epoll内核源码详解(转 作者:赛罗·奥特曼 来源:牛客网)

       发现自己发的一篇面经后,很多小伙伴向我索要epoll的内核源码实现,那我就在牛客网发下这源码还有自己总结的流程.

另外 网上很多博客说epoll使用了共享内存,这个是完全错误的 ,可以阅读源码,会发现完全没有使用共享内存的任何api,
而是 使用了copy_from_user跟__put_user进行内核跟用户虚拟空间数据交互.

 
  1  *  fs/eventpoll.c (Efficient event retrieval implementation)
  2  *  Copyright (C) 2001,...,2009  Davide Libenzi
  3  *
  4  *  This program is free software; you can redistribute it and/or modify
  5  *  it under the terms of the GNU General Public License as published by
  6  *  the Free Software Foundation; either version 2 of the License, or
  7  *  (at your option) any later version.
  8  *
  9  *  Davide Libenzi <davidel@xmailserver.org>
 10  *
 11  */
 12 /*
 13  * 在深入了解epoll的实现之前, 先来了解内核的3个方面.
 14  * 1. 等待队列 waitqueue
 15  * 我们简单解释一下等待队列:
 16  * 队列头(wait_queue_head_t)往往是资源生产者,
 17  * 队列成员(wait_queue_t)往往是资源消费者,
 18  * 当头的资源ready后, 会逐个执行每个成员指定的回调函数,
 19  * 来通知它们资源已经ready了, 等待队列大致就这个意思.
 20  * 2. 内核的poll机制
 21  * 被Poll的fd, 必须在实现上支持内核的Poll技术,
 22  * 比如fd是某个字符设备,或者是个socket, 它必须实现
 23  * file_operations中的poll操作, 给自己分配有一个等待队列头.
 24  * 主动poll fd的某个进程必须分配一个等待队列成员, 添加到
 25  * fd的对待队列里面去, 并指定资源ready时的回调函数.
 26  * 用socket做例子, 它必须有实现一个poll操作, 这个Poll是
 27  * 发起轮询的代码必须主动调用的, 该函数中必须调用poll_wait(),
 28  * poll_wait会将发起者作为等待队列成员加入到socket的等待队列中去.
 29  * 这样socket发生状态变化时可以通过队列头逐个通知所有关心它的进程.
 30  * 这一点必须很清楚的理解, 否则会想不明白epoll是如何
 31  * 得知fd的状态发生变化的.
 32  * 3. epollfd本身也是个fd, 所以它本身也可以被epoll,
 33  * 可以猜测一下它是不是可以无限嵌套epoll下去...
 34  *
 35  * epoll基本上就是使用了上面的1,2点来完成.
 36  * 可见epoll本身并没有给内核引入什么特别复杂或者高深的技术,
 37  * 只不过是已有功能的重新组合, 达到了超过select的效果.
 38  */
 39 /*
 40  * 相关的其它内核知识:
 41  * 1. fd我们知道是文件描述符, 在内核态, 与之对应的是struct file结构,
 42  * 可以看作是内核态的文件描述符.
 43  * 2. spinlock, 自旋锁, 必须要非常小心使用的锁,
 44  * 尤其是调用spin_lock_irqsave()的时候, 中断关闭, 不会发生进程调度,
 45  * 被保护的资源其它CPU也无法访问. 这个锁是很强力的, 所以只能锁一些
 46  * 非常轻量级的操作.
 47  * 3. 引用计数在内核中是非常重要的概念,
 48  * 内核代码里面经常有些release, free释放资源的函数几乎不加任何锁,
 49  * 这是因为这些函数往往是在对象的引用计数变成0时被调用,
 50  * 既然没有进程在使用在这些对象, 自然也不需要加锁.
 51  * struct file 是持有引用计数的.
 52  */
 53 /* --- epoll相关的数据结构 --- */
 54 /*
 55  * This structure is stored inside the "private_data" member of the file
 56  * structure and rapresent the main data sructure for the eventpoll
 57  * interface.
 58  */
 59 /* 每创建一个epollfd, 内核就会分配一个eventpoll与之对应, 可以说是
 60  * 内核态的epollfd. */
 61 struct eventpoll {
 62     /* Protect the this structure access */
 63     spinlock_t lock;
 64     /*
 65      * This mutex is used to ensure that files are not removed
 66      * while epoll is using them. This is held during the event
 67      * collection loop, the file cleanup path, the epoll file exit
 68      * code and the ctl operations.
 69      */
 70     /* 添加, 修改或者删除监听fd的时候, 以及epoll_wait返回, 向用户空间
 71      * 传递数据时都会持有这个互斥锁, 所以在用户空间可以放心的在多个线程
 72      * 中同时执行epoll相关的操作, 内核级已经做了保护. */
 73     struct mutex mtx;
 74     /* Wait queue used by sys_epoll_wait() */
 75     /* 调用epoll_wait()时, 我们就是"睡"在了这个等待队列上... */
 76     wait_queue_head_t wq;
 77     /* Wait queue used by file->poll() */
 78     /* 这个用于epollfd本事被poll的时候... */
 79     wait_queue_head_t poll_wait;
 80     /* List of ready file descriptors */
 81     /* 所有已经ready的epitem都在这个链表里面 */
 82     struct list_head rdllist;
 83     /* RB tree root used to store monitored fd structs */
 84     /* 所有要监听的epitem都在这里 */
 85     struct rb_root rbr;
 86     /*
 87         这是一个单链表链接着所有的struct epitem当event转移到用户空间时
 88      */
 89      * This is a single linked list that chains all the "struct epitem" that
 90      * happened while transfering ready events to userspace w/out
 91      * holding ->lock.
 92      */
 93     struct epitem *ovflist;
 94     /* The user that created the eventpoll descriptor */
 95     /* 这里保存了一些用户变量, 比如fd监听数量的最大值等等 */
 96     struct user_struct *user;
 97 };
 98 /*
 99  * Each file descriptor added to the eventpoll interface will
100  * have an entry of this type linked to the "rbr" RB tree.
101  */
102 /* epitem 表示一个被监听的fd */
103 struct epitem {
104     /* RB tree node used to link this structure to the eventpoll RB tree */
105     /* rb_node, 当使用epoll_ctl()将一批fds加入到某个epollfd时, 内核会分配
106      * 一批的epitem与fds们对应, 而且它们以rb_tree的形式组织起来, tree的root
107      * 保存在epollfd, 也就是struct eventpoll中.
108      * 在这里使用rb_tree的原因我认为是提高查找,插入以及删除的速度.
109      * rb_tree对以上3个操作都具有O(lgN)的时间复杂度 */
110     struct rb_node rbn;
111     /* List header used to link this structure to the eventpoll ready list */
112     /* 链表节点, 所有已经ready的epitem都会被链到eventpoll的rdllist中 */
113     struct list_head rdllink;
114     /*
115      * Works together "struct eventpoll"->ovflist in keeping the
116      * single linked chain of items.
117      */
118     /* 这个在代码中再解释... */
119     struct epitem *next;
120     /* The file descriptor information this item refers to */
121     /* epitem对应的fd和struct file */
122     struct epoll_filefd ffd;
123     /* Number of active wait queue attached to poll operations */
124     int nwait;
125     /* List containing poll wait queues */
126     struct list_head pwqlist;
127     /* The "container" of this item */
128     /* 当前epitem属于哪个eventpoll */
129     struct eventpoll *ep;
130     /* List header used to link this item to the "struct file" items list */
131     struct list_head fllink;
132     /* The structure that describe the interested events and the source fd */
133     /* 当前的epitem关系哪些events, 这个数据是调用epoll_ctl时从用户态传递过来 */
134     struct epoll_event event;
135 };
136 struct epoll_filefd {
137     struct file *file;
138     int fd;
139 };
140 /* poll所用到的钩子Wait structure used by the poll hooks */
141 struct eppoll_entry {
142     /* List header used to link this structure to the "struct epitem" */
143     struct list_head llink;
144     /* The "base" pointer is set to the container "struct epitem" */
145     struct epitem *base;
146     /*
147      * Wait queue item that will be linked to the target file wait
148      * queue head.
149      */
150     wait_queue_t wait;
151     /* The wait queue head that linked the "wait" wait queue item */
152     wait_queue_head_t *whead;
153 };
154 /* Wrapper struct used by poll queueing */
155 struct ep_pqueue {
156     poll_table pt;
157     struct epitem *epi;
158 };
159 /* Used by the ep_send_events() function as callback private data */
160 struct ep_send_events_data {
161     int maxevents;
162     struct epoll_event __user *events;
163 };
164  
165 /* --- 代码注释 --- */
166 /* 你没看错, 这就是epoll_create()的真身, 基本啥也不干直接调用epoll_create1了,
167  * 另外你也可以发现, size这个参数其实是没有任何用处的... */
168 SYSCALL_DEFINE1(epoll_create, int, size)
169 {
170         if (size <= 0)
171                 return -EINVAL;
172         return sys_epoll_create1(0);
173 }
174 /* 这才是真正的epoll_create啊~~ */
175 SYSCALL_DEFINE1(epoll_create1, int, flags)
176 {
177     int error;
178     struct eventpoll *ep = NULL;//主描述符
179     /* Check the EPOLL_* constant for consistency.  */
180     /* 这句没啥用处... */
181     BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
182     /* 对于epoll来讲, 目前唯一有效的FLAG就是CLOEXEC */
183     if (flags & ~EPOLL_CLOEXEC)
184         return -EINVAL;
185     /*
186      * Create the internal data structure ("struct eventpoll").
187      */
188     /* 分配一个struct eventpoll, 分配和初始化细节我们随后深聊~ */
189     error = ep_alloc(&ep);
190     if (error < 0)
191         return error;
192     /*
193      * Creates all the items needed to setup an eventpoll file. That is,
194      * a file structure and a free file descriptor.
195      */
196     /* 这里是创建一个匿名fd, 说起来就话长了...长话短说:
197      * epollfd本身并不存在一个真正的文件与之对应, 所以内核需要创建一个
198      * "虚拟"的文件, 并为之分配真正的struct file结构, 而且有真正的fd.
199      * 这里2个参数比较关键:
200      * eventpoll_fops, fops就是file operations, 就是当你对这个文件(这里是虚拟的)进行操作(比如读)时,
201      * fops里面的函数指针指向真正的操作实现, 类似C++里面虚函数和子类的概念.
202      * epoll只实现了poll和release(就是close)操作, 其它文件系统操作都有VFS全权处理了.
203      * ep, ep就是struct epollevent, 它会作为一个私有数据保存在struct file的private指针里面.
204      * 其实说白了, 就是为了能通过fd找到struct file, 通过struct file能找到eventpoll结构.
205      * 如果懂一点Linux下字符设备驱动开发, 这里应该是很好理解的,
206      * 推荐阅读 <Linux device driver 3rd>
207      */
208     error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
209                  O_RDWR | (flags & O_CLOEXEC));
210     if (error < 0)
211         ep_free(ep);
212     return error;
213 }
214 /*
215 * 创建好epollfd后, 接下来我们要往里面添加fd咯
216 * 来看epoll_ctl
217 * epfd 就是epollfd
218 * op ADD,MOD,DEL
219 * fd 需要监听的描述符
220 * event 我们关心的events
221 */
222 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
223         struct epoll_event __user *, event)
224 {
225     int error;
226     struct file *file, *tfile;
227     struct eventpoll *ep;
228     struct epitem *epi;
229     struct epoll_event epds;
230     error = -EFAULT;
231     /*
232      * 错误处理以及从用户空间将epoll_event结构copy到内核空间.
233      */
234     if (ep_op_has_event(op) &&
235         copy_from_user(&epds, event, sizeof(struct epoll_event)))
236         goto error_return;
237     /* Get the "struct file *" for the eventpoll file */
238     /* 取得struct file结构, epfd既然是真正的fd, 那么内核空间
239      * 就会有与之对于的一个struct file结构
240      * 这个结构在epoll_create1()中, 由函数anon_inode_getfd()分配 */
241     error = -EBADF;
242     file = fget(epfd);
243     if (!file)
244         goto error_return;
245     /* Get the "struct file *" for the target file */
246     /* 我们需要监听的fd, 它当然也有个struct file结构, 上下2个不要搞混了哦 */
247     tfile = fget(fd);
248     if (!tfile)
249         goto error_fput;
250     /* The target file descriptor must support poll */
251     error = -EPERM;
252     /* 如果监听的文件不支持poll, 那就没辙了.
253      * 你知道什么情况下, 文件会不支持poll吗?
254      */
255     if (!tfile->f_op || !tfile->f_op->poll)
256         goto error_tgt_fput;
257     /*
258      * We have to check that the file structure underneath the file descriptor
259      * the user passed to us _is_ an eventpoll file. And also we do not permit
260      * adding an epoll file descriptor inside itself.
261      */
262     error = -EINVAL;
263     /* epoll不能自己监听自己... */
264     if (file == tfile || !is_file_epoll(file))
265         goto error_tgt_fput;
266     /*
267      * At this point it is safe to assume that the "private_data" contains
268      * our own data structure.
269      */
270     /* 取到我们的eventpoll结构, 来自与epoll_create1()中的分配 */
271     ep = file->private_data;
272     /* 接下来的操作有可能修改数据结构内容, 锁之~ */
273     mutex_lock(&ep->mtx);
274     /*
275      * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
276      * above, we can be sure to be able to use the item looked up by
277      * ep_find() till we release the mutex.
278      */
279     /* 对于每一个监听的fd, 内核都有分配一个epitem结构,
280      * 而且我们也知道, epoll是不允许重复添加fd的,
281      * 所以我们首先查找该fd是不是已经存在了.
282      * ep_find()其实就是RBTREE查找, 跟C++STL的map差不多一回事, O(lgn)的时间复杂度.
283      */
284     epi = ep_find(ep, tfile, fd);
285     error = -EINVAL;
286     switch (op) {
287         /* 首先我们关心添加 */
288     case EPOLL_CTL_ADD:
289         if (!epi) {
290             /* 之前的find没有找到有效的epitem, 证明是第一次插入, 接受!
291              * 这里我们可以知道, POLLERR和POLLHUP事件内核总是会关心的
292              * */
293             epds.events |= POLLERR | POLLHUP;
294             /* rbtree插入, 详情见ep_insert()的分析
295              * 其实我觉得这里有insert的话, 之前的find应该
296              * 是可以省掉的... */
297             error = ep_insert(ep, &epds, tfile, fd);
298         } else
299             /* 找到了!? 重复添加! */
300             error = -EEXIST;
301         break;
302         /* 删除和修改操作都比较简单 */
303     case EPOLL_CTL_DEL:
304         if (epi)
305             error = ep_remove(ep, epi);
306         else
307             error = -ENOENT;
308         break;
309     case EPOLL_CTL_MOD:
310         if (epi) {
311             epds.events |= POLLERR | POLLHUP;
312             error = ep_modify(ep, epi, &epds);
313         } else
314             error = -ENOENT;
315         break;
316     }
317     mutex_unlock(&ep->mtx);
318 error_tgt_fput:
319     fput(tfile);
320 error_fput:
321     fput(file);
322 error_return:
323     return error;
324 }
325 /* 分配一个eventpoll结构 */
326 static int ep_alloc(struct eventpoll **pep)
327 {
328     int error;
329     struct user_struct *user;
330     struct eventpoll *ep;
331     /* 获取当前用户的一些信息, 比如是不是root啦, 最大监听fd数目啦 */
332     user = get_current_user();
333     error = -ENOMEM;
334     ep = kzalloc(sizeof(*ep), GFP_KERNEL);
335     if (unlikely(!ep))
336         goto free_uid;
337     /* 这些都是初始化啦 */
338     spin_lock_init(&ep->lock);
339     mutex_init(&ep->mtx);
340     init_waitqueue_head(&ep->wq);//初始化自己睡在的等待队列
341     init_waitqueue_head(&ep->poll_wait);//初始化
342     INIT_LIST_HEAD(&ep->rdllist);//初始化就绪链表
343     ep->rbr = RB_ROOT;
344     ep->ovflist = EP_UNACTIVE_PTR;
345     ep->user = user;
346     *pep = ep;
347     return 0;
348 free_uid:
349     free_uid(user);
350     return error;
351 }
352 /*
353  * Must be called with "mtx" held.
354  */
355 /*
356  * ep_insert()在epoll_ctl()中被调用, 完成往epollfd里面添加一个监听fd的工作
357  * tfile是fd在内核态的struct file结构
358  */
359 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
360              struct file *tfile, int fd)
361 {
362     int error, revents, pwake = 0;
363     unsigned long flags;
364     struct epitem *epi;
365     struct ep_pqueue epq;
366     /* 查看是否达到当前用户的最大监听数 */
367     if (unlikely(atomic_read(&ep->user->epoll_watches) >=
368              max_user_watches))
369         return -ENOSPC;
370     /* 从著名的slab中分配一个epitem */
371     if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
372         return -ENOMEM;
373     /* Item initialization follow here ... */
374     /* 这些都是相关成员的初始化... */
375     INIT_LIST_HEAD(&epi->rdllink);
376     INIT_LIST_HEAD(&epi->fllink);
377     INIT_LIST_HEAD(&epi->pwqlist);
378     epi->ep = ep;
379     /* 这里保存了我们需要监听的文件fd和它的file结构 */
380     ep_set_ffd(&epi->ffd, tfile, fd);
381     epi->event = *event;
382     epi->nwait = 0;
383     /* 这个指针的初值不是NULL哦... */
384     epi->next = EP_UNACTIVE_PTR;
385     /* Initialize the poll table using the queue callback */
386     /* 好, 我们终于要进入到poll的正题了 */
387     epq.epi = epi;
388     /* 初始化一个poll_table
389      * 其实就是指定调用poll_wait(注意不是epoll_wait!!!)时的回调函数,和我们关心哪些events,
390      * ep_ptable_queue_proc()就是我们的回调啦, 初值是所有event都关心 */
391     init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
392     /*
393      * Attach the item to the poll hooks and get current event bits.
394      * We can safely use the file* here because its usage count has
395      * been increased by the caller of this function. Note that after
396      * this operation completes, the poll callback can start hitting
397      * the new item.
398      */
399     /* 这一部很关键, 也比较难懂, 完全是内核的poll机制导致的...
400      * 首先, f_op->poll()一般来说只是个wrapper, 它会调用真正的poll实现,
401      * 拿UDP的socket来举例, 这里就是这样的调用流程: f_op->poll(), sock_poll(),
402      * udp_poll(), datagram_poll(), sock_poll_wait(), 最后调用到我们上面指定的
403      * ep_ptable_queue_proc()这个回调函数...(好深的调用路径...).
404      * 完成这一步, 我们的epitem就跟这个socket关联起来了, 当它有状态变化时,
405      * 会通过ep_poll_callback()来通知.
406      * 最后, 这个函数还会查询当前的fd是不是已经有啥event已经ready了, 有的话
407      * 会将event返回. */
408     revents = tfile->f_op->poll(tfile, &epq.pt);
409     /*
410      * We have to check if something went wrong during the poll wait queue
411      * install process. Namely an allocation for a wait queue failed due
412      * high memory pressure.
413      */
414     error = -ENOMEM;
415     if (epi->nwait < 0)
416         goto error_unregister;
417     /* Add the current item to the list of active epoll hook for this file */
418     /* 这个就是每个文件会将所有监听自己的epitem链起来 */
419     spin_lock(&tfile->f_lock);
420     list_add_tail(&epi->fllink, &tfile->f_ep_links);
421     spin_unlock(&tfile->f_lock);
422     /*
423      * Add the current item to the RB tree. All RB tree operations are
424      * protected by "mtx", and ep_insert() is called with "mtx" held.
425      */
426     /* 都搞定后, 将epitem插入到对应的eventpoll中去 */
427     ep_rbtree_insert(ep, epi);
428     /* We have to drop the new item inside our item list to keep track of it */
429     spin_lock_irqsave(&ep->lock, flags);
430     /* If the file is already "ready" we drop it inside the ready list */
431     /* 到达这里后, 如果我们监听的fd已经有事件发生, 那就要处理一下 */
432     if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
433         /* 将当前的epitem加入到ready list中去 */
434         list_add_tail(&epi->rdllink, &ep->rdllist);
435         /* Notify waiting tasks that events are available */
436         /* 谁在epoll_wait, 就唤醒它... */
437         if (waitqueue_active(&ep->wq))
438             wake_up_locked(&ep->wq);
439         /* 谁在epoll当前的epollfd, 也唤醒它... */
440         if (waitqueue_active(&ep->poll_wait))
441             pwake++;
442     }
443     spin_unlock_irqrestore(&ep->lock, flags);
444     atomic_inc(&ep->user->epoll_watches);
445     /* We have to call this outside the lock */
446     if (pwake)
447         ep_poll_safewake(&ep->poll_wait);
448     return 0;
449 error_unregister:
450     ep_unregister_pollwait(ep, epi);
451     /*
452      * We need to do this because an event could have been arrived on some
453      * allocated wait queue. Note that we don't care about the ep->ovflist
454      * list, since that is used/cleaned only inside a section bound by "mtx".
455      * And ep_insert() is called with "mtx" held.
456      */
457     spin_lock_irqsave(&ep->lock, flags);
458     if (ep_is_linked(&epi->rdllink))
459         list_del_init(&epi->rdllink);
460     spin_unlock_irqrestore(&ep->lock, flags);
461     kmem_cache_free(epi_cache, epi);
462     return error;
463 }
464 /*
465  * This is the callback that is used to add our wait queue to the
466  * target file wakeup lists.
467  */
468 /*
469  * 该函数在调用f_op->poll()时会被调用.
470  * 也就是epoll主动poll某个fd时, 用来将epitem与指定的fd关联起来的.
471  * 关联的办法就是使用等待队列(waitqueue)
472  */
473 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
474                  poll_table *pt)
475 {
476     struct epitem *epi = ep_item_from_epqueue(pt);
477     struct eppoll_entry *pwq;
478     if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
479         /* 初始化等待队列, 指定ep_poll_callback为唤醒时的回调函数,
480          * 当我们监听的fd发生状态改变时, 也就是队列头被唤醒时,
481          * 指定的回调函数将会被调用. */
482         init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
483         pwq->whead = whead;
484         pwq->base = epi;
485         /* 将刚分配的等待队列成员加入到头中, 头是由fd持有的 */
486         add_wait_queue(whead, &pwq->wait);
487         list_add_tail(&pwq->llink, &epi->pwqlist);
488         /* nwait记录了当前epitem加入到了多少个等待队列中,
489          * 我认为这个值最大也只会是1... */
490         epi->nwait++;
491     } else {
492         /* We have to signal that an error occurred */
493         epi->nwait = -1;
494     }
495 }
496 /*
497  * This is the callback that is passed to the wait queue wakeup
498  * machanism. It is called by the stored file descriptors when they
499  * have events to report.
500  */
501 /*
502  * 这个是关键性的回调函数, 当我们监听的fd发生状态改变时, 它会被调用.
503  * 参数key被当作一个unsigned long整数使用, 携带的是events.
504  */
505 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
506 {
507     int pwake = 0;
508     unsigned long flags;
509     struct epitem *epi = ep_item_from_wait(wait);//从等待队列获取epitem.需要知道哪个进程挂载到这个设备
510     struct eventpoll *ep = epi->ep;//获取
511     spin_lock_irqsave(&ep->lock, flags);
512     /*
513      * If the event mask does not contain any poll(2) event, we consider the
514      * descriptor to be disabled. This condition is likely the effect of the
515      * EPOLLONESHOT bit that disables the descriptor when an event is received,
516      * until the next EPOLL_CTL_MOD will be issued.
517      */
518     if (!(epi->event.events & ~EP_PRIVATE_BITS))
519         goto out_unlock;
520     /*
521      * Check the events coming with the callback. At this stage, not
522      * every device reports the events in the "key" parameter of the
523      * callback. We need to be able to handle both cases here, hence the
524      * test for "key" != NULL before the event match test.
525      */
526     /* 没有我们关心的event... */
527     if (key && !((unsigned long) key & epi->event.events))
528         goto out_unlock;
529     /*
530      * If we are trasfering events to userspace, we can hold no locks
531      * (because we're accessing user memory, and because of linux f_op->poll()
532      * semantics). All the events that happens during that period of time are
533      * chained in ep->ovflist and requeued later on.
534      */
535     /*
536      * 这里看起来可能有点费解, 其实干的事情比较简单:
537      * 如果该callback被调用的同时, epoll_wait()已经返回了,
538      * 也就是说, 此刻应用程序有可能已经在循环获取events,
539      * 这种情况下, 内核将此刻发生event的epitem用一个单独的链表
540      * 链起来, 不发给应用程序, 也不丢弃, 而是在下一次epoll_wait
541      * 时返回给用户.
542      */
543     if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
544         if (epi->next == EP_UNACTIVE_PTR) {
545             epi->next = ep->ovflist;
546             ep->ovflist = epi;
547         }
548         goto out_unlock;
549     }
550     /* If this file is already in the ready list we exit soon */
551     /* 将当前的epitem放入ready list */
552     if (!ep_is_linked(&epi->rdllink))
553         list_add_tail(&epi->rdllink, &ep->rdllist);
554     /*
555      * Wake up ( if active ) both the eventpoll wait list and the ->poll()
556      * wait list.
557      */
558     /* 唤醒epoll_wait... */
559     if (waitqueue_active(&ep->wq))
560         wake_up_locked(&ep->wq);
561     /* 如果epollfd也在被poll, 那就唤醒队列里面的所有成员. */
562     if (waitqueue_active(&ep->poll_wait))
563         pwake++;
564 out_unlock:
565     spin_unlock_irqrestore(&ep->lock, flags);
566     /* We have to call this outside the lock */
567     if (pwake)
568         ep_poll_safewake(&ep->poll_wait);
569     return 1;
570 }
571 /*
572  * Implement the event wait interface for the eventpoll file. It is the kernel
573  * part of the user space epoll_wait(2).
574  */
575 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
576         int, maxevents, int, timeout)
577 {
578     int error;
579     struct file *file;
580     struct eventpoll *ep;
581     /* The maximum number of event must be greater than zero */
582     if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
583         return -EINVAL;
584     /* Verify that the area passed by the user is writeable */
585     /* 这个地方有必要说明一下:
586      * 内核对应用程序采取的策略是"绝对不信任",
587      * 所以内核跟应用程序之间的数据交互大都是copy, 不允许(也时候也是不能...)指针引用.
588      * epoll_wait()需要内核返回数据给用户空间, 内存由用户程序提供,
589      * 所以内核会用一些手段来验证这一段内存空间是不是有效的.
590      */
591     if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
592         error = -EFAULT;
593         goto error_return;
594     }
595     /* Get the "struct file *" for the eventpoll file */
596     error = -EBADF;
597     /* 获取epollfd的struct file, epollfd也是文件嘛 */
598     file = fget(epfd);
599     if (!file)
600         goto error_return;
601     /*
602      * We have to check that the file structure underneath the fd
603      * the user passed to us _is_ an eventpoll file.
604      */
605     error = -EINVAL;
606     /* 检查一下它是不是一个真正的epollfd... */
607     if (!is_file_epoll(file))
608         goto error_fput;
609     /*
610      * At this point it is safe to assume that the "private_data" contains
611      * our own data structure.
612      */
613     /* 获取eventpoll结构 */
614     ep = file->private_data;
615     /* Time to fish for events ... */
616     /* OK, 睡觉, 等待事件到来~~ */
617     error = ep_poll(ep, events, maxevents, timeout);
618 error_fput:
619     fput(file);
620 error_return:
621     return error;
622 }
623 /* 这个函数真正将执行epoll_wait的进程带入睡眠状态... */
624 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
625            int maxevents, long timeout)
626 {
627     int res, eavail;
628     unsigned long flags;
629     long jtimeout;
630     wait_queue_t wait;//等待队列
631     /*
632      * Calculate the timeout by checking for the "infinite" value (-1)
633      * and the overflow condition. The passed timeout is in milliseconds,
634      * that why (t * HZ) / 1000.
635      */
636     /* 计算睡觉时间, 毫秒要转换为HZ */
637     jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
638         MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
639 retry:
640     spin_lock_irqsave(&ep->lock, flags);
641     res = 0;
642     /* 如果ready list不为空, 就不睡了, 直接干活... */
643     if (list_empty(&ep->rdllist)) {
644         /*
645          * We don't have any available event to return to the caller.
646          * We need to sleep here, and we will be wake up by
647          * ep_poll_callback() when events will become available.
648          */
649         /* OK, 初始化一个等待队列, 准备直接把自己挂起,
650          * 注意current是一个宏, 代表当前进程 */
651         init_waitqueue_entry(&wait, current);//初始化等待队列,wait表示当前进程
652         __add_wait_queue_exclusive(&ep->wq, &wait);//挂载到ep结构的等待队列
653         for (;;) {
654             /*
655              * We don't want to sleep if the ep_poll_callback() sends us
656              * a wakeup in between. That's why we set the task state
657              * to TASK_INTERRUPTIBLE before doing the checks.
658              */
659             /* 将当前进程设置位睡眠, 但是可以被信号唤醒的状态,
660              * 注意这个设置是"将来时", 我们此刻还没睡! */
661             set_current_state(TASK_INTERRUPTIBLE);
662             /* 如果这个时候, ready list里面有成员了,
663              * 或者睡眠时间已经过了, 就直接不睡了... */
664             if (!list_empty(&ep->rdllist) || !jtimeout)
665                 break;
666             /* 如果有信号产生, 也起床... */
667             if (signal_pending(current)) {
668                 res = -EINTR;
669                 break;
670             }
671             /* 啥事都没有,解锁, 睡觉... */
672             spin_unlock_irqrestore(&ep->lock, flags);
673             /* jtimeout这个时间后, 会被唤醒,
674              * ep_poll_callback()如果此时被调用,
675              * 那么我们就会直接被唤醒, 不用等时间了...
676              * 再次强调一下ep_poll_callback()的调用时机是由被监听的fd
677              * 的具体实现, 比如socket或者某个设备驱动来决定的,
678              * 因为等待队列头是他们持有的, epoll和当前进程
679              * 只是单纯的等待...
680              **/
681             jtimeout = schedule_timeout(jtimeout);//睡觉
682             spin_lock_irqsave(&ep->lock, flags);
683         }
684         __remove_wait_queue(&ep->wq, &wait);
685         /* OK 我们醒来了... */
686         set_current_state(TASK_RUNNING);
687     }
688     /* Is it worth to try to dig for events ? */
689     eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
690     spin_unlock_irqrestore(&ep->lock, flags);
691     /*
692      * Try to transfer events to user space. In case we get 0 events and
693      * there's still timeout left over, we go trying again in search of
694      * more luck.
695      */
696     /* 如果一切正常, 有event发生, 就开始准备数据copy给用户空间了... */
697     if (!res && eavail &&
698         !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
699         goto retry;
700     return res;
701 }
702 /* 这个简单, 我们直奔下一个... */
703 static int ep_send_events(struct eventpoll *ep,
704               struct epoll_event __user *events, int maxevents)
705 {
706     struct ep_send_events_data esed;
707     esed.maxevents = maxevents;
708     esed.events = events;
709     return ep_scan_ready_list(ep, ep_send_events_proc, &esed);
710 }
711 /**
712  * ep_scan_ready_list - Scans the ready list in a way that makes possible for
713  *                      the scan code, to call f_op->poll(). Also allows for
714  *                      O(NumReady) performance.
715  *
716  * @ep: Pointer to the epoll private data structure.
717  * @sproc: Pointer to the scan callback.
718  * @priv: Private opaque data passed to the @sproc callback.
719  *
720  * Returns: The same integer error code returned by the @sproc callback.
721  */
722 static int ep_scan_ready_list(struct eventpoll *ep,
723                   int (*sproc)(struct eventpoll *,
724                        struct list_head *, void *),
725                   void *priv)
726 {
727     int error, pwake = 0;
728     unsigned long flags;
729     struct epitem *epi, *nepi;
730     LIST_HEAD(txlist);
731     /*
732      * We need to lock this because we could be hit by
733      * eventpoll_release_file() and epoll_ctl().
734      */
735     mutex_lock(&ep->mtx);
736     /*
737      * Steal the ready list, and re-init the original one to the
738      * empty list. Also, set ep->ovflist to NULL so that events
739      * happening while looping w/out locks, are not lost. We cannot
740      * have the poll callback to queue directly on ep->rdllist,
741      * because we want the "sproc" callback to be able to do it
742      * in a lockless way.
743      */
744     spin_lock_irqsave(&ep->lock, flags);
745     /* 这一步要注意, 首先, 所有监听到events的epitem都链到rdllist上了,
746      * 但是这一步之后, 所有的epitem都转移到了txlist上, 而rdllist被清空了,
747      * 要注意哦, rdllist已经被清空了! */
748     list_splice_init(&ep->rdllist, &txlist);
749     /* ovflist, 在ep_poll_callback()里面我解释过, 此时此刻我们不希望
750      * 有新的event加入到ready list中了, 保存后下次再处理... */
751     ep->ovflist = NULL;
752     spin_unlock_irqrestore(&ep->lock, flags);
753     /*
754      * Now call the callback function.
755      */
756     /* 在这个回调函数里面处理每个epitem
757      * sproc 就是 ep_send_events_proc, 下面会注释到. */
758     error = (*sproc)(ep, &txlist, priv);
759     spin_lock_irqsave(&ep->lock, flags);
760     /*
761      * During the time we spent inside the "sproc" callback, some
762      * other events might have been queued by the poll callback.
763      * We re-insert them inside the main ready-list here.
764      */
765     /* 现在我们来处理ovflist, 这些epitem都是我们在传递数据给用户空间时
766      * 监听到了事件. */
767     for (nepi = ep->ovflist; (epi = nepi) != NULL;
768          nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
769         /*
770          * We need to check if the item is already in the list.
771          * During the "sproc" callback execution time, items are
772          * queued into ->ovflist but the "txlist" might already
773          * contain them, and the list_splice() below takes care of them.
774          */
775         /* 将这些直接放入readylist */
776         if (!ep_is_linked(&epi->rdllink))
777             list_add_tail(&epi->rdllink, &ep->rdllist);
778     }
779     /*
780      * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
781      * releasing the lock, events will be queued in the normal way inside
782      * ep->rdllist.
783      */
784     ep->ovflist = EP_UNACTIVE_PTR;
785     /*
786      * Quickly re-inject items left on "txlist".
787      */
788     /* 上一次没有处理完的epitem, 重新插入到ready list */
789     list_splice(&txlist, &ep->rdllist);
790     /* ready list不为空, 直接唤醒... */
791     if (!list_empty(&ep->rdllist)) {
792         /*
793          * Wake up (if active) both the eventpoll wait list and
794          * the ->poll() wait list (delayed after we release the lock).
795          */
796         if (waitqueue_active(&ep->wq))
797             wake_up_locked(&ep->wq);
798         if (waitqueue_active(&ep->poll_wait))
799             pwake++;
800     }
801     spin_unlock_irqrestore(&ep->lock, flags);
802     mutex_unlock(&ep->mtx);
803     /* We have to call this outside the lock */
804     if (pwake)
805         ep_poll_safewake(&ep->poll_wait);
806     return error;
807 }
808 /* 该函数作为callbakc在ep_scan_ready_list()中被调用
809  * head是一个链表, 包含了已经ready的epitem,
810  * 这个不是eventpoll里面的ready list, 而是上面函数中的txlist.
811  */
812 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
813                    void *priv)
814 {
815     struct ep_send_events_data *esed = priv;
816     int eventcnt;
817     unsigned int revents;
818     struct epitem *epi;
819     struct epoll_event __user *uevent;
820     /*
821      * We can loop without lock because we are passed a task private list.
822      * Items cannot vanish during the loop because ep_scan_ready_list() is
823      * holding "mtx" during this call.
824      */
825     /* 扫描整个链表... */
826     for (eventcnt = 0, uevent = esed->events;
827          !list_empty(head) && eventcnt < esed->maxevents;) {
828         /* 取出第一个成员 */
829         epi = list_first_entry(head, struct epitem, rdllink);
830         /* 然后从链表里面移除 */
831         list_del_init(&epi->rdllink);
832         /* 读取events,
833          * 注意events我们ep_poll_callback()里面已经取过一次了, 为啥还要再取?
834          * 1. 我们当然希望能拿到此刻的最新数据, events是会变的~
835          * 2. 不是所有的poll实现, 都通过等待队列传递了events, 有可能某些驱动压根没传
836          * 必须主动去读取. */
837         revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
838             epi->event.events;
839         if (revents) {
840             /* 将当前的事件和用户传入的数据都copy给用户空间,
841              * 就是epoll_wait()后应用程序能读到的那一堆数据. */
842             if (__put_user(revents, &uevent->events) ||
843                 __put_user(epi->event.data, &uevent->data)) {
844                 list_add(&epi->rdllink, head);
845                 return eventcnt ? eventcnt : -EFAULT;
846             }
847             eventcnt++;
848             uevent++;
849             if (epi->event.events & EPOLLONESHOT)
850                 epi->event.events &= EP_PRIVATE_BITS;
851             else if (!(epi->event.events & EPOLLET)) {
852                 /* 嘿嘿, EPOLLET和非ET的区别就在这一步之差呀~
853                  * 如果是ET, epitem是不会再进入到readly list,
854                  * 除非fd再次发生了状态改变, ep_poll_callback被调用.
855                  * 如果是非ET, 不管你还有没有有效的事件或者数据,
856                  * 都会被重新插入到ready list, 再下一次epoll_wait
857                  * 时, 会立即返回, 并通知给用户空间. 当然如果这个
858                  * 被监听的fds确实没事件也没数据了, epoll_wait会返回一个0,
859                  * 空转一次.
860                  */
861                 list_add_tail(&epi->rdllink, &ep->rdllist);
862             }
863         }
864     }
865     return eventcnt;
866 }
867 /* ep_free在epollfd被close时调用,
868  * 释放一些资源而已, 比较简单 */
869 static void ep_free(struct eventpoll *ep)
870 {
871     struct rb_node *rbp;
872     struct epitem *epi;
873     /* We need to release all tasks waiting for these file */
874     if (waitqueue_active(&ep->poll_wait))
875         ep_poll_safewake(&ep->poll_wait);
876     /*
877      * We need to lock this because we could be hit by
878      * eventpoll_release_file() while we're freeing the "struct eventpoll".
879      * We do not need to hold "ep->mtx" here because the epoll file
880      * is on the way to be removed and no one has references to it
881      * anymore. The only hit might come from eventpoll_release_file() but
882      * holding "epmutex" is sufficent here.
883      */
884     mutex_lock(&epmutex);
885     /*
886      * Walks through the whole tree by unregistering poll callbacks.
887      */
888     for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
889         epi = rb_entry(rbp, struct epitem, rbn);
890         ep_unregister_pollwait(ep, epi);
891     }
892     /*
893      * Walks through the whole tree by freeing each "struct epitem". At this
894      * point we are sure no poll callbacks will be lingering around, and also by
895      * holding "epmutex" we can be sure that no file cleanup code will hit
896      * us during this operation. So we can avoid the lock on "ep->lock".
897      */
898     /* 之所以在关闭epollfd之前不需要调用epoll_ctl移除已经添加的fd,
899      * 是因为这里已经做了... */
900     while ((rbp = rb_first(&ep->rbr)) != NULL) {
901         epi = rb_entry(rbp, struct epitem, rbn);
902         ep_remove(ep, epi);
903     }
904     mutex_unlock(&epmutex);
905     mutex_destroy(&ep->mtx);
906     free_uid(ep->user);
907     kfree(ep);
908 }
909 /* File callbacks that implement the eventpoll file behaviour */
910 static const struct file_operations eventpoll_fops = {
911     .release    = ep_eventpoll_release,
912     .poll       = ep_eventpoll_poll
913 };
914 /* Fast test to see if the file is an evenpoll file */
915 static inline int is_file_epoll(struct file *f)
916 {
917     return f->f_op == &eventpoll_fops;
918 }
919 /* OK, eventpoll我认为比较重要的函数都注释完了... */

 

epoll_create
从slab缓存中创建一个eventpoll对象,并且创建一个匿名的fd跟fd对应的file对象,
而eventpoll对象保存在struct file结构的private指针中,并且返回,
该fd对应的file operations只是实现了poll跟release操作

创建eventpoll对象的初始化操作
获取当前用户信息,是不是root,最大监听fd数目等并且保存到eventpoll对象中
初始化等待队列,初始化就绪链表,初始化红黑树的头结点

epoll_ctl操作
将epoll_event结构拷贝到内核空间中
并且判断加入的fd是否支持poll结构(epoll,poll,selectI/O多路复用必须支持poll操作).
并且从epfd->file->privatedata获取event_poll对象,根据op区分是添加删除还是修改,
首先在eventpoll结构中的红黑树查找是否已经存在了相对应的fd,没找到就支持插入操作,否则报重复的错误.
相对应的修改,删除比较简单就不啰嗦了

插入操作时,会创建一个与fd对应的epitem结构,并且初始化相关成员,比如保存监听的fd跟file结构之类的
重要的是指定了调用poll_wait时的回调函数用于数据就绪时唤醒进程,(其内部,初始化设备的等待队列,将该进程注册到等待队列)完成这一步, 我们的epitem就跟这个socket关联起来了, 当它有状态变化时,
会通过ep_poll_callback()来通知.
最后调用加入的fd的file operation->poll函数(最后会调用poll_wait操作)用于完成注册操作.
最后将epitem结构添加到红黑树中

epoll_wait操作
计算睡眠时间(如果有),判断eventpoll对象的链表是否为空,不为空那就干活不睡明.并且初始化一个等待队列,把自己挂上去,设置自己的进程状态
为可睡眠状态.判断是否有信号到来(有的话直接被中断醒来,),如果啥事都没有那就调用schedule_timeout进行睡眠,如果超时或者被唤醒,首先从自己初始化的等待队列删除
,然后开始拷贝资源给用户空间了
拷贝资源则是先把就绪事件链表转移到中间链表,然后挨个遍历拷贝到用户空间,
并且挨个判断其是否为水平触发,是的话再次插入到就绪链表

posted on 2018-01-12 09:27 wsw_seu 阅读(...) 评论(...) 编辑 收藏

导航