linux 系统调用epoll_create epoll_ctl epoll_wait

epoll是Linux内核为处理大批量句柄而作了改进的poll，是Linux下多路复用IO接口select/poll的增强版本，它能显著减少程序在大量并发连接中只有少量活跃的情况下的系统CPU利用率。
epoll主要涉及epoll_create epoll_ctl epoll_wait三个基本系统调用，及增加功能的epoll_create1,epoll_pwait

 

epoll相对poll/select比较：
1.poll/select每次调用都要将文件描述符及事件由用户空间复制到内核空间；epoll只需要在epoll_ctl中一次复制并保存在内核中
2.poll/select调用被阻塞时都会添加waiter到文件描述符表中所有文件的等待队列中，poll/select退出时再从等待队列中删除；epoll只需在epoll_ctl中将waiter添加到文件的等待队列中，直到显示被删除(close,epoll_ctl)
3.poll/select每次调用都会轮询文件描述符表中所有文件是否有事件发生；而epoll_wait只是从readylist链表中的取文件描述符及发生事件(有事件发生时内核调用epoll_ctl注册的回调函数将文件描述符及事件添加到readylist中)，没有时阻塞epoll_wait
由以上3点可以看出epoll在大量并发连接且只有少量活跃时相对poll/select有相当大的性能优势

 

下面讨论epoll的实现，内核版本2.6.32.60

I.数据结构
i.eventpoll
epoll接口的主要数据结构

/* fs/eventpoll.c */
 165 /*
 166  * This structure is stored inside the "private_data" member of the file
 167  * structure and rapresent the main data sructure for the eventpoll
 168  * interface.
 169  */
 170 struct eventpoll {
 171         /* Protect the this structure access */
 172         spinlock_t lock;
 173 
 174         /*
 175          * This mutex is used to ensure that files are not removed
 176          * while epoll is using them. This is held during the event
 177          * collection loop, the file cleanup path, the epoll file exit
 178          * code and the ctl operations.
 179          */
 180         struct mutex mtx;
 181 
 182         /* Wait queue used by sys_epoll_wait() */
 183         wait_queue_head_t wq;
 184 
 185         /* Wait queue used by file->poll() */
 186         wait_queue_head_t poll_wait;
 187 
 188         /* List of ready file descriptors */
 189         struct list_head rdllist;
 190 
 191         /* RB tree root used to store monitored fd structs */
 192         struct rb_root rbr;
 193 
 194         /*
 195          * This is a single linked list that chains all the "struct epitem" that
 196          * happened while transfering ready events to userspace w/out
 197          * holding ->lock.
 198          */
 199         struct epitem *ovflist;
 200 
 201         /* The user that created the eventpoll descriptor */
 202         struct user_struct *user;
 203 
 204         struct file *file;
 205 
 206         /* used to optimize loop detection check */
 207         int visited;
 208         struct list_head visited_list_link;
 209 };

wq：用于记录epoll_wait系统调用的waiter
poll_wait：用于记录文件poll操作的waiter(由epoll_create创建的文件)
rdllist：已经ready的文件描述符链表
rbr：红黑树树根，用于文件描述符及事件的存储及快速查找
ovflist：文件描述符及事件由内核空间复制到用户空间时，期间出现的ready文件描述符及事件记录在该链表

 

ii.epitem
每个添加到eventpoll接口的文件都有对应的epitem，记录相应的事件信息等

 129 /*
 130  * Each file descriptor added to the eventpoll interface will
 131  * have an entry of this type linked to the "rbr" RB tree.
 132  */
 133 struct epitem {
 134         /* RB tree node used to link this structure to the eventpoll RB tree */
 135         struct rb_node rbn;
 136 
 137         /* List header used to link this structure to the eventpoll ready list */
 138         struct list_head rdllink;
 139 
 140         /*
 141          * Works together "struct eventpoll"->ovflist in keeping the
 142          * single linked chain of items.
 143          */
 144         struct epitem *next;
 145 
 146         /* The file descriptor information this item refers to */
 147         struct epoll_filefd ffd;
 148 
 149         /* Number of active wait queue attached to poll operations */
 150         int nwait;
 151 
 152         /* List containing poll wait queues */
 153         struct list_head pwqlist;
 154 
 155         /* The "container" of this item */
 156         struct eventpoll *ep;
 157 
 158         /* List header used to link this item to the "struct file" items list */
 159         struct list_head fllink;
 160 
 161         /* The structure that describe the interested events and the source fd */
 162         struct epoll_event event;
 163 };

rbn：将epitem添加到eventpoll红黑树中
rdllink：将epitem添加到eventpoll就绪链表中
next：将epitem添加到ovflist链表中
ffd：表示添加到eventpoll的文件
nwait：wait queue大小
pwqlist：wait queue链表
ep：epitem所属的eventpoll

 

iii.eppoll_entry

 211 /* Wait structure used by the poll hooks */
 212 struct eppoll_entry {
 213         /* List header used to link this structure to the "struct epitem" */
 214         struct list_head llink;
 215 
 216         /* The "base" pointer is set to the container "struct epitem" */
 217         struct epitem *base;
 218 
 219         /*
 220          * Wait queue item that will be linked to the target file wait
 221          * queue head.
 222          */
 223         wait_queue_t wait;
 224 
 225         /* The wait queue head that linked the "wait" wait queue item */
 226         wait_queue_head_t *whead;
 227 };

llink：链入epitem的pwqlist
base：指向epitem
wait：链入目标文件的等待队列
whead：目标文件的等待等待队列

 

iv.数据结构关系图

II.epoll_create

1456 /*
1457  * Open an eventpoll file descriptor.
1458  */
1459 SYSCALL_DEFINE1(epoll_create1, int, flags)
1460 {
1461         int error, fd;
1462         struct eventpoll *ep = NULL;
1463         struct file *file;
1464 
1465         /* Check the EPOLL_* constant for consistency.  */
1466         BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1467 
1468         if (flags & ~EPOLL_CLOEXEC)
1469                 return -EINVAL;
1470         /*
1471          * Create the internal data structure ("struct eventpoll").
1472          */
1473         error = ep_alloc(&ep);
1474         if (error < 0)
1475                 return error;
1476         /*
1477          * Creates all the items needed to setup an eventpoll file. That is,
1478          * a file structure and a free file descriptor.
1479          */
1480         fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1481         if (fd < 0) {
1482                 error = fd;
1483                 goto out_free_ep;
1484         }
1485         file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1486                                  O_RDWR | (flags & O_CLOEXEC));
1487         if (IS_ERR(file)) {
1488                 error = PTR_ERR(file);
1489                 goto out_free_fd;
1490         }
1491         fd_install(fd, file);
1492         ep->file = file;
1493         return fd;
1494 
1495 out_free_fd:
1496         put_unused_fd(fd);
1497 out_free_ep:
1498         ep_free(ep);
1499         return error;
1500 }
1501 
1502 SYSCALL_DEFINE1(epoll_create, int, size)
1503 {
1504         if (size <= 0)
1505                 return -EINVAL;
1506 
1507         return sys_epoll_create1(0);
1508 }

1.参数检查，size/flags
2.分配eventpoll并初始化
3.分配eventpoll文件描述符
4.创建文件对象，操作是eventpoll_fops
5.文件对象与文件描述符关联
6.返回eventpoll文件描述符
由以上可以看出epoll_create中的size在该版本中没有用

 

i.ep_eventpoll_poll

ep_eventpoll_poll是eventpoll文件的poll操作，主要用于将wait添加到文件的等待队列中并返回POLL事件

 686 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
 687                                void *priv)
 688 {
 689         struct epitem *epi, *tmp;
 690 
 691         list_for_each_entry_safe(epi, tmp, head, rdllink) {
 692                 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
 693                     epi->event.events)
 694                         return POLLIN | POLLRDNORM;
 695                 else {
 696                         /*
 697                          * Item has been dropped into the ready list by the poll
 698                          * callback, but it's not actually ready, as far as
 699                          * caller requested events goes. We can remove it here.
 700                          */
 701                         list_del_init(&epi->rdllink);
 702                 }
 703         }
 704 
 705         return 0;
 706 }
 707 
 708 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
 709 {
 710         return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
 711 }
 712 
 713 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
 714 {
 715         int pollflags;
 716         struct eventpoll *ep = file->private_data;
 717 
 718         /* Insert inside our poll wait queue */
 719         poll_wait(file, &ep->poll_wait, wait);
 720 
 721         /*
 722          * Proceed to find out if wanted events are really available inside
 723          * the ready list. This need to be done under ep_call_nested()
 724          * supervision, since the call to f_op->poll() done on listed files
 725          * could re-enter here.
 726          */
 727         pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
 728                                    ep_poll_readyevents_proc, ep, ep, current);
 729 
 730         return pollflags != -1 ? pollflags : 0;
 731 }

1.将wait添加到eventpoll结构poll_wait域表示的等待队列中；用于rdllist中有描述符时调用wait的回调者
2.根据eventpoll就绪链表中的文件是否满足POLL事件，返回eventpoll的自己的POLL事件；即如果注册到eventpoll接口中的文件满足POLL事件，则说明eventpoll自己是可以做POLLIN|POLLRDNORM相应的操作

 

ii.ep_eventpoll_release

 631 static void ep_free(struct eventpoll *ep)
 632 {
 633         struct rb_node *rbp;
 634         struct epitem *epi;
 635 
 636         /* We need to release all tasks waiting for these file */
 637         if (waitqueue_active(&ep->poll_wait))
 638                 ep_poll_safewake(&ep->poll_wait);
 639 
 640         /*
 641          * We need to lock this because we could be hit by
 642          * eventpoll_release_file() while we're freeing the "struct eventpoll".
 643          * We do not need to hold "ep->mtx" here because the epoll file
 644          * is on the way to be removed and no one has references to it
 645          * anymore. The only hit might come from eventpoll_release_file() but
 646          * holding "epmutex" is sufficent here.
 647          */
 648         mutex_lock(&epmutex);
 649 
 650         /*
 651          * Walks through the whole tree by unregistering poll callbacks.
 652          */
 653         for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
 654                 epi = rb_entry(rbp, struct epitem, rbn);
 655 
 656                 ep_unregister_pollwait(ep, epi);
 657         }
 658 
 659         /*
 660          * Walks through the whole tree by freeing each "struct epitem". At this
 661          * point we are sure no poll callbacks will be lingering around, and also by
 662          * holding "epmutex" we can be sure that no file cleanup code will hit
 663          * us during this operation. So we can avoid the lock on "ep->lock".
 664          */
 665         while ((rbp = rb_first(&ep->rbr)) != NULL) {
 666                 epi = rb_entry(rbp, struct epitem, rbn);
 667                 ep_remove(ep, epi);
 668         }
 669 
 670         mutex_unlock(&epmutex);
 671         mutex_destroy(&ep->mtx);
 672         free_uid(ep->user);
 673         kfree(ep);
 674 }
 675 
 676 static int ep_eventpoll_release(struct inode *inode, struct file *file)
 677 {
 678         struct eventpoll *ep = file->private_data;
 679 
 680         if (ep)
 681                 ep_free(ep);
 682 
 683         return 0;
 684 }

关闭eventpoll文件描述符时VFS会触发ep_eventpoll_release，主要用于释放eventpoll相应的资源
1.唤醒poll_wait中的等待者
2.将注册到目标文件的poll回调函数，从目标文件的等待者队列删除
3.释放epitem
4.释放mutex锁
5.释放user
6.回收eventpoll结构所占用的内存

 

 

III.epoll_ctl

1510 /*
1511  * The following function implements the controller interface for
1512  * the eventpoll file that enables the insertion/removal/change of
1513  * file descriptors inside the interest set.
1514  */
1515 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1516                 struct epoll_event __user *, event)
1517 {
1518         int error;
1519         int did_lock_epmutex = 0;
1520         struct file *file, *tfile;
1521         struct eventpoll *ep;
1522         struct epitem *epi;
1523         struct epoll_event epds;
1524 
1525         error = -EFAULT;
1526         if (ep_op_has_event(op) &&
1527             copy_from_user(&epds, event, sizeof(struct epoll_event)))
1528                 goto error_return;
1529 
1530         /* Get the "struct file *" for the eventpoll file */
1531         error = -EBADF;
1532         file = fget(epfd);
1533         if (!file)
1534                 goto error_return;
1535 
1536         /* Get the "struct file *" for the target file */
1537         tfile = fget(fd);
1538         if (!tfile)
1539                 goto error_fput;
1540 
1541         /* The target file descriptor must support poll */
1542         error = -EPERM;
1543         if (!tfile->f_op || !tfile->f_op->poll)
1544                 goto error_tgt_fput;
1545 
1546         /*
1547          * We have to check that the file structure underneath the file descriptor
1548          * the user passed to us _is_ an eventpoll file. And also we do not permit
1549          * adding an epoll file descriptor inside itself.
1550          */
1551         error = -EINVAL;
1552         if (file == tfile || !is_file_epoll(file))
1553                 goto error_tgt_fput;
1554 
1555         /*
1556          * At this point it is safe to assume that the "private_data" contains
1557          * our own data structure.
1558          */
1559         ep = file->private_data;
1560 
1561         /*
1562          * When we insert an epoll file descriptor, inside another epoll file
1563          * descriptor, there is the change of creating closed loops, which are
1564          * better be handled here, than in more critical paths. While we are
1565          * checking for loops we also determine the list of files reachable
1566          * and hang them on the tfile_check_list, so we can check that we
1567          * haven't created too many possible wakeup paths.
1568          *
1569          * We need to hold the epmutex across both ep_insert and ep_remove
1570          * b/c we want to make sure we are looking at a coherent view of
1571          * epoll network.
1572          */
1573         if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
1574                 mutex_lock(&epmutex);
1575                 did_lock_epmutex = 1;
1576         }
1577         if (op == EPOLL_CTL_ADD) {
1578                 if (is_file_epoll(tfile)) {
1579                         error = -ELOOP;
1580                         if (ep_loop_check(ep, tfile) != 0) {
1581                                 clear_tfile_check_list();
1582                                 goto error_tgt_fput;
1583                         }
1584                 } else
1585                         list_add(&tfile->f_tfile_llink, &tfile_check_list);
1586         }
1587 
1588         mutex_lock_nested(&ep->mtx, 0);
1589 
1590         /*
1591          * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1592          * above, we can be sure to be able to use the item looked up by
1593          * ep_find() till we release the mutex.
1594          */
1595         epi = ep_find(ep, tfile, fd);
1596 
1597         error = -EINVAL;
1598         switch (op) {
1599         case EPOLL_CTL_ADD:
1600                 if (!epi) {
1601                         epds.events |= POLLERR | POLLHUP;
1602                         error = ep_insert(ep, &epds, tfile, fd);
1603                 } else
1604                         error = -EEXIST;
1605                 clear_tfile_check_list();
1606                 break;
1607         case EPOLL_CTL_DEL:
1608                 if (epi)
1609                         error = ep_remove(ep, epi);
1610                 else
1611                         error = -ENOENT;
1612                 break;
1613         case EPOLL_CTL_MOD:
1614                 if (epi) {
1615                         epds.events |= POLLERR | POLLHUP;
1616                         error = ep_modify(ep, epi, &epds);
1617                 } else
1618                         error = -ENOENT;
1619                 break;
1620         }
1621         mutex_unlock(&ep->mtx);
1622 
1623 error_tgt_fput:
1624         if (did_lock_epmutex)
1625                 mutex_unlock(&epmutex);
1626 
1627         fput(tfile);
1628 error_fput:
1629         fput(file);
1630 error_return:
1631 
1632         return error;
1633 }

1.参数检查
2.根据目标文件对象及目标文件描述符查找epitem
3.根据op操作，添加epitem/删除epitem/修改eptime

 

i.ep_insert

1056 /*
1057  * Must be called with "mtx" held.
1058  */
1059 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1060                      struct file *tfile, int fd)
1061 {
1062         int error, revents, pwake = 0;
1063         unsigned long flags;
1064         struct epitem *epi;
1065         struct ep_pqueue epq;
1066 
1067         if (unlikely(atomic_read(&ep->user->epoll_watches) >=
1068                      max_user_watches))
1069                 return -ENOSPC;
1070         if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1071                 return -ENOMEM;
1072 
1073         /* Item initialization follow here ... */
1074         INIT_LIST_HEAD(&epi->rdllink);
1075         INIT_LIST_HEAD(&epi->fllink);
1076         INIT_LIST_HEAD(&epi->pwqlist);
1077         epi->ep = ep;
1078         ep_set_ffd(&epi->ffd, tfile, fd);
1079         epi->event = *event;
1080         epi->nwait = 0;
1081         epi->next = EP_UNACTIVE_PTR;
1082 
1083         /* Initialize the poll table using the queue callback */
1084         epq.epi = epi;
1085         init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1086 
1087         /*
1088          * Attach the item to the poll hooks and get current event bits.
1089          * We can safely use the file* here because its usage count has
1090          * been increased by the caller of this function. Note that after
1091          * this operation completes, the poll callback can start hitting
1092          * the new item.
1093          */
1094         revents = tfile->f_op->poll(tfile, &epq.pt);
1095 
1096         /*
1097          * We have to check if something went wrong during the poll wait queue
1098          * install process. Namely an allocation for a wait queue failed due
1099          * high memory pressure.
1100          */
1101         error = -ENOMEM;
1102         if (epi->nwait < 0)
1103                 goto error_unregister;
1104 
1105         /* Add the current item to the list of active epoll hook for this file */
1106         spin_lock(&tfile->f_lock);
1107         list_add_tail(&epi->fllink, &tfile->f_ep_links);
1108         spin_unlock(&tfile->f_lock);
1109 
1110         /*
1111          * Add the current item to the RB tree. All RB tree operations are
1112          * protected by "mtx", and ep_insert() is called with "mtx" held.
1113          */
1114         ep_rbtree_insert(ep, epi);
1115 
1116         /* now check if we've created too many backpaths */
1117         error = -EINVAL;
1118         if (reverse_path_check())
1119                 goto error_remove_epi;
1120 
1121         /* We have to drop the new item inside our item list to keep track of it */
1122         spin_lock_irqsave(&ep->lock, flags);
1123 
1124         /* If the file is already "ready" we drop it inside the ready list */
1125         if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1126                 list_add_tail(&epi->rdllink, &ep->rdllist);
1127 
1128                 /* Notify waiting tasks that events are available */
1129                 if (waitqueue_active(&ep->wq))
1130                         wake_up_locked(&ep->wq);
1131                 if (waitqueue_active(&ep->poll_wait))
1132                         pwake++;
1133         }
1134 
1135         spin_unlock_irqrestore(&ep->lock, flags);
1136 
1137         atomic_inc(&ep->user->epoll_watches);
1138 
1139         /* We have to call this outside the lock */
1140         if (pwake)
1141                 ep_poll_safewake(&ep->poll_wait);
1142 
1143         return 0;
1144 
1145 error_remove_epi:
1146         spin_lock(&tfile->f_lock);
1147         if (ep_is_linked(&epi->fllink))
1148                 list_del_init(&epi->fllink);
1149         spin_unlock(&tfile->f_lock);
1150 
1151         rb_erase(&epi->rbn, &ep->rbr);
1152 
1153 error_unregister:
1154         ep_unregister_pollwait(ep, epi);
1155 
1156         /*
1157          * We need to do this because an event could have been arrived on some
1158          * allocated wait queue. Note that we don't care about the ep->ovflist
1159          * list, since that is used/cleaned only inside a section bound by "mtx".
1160          * And ep_insert() is called with "mtx" held.
1161          */
1162         spin_lock_irqsave(&ep->lock, flags);
1163         if (ep_is_linked(&epi->rdllink))
1164                 list_del_init(&epi->rdllink);
1165         spin_unlock_irqrestore(&ep->lock, flags);
1166 
1167         kmem_cache_free(epi_cache, epi);
1168 
1169         return error;
1170 }

1.分配epitem并根据目标文件初始化
2.初始化poll回调函数并添加到目标文件的等待队列中
3.将epitem链入目标文件的f_ep_links中
4.将epitem添加到eventpoll红黑树中
5.如果目标文件有相应的事件发生，则将epitem添加到eventpoll的就绪链表中，并唤醒等待进程epoll_wait/poll

 917 /*
 918  * This is the callback that is used to add our wait queue to the
 919  * target file wakeup lists.
 920  */
 921 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
 922                                  poll_table *pt)
 923 {
 924         struct epitem *epi = ep_item_from_epqueue(pt);
 925         struct eppoll_entry *pwq;
 926 
 927         if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
 928                 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
 929                 pwq->whead = whead;
 930                 pwq->base = epi;
 931                 add_wait_queue(whead, &pwq->wait);
 932                 list_add_tail(&pwq->llink, &epi->pwqlist);
 933                 epi->nwait++;
 934         } else {
 935                 /* We have to signal that an error occurred */
 936                 epi->nwait = -1;
 937         }
 938 }

1.分配eppoll_entry
2.初始化等待者回调函数为ep_poll_callback，并添加到目标文件的等待队列中
2.建立eppoll_entry与epitem的关联

 837 /*
 838  * This is the callback that is passed to the wait queue wakeup
 839  * machanism. It is called by the stored file descriptors when they
 840  * have events to report.
 841  */
 842 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
 843 {
 844         int pwake = 0;
 845         unsigned long flags;
 846         struct epitem *epi = ep_item_from_wait(wait);
 847         struct eventpoll *ep = epi->ep;
 848 
 849         if ((unsigned long)key & POLLFREE) {
 850                 ep_pwq_from_wait(wait)->whead = NULL;
 851                 /*
 852                  * whead = NULL above can race with ep_remove_wait_queue()
 853                  * which can do another remove_wait_queue() after us, so we
 854                  * can't use __remove_wait_queue(). whead->lock is held by
 855                  * the caller.
 856                  */
 857                 list_del_init(&wait->task_list);
 858         }
 859 
 860         spin_lock_irqsave(&ep->lock, flags);
 861 
 862         /*
 863          * If the event mask does not contain any poll(2) event, we consider the
 864          * descriptor to be disabled. This condition is likely the effect of the
 865          * EPOLLONESHOT bit that disables the descriptor when an event is received,
 866          * until the next EPOLL_CTL_MOD will be issued.
 867          */
 868         if (!(epi->event.events & ~EP_PRIVATE_BITS))
 869                 goto out_unlock;
 870 
 871         /*
 872          * Check the events coming with the callback. At this stage, not
 873          * every device reports the events in the "key" parameter of the
 874          * callback. We need to be able to handle both cases here, hence the
 875          * test for "key" != NULL before the event match test.
 876          */
 877         if (key && !((unsigned long) key & epi->event.events))
 878                 goto out_unlock;
 879 
 880         /*
 881          * If we are trasfering events to userspace, we can hold no locks
 882          * (because we're accessing user memory, and because of linux f_op->poll()
 883          * semantics). All the events that happens during that period of time are
 884          * chained in ep->ovflist and requeued later on.
 885          */
 886         if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
 887                 if (epi->next == EP_UNACTIVE_PTR) {
 888                         epi->next = ep->ovflist;
 889                         ep->ovflist = epi;
 890                 }
 891                 goto out_unlock;
 892         }
 893 
 894         /* If this file is already in the ready list we exit soon */
 895         if (!ep_is_linked(&epi->rdllink))
 896                 list_add_tail(&epi->rdllink, &ep->rdllist);
 897 
 898         /*
 899          * Wake up ( if active ) both the eventpoll wait list and the ->poll()
 900          * wait list.
 901          */
 902         if (waitqueue_active(&ep->wq))
 903                 wake_up_locked(&ep->wq);
 904         if (waitqueue_active(&ep->poll_wait))
 905                 pwake++;
 906 
 907 out_unlock:
 908         spin_unlock_irqrestore(&ep->lock, flags);
 909 
 910         /* We have to call this outside the lock */
 911         if (pwake)
 912                 ep_poll_safewake(&ep->poll_wait);
 913 
 914         return 1;
 915 }

1.如果事件有POLLFREE，则将等待者从目标文件的等待队列中删除
2.如果事件只有EPOLLONESHOT | EPOLLET，则直接退出；主要用于EPOLLONESHOT时清空事件中的其它位
3.如果正在复制就绪链表中的文件描述符及事件到用户空间，且epitem不在ovflist链表中，则将epitem添加到ovflist链表中，并退出
4.如果epitem不在就绪链表中，将epitem添加到就绪链表中
5.唤醒等待队列
注：
ep_ptable_queue_proc是文件poll操作的回调函数
ep_poll_callback是有事件发生时唤醒等待队列的回调函数

 

ii.ep_remove

 591 /*
 592  * Removes a "struct epitem" from the eventpoll RB tree and deallocates
 593  * all the associated resources. Must be called with "mtx" held.
 594  */
 595 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
 596 {
 597         unsigned long flags;
 598         struct file *file = epi->ffd.file;
 599 
 600         /*
 601          * Removes poll wait queue hooks. We _have_ to do this without holding
 602          * the "ep->lock" otherwise a deadlock might occur. This because of the
 603          * sequence of the lock acquisition. Here we do "ep->lock" then the wait
 604          * queue head lock when unregistering the wait queue. The wakeup callback
 605          * will run by holding the wait queue head lock and will call our callback
 606          * that will try to get "ep->lock".
 607          */
 608         ep_unregister_pollwait(ep, epi);
 609 
 610         /* Remove the current item from the list of epoll hooks */
 611         spin_lock(&file->f_lock);
 612         if (ep_is_linked(&epi->fllink))
 613                 list_del_init(&epi->fllink);
 614         spin_unlock(&file->f_lock);
 615 
 616         rb_erase(&epi->rbn, &ep->rbr);
 617 
 618         spin_lock_irqsave(&ep->lock, flags);
 619         if (ep_is_linked(&epi->rdllink))
 620                 list_del_init(&epi->rdllink);
 621         spin_unlock_irqrestore(&ep->lock, flags);
 622 
 623         /* At this point it is safe to free the eventpoll item */
 624         kmem_cache_free(epi_cache, epi);
 625 
 626         atomic_dec(&ep->user->epoll_watches);
 627 
 628         return 0;
 629 }

1.删除epitem的等待者
2.将epitem从目标文件的fllink链表中移除
3.将epitem从红黑树中移除
4.将epitem从就绪链表中移除
5.回收epitem的内存
6.归还epoll计数

 

iii.ep_modify

1172 /*
1173  * Modify the interest event mask by dropping an event if the new mask
1174  * has a match in the current file status. Must be called with "mtx" held.
1175  */
1176 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1177 {
1178         int pwake = 0;
1179         unsigned int revents;
1180 
1181         /*
1182          * Set the new event interest mask before calling f_op->poll();
1183          * otherwise we might miss an event that happens between the
1184          * f_op->poll() call and the new event set registering.
1185          */
1186         epi->event.events = event->events;
1187         epi->event.data = event->data; /* protected by mtx */
1188 
1189         /*
1190          * Get current event bits. We can safely use the file* here because
1191          * its usage count has been increased by the caller of this function.
1192          */
1193         revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1194 
1195         /*
1196          * If the item is "hot" and it is not registered inside the ready
1197          * list, push it inside.
1198          */
1199         if (revents & event->events) {
1200                 spin_lock_irq(&ep->lock);
1201                 if (!ep_is_linked(&epi->rdllink)) {
1202                         list_add_tail(&epi->rdllink, &ep->rdllist);
1203 
1204                         /* Notify waiting tasks that events are available */
1205                         if (waitqueue_active(&ep->wq))
1206                                 wake_up_locked(&ep->wq);
1207                         if (waitqueue_active(&ep->poll_wait))
1208                                 pwake++;
1209                 }
1210                 spin_unlock_irq(&ep->lock);
1211         }
1212 
1213         /* We have to call this outside the lock */
1214         if (pwake)
1215                 ep_poll_safewake(&ep->poll_wait);
1216 
1217         return 0;
1218 }

1.修改epitem中的事件及数据
2.查询目标文件的POLL事件
3.如果目标文件的POLL事件有我们关心的事件，且不在就绪链表中，则将epitem添加到就绪链表并唤醒等待者

 

 

IV.epoll_wait

1635 /*
1636  * Implement the event wait interface for the eventpoll file. It is the kernel
1637  * part of the user space epoll_wait(2).
1638  */
1639 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1640                 int, maxevents, int, timeout)
1641 {
1642         int error;
1643         struct file *file;
1644         struct eventpoll *ep;
1645 
1646         /* The maximum number of event must be greater than zero */
1647         if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1648                 return -EINVAL;
1649 
1650         /* Verify that the area passed by the user is writeable */
1651         if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1652                 error = -EFAULT;
1653                 goto error_return;
1654         }
1655 
1656         /* Get the "struct file *" for the eventpoll file */
1657         error = -EBADF;
1658         file = fget(epfd);
1659         if (!file)
1660                 goto error_return;
1661 
1662         /*
1663          * We have to check that the file structure underneath the fd
1664          * the user passed to us _is_ an eventpoll file.
1665          */
1666         error = -EINVAL;
1667         if (!is_file_epoll(file))
1668                 goto error_fput;
1669 
1670         /*
1671          * At this point it is safe to assume that the "private_data" contains
1672          * our own data structure.
1673          */
1674         ep = file->private_data;
1675 
1676         /* Time to fish for events ... */
1677         error = ep_poll(ep, events, maxevents, timeout);
1678 
1679 error_fput:
1680         fput(file);
1681 error_return:
1682 
1683         return error;
1684 }

1.参数检查
2.获取eventpoll
3.调用ep_poll

1290 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1291                    int maxevents, long timeout)
1292 {
1293         int res, eavail;
1294         unsigned long flags;
1295         long jtimeout;
1296         wait_queue_t wait;
1297 
1298         /*
1299          * Calculate the timeout by checking for the "infinite" value (-1)
1300          * and the overflow condition. The passed timeout is in milliseconds,
1301          * that why (t * HZ) / 1000.
1302          */
1303         jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1304                 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1305 
1306 retry:
1307         spin_lock_irqsave(&ep->lock, flags);
1308 
1309         res = 0;
1310         if (list_empty(&ep->rdllist)) {
1311                 /*
1312                  * We don't have any available event to return to the caller.
1313                  * We need to sleep here, and we will be wake up by
1314                  * ep_poll_callback() when events will become available.
1315                  */
1316                 init_waitqueue_entry(&wait, current);
1317                 wait.flags |= WQ_FLAG_EXCLUSIVE;
1318                 __add_wait_queue(&ep->wq, &wait);
1319 
1320                 for (;;) {
1321                         /*
1322                          * We don't want to sleep if the ep_poll_callback() sends us
1323                          * a wakeup in between. That's why we set the task state
1324                          * to TASK_INTERRUPTIBLE before doing the checks.
1325                          */
1326                         set_current_state(TASK_INTERRUPTIBLE);
1327                         if (!list_empty(&ep->rdllist) || !jtimeout)
1328                                 break;
1329                         if (signal_pending(current)) {
1330                                 res = -EINTR;
1331                                 break;
1332                         }
1333 
1334                         spin_unlock_irqrestore(&ep->lock, flags);
1335                         jtimeout = schedule_timeout(jtimeout);
1336                         spin_lock_irqsave(&ep->lock, flags);
1337                 }
1338                 __remove_wait_queue(&ep->wq, &wait);
1339 
1340                 set_current_state(TASK_RUNNING);
1341         }
1342         /* Is it worth to try to dig for events ? */
1343         eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
1344 
1345         spin_unlock_irqrestore(&ep->lock, flags);
1346 
1347         /*
1348          * Try to transfer events to user space. In case we get 0 events and
1349          * there's still timeout left over, we go trying again in search of
1350          * more luck.
1351          */
1352         if (!res && eavail &&
1353             !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1354                 goto retry;
1355 
1356         return res;
1357 }

1.计算超时时间
2.如果就绪链表为空
  A.将进程添加到epoll_wait等待队列中
  B.阻塞epoll_wait，直到就绪链表非空或有信号需要处理
  C.将进程从epoll_wait等待队列中移除，恢复进程的执行
3.如果就绪链表或ovflist链表非空，则将文件描述符及事件复制到用户空间的epoll_wait事件参数中

1220 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1221                                void *priv)
1222 {
1223         struct ep_send_events_data *esed = priv;
1224         int eventcnt;
1225         unsigned int revents;
1226         struct epitem *epi;
1227         struct epoll_event __user *uevent;
1228 
1229         /*
1230          * We can loop without lock because we are passed a task private list.
1231          * Items cannot vanish during the loop because ep_scan_ready_list() is
1232          * holding "mtx" during this call.
1233          */
1234         for (eventcnt = 0, uevent = esed->events;
1235              !list_empty(head) && eventcnt < esed->maxevents;) {
1236                 epi = list_first_entry(head, struct epitem, rdllink);
1237 
1238                 list_del_init(&epi->rdllink);
1239 
1240                 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1241                         epi->event.events;
1242 
1243                 /*
1244                  * If the event mask intersect the caller-requested one,
1245                  * deliver the event to userspace. Again, ep_scan_ready_list()
1246                  * is holding "mtx", so no operations coming from userspace
1247                  * can change the item.
1248                  */
1249                 if (revents) {
1250                         if (__put_user(revents, &uevent->events) ||
1251                             __put_user(epi->event.data, &uevent->data)) {
1252                                 list_add(&epi->rdllink, head);
1253                                 return eventcnt ? eventcnt : -EFAULT;
1254                         }
1255                         eventcnt++;
1256                         uevent++;
1257                         if (epi->event.events & EPOLLONESHOT)
1258                                 epi->event.events &= EP_PRIVATE_BITS;
1259                         else if (!(epi->event.events & EPOLLET)) {
1260                                 /*
1261                                  * If this file has been added with Level
1262                                  * Trigger mode, we need to insert back inside
1263                                  * the ready list, so that the next call to
1264                                  * epoll_wait() will check again the events
1265                                  * availability. At this point, noone can insert
1266                                  * into ep->rdllist besides us. The epoll_ctl()
1267                                  * callers are locked out by
1268                                  * ep_scan_ready_list() holding "mtx" and the
1269                                  * poll callback will queue them in ep->ovflist.
1270                                  */
1271                                 list_add_tail(&epi->rdllink, &ep->rdllist);
1272                         }
1273                 }
1274         }
1275 
1276         return eventcnt;
1277 }
1278 
1279 static int ep_send_events(struct eventpoll *ep,
1280                           struct epoll_event __user *events, int maxevents)
1281 {
1282         struct ep_send_events_data esed;
1283 
1284         esed.maxevents = maxevents;
1285         esed.events = events;
1286 
1287         return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
1288 }
1289 

1.遍历就绪链表，并取出epitem
  A.调用目标文件的poll操作，取出POLL事件；
  B.如果目标文件的POLL事件与关心的POLL事件有交集，则将文件描述符及事件复制到用户空间的epoll_wait事件参数中
  C.如果事件是EPOLLONESHOT的，则清除关心POLL事件位；如果事件是非边沿触发，即水平触发，则将epitem重新放回到就绪链表中，不再就绪的epitem会在下次epoll_wait时移除
2.直到就绪链表为空或事件已经到达最大值
3.返回事件数

 

V.epoll_pwait
epoll_pwait是在epoll_wait之上添加了信号屏蔽功能

1688 /*
1689  * Implement the event wait interface for the eventpoll file. It is the kernel
1690  * part of the user space epoll_pwait(2).
1691  */
1692 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1693                 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1694                 size_t, sigsetsize)
1695 {
1696         int error;
1697         sigset_t ksigmask, sigsaved;
1698 
1699         /*
1700          * If the caller wants a certain signal mask to be set during the wait,
1701          * we apply it here.
1702          */
1703         if (sigmask) {
1704                 if (sigsetsize != sizeof(sigset_t))
1705                         return -EINVAL;
1706                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1707                         return -EFAULT;
1708                 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1709                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1710         }
1711 
1712         error = sys_epoll_wait(epfd, events, maxevents, timeout);
1713 
1714         /*
1715          * If we changed the signal mask, we need to restore the original one.
1716          * In case we've got a signal while waiting, we do not restore the
1717          * signal mask yet, and we allow do_signal() to deliver the signal on
1718          * the way back to userspace, before the signal mask is restored.
1719          */
1720         if (sigmask) {
1721                 if (error == -EINTR) {
1722                         memcpy(¤t->saved_sigmask, &sigsaved,
1723                                sizeof(sigsaved));
1724                         set_restore_sigmask();
1725                 } else
1726                         sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1727         }
1728 
1729         return error;
1730 }

1.如果有屏蔽信号，则将屏蔽信号掩码复制到内核空间；由于SIGKILL/SIGSTOP是不可屏蔽的，将它们从掩码中移除；设置信号阻塞掩码，并保存现有的信号阻塞掩码
2.调用epoll_wait
3.恢复信号阻塞掩码