ST_HIDDEN int _st_epoll_pollset_add(struct pollfd *pds, int npds)
{
struct epoll_event ev;
int i, fd;
int old_events, events, op;
/* Do as many checks as possible up front */
for (i = 0; i < npds; i++) {
fd = pds[i].fd;
if (fd < 0 || !pds[i].events || (pds[i].events & ~(POLLIN | POLLOUT | POLLPRI))) {
errno = EINVAL;
return -1;
}
if (fd >= _st_epoll_data->fd_data_size && _st_epoll_fd_data_expand(fd) < 0) {
return -1;
}
}
for (i = 0; i < npds; i++) {
fd = pds[i].fd;
old_events = _ST_EPOLL_EVENTS(fd);
if (pds[i].events & POLLIN) {
_ST_EPOLL_READ_CNT(fd)++;
}
if (pds[i].events & POLLOUT) {
_ST_EPOLL_WRITE_CNT(fd)++;
}
if (pds[i].events & POLLPRI) {
_ST_EPOLL_EXCEP_CNT(fd)++;
}
events = _ST_EPOLL_EVENTS(fd);
if (events != old_events) {
op = old_events ? EPOLL_CTL_MOD : EPOLL_CTL_ADD;
ev.events = events;
ev.data.fd = fd;
if (epoll_ctl(_st_epoll_data->epfd, op, fd, &ev) < 0 && (op != EPOLL_CTL_ADD || errno != EEXIST)) {
break;
}
if (op == EPOLL_CTL_ADD) {
_st_epoll_data->evtlist_cnt++;
if (_st_epoll_data->evtlist_cnt > _st_epoll_data->evtlist_size) {
_st_epoll_evtlist_expand();
}
}
}
}
if (i < npds) {
/* Error */
int err = errno;
/* Unroll the state */
_st_epoll_pollset_del(pds, i + 1);
errno = err;
return -1;
}
return 0;
}
ST_HIDDEN void _st_epoll_pollset_del(struct pollfd *pds, int npds)
{
struct epoll_event ev;
struct pollfd *pd;
struct pollfd *epd = pds + npds;
int old_events, events, op;
/*
* It's more or less OK if deleting fails because a descriptor
* will either be closed or deleted in dispatch function after
* it fires.
*/
for (pd = pds; pd < epd; pd++) {
old_events = _ST_EPOLL_EVENTS(pd->fd);
if (pd->events & POLLIN) {
_ST_EPOLL_READ_CNT(pd->fd)--;
}
if (pd->events & POLLOUT) {
_ST_EPOLL_WRITE_CNT(pd->fd)--;
}
if (pd->events & POLLPRI) {
_ST_EPOLL_EXCEP_CNT(pd->fd)--;
}
events = _ST_EPOLL_EVENTS(pd->fd);
/*
* The _ST_EPOLL_REVENTS check below is needed so we can use
* this function inside dispatch(). Outside of dispatch()
* _ST_EPOLL_REVENTS is always zero for all descriptors.
*/
if (events != old_events && _ST_EPOLL_REVENTS(pd->fd) == 0) {
op = events ? EPOLL_CTL_MOD : EPOLL_CTL_DEL;
ev.events = events;
ev.data.fd = pd->fd;
if (epoll_ctl(_st_epoll_data->epfd, op, pd->fd, &ev) == 0 && op == EPOLL_CTL_DEL) {
_st_epoll_data->evtlist_cnt--;
}
}
}
}
ST_HIDDEN void _st_epoll_dispatch(void)
{
st_utime_t min_timeout;
_st_clist_t *q;
_st_pollq_t *pq;
struct pollfd *pds, *epds;
struct epoll_event ev;
int timeout, nfd, i, osfd, notify;
int events, op;
short revents;
if (_ST_SLEEPQ == NULL) {
timeout = -1;
} else {
min_timeout = (_ST_SLEEPQ->due <= _ST_LAST_CLOCK) ? 0 :
(_ST_SLEEPQ->due - _ST_LAST_CLOCK);
timeout = (int) (min_timeout / 1000);
}
if (_st_epoll_data->pid != getpid()) {
/* We probably forked, reinitialize epoll set */
close(_st_epoll_data->epfd);
_st_epoll_data->epfd = epoll_create(_st_epoll_data->fd_hint);
if (_st_epoll_data->epfd < 0) {
/* There is nothing we can do here, will retry later */
return;
}
fcntl(_st_epoll_data->epfd, F_SETFD, FD_CLOEXEC);
_st_epoll_data->pid = getpid();
/* Put all descriptors on ioq into new epoll set */
memset(_st_epoll_data->fd_data, 0,
_st_epoll_data->fd_data_size * sizeof(_epoll_fd_data_t));
_st_epoll_data->evtlist_cnt = 0;
for (q = _ST_IOQ.next; q != &_ST_IOQ; q = q->next) {
pq = _ST_POLLQUEUE_PTR(q);
_st_epoll_pollset_add(pq->pds, pq->npds);
}
}
/* Check for I/O operations */
nfd = epoll_wait(_st_epoll_data->epfd, _st_epoll_data->evtlist,
_st_epoll_data->evtlist_size, timeout);
if (nfd > 0) {
for (i = 0; i < nfd; i++) {
osfd = _st_epoll_data->evtlist[i].data.fd;
_ST_EPOLL_REVENTS(osfd) = _st_epoll_data->evtlist[i].events;
if (_ST_EPOLL_REVENTS(osfd) & (EPOLLERR | EPOLLHUP)) {
/* Also set I/O bits on error */
_ST_EPOLL_REVENTS(osfd) |= _ST_EPOLL_EVENTS(osfd);
}
}
for (q = _ST_IOQ.next; q != &_ST_IOQ; q = q->next) {
pq = _ST_POLLQUEUE_PTR(q);
notify = 0;
epds = pq->pds + pq->npds;
for (pds = pq->pds; pds < epds; pds++) {
if (_ST_EPOLL_REVENTS(pds->fd) == 0) {
pds->revents = 0;
continue;
}
osfd = pds->fd;
events = pds->events;
revents = 0;
if ((events & POLLIN) && (_ST_EPOLL_REVENTS(osfd) & EPOLLIN))
revents |= POLLIN;
if ((events & POLLOUT) && (_ST_EPOLL_REVENTS(osfd) & EPOLLOUT))
revents |= POLLOUT;
if ((events & POLLPRI) && (_ST_EPOLL_REVENTS(osfd) & EPOLLPRI))
revents |= POLLPRI;
if (_ST_EPOLL_REVENTS(osfd) & EPOLLERR)
revents |= POLLERR;
if (_ST_EPOLL_REVENTS(osfd) & EPOLLHUP)
revents |= POLLHUP;
pds->revents = revents;
if (revents) {
notify = 1;
}
}
if (notify) {
ST_REMOVE_LINK(&pq->links);
pq->on_ioq = 0;
/*
* Here we will only delete/modify descriptors that
* didn't fire (see comments in _st_epoll_pollset_del()).
*/
_st_epoll_pollset_del(pq->pds, pq->npds);
if (pq->thread->flags & _ST_FL_ON_SLEEPQ)
_ST_DEL_SLEEPQ(pq->thread);
pq->thread->state = _ST_ST_RUNNABLE;
_ST_ADD_RUNQ(pq->thread);
}
}
for (i = 0; i < nfd; i++) {
/* Delete/modify descriptors that fired */
osfd = _st_epoll_data->evtlist[i].data.fd;
_ST_EPOLL_REVENTS(osfd) = 0;
events = _ST_EPOLL_EVENTS(osfd);
op = events ? EPOLL_CTL_MOD : EPOLL_CTL_DEL;
ev.events = events;
ev.data.fd = osfd;
if (epoll_ctl(_st_epoll_data->epfd, op, osfd, &ev) == 0 &&
op == EPOLL_CTL_DEL) {
_st_epoll_data->evtlist_cnt--;
}
}
}
}
// epoll 事件分发
ST_HIDDEN void _st_epoll_dispatch(void)
{
st_utime_t min_timeout;
_st_clist_t *q;
_st_pollq_t *pq;
struct pollfd *pds, *epds;
struct epoll_event ev;
int timeout, nfd, i, osfd, notify;
int events, op;
short revents;
if (_ST_SLEEPQ == NULL) {
timeout = -1;
} else {
// 获取最早睡眠的线程还要睡多久（用于epoll超时）
min_timeout = (_ST_SLEEPQ->due <= _ST_LAST_CLOCK) ? 0 :
(_ST_SLEEPQ->due - _ST_LAST_CLOCK);
timeout = (int) (min_timeout / 1000);
}
// fork 出的子进程，则重新获取 epoll fd
if (_st_epoll_data->pid != getpid()) {
/* We probably forked, reinitialize epoll set */
close(_st_epoll_data->epfd);
_st_epoll_data->epfd = epoll_create(_st_epoll_data->fd_hint);
if (_st_epoll_data->epfd < 0) {
/* There is nothing we can do here, will retry later */
return;
}
// exec 调用时关闭
fcntl(_st_epoll_data->epfd, F_SETFD, FD_CLOEXEC);
_st_epoll_data->pid = getpid();
/* Put all descriptors on ioq into new epoll set */
memset(_st_epoll_data->fd_data, 0,
_st_epoll_data->fd_data_size * sizeof(_epoll_fd_data_t));
_st_epoll_data->evtlist_cnt = 0;
// 将 io 队列的 net fd 都加入事件系统
for (q = _ST_IOQ.next; q != &_ST_IOQ; q = q->next) {
pq = _ST_POLLQUEUE_PTR(q);
_st_epoll_pollset_add(pq->pds, pq->npds);
}
}
/* Check for I/O operations */
nfd = epoll_wait(_st_epoll_data->epfd, _st_epoll_data->evtlist,
_st_epoll_data->evtlist_size, timeout);
if (nfd > 0) {
for (i = 0; i < nfd; i++) {
osfd = _st_epoll_data->evtlist[i].data.fd;
_ST_EPOLL_REVENTS(osfd) = _st_epoll_data->evtlist[i].events;
if (_ST_EPOLL_REVENTS(osfd) & (EPOLLERR | EPOLLHUP)) {
/* Also set I/O bits on error */
_ST_EPOLL_REVENTS(osfd) |= _ST_EPOLL_EVENTS(osfd);
}
}
// ####
// 依次从 io 队列取出每个线程等待 io 的 pollfd，pq 是一个线程中加入的所有文件描述符
for (q = _ST_IOQ.next; q != &_ST_IOQ; q = q->next) {
pq = _ST_POLLQUEUE_PTR(q);
notify = 0;
epds = pq->pds + pq->npds;
// 遍历每个线程 pollfd ,获取对应的事件，然后从 io 队列中移除；如果 pollfd 上
// 有事件发生，则将其从 io 队列移除，没有事件的继续在 io 队列等待
for (pds = pq->pds; pds < epds; pds++) {
if (_ST_EPOLL_REVENTS(pds->fd) == 0) {
pds->revents = 0;
continue;
}
osfd = pds->fd;
events = pds->events;
revents = 0;
if ((events & POLLIN) && (_ST_EPOLL_REVENTS(osfd) & EPOLLIN))
revents |= POLLIN;
if ((events & POLLOUT) && (_ST_EPOLL_REVENTS(osfd) & EPOLLOUT))
revents |= POLLOUT;
if ((events & POLLPRI) && (_ST_EPOLL_REVENTS(osfd) & EPOLLPRI))
revents |= POLLPRI;
if (_ST_EPOLL_REVENTS(osfd) & EPOLLERR)
revents |= POLLERR;
if (_ST_EPOLL_REVENTS(osfd) & EPOLLHUP)
revents |= POLLHUP;
pds->revents = revents;
// 判断是不是有事件发生
if (revents) {
notify = 1;
}
}
if (notify) {
// 将有时间发生的描述符从 io 队列移除
ST_REMOVE_LINK(&pq->links);
pq->on_ioq = 0;
/*
* Here we will only delete/modify descriptors that
* didn't fire (see comments in _st_epoll_pollset_del()).
*/
// 将当前线程已处理完事件的文件描述符从事件系统中移除
_st_epoll_pollset_del(pq->pds, pq->npds);
// 如果当前线程处于睡眠则从睡眠队列移除，然后将当前线程加入运行队列
if (pq->thread->flags & _ST_FL_ON_SLEEPQ)
_ST_DEL_SLEEPQ(pq->thread);
pq->thread->state = _ST_ST_RUNNABLE;
_ST_ADD_RUNQ(pq->thread);
}
}
// 修改剩下还有等待事件的描述符
for (i = 0; i < nfd; i++) {
/* Delete/modify descriptors that fired */
osfd = _st_epoll_data->evtlist[i].data.fd;
_ST_EPOLL_REVENTS(osfd) = 0;
events = _ST_EPOLL_EVENTS(osfd);
op = events ? EPOLL_CTL_MOD : EPOLL_CTL_DEL;
ev.events = events;
ev.data.fd = osfd;
if (epoll_ctl(_st_epoll_data->epfd, op, osfd, &ev) == 0 &&
op == EPOLL_CTL_DEL) {
_st_epoll_data->evtlist_cnt--;
}
}
}
}