alloc memory", loop && async_watcher));
ev_async_init(async_watcher, async_cb);
ev_async_start(loop, async_watcher);
async_watcher->data = loop;
pthread_create(&pid, NULL, routine, async_watcher);
ev_run(loop, 0);
/* 资源回收 */
ev_loop_destroy(loop);
free(async_watcher);
}
ev_async_init(ev,cb)
- ev:ev_async
- cb:回调函数
7. ev_prepare/ev_idle
libev每次loop收集各种事件之前都会先调用ev_prepare的回调函数(如果有的话), 如果存在比ev_idle优先级更高的监听有事件待处理, 则ev_idle的事件不会处理, 如存在优先级1,2的事件待处理, 则优先级为1的ev_idle的事件不会被处理, 只有在优先级1,2的所有事件都处理完后才会把ev_idle的事件添加到带处理的事件队列中去.
static void idle_cb(struct ev_loop *loop, ev_idle *watcher, int revents)
{
fprintf(stdout, "no one has higher priority than me now\n");
ev_idle_stop(loop, watcher);
}
static void prepare_cb(struct ev_loop *loop, ev_prepare *watcher, int revents)
{
fprintf(stdout, "prepare_cb\n");
}
static void ev_test()
{
struct ev_loop *loop;
ev_io *io_watcher;
ev_idle *idle_watcher;
ev_prepare *prepare_watcher;
loop = ev_default_loop(0);
prepare_watcher = (ev_prepare *)calloc(1, sizeof(*prepare_watcher));
idle_watcher = (ev_idle *)calloc(1, sizeof(*idle_watcher));
io_watcher = (ev_io *)calloc(1, sizeof(*io_watcher));
assert(("can not alloc memory", loop && prepare_watcher && io_watcher && idle_watcher));
ev_io_init(io_watcher, io_cb, STDIN_FILENO, EV_READ);
ev_prepare_init(prepare_watcher, prepare_cb);
ev_idle_init(idle_watcher, idle_cb);
ev_prepare_start(loop, prepare_watcher);
ev_io_start(loop, io_watcher);
ev_idle_start(loop, idle_watcher);
ev_run(loop, 0);
ev_loop_destroy(loop);
free(io_watcher);
free(idle_watcher);
free(prepare_watcher);
}
ev_prepare(ev, cb)/ev_idle(ev, cb)
- ev:ev_prepare/ev_idle
- cb:回调函数
可以看到每次输入前都先有ev_prepare的回调,只有不存在优先级别idle高的时间待处理时才会处理idle的回调。
8. ev_stat
不建议使用,还不如用个定时器自己去检测文件是否改动
9. ev_fork
注册ev_fork,在libev自动检测到fork调用(开启了EVFLAG_FORKCHECK),或者用户调用ev_loop_fork()通知libev有fork调用时ev_fork回调被触发
10. ev_cleanup
注册ev_cleanup的watcher,在libev销毁时调用ev_cleanup的回调,用来做一些清理工作
11. 简单回显服务器
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <strings.h>
#include <assert.h>
#include <fcntl.h>
#include <unistd.h>
#include <errno.h>
#include <sys/socket.h>
#include <arpa/inet.h>
#include <ev.h>
/* client number limitation */
#define MAX_CLIENTS 1000
/* message length limitation */
#define MAX_MESSAGE_LEN (256)
#define err_message(msg) \
do {perror(msg); exit(EXIT_FAILURE);} while(0)
/* record the number of clients */
static int client_number;
static int create_serverfd(char const *addr, uint16_t u16port)
{
int fd;
struct sockaddr_in server;
fd = socket(AF_INET, SOCK_STREAM, 0);
if (fd < 0) err_message("socket err\n");
server.sin_family = AF_INET;
server.sin_port = htons(u16port);
inet_pton(AF_INET, addr, &server.sin_addr);
if (bind(fd, (struct sockaddr *)&server, sizeof(server)) < 0) err_message("bind err\n");
if (listen(fd, 10) < 0) err_message("listen err\n");
return fd;
}
static void read_cb(EV_P_ ev_io *watcher, int revents)
{
ssize_t ret;
char buf[MAX_MESSAGE_LEN] = {0};
ret = recv(watcher->fd, buf, sizeof(buf) - 1, MSG_DONTWAIT);
if (ret > 0) {
write(watcher->fd, buf, ret);
} else if ((ret < 0) && (errno == EAGAIN || errno == EWOULDBLOCK)) {
return;
} else {
fprintf(stdout, "client closed (fd=%d)\n", watcher->fd);
--client_number;
ev_io_stop(EV_A_ watcher);
close(watcher->fd);
free(watcher);
}
}
static void accept_cb(EV_P_ ev_io *watcher, int revents)
{
int connfd;
ev_io *client;
connfd = accept(watcher->fd, NULL, NULL);
if (connfd > 0) {
if (++client_number > MAX_CLIENTS) {
close(watcher->fd);
} else {
client = calloc(1, sizeof(*client));
ev_io_init(client, read_cb, connfd, EV_READ);
ev_io_start(EV_A_ client);
}
} else if ((connfd < 0) && (errno == EAGAIN || errno == EWOULDBLOCK)) {
return;
} else {
close(watcher->fd);
ev_break(EV_A_ EVBREAK_ALL);
/* this will lead main to exit, no need to free watchers of clients */
}
}
static void start_server(char const *addr, uint16_t u16port)
{
int fd;
#ifdef EV_MULTIPLICITY
struct ev_loop *loop;
#else
int loop;
#endif
ev_io *watcher;
fd = create_serverfd(addr, u16port);
loop = ev_default_loop(EVFLAG_NOENV);
watcher = calloc(1, sizeof(*watcher));
assert(("can not alloc memory\n", loop && watcher));
/* set nonblock flag */
fcntl(fd, F_SETFL, fcntl(fd, F_GETFL, 0) | O_NONBLOCK);
ev_io_init(watcher, accept_cb, fd, EV_READ);
ev_io_start(EV_A_ watcher);
ev_run(EV_A_ 0);
ev_loop_destroy(EV_A);
free(watcher);
}
static void signal_handler(int signo)
{
switch (signo) {
case SIGPIPE:
break;
default:
// unreachable
break;
}
}
int main(void)
{
signal(SIGPIPE, signal_handler);
start_server("127.0.0.1", 10009);
return 0;
}
客户端:
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <strings.h>
#include <assert.h>
#include <fcntl.h>
#include <unistd.h>
#include <errno.h>
#include <sys/socket.h>
#include <arpa/inet.h>
#include <pthread.h>
#define err_message(msg) \
do {perror(msg); exit(EXIT_FAILURE);} while(0)
static int create_clientfd(char const *addr, uint16_t u16port)
{
int fd;
struct sockaddr_in server;
fd = socket(AF_INET, SOCK_STREAM, 0);
if (fd < 0) err_message("socket err\n");
server.sin_family = AF_INET;
server.sin_port = htons(u16port);
inet_pton(AF_INET, addr, &server.sin_addr);
if (connect(fd, (struct sockaddr *)&server, sizeof(server)) < 0) perror("connect err\n");
return fd;
}
static void *routine(void *args)
{
int fd;
char buf[128];
fd = create_clientfd("127.0.0.1", 10009);
for (; ;) {
write(fd, "Hello", strlen("hello"));
memset(buf, '\0', sizeof(buf));
read(fd, buf, sizeof(buf) - 1);
fprintf(stdout, "pthreadid:%ld %s\n", pthread_self(), buf);
usleep(100 * 1000);
}
}
int main(void)
{
pthread_t pids[4];
for (int i = 0; i < sizeof(pids)/sizeof(pthread_t); ++i) {
pthread_create(pids + i, NULL, routine, 0);
}
for (int i = 0; i < sizeof(pids)/sizeof(pthread_t); ++i) {
pthread_join(pids[i], 0);
}
return 0;
}
Makefile
all:server client
server_src += \
server.c
server_obj := $(patsubst %.c, %.o, $(server_src))
client_src += \
client.c
client_obj:= $(patsubst %.c, %.o, $(client_src))
CC := clang
CFLAGS += -Wall -fPIC
server:$(server_obj)
$(CC) -o $@ $^ -lev
%.o:%.c
$(CC) -o $@ -c $< $(CFLAGS)
client:$(client_obj)
$(CC) -o $@ $^ -lpthread
%.o:%.c
$(CC) -o $@ -c $< $(CFLAGS)
.PHONY:clean all
clean:
@rm -rf server client *.o