linux kernel中epoll的设计和实现

http://www.pagefault.info/?p=264 

原创文章，转载请注明： 转载自pagefault

本文链接地址: linux kernel中epoll的设计和实现

这里就不贴源码了，源码分析的话，网上一大堆，我这里只是简要的描述下epoll的实现和一些关键的代码片段。

相关的文件在 fs/eventpoll.c中,我看的是2.6.38的内核代码.

1 epoll在创建的时候会调用anon_inode_getfd新建一个file instance，也就是epoll可以看成一个文件。因此我们可以看到epoll_create会返回一个fd.

1

2

error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,

             O_RDWR | (flags & O_CLOEXEC));

2 epoll所管理的所有的句柄都是放在一个大的结构eventpoll(红黑树)中,而这个结构是保存在file 的private_data域中的(因为epoll本身就是一个文件).这样每次通过epoll fd就可以直接得到eventpoll.

1

2

3

4

5

    file = fget(epfd);

    /* Get the "struct file *" for the target file */

    tfile = fget(fd);

...................................

    ep = file->private_data;

3 每一个加入到epoll监听的句柄(也就是红黑树的一个节点)都是一个epitem.它包含了一个 struct eventpoll *ep，也就是它所属于的eventpoll(epoll实例).

1

2

3

4

5

6

7

8

9

10

11

12

if(!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))

    return-ENOMEM;

 

/* Item initialization follow here ... */

INIT_LIST_HEAD(&epi->rdllink);

INIT_LIST_HEAD(&epi->fllink);

INIT_LIST_HEAD(&epi->pwqlist);

epi->ep = ep;

ep_set_ffd(&epi->ffd, tfile, fd);

epi->event = *event;

epi->nwait = 0;

epi->next = EP_UNACTIVE_PTR;

4 在eventpoll中包含两个wait queue，一个是被epoll_wait使用的(wq)，一个是被file->poll使用的(poll_wait).这两个都是属于eventpoll.而在epitem也有一个wait queue(pwqlist)，这个queue是fd私有的wait queue，所以它是保存在epitem中的。

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

structepitem {

    /* RB tree node used to link this structure to the eventpoll RB tree */

    structrb_node rbn;

................................................

    /* List containing poll wait queues */

    structlist_head pwqlist;

 

    /* The "container" of this item */

    structeventpoll *ep;

.............................................

};

 

structeventpoll {

    /* Protect the this structure access */

    spinlock_t lock;

    /*

     * This mutex is used to ensure that files are not removed

     * while epoll is using them. This is held during the event

     * collection loop, the file cleanup path, the epoll file exit

     * code and the ctl operations.

     */

    structmutex mtx;

 

    /* Wait queue used by sys_epoll_wait() */

    wait_queue_head_t wq;

 

    /* Wait queue used by file->poll() */

    wait_queue_head_t poll_wait;

 

    /* List of ready file descriptors */

    structlist_head rdllist;

 

    /* RB tree root used to store monitored fd structs */

    structrb_root rbr;

 

    /*

     * This is a single linked list that chains all the "struct epitem" that

     * happened while transfering ready events to userspace w/out

     * holding ->lock.

     */

    structepitem *ovflist;

.................................................

};

5 当我们添加一个句柄到一个epoll fd的时候，默认是会包含POLLERR和POLLHUP事件.并将epitem插入到红黑树中(eventpoll).然后会初始化一个poll_table,然后设置它的回调函数为ep_poll_callback，紧接着调用file->poll,如果是socket，则会调用tcp_poll,这个函数将会调用ep_poll_callback.

1

2

3

4

5

6

7

8

9

    switch(op) {

    caseEPOLL_CTL_ADD:

        if(!epi) {

            epds.events |= POLLERR | POLLHUP;

            error = ep_insert(ep, &epds, tfile, fd);

        }else

            error = -EEXIST;

        break;

.............................................

6 ep_poll_callback这个函数主要是绑定epitem的wait queue的回调为ep_poll_callback.也就是对应fd如果有事件，则就会调用ep_poll_callback。

7 epoll中保存了一个read list(rdllist)，所有的已经有通知事件的句柄，都会放到这个list中，而对应的操作就是在ep_poll_callback中，在ep_poll_callback中主要就是由wait queue指针来取得对应的epitem，然后再取得eventpoll，并将这个epitem加入到ready list，唤醒epoll_wait(wq).这里可以看到由于一个句柄只会对应一个epitem，所以在rdllist中，也不会有重复的epitem，在ep_poll_callback会判断是否rdllist中是否已经包含了将要插入的epitem，如果包含，则直接返回.

1

2

3

4

5

6

7

8

staticint ep_poll_callback(wait_queue_t *wait, unsigned mode, intsync, void*key)

{

    structepitem *epi = ep_item_from_wait(wait);

    structeventpoll *ep = epi->ep;

...............................................

    /* If this file is already in the ready list we exit soon */

    if(!ep_is_linked(&epi->rdllink))

        list_add_tail(&epi->rdllink, &ep->rdllist);

8 当系统调用epoll_wait,如果ready list(rdllist)为空，则休眠等待被唤醒。当被唤醒之后(第7条),将rdllist复制(指针)到一个新的list(主要是针对LT),然后调用ep_send_events_proc对这个新的list进行遍历(对应的会从rdllist中删除这个epitem).遍历完毕后，最终会返回给用户对应的数据.

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

if(list_empty(&ep->rdllist)) {

    /*

     * We don't have any available event to return to the caller.

     * We need to sleep here, and we will be wake up by

     * ep_poll_callback() when events will become available.

     */

    init_waitqueue_entry(&wait, current);

    __add_wait_queue_exclusive(&ep->wq, &wait);

 

    for(;;) {

        /*

         * We don't want to sleep if the ep_poll_callback() sends us

         * a wakeup in between. That's why we set the task state

         * to TASK_INTERRUPTIBLE before doing the checks.

         */

        set_current_state(TASK_INTERRUPTIBLE);

        if(!list_empty(&ep->rdllist) || timed_out)

            break;

        if(signal_pending(current)) {

            res = -EINTR;

            break;

        }

 

        spin_unlock_irqrestore(&ep->lock, flags);

        if(!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))

            timed_out = 1;

 

        spin_lock_irqsave(&ep->lock, flags);

    }

    __remove_wait_queue(&ep->wq, &wait);

 

    set_current_state(TASK_RUNNING);

}

9 LT和ET的区别是在ep_send_events_proc中处理的，如果是LT，不但会将对应的数据返回给用户，并且会将当前的epitem再次加入到rdllist中。这样子，如果下次再次被唤醒就会给用户空间再次返回事件.

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

staticint ep_send_events_proc(struct eventpoll *ep, struct list_head *head,

                   void*priv)

{

....................................................

            if(__put_user(revents, &uevent->events) ||

                __put_user(epi->event.data, &uevent->data)) {

                list_add(&epi->rdllink, head);

                returneventcnt ? eventcnt : -EFAULT;

            }

            eventcnt++;

            uevent++;

            if(epi->event.events & EPOLLONESHOT)

                epi->event.events &= EP_PRIVATE_BITS;

            elseif (!(epi->event.events & EPOLLET)) {

                /*

                 * If this file has been added with Level

                 * Trigger mode, we need to insert back inside

                 * the ready list, so that the next call to

                 * epoll_wait() will check again the events

                 * availability. At this point, noone can insert

                 * into ep->rdllist besides us. The epoll_ctl()

                 * callers are locked out by

                 * ep_scan_ready_list() holding "mtx" and the

                 * poll callback will queue them in ep->ovflist.

                 */

                list_add_tail(&epi->rdllink, &ep->rdllist);

            }

..............................

}

10 eventpol还有一个list，叫做ovflist，主要是解决当内核在传输数据给用户空间(ep_send_events_proc)时的锁(eventpoll->mtx)，此时epoll就是将这个时候传递上来的事件保存到ovflist中。

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

staticint ep_poll_callback(wait_queue_t *wait, unsigned mode, intsync, void*key)

{

......................................

    /*

     * If we are trasfering events to userspace, we can hold no locks

     * (because we're accessing user memory, and because of linux f_op->poll()

     * semantics). All the events that happens during that period of time are

     * chained in ep->ovflist and requeued later on.

     */

    if(unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {

        if(epi->next == EP_UNACTIVE_PTR) {

            epi->next = ep->ovflist;

            ep->ovflist = epi;

        }

        gotoout_unlock;

    }

.........................

}

11 当从ep_send_events_proc返回(ep_scan_ready_list)后，会遍历ovflist，然后将ready的epitem保存到rdllist,以便与下次再次被唤醒时进行操作.

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

staticint ep_scan_ready_list(struct eventpoll *ep,

                  int(*sproc)(structeventpoll *,

                       structlist_head *, void*),

                  void*priv)

{

...........................................

    /*

     * During the time we spent inside the "sproc" callback, some

     * other events might have been queued by the poll callback.

     * We re-insert them inside the main ready-list here.

     */

    for(nepi = ep->ovflist; (epi = nepi) != NULL;

         nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {

        /*

         * We need to check if the item is already in the list.

         * During the "sproc" callback execution time, items are

         * queued into ->ovflist but the "txlist" might already

         * contain them, and the list_splice() below takes care of them.

         */

        if(!ep_is_linked(&epi->rdllink))

            list_add_tail(&epi->rdllink, &ep->rdllist);

    }

........................

}