nginx epoll详解

nginx epoll 事件模型

nginx做为一个异步高效的事件驱动型web服务器，在linux平台中当系统支持epoll时nginx默认采用epoll来高效的处理事件。nginx中使用ngx_event_t结构来表示一个事件，先介绍下ngx_event_t结构体中成员的含义：

struct ngx_event_s {
void *data; //与事件关联的对象，常指向事件所在的ngx_connection_t连接对象
unsigned write:1; //可写标识位，1表示对应的tcp连接是可写的
unsigned accept:1;// 1表示对应的连接是处于监听状态的连接，即可接收新的连接
/* used to detect the stale events in kqueue, rtsig, and epoll */
unsigned instance:1; //可来区分事件是否已过期
/*
* the event was passed or would be passed to a kernel;
* in aio mode - operation was posted.
*/
unsigned active:1;// 1表示事件活跃，即事件已添加到epoll中
unsigned disabled:1;//epoll中不使用该标识位
/* the ready event; in aio mode 0 means that no operation can be posted */
unsigned ready:1; //事件已就绪(即可读或可写)
unsigned oneshot:1;//epoll不使用该标识位
/* aio operation is complete */
unsigned complete:1;//aio中使用，表示事件对应的aio异步操作已完成(io_getevents函数已成功返回)
unsigned eof:1;// 1表示当前处理的字符流已完成，如调用recv读取连接数据时返回0，此时置该标识位为1
unsigned error:1;// 1表示事件处理过程中发生错误
unsigned timedout:1; //事件是否超时，1:表示超时。超时后事件对应的请求不需再被处理(对于http模块来说事件超时后直接关闭请求)
unsigned timer_set:1; //为1时表示这个事件在定时器中
unsigned delayed:1;// 1表示需延迟处理该事件，常用于限速功能中
unsigned deferred_accept:1;//延迟接收接连，即当连接中收到对象发送的数据后才真正建立连接
/* the pending eof reported by kqueue, epoll or in aio chain operation */
unsigned pending_eof:1;// 1表示TCP连接对向关闭读端，即epoll返回EPOLLRDHUP
#if !(NGX_THREADS)
unsigned posted_ready:1;//该标识位在1.5.5版本源码中只在ngx_epoll_process_events函数中有置位，其它地方并没有用到
#endif
#if (NGX_WIN32)
/* setsockopt(SO_UPDATE_ACCEPT_CONTEXT) was successful */
unsigned accept_context_updated:1;
#endif
#if (NGX_HAVE_KQUEUE)
unsigned kq_vnode:1;
/* the pending errno reported by kqueue */
int kq_errno;
#endif
/*
* kqueue only:
* accept: number of sockets that wait to be accepted
* read: bytes to read when event is ready
* or lowat when event is set with NGX_LOWAT_EVENT flag
* write: available space in buffer when event is ready
* or lowat when event is set with NGX_LOWAT_EVENT flag
*
* iocp: TODO
*
* otherwise:
* accept: 1 if accept many, 0 otherwise
*/
#if (NGX_HAVE_KQUEUE) || (NGX_HAVE_IOCP)
int available;
#else
unsigned available:1;// 1表示每次调用accept时尽可能多的接收TCP连接，与multi_accept配置项对应
#endif
ngx_event_handler_pt handler; // 事件产生后的回调函数句柄
#if (NGX_HAVE_AIO)
#if (NGX_HAVE_IOCP)
ngx_event_ovlp_t ovlp;
#else
struct aiocb aiocb;
#endif
#endif
ngx_uint_t index; //epoll中不使用
ngx_log_t *log; //ngx_log_t对象
ngx_rbtree_node_t timer;
unsigned closed:1; // 1表示事件已关闭
/* to test on worker exit */
unsigned channel:1;// 只在ngx_add_channel_event函数中有置位，其它地方没用到
unsigned resolver:1; // resolver功能中使用?
#if (NGX_THREADS)
unsigned locked:1;
unsigned posted_ready:1;
unsigned posted_timedout:1;
unsigned posted_eof:1;
#if (NGX_HAVE_KQUEUE)
/* the pending errno reported by kqueue */
int posted_errno;
#endif
#if (NGX_HAVE_KQUEUE) || (NGX_HAVE_IOCP)
int posted_available;
#else
unsigned posted_available:1;
#endif
ngx_atomic_t *lock;
ngx_atomic_t *own_lock;
#endif
/* the links of the posted queue */
ngx_event_t *next;
ngx_event_t **prev;
#if 0
/* the threads support */
/*
* the event thread context, we store it here
* if $(CC) does not understand __thread declaration
* and pthread_getspecific() is too costly
*/
void *thr_ctx;
#if (NGX_EVENT_T_PADDING)
/* event should not cross cache line in SMP */
uint32_t padding[NGX_EVENT_T_PADDING];
#endif
#endif
};
#if (NGX_HAVE_FILE_AIO)
struct ngx_event_aio_s {
void *data;
ngx_event_handler_pt handler;
ngx_file_t *file;
ngx_fd_t fd;
#if (NGX_HAVE_EVENTFD)
int64_t res;
#if (NGX_TEST_BUILD_EPOLL)
ngx_err_t err;
size_t nbytes;
#endif
#else
ngx_err_t err;
size_t nbytes;
#endif
#if (NGX_HAVE_AIO_SENDFILE)
off_t last_offset;
#endif
ngx_aiocb_t aiocb;
ngx_event_t event;
};
#endif

nginx中使用ngx_epoll_module模块来封装epoll机制处理事件，ngx_epoll_module模块只对两个配置项感兴趣，其ngx_command_t结构如下：

static ngx_command_t ngx_epoll_commands[] = {
{
/***epoll_events配置项表示epoll_wait函数每次最多返回多少个事件，在ngx_epoll_init函数中
会预先分配epoll_events配置项指定的epoll_event结构个数**/
ngx_string("epoll_events"),
NGX_EVENT_CONF|NGX_CONF_TAKE1,
ngx_conf_set_num_slot,
0,
offsetof(ngx_epoll_conf_t, events),
NULL },
{
/***worker_aio_requests配置项表示创建的aio context能并发处理异步事件的个数，即io_setup函数的第一个参数***/
ngx_string("worker_aio_requests"),
NGX_EVENT_CONF|NGX_CONF_TAKE1,
ngx_conf_set_num_slot,
0,
offsetof(ngx_epoll_conf_t, aio_requests),
NULL },
ngx_null_command
};

ngx_epoll_module的ngx_event_module_t结构如下：

ngx_event_module_t ngx_epoll_module_ctx = {
&epoll_name,
ngx_epoll_create_conf, /* create configuration */
ngx_epoll_init_conf, /* init configuration */
{
//向epoll中添加事件时调用
ngx_epoll_add_event, /* add an event */
//从epoll中删除事件时调用
ngx_epoll_del_event, /* delete an event */
/***epoll中不存在enable/disable事件的情况，这里默认设置成添加/删除事件的函数***/
ngx_epoll_add_event, /* enable an event */
ngx_epoll_del_event, /* disable an event */
//向epoll中添加tcp连接时调用，每个tcp连接对象一个读事件和一个写事件
ngx_epoll_add_connection, /* add an connection */
//从epoll中删除事件时调用
ngx_epoll_del_connection, /* delete an connection */
NULL, /* process the changes */
// epoll 事件处理函数
ngx_epoll_process_events, /* process the events */
//epoll模块初始化函数
ngx_epoll_init, /* init the events */
//epoll模块清理函数只在多线程模型中被调用
ngx_epoll_done, /* done the events */
}
};

ngx_epoll_create_conf在配置项解析前调用用来初始化配置结构，ngx_epoll_init_conf函数在配置项解析完后调用，如果配置文件是不存在epoll_events或worker_aio_requests配置项，默认将epoll_events设置为512，worker_aio_requests设置为32。ngx_epoll_module_ctx结构体中后十个函数对应于ngx_event_actions_t结构，它是事件模块独有的结构。ngx_epoll_init函数在什么时候被调用呢，它在nginx启动过程中每个worker进程启动后被调用(由ngx_event_core_module的ngx_event_process_init函数调用)。

ngx_epoll_module源码分析

ngx_epoll_init函数：

static ngx_int_t
ngx_epoll_init(ngx_cycle_t *cycle, ngx_msec_t timer)
{
ngx_epoll_conf_t *epcf;
// 获取ngx_epoll_module模块存放配置项的结构
epcf = ngx_event_get_conf(cycle->conf_ctx, ngx_epoll_module);
if (ep == -1) {
// 创建epoll，成功返回描述符，失败返回-1
ep = epoll_create(cycle->connection_n / 2);
if (ep == -1) {
ngx_log_error(NGX_LOG_EMERG, cycle->log, ngx_errno,
"epoll_create() failed");
return NGX_ERROR;
}
/***如果系统支持aio , 这里初始化aio***/
#if (NGX_HAVE_FILE_AIO)
ngx_epoll_aio_init(cycle, epcf);
#endif
}
/***预分配events个epoll_event结构, epcf->events由epoll_events配置项指定，默认为512***/
if (nevents < epcf->events) {
if (event_list) {
ngx_free(event_list);
}
event_list = ngx_alloc(sizeof(struct epoll_event) * epcf->events,
cycle->log);
if (event_list == NULL) {
return NGX_ERROR;
}
}
nevents = epcf->events;
//指定I/O读写的方法
ngx_io = ngx_os_io;
// 设置ngx_event_actions接口，后续通过ngx_event_actions来调用epoll模块中的方法
ngx_event_actions = ngx_epoll_module_ctx.actions;
/***nginx使用epoll事件模型时NGX_HAVE_CLEAR_EVENT宏被定义, NGX_USE_CLEAR_EVENT宏表示使用epoll的ET模式***/
#if (NGX_HAVE_CLEAR_EVENT)
ngx_event_flags = NGX_USE_CLEAR_EVENT
#else
ngx_event_flags = NGX_USE_LEVEL_EVENT
#endif
|NGX_USE_GREEDY_EVENT
|NGX_USE_EPOLL_EVENT;
return NGX_OK;
}

ngx_epoll_add_event函数：

static ngx_int_t
ngx_epoll_add_event(ngx_event_t *ev, ngx_int_t event, ngx_uint_t flags)
{
int op;
uint32_t events, prev;
ngx_event_t *e;
ngx_connection_t *c;
struct epoll_event ee;
//获取事件关联的连接
c = ev->data;
events = (uint32_t) event;
/***根据event参数判断当前是添加读事件还是写事件***/
if (event == NGX_READ_EVENT) {
e = c->write;
prev = EPOLLOUT;
#if (NGX_READ_EVENT != EPOLLIN|EPOLLRDHUP)
events = EPOLLIN|EPOLLRDHUP;
#endif
} else {
e = c->read;
prev = EPOLLIN|EPOLLRDHUP;
#if (NGX_WRITE_EVENT != EPOLLOUT)
events = EPOLLOUT;
#endif
}
/***如果当前需添加读事件，就通过active标识判断读事件所关联的连接对应的写事件是否活跃(
活跃表示事件已添加到epoll中)。***/
if (e->active) {
op = EPOLL_CTL_MOD;
events |= prev;
} else {
op = EPOLL_CTL_ADD;
}
//将flags参数加入到epoll标志中
ee.events = events | (uint32_t) flags;
/*** ptr存储事件关联的连接对象(ngx_connection_t*)及事件过期比特位，
linux平台中任何对象的地址最低位必定为零***/
ee.data.ptr = (void *) ((uintptr_t) c | ev->instance);
ngx_log_debug3(NGX_LOG_DEBUG_EVENT, ev->log, 0,
"epoll add event: fd:%d op:%d ev:%08XD",
c->fd, op, ee.events);
//向epoll中添加事件
if (epoll_ctl(ep, op, c->fd, &ee) == -1) {
ngx_log_error(NGX_LOG_ALERT, ev->log, ngx_errno,
"epoll_ctl(%d, %d) failed", op, c->fd);
return NGX_ERROR;
}
//标识事件活跃
ev->active = 1;
#if 0
ev->oneshot = (flags & NGX_ONESHOT_EVENT) ? 1 : 0;
#endif
return NGX_OK;
}

ngx_epoll_del_event函数：

static ngx_int_t
ngx_epoll_del_event(ngx_event_t *ev, ngx_int_t event, ngx_uint_t flags)
{
int op;
uint32_t prev;
ngx_event_t *e;
ngx_connection_t *c;
struct epoll_event ee;
/*
* when the file descriptor is closed, the epoll automatically deletes
* it from its queue, so we do not need to delete explicitly the event
* before the closing the file descriptor
*/
/***上面的注释说得很清楚了，当文件描述符被关闭后，epoll会自动将其删除。***/
if (flags & NGX_CLOSE_EVENT) {
ev->active = 0;
return NGX_OK;
}
//获取事件关联的连接
c = ev->data;
/***根据event参数判断当前是删除读事件还是写事件***/
if (event == NGX_READ_EVENT) {
e = c->write;
prev = EPOLLOUT;
} else {
e = c->read;
prev = EPOLLIN|EPOLLRDHUP;
}
/***参考ngx_epoll_add_event函数***/
if (e->active) {
op = EPOLL_CTL_MOD;
ee.events = prev | (uint32_t) flags;
ee.data.ptr = (void *) ((uintptr_t) c | ev->instance);
} else {
op = EPOLL_CTL_DEL;
ee.events = 0;
ee.data.ptr = NULL;
}
ngx_log_debug3(NGX_LOG_DEBUG_EVENT, ev->log, 0,
"epoll del event: fd:%d op:%d ev:%08XD",
c->fd, op, ee.events);
//从epoll中删除事件
if (epoll_ctl(ep, op, c->fd, &ee) == -1) {
ngx_log_error(NGX_LOG_ALERT, ev->log, ngx_errno,
"epoll_ctl(%d, %d) failed", op, c->fd);
return NGX_ERROR;
}
//清除事件活跃标识
ev->active = 0;
return NGX_OK;
}

ngx_epoll_add_connection及ngx_epoll_del_connection函数

这两个函数的实现很简单，也是通过调用epoll_ctl添加事件，只是会同时将读/写事件一起添加进epoll，这里不再列出源码。

ngx_epoll_process_events函数：

static ngx_int_t
ngx_epoll_process_events(ngx_cycle_t *cycle, ngx_msec_t timer, ngx_uint_t flags)
{
int events;
uint32_t revents;
ngx_int_t instance, i;
ngx_uint_t level;
ngx_err_t err;
ngx_event_t *rev, *wev, **queue;
ngx_connection_t *c;
/* NGX_TIMER_INFINITE == INFTIM */
ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
"epoll timer: %M", timer);
//调用epoll_wait获取已准备就绪的事件
events = epoll_wait(ep, event_list, (int) nevents, timer);
err = (events == -1) ? ngx_errno : 0;
/***NGX_UPDATE_TIME标识在没有设置timer_resolution配置项时有效表示每次调用epoll_wait函数返回会都更新时间。
ngx_event_timer_alarm变量在设置timer_resolution配置项时有效，每间隔timer_resolution配置项参数值就会设置
ngx_event_timer_alarm变量为1表示需更新时间。***/
if (flags & NGX_UPDATE_TIME || ngx_event_timer_alarm) {
ngx_time_update();
}
//err为非零指示epoll_wait失败
if (err) {
if (err == NGX_EINTR) {
if (ngx_event_timer_alarm) {
ngx_event_timer_alarm = 0;
return NGX_OK;
}
level = NGX_LOG_INFO;
} else {
level = NGX_LOG_ALERT;
}
ngx_log_error(level, cycle->log, err, "epoll_wait() failed");
return NGX_ERROR;
}
if (events == 0) {
if (timer != NGX_TIMER_INFINITE) {
return NGX_OK;
}
ngx_log_error(NGX_LOG_ALERT, cycle->log, 0,
"epoll_wait() returned no events without timeout");
return NGX_ERROR;
}
//仅在多线程环境下此锁才有效
ngx_mutex_lock(ngx_posted_events_mutex);
/***循环处理已就绪的事件***/
for (i = 0; i < events; i++) {
//获取事件关联的连接对象，对象地址最低位保存有在事件添加时设置的事件过期位
c = event_list[i].data.ptr;
//取事件过期位
instance = (uintptr_t) c & 1;
//屏蔽掉连接对象的最低位
c = (ngx_connection_t *) ((uintptr_t) c & (uintptr_t) ~1);
rev = c->read;
/***同一条连接的读/写事件的instance位值相同，由于下面先处理读事件这里通过读事件
的过期位来判断连接是否过期，当fd为-1时也表示连接过期。***/
if (c->fd == -1 || rev->instance != instance) {
/*
* the stale event from a file descriptor
* that was just closed in this iteration
*/
ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
"epoll: stale event %p", c);
continue;
}
//获取连接已就绪的事件类型
revents = event_list[i].events;
ngx_log_debug3(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
"epoll: fd:%d ev:%04XD d:%p",
c->fd, revents, event_list[i].data.ptr);
/***连接出现错误,EPOLLHUP标识表示收到RST报文。检测到这两种类型时 tcp连接中可能还有
数据未被读取***/
if (revents & (EPOLLERR|EPOLLHUP)) {
ngx_log_debug2(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
"epoll_wait() error on fd:%d ev:%04XD",
c->fd, revents);
}
#if 0
if (revents & ~(EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP)) {
ngx_log_error(NGX_LOG_ALERT, cycle->log, 0,
"strange epoll_wait() events fd:%d ev:%04XD",
c->fd, revents);
}
#endif
/***如果连接发生错误但未置EPOLLIN及EPOLLOUT，这时我们加上EPOLLIN和EPOLLOUT，在调用读/写事件的
回调函数时就会知道为什么出现错误。如果不加EPOLLIN和EPOLLOUT，后面就没法调用读/写事件的
回调函数也就无法处理该连接了。***/
if ((revents & (EPOLLERR|EPOLLHUP))
&& (revents & (EPOLLIN|EPOLLOUT)) == 0)
{
/*
* if the error events were returned without EPOLLIN or EPOLLOUT,
* then add these flags to handle the events at least in one
* active handler
*/
revents |= EPOLLIN|EPOLLOUT;
}
/***连接可读且活跃***/
if ((revents & EPOLLIN) && rev->active) {
#if (NGX_HAVE_EPOLLRDHUP)
//EPOLLRDHUP表示连接对方关闭了读端
if (revents & EPOLLRDHUP) {
rev->pending_eof = 1;
}
#endif
//NGX_POST_THREAD_EVENTS宏末被使用
if ((flags & NGX_POST_THREAD_EVENTS) && !rev->accept) {
rev->posted_ready = 1;
} else {
//标识事件已就绪
rev->ready = 1;
}
/***NGX_POST_EVENTS表示事件需要延后处理，这里根据accept标识位将事件加入到相应队列中***/
if (flags & NGX_POST_EVENTS) {
queue = (ngx_event_t **) (rev->accept ?
&ngx_posted_accept_events : &ngx_posted_events);
ngx_locked_post_event(rev, queue);
} else {
//调用事件的回调函数
rev->handler(rev);
}
}
wev = c->write;
/***连接可写且活跃***/
if ((revents & EPOLLOUT) && wev->active) {
//重新检查事件是否过期，因为在处理读事件过程中该事件可能已结束。
if (c->fd == -1 || wev->instance != instance) {
/*
* the stale event from a file descriptor
* that was just closed in this iteration
*/
ngx_log_debug1(NGX_LOG_DEBUG_EVENT, cycle->log, 0,
"epoll: stale event %p", c);
continue;
}
if (flags & NGX_POST_THREAD_EVENTS) {
wev->posted_ready = 1;
} else {
wev->ready = 1;
}
if (flags & NGX_POST_EVENTS) {
ngx_locked_post_event(wev, &ngx_posted_events);
} else {
wev->handler(wev);
}
}
}
ngx_mutex_unlock(ngx_posted_events_mutex);
return NGX_OK;
}