2 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * Davide Libenzi <davidel@xmailserver.org>
14 #include <linux/init.h>
15 #include <linux/kernel.h>
16 #include <linux/sched.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <asm/uaccess.h>
37 #include <asm/system.h>
40 #include <asm/atomic.h>
44 * There are three level of locking required by epoll :
48 * 3) ep->lock (spinlock)
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a spinlock (ep->lock) because we manipulate objects
52 * from inside the poll callback, that might be triggered from
53 * a wake_up() that in turn might be called from IRQ context.
54 * So we can't sleep inside the poll callback and hence we need
55 * a spinlock. During the event transfer loop (from kernel to
56 * user space) we could end up sleeping due a copy_to_user(), so
57 * we need a lock that will allow us to sleep. This lock is a
58 * mutex (ep->mtx). It is acquired during the event transfer loop,
59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * This mutex is acquired by ep_free() during the epoll file
63 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
66 * It is possible to drop the "ep->mtx" and to use the global
67 * mutex "epmutex" (together with "ep->lock") to have it working,
68 * but having "ep->mtx" will make the interface more scalable.
69 * Events that require holding "epmutex" are very rare, while for
70 * normal operations the epoll private "ep->mtx" will guarantee
71 * a better scalability.
74 /* Epoll private bits inside the event mask */
75 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
77 /* Maximum number of nesting allowed inside epoll sets */
78 #define EP_MAX_NESTS 4
80 /* Maximum msec timeout value storeable in a long int */
81 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
83 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
85 #define EP_UNACTIVE_PTR ((void *) -1L)
87 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
95 * Structure used to track possible nested calls, for too deep recursions
98 struct nested_call_node {
99 struct list_head llink;
105 * This structure is used as collector for nested calls, to check for
106 * maximum recursion dept and loop cycles.
108 struct nested_calls {
109 struct list_head tasks_call_list;
114 * Each file descriptor added to the eventpoll interface will
115 * have an entry of this type linked to the "rbr" RB tree.
118 /* RB tree node used to link this structure to the eventpoll RB tree */
121 /* List header used to link this structure to the eventpoll ready list */
122 struct list_head rdllink;
125 * Works together "struct eventpoll"->ovflist in keeping the
126 * single linked chain of items.
130 /* The file descriptor information this item refers to */
131 struct epoll_filefd ffd;
133 /* Number of active wait queue attached to poll operations */
136 /* List containing poll wait queues */
137 struct list_head pwqlist;
139 /* The "container" of this item */
140 struct eventpoll *ep;
142 /* List header used to link this item to the "struct file" items list */
143 struct list_head fllink;
145 /* The structure that describe the interested events and the source fd */
146 struct epoll_event event;
150 * This structure is stored inside the "private_data" member of the file
151 * structure and rapresent the main data sructure for the eventpoll
155 /* Protect the this structure access */
159 * This mutex is used to ensure that files are not removed
160 * while epoll is using them. This is held during the event
161 * collection loop, the file cleanup path, the epoll file exit
162 * code and the ctl operations.
166 /* Wait queue used by sys_epoll_wait() */
167 wait_queue_head_t wq;
169 /* Wait queue used by file->poll() */
170 wait_queue_head_t poll_wait;
172 /* List of ready file descriptors */
173 struct list_head rdllist;
175 /* RB tree root used to store monitored fd structs */
179 * This is a single linked list that chains all the "struct epitem" that
180 * happened while transfering ready events to userspace w/out
183 struct epitem *ovflist;
185 /* The user that created the eventpoll descriptor */
186 struct user_struct *user;
189 /* Wait structure used by the poll hooks */
190 struct eppoll_entry {
191 /* List header used to link this structure to the "struct epitem" */
192 struct list_head llink;
194 /* The "base" pointer is set to the container "struct epitem" */
198 * Wait queue item that will be linked to the target file wait
203 /* The wait queue head that linked the "wait" wait queue item */
204 wait_queue_head_t *whead;
207 /* Wrapper struct used by poll queueing */
213 /* Used by the ep_send_events() function as callback private data */
214 struct ep_send_events_data {
216 struct epoll_event __user *events;
220 * Configuration options available inside /proc/sys/fs/epoll/
222 /* Maximum number of epoll watched descriptors, per user */
223 static int max_user_watches __read_mostly;
226 * This mutex is used to serialize ep_free() and eventpoll_release_file().
228 static DEFINE_MUTEX(epmutex);
230 /* Used for safe wake up implementation */
231 static struct nested_calls poll_safewake_ncalls;
233 /* Used to call file's f_op->poll() under the nested calls boundaries */
234 static struct nested_calls poll_readywalk_ncalls;
236 /* Slab cache used to allocate "struct epitem" */
237 static struct kmem_cache *epi_cache __read_mostly;
239 /* Slab cache used to allocate "struct eppoll_entry" */
240 static struct kmem_cache *pwq_cache __read_mostly;
244 #include <linux/sysctl.h>
248 ctl_table epoll_table[] = {
250 .procname = "max_user_watches",
251 .data = &max_user_watches,
252 .maxlen = sizeof(int),
254 .proc_handler = &proc_dointvec_minmax,
259 #endif /* CONFIG_SYSCTL */
262 /* Setup the structure that is used as key for the RB tree */
263 static inline void ep_set_ffd(struct epoll_filefd *ffd,
264 struct file *file, int fd)
270 /* Compare RB tree keys */
271 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
272 struct epoll_filefd *p2)
274 return (p1->file > p2->file ? +1:
275 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
278 /* Tells us if the item is currently linked */
279 static inline int ep_is_linked(struct list_head *p)
281 return !list_empty(p);
284 /* Get the "struct epitem" from a wait queue pointer */
285 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
287 return container_of(p, struct eppoll_entry, wait)->base;
290 /* Get the "struct epitem" from an epoll queue wrapper */
291 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
293 return container_of(p, struct ep_pqueue, pt)->epi;
296 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
297 static inline int ep_op_has_event(int op)
299 return op != EPOLL_CTL_DEL;
302 /* Initialize the poll safe wake up structure */
303 static void ep_nested_calls_init(struct nested_calls *ncalls)
305 INIT_LIST_HEAD(&ncalls->tasks_call_list);
306 spin_lock_init(&ncalls->lock);
310 * ep_call_nested - Perform a bound (possibly) nested call, by checking
311 * that the recursion limit is not exceeded, and that
312 * the same nested call (by the meaning of same cookie) is
315 * @ncalls: Pointer to the nested_calls structure to be used for this call.
316 * @max_nests: Maximum number of allowed nesting calls.
317 * @nproc: Nested call core function pointer.
318 * @priv: Opaque data to be passed to the @nproc callback.
319 * @cookie: Cookie to be used to identify this nested call.
321 * Returns: Returns the code returned by the @nproc callback, or -1 if
322 * the maximum recursion limit has been exceeded.
324 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
325 int (*nproc)(void *, void *, int), void *priv,
328 int error, call_nests = 0;
330 int this_cpu = get_cpu();
331 struct list_head *lsthead = &ncalls->tasks_call_list;
332 struct nested_call_node *tncur;
333 struct nested_call_node tnode;
335 spin_lock_irqsave(&ncalls->lock, flags);
338 * Try to see if the current task is already inside this wakeup call.
339 * We use a list here, since the population inside this set is always
342 list_for_each_entry(tncur, lsthead, llink) {
343 if (tncur->cpu == this_cpu &&
344 (tncur->cookie == cookie || ++call_nests > max_nests)) {
346 * Ops ... loop detected or maximum nest level reached.
347 * We abort this wake by breaking the cycle itself.
354 /* Add the current task and cookie to the list */
355 tnode.cpu = this_cpu;
356 tnode.cookie = cookie;
357 list_add(&tnode.llink, lsthead);
359 spin_unlock_irqrestore(&ncalls->lock, flags);
361 /* Call the nested function */
362 error = (*nproc)(priv, cookie, call_nests);
364 /* Remove the current task from the list */
365 spin_lock_irqsave(&ncalls->lock, flags);
366 list_del(&tnode.llink);
368 spin_unlock_irqrestore(&ncalls->lock, flags);
374 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
376 wake_up_nested((wait_queue_head_t *) cookie, 1 + call_nests);
381 * Perform a safe wake up of the poll wait list. The problem is that
382 * with the new callback'd wake up system, it is possible that the
383 * poll callback is reentered from inside the call to wake_up() done
384 * on the poll wait queue head. The rule is that we cannot reenter the
385 * wake up code from the same task more than EP_MAX_NESTS times,
386 * and we cannot reenter the same wait queue head at all. This will
387 * enable to have a hierarchy of epoll file descriptor of no more than
390 static void ep_poll_safewake(wait_queue_head_t *wq)
392 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
393 ep_poll_wakeup_proc, NULL, wq);
397 * This function unregister poll callbacks from the associated file descriptor.
398 * Since this must be called without holding "ep->lock" the atomic exchange trick
399 * will protect us from multiple unregister.
401 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
404 struct list_head *lsthead = &epi->pwqlist;
405 struct eppoll_entry *pwq;
407 /* This is called without locks, so we need the atomic exchange */
408 nwait = xchg(&epi->nwait, 0);
411 while (!list_empty(lsthead)) {
412 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
414 list_del_init(&pwq->llink);
415 remove_wait_queue(pwq->whead, &pwq->wait);
416 kmem_cache_free(pwq_cache, pwq);
422 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
423 * the scan code, to call f_op->poll(). Also allows for
424 * O(NumReady) performance.
426 * @ep: Pointer to the epoll private data structure.
427 * @sproc: Pointer to the scan callback.
428 * @priv: Private opaque data passed to the @sproc callback.
430 * Returns: The same integer error code returned by the @sproc callback.
432 static int ep_scan_ready_list(struct eventpoll *ep,
433 int (*sproc)(struct eventpoll *,
434 struct list_head *, void *),
437 int error, pwake = 0;
439 struct epitem *epi, *nepi;
443 * We need to lock this because we could be hit by
444 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
446 mutex_lock(&ep->mtx);
449 * Steal the ready list, and re-init the original one to the
450 * empty list. Also, set ep->ovflist to NULL so that events
451 * happening while looping w/out locks, are not lost. We cannot
452 * have the poll callback to queue directly on ep->rdllist,
453 * because we want the "sproc" callback to be able to do it
456 spin_lock_irqsave(&ep->lock, flags);
457 list_splice_init(&ep->rdllist, &txlist);
459 spin_unlock_irqrestore(&ep->lock, flags);
462 * Now call the callback function.
464 error = (*sproc)(ep, &txlist, priv);
466 spin_lock_irqsave(&ep->lock, flags);
468 * During the time we spent inside the "sproc" callback, some
469 * other events might have been queued by the poll callback.
470 * We re-insert them inside the main ready-list here.
472 for (nepi = ep->ovflist; (epi = nepi) != NULL;
473 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
475 * We need to check if the item is already in the list.
476 * During the "sproc" callback execution time, items are
477 * queued into ->ovflist but the "txlist" might already
478 * contain them, and the list_splice() below takes care of them.
480 if (!ep_is_linked(&epi->rdllink))
481 list_add_tail(&epi->rdllink, &ep->rdllist);
484 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
485 * releasing the lock, events will be queued in the normal way inside
488 ep->ovflist = EP_UNACTIVE_PTR;
491 * Quickly re-inject items left on "txlist".
493 list_splice(&txlist, &ep->rdllist);
495 if (!list_empty(&ep->rdllist)) {
497 * Wake up (if active) both the eventpoll wait list and
498 * the ->poll() wait list (delayed after we release the lock).
500 if (waitqueue_active(&ep->wq))
501 wake_up_locked(&ep->wq);
502 if (waitqueue_active(&ep->poll_wait))
505 spin_unlock_irqrestore(&ep->lock, flags);
507 mutex_unlock(&ep->mtx);
509 /* We have to call this outside the lock */
511 ep_poll_safewake(&ep->poll_wait);
517 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
518 * all the associated resources. Must be called with "mtx" held.
520 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
523 struct file *file = epi->ffd.file;
526 * Removes poll wait queue hooks. We _have_ to do this without holding
527 * the "ep->lock" otherwise a deadlock might occur. This because of the
528 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
529 * queue head lock when unregistering the wait queue. The wakeup callback
530 * will run by holding the wait queue head lock and will call our callback
531 * that will try to get "ep->lock".
533 ep_unregister_pollwait(ep, epi);
535 /* Remove the current item from the list of epoll hooks */
536 spin_lock(&file->f_lock);
537 if (ep_is_linked(&epi->fllink))
538 list_del_init(&epi->fllink);
539 spin_unlock(&file->f_lock);
541 rb_erase(&epi->rbn, &ep->rbr);
543 spin_lock_irqsave(&ep->lock, flags);
544 if (ep_is_linked(&epi->rdllink))
545 list_del_init(&epi->rdllink);
546 spin_unlock_irqrestore(&ep->lock, flags);
548 /* At this point it is safe to free the eventpoll item */
549 kmem_cache_free(epi_cache, epi);
551 atomic_dec(&ep->user->epoll_watches);
556 static void ep_free(struct eventpoll *ep)
561 /* We need to release all tasks waiting for these file */
562 if (waitqueue_active(&ep->poll_wait))
563 ep_poll_safewake(&ep->poll_wait);
566 * We need to lock this because we could be hit by
567 * eventpoll_release_file() while we're freeing the "struct eventpoll".
568 * We do not need to hold "ep->mtx" here because the epoll file
569 * is on the way to be removed and no one has references to it
570 * anymore. The only hit might come from eventpoll_release_file() but
571 * holding "epmutex" is sufficent here.
573 mutex_lock(&epmutex);
576 * Walks through the whole tree by unregistering poll callbacks.
578 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
579 epi = rb_entry(rbp, struct epitem, rbn);
581 ep_unregister_pollwait(ep, epi);
585 * Walks through the whole tree by freeing each "struct epitem". At this
586 * point we are sure no poll callbacks will be lingering around, and also by
587 * holding "epmutex" we can be sure that no file cleanup code will hit
588 * us during this operation. So we can avoid the lock on "ep->lock".
590 while ((rbp = rb_first(&ep->rbr)) != NULL) {
591 epi = rb_entry(rbp, struct epitem, rbn);
595 mutex_unlock(&epmutex);
596 mutex_destroy(&ep->mtx);
601 static int ep_eventpoll_release(struct inode *inode, struct file *file)
603 struct eventpoll *ep = file->private_data;
611 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
614 struct epitem *epi, *tmp;
616 list_for_each_entry_safe(epi, tmp, head, rdllink) {
617 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
619 return POLLIN | POLLRDNORM;
622 * Item has been dropped into the ready list by the poll
623 * callback, but it's not actually ready, as far as
624 * caller requested events goes. We can remove it here.
626 list_del_init(&epi->rdllink);
633 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
635 return ep_scan_ready_list(priv, ep_read_events_proc, NULL);
638 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
641 struct eventpoll *ep = file->private_data;
643 /* Insert inside our poll wait queue */
644 poll_wait(file, &ep->poll_wait, wait);
647 * Proceed to find out if wanted events are really available inside
648 * the ready list. This need to be done under ep_call_nested()
649 * supervision, since the call to f_op->poll() done on listed files
650 * could re-enter here.
652 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
653 ep_poll_readyevents_proc, ep, ep);
655 return pollflags != -1 ? pollflags : 0;
658 /* File callbacks that implement the eventpoll file behaviour */
659 static const struct file_operations eventpoll_fops = {
660 .release = ep_eventpoll_release,
661 .poll = ep_eventpoll_poll
664 /* Fast test to see if the file is an evenpoll file */
665 static inline int is_file_epoll(struct file *f)
667 return f->f_op == &eventpoll_fops;
671 * This is called from eventpoll_release() to unlink files from the eventpoll
672 * interface. We need to have this facility to cleanup correctly files that are
673 * closed without being removed from the eventpoll interface.
675 void eventpoll_release_file(struct file *file)
677 struct list_head *lsthead = &file->f_ep_links;
678 struct eventpoll *ep;
682 * We don't want to get "file->f_lock" because it is not
683 * necessary. It is not necessary because we're in the "struct file"
684 * cleanup path, and this means that noone is using this file anymore.
685 * So, for example, epoll_ctl() cannot hit here since if we reach this
686 * point, the file counter already went to zero and fget() would fail.
687 * The only hit might come from ep_free() but by holding the mutex
688 * will correctly serialize the operation. We do need to acquire
689 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
690 * from anywhere but ep_free().
692 * Besides, ep_remove() acquires the lock, so we can't hold it here.
694 mutex_lock(&epmutex);
696 while (!list_empty(lsthead)) {
697 epi = list_first_entry(lsthead, struct epitem, fllink);
700 list_del_init(&epi->fllink);
701 mutex_lock(&ep->mtx);
703 mutex_unlock(&ep->mtx);
706 mutex_unlock(&epmutex);
709 static int ep_alloc(struct eventpoll **pep)
712 struct user_struct *user;
713 struct eventpoll *ep;
715 user = get_current_user();
717 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
721 spin_lock_init(&ep->lock);
722 mutex_init(&ep->mtx);
723 init_waitqueue_head(&ep->wq);
724 init_waitqueue_head(&ep->poll_wait);
725 INIT_LIST_HEAD(&ep->rdllist);
727 ep->ovflist = EP_UNACTIVE_PTR;
740 * Search the file inside the eventpoll tree. The RB tree operations
741 * are protected by the "mtx" mutex, and ep_find() must be called with
744 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
748 struct epitem *epi, *epir = NULL;
749 struct epoll_filefd ffd;
751 ep_set_ffd(&ffd, file, fd);
752 for (rbp = ep->rbr.rb_node; rbp; ) {
753 epi = rb_entry(rbp, struct epitem, rbn);
754 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
769 * This is the callback that is passed to the wait queue wakeup
770 * machanism. It is called by the stored file descriptors when they
771 * have events to report.
773 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
777 struct epitem *epi = ep_item_from_wait(wait);
778 struct eventpoll *ep = epi->ep;
780 spin_lock_irqsave(&ep->lock, flags);
783 * If the event mask does not contain any poll(2) event, we consider the
784 * descriptor to be disabled. This condition is likely the effect of the
785 * EPOLLONESHOT bit that disables the descriptor when an event is received,
786 * until the next EPOLL_CTL_MOD will be issued.
788 if (!(epi->event.events & ~EP_PRIVATE_BITS))
792 * If we are trasfering events to userspace, we can hold no locks
793 * (because we're accessing user memory, and because of linux f_op->poll()
794 * semantics). All the events that happens during that period of time are
795 * chained in ep->ovflist and requeued later on.
797 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
798 if (epi->next == EP_UNACTIVE_PTR) {
799 epi->next = ep->ovflist;
805 /* If this file is already in the ready list we exit soon */
806 if (!ep_is_linked(&epi->rdllink))
807 list_add_tail(&epi->rdllink, &ep->rdllist);
810 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
813 if (waitqueue_active(&ep->wq))
814 wake_up_locked(&ep->wq);
815 if (waitqueue_active(&ep->poll_wait))
819 spin_unlock_irqrestore(&ep->lock, flags);
821 /* We have to call this outside the lock */
823 ep_poll_safewake(&ep->poll_wait);
829 * This is the callback that is used to add our wait queue to the
830 * target file wakeup lists.
832 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
835 struct epitem *epi = ep_item_from_epqueue(pt);
836 struct eppoll_entry *pwq;
838 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
839 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
842 add_wait_queue(whead, &pwq->wait);
843 list_add_tail(&pwq->llink, &epi->pwqlist);
846 /* We have to signal that an error occurred */
851 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
854 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
859 epic = rb_entry(parent, struct epitem, rbn);
860 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
862 p = &parent->rb_right;
864 p = &parent->rb_left;
866 rb_link_node(&epi->rbn, parent, p);
867 rb_insert_color(&epi->rbn, &ep->rbr);
871 * Must be called with "mtx" held.
873 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
874 struct file *tfile, int fd)
876 int error, revents, pwake = 0;
879 struct ep_pqueue epq;
881 if (unlikely(atomic_read(&ep->user->epoll_watches) >=
884 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
887 /* Item initialization follow here ... */
888 INIT_LIST_HEAD(&epi->rdllink);
889 INIT_LIST_HEAD(&epi->fllink);
890 INIT_LIST_HEAD(&epi->pwqlist);
892 ep_set_ffd(&epi->ffd, tfile, fd);
895 epi->next = EP_UNACTIVE_PTR;
897 /* Initialize the poll table using the queue callback */
899 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
902 * Attach the item to the poll hooks and get current event bits.
903 * We can safely use the file* here because its usage count has
904 * been increased by the caller of this function. Note that after
905 * this operation completes, the poll callback can start hitting
908 revents = tfile->f_op->poll(tfile, &epq.pt);
911 * We have to check if something went wrong during the poll wait queue
912 * install process. Namely an allocation for a wait queue failed due
913 * high memory pressure.
917 goto error_unregister;
919 /* Add the current item to the list of active epoll hook for this file */
920 spin_lock(&tfile->f_lock);
921 list_add_tail(&epi->fllink, &tfile->f_ep_links);
922 spin_unlock(&tfile->f_lock);
925 * Add the current item to the RB tree. All RB tree operations are
926 * protected by "mtx", and ep_insert() is called with "mtx" held.
928 ep_rbtree_insert(ep, epi);
930 /* We have to drop the new item inside our item list to keep track of it */
931 spin_lock_irqsave(&ep->lock, flags);
933 /* If the file is already "ready" we drop it inside the ready list */
934 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
935 list_add_tail(&epi->rdllink, &ep->rdllist);
937 /* Notify waiting tasks that events are available */
938 if (waitqueue_active(&ep->wq))
939 wake_up_locked(&ep->wq);
940 if (waitqueue_active(&ep->poll_wait))
944 spin_unlock_irqrestore(&ep->lock, flags);
946 atomic_inc(&ep->user->epoll_watches);
948 /* We have to call this outside the lock */
950 ep_poll_safewake(&ep->poll_wait);
955 ep_unregister_pollwait(ep, epi);
958 * We need to do this because an event could have been arrived on some
959 * allocated wait queue. Note that we don't care about the ep->ovflist
960 * list, since that is used/cleaned only inside a section bound by "mtx".
961 * And ep_insert() is called with "mtx" held.
963 spin_lock_irqsave(&ep->lock, flags);
964 if (ep_is_linked(&epi->rdllink))
965 list_del_init(&epi->rdllink);
966 spin_unlock_irqrestore(&ep->lock, flags);
968 kmem_cache_free(epi_cache, epi);
974 * Modify the interest event mask by dropping an event if the new mask
975 * has a match in the current file status. Must be called with "mtx" held.
977 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
980 unsigned int revents;
984 * Set the new event interest mask before calling f_op->poll(), otherwise
985 * a potential race might occur. In fact if we do this operation inside
986 * the lock, an event might happen between the f_op->poll() call and the
987 * new event set registering.
989 epi->event.events = event->events;
992 * Get current event bits. We can safely use the file* here because
993 * its usage count has been increased by the caller of this function.
995 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
997 spin_lock_irqsave(&ep->lock, flags);
999 /* Copy the data member from inside the lock */
1000 epi->event.data = event->data;
1003 * If the item is "hot" and it is not registered inside the ready
1004 * list, push it inside.
1006 if (revents & event->events) {
1007 if (!ep_is_linked(&epi->rdllink)) {
1008 list_add_tail(&epi->rdllink, &ep->rdllist);
1010 /* Notify waiting tasks that events are available */
1011 if (waitqueue_active(&ep->wq))
1012 wake_up_locked(&ep->wq);
1013 if (waitqueue_active(&ep->poll_wait))
1017 spin_unlock_irqrestore(&ep->lock, flags);
1019 /* We have to call this outside the lock */
1021 ep_poll_safewake(&ep->poll_wait);
1026 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1029 struct ep_send_events_data *esed = priv;
1031 unsigned int revents;
1033 struct epoll_event __user *uevent;
1036 * We can loop without lock because we are passed a task private list.
1037 * Items cannot vanish during the loop because ep_scan_ready_list() is
1038 * holding "mtx" during this call.
1040 for (eventcnt = 0, uevent = esed->events;
1041 !list_empty(head) && eventcnt < esed->maxevents;) {
1042 epi = list_first_entry(head, struct epitem, rdllink);
1044 list_del_init(&epi->rdllink);
1046 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1050 * If the event mask intersect the caller-requested one,
1051 * deliver the event to userspace. Again, ep_scan_ready_list()
1052 * is holding "mtx", so no operations coming from userspace
1053 * can change the item.
1056 if (__put_user(revents, &uevent->events) ||
1057 __put_user(epi->event.data, &uevent->data))
1058 return eventcnt ? eventcnt : -EFAULT;
1061 if (epi->event.events & EPOLLONESHOT)
1062 epi->event.events &= EP_PRIVATE_BITS;
1063 else if (!(epi->event.events & EPOLLET)) {
1065 * If this file has been added with Level
1066 * Trigger mode, we need to insert back inside
1067 * the ready list, so that the next call to
1068 * epoll_wait() will check again the events
1069 * availability. At this point, noone can insert
1070 * into ep->rdllist besides us. The epoll_ctl()
1071 * callers are locked out by
1072 * ep_scan_ready_list() holding "mtx" and the
1073 * poll callback will queue them in ep->ovflist.
1075 list_add_tail(&epi->rdllink, &ep->rdllist);
1083 static int ep_send_events(struct eventpoll *ep,
1084 struct epoll_event __user *events, int maxevents)
1086 struct ep_send_events_data esed;
1088 esed.maxevents = maxevents;
1089 esed.events = events;
1091 return ep_scan_ready_list(ep, ep_send_events_proc, &esed);
1094 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1095 int maxevents, long timeout)
1098 unsigned long flags;
1103 * Calculate the timeout by checking for the "infinite" value (-1)
1104 * and the overflow condition. The passed timeout is in milliseconds,
1105 * that why (t * HZ) / 1000.
1107 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1108 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1111 spin_lock_irqsave(&ep->lock, flags);
1114 if (list_empty(&ep->rdllist)) {
1116 * We don't have any available event to return to the caller.
1117 * We need to sleep here, and we will be wake up by
1118 * ep_poll_callback() when events will become available.
1120 init_waitqueue_entry(&wait, current);
1121 wait.flags |= WQ_FLAG_EXCLUSIVE;
1122 __add_wait_queue(&ep->wq, &wait);
1126 * We don't want to sleep if the ep_poll_callback() sends us
1127 * a wakeup in between. That's why we set the task state
1128 * to TASK_INTERRUPTIBLE before doing the checks.
1130 set_current_state(TASK_INTERRUPTIBLE);
1131 if (!list_empty(&ep->rdllist) || !jtimeout)
1133 if (signal_pending(current)) {
1138 spin_unlock_irqrestore(&ep->lock, flags);
1139 jtimeout = schedule_timeout(jtimeout);
1140 spin_lock_irqsave(&ep->lock, flags);
1142 __remove_wait_queue(&ep->wq, &wait);
1144 set_current_state(TASK_RUNNING);
1146 /* Is it worth to try to dig for events ? */
1147 eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
1149 spin_unlock_irqrestore(&ep->lock, flags);
1152 * Try to transfer events to user space. In case we get 0 events and
1153 * there's still timeout left over, we go trying again in search of
1156 if (!res && eavail &&
1157 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1164 * Open an eventpoll file descriptor.
1166 SYSCALL_DEFINE1(epoll_create1, int, flags)
1169 struct eventpoll *ep = NULL;
1171 /* Check the EPOLL_* constant for consistency. */
1172 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1174 if (flags & ~EPOLL_CLOEXEC)
1177 * Create the internal data structure ("struct eventpoll").
1179 error = ep_alloc(&ep);
1183 * Creates all the items needed to setup an eventpoll file. That is,
1184 * a file structure and a free file descriptor.
1186 error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1194 SYSCALL_DEFINE1(epoll_create, int, size)
1199 return sys_epoll_create1(0);
1203 * The following function implements the controller interface for
1204 * the eventpoll file that enables the insertion/removal/change of
1205 * file descriptors inside the interest set.
1207 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1208 struct epoll_event __user *, event)
1211 struct file *file, *tfile;
1212 struct eventpoll *ep;
1214 struct epoll_event epds;
1217 if (ep_op_has_event(op) &&
1218 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1221 /* Get the "struct file *" for the eventpoll file */
1227 /* Get the "struct file *" for the target file */
1232 /* The target file descriptor must support poll */
1234 if (!tfile->f_op || !tfile->f_op->poll)
1235 goto error_tgt_fput;
1238 * We have to check that the file structure underneath the file descriptor
1239 * the user passed to us _is_ an eventpoll file. And also we do not permit
1240 * adding an epoll file descriptor inside itself.
1243 if (file == tfile || !is_file_epoll(file))
1244 goto error_tgt_fput;
1247 * At this point it is safe to assume that the "private_data" contains
1248 * our own data structure.
1250 ep = file->private_data;
1252 mutex_lock(&ep->mtx);
1255 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1256 * above, we can be sure to be able to use the item looked up by
1257 * ep_find() till we release the mutex.
1259 epi = ep_find(ep, tfile, fd);
1265 epds.events |= POLLERR | POLLHUP;
1267 error = ep_insert(ep, &epds, tfile, fd);
1273 error = ep_remove(ep, epi);
1279 epds.events |= POLLERR | POLLHUP;
1280 error = ep_modify(ep, epi, &epds);
1285 mutex_unlock(&ep->mtx);
1297 * Implement the event wait interface for the eventpoll file. It is the kernel
1298 * part of the user space epoll_wait(2).
1300 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1301 int, maxevents, int, timeout)
1305 struct eventpoll *ep;
1307 /* The maximum number of event must be greater than zero */
1308 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1311 /* Verify that the area passed by the user is writeable */
1312 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1317 /* Get the "struct file *" for the eventpoll file */
1324 * We have to check that the file structure underneath the fd
1325 * the user passed to us _is_ an eventpoll file.
1328 if (!is_file_epoll(file))
1332 * At this point it is safe to assume that the "private_data" contains
1333 * our own data structure.
1335 ep = file->private_data;
1337 /* Time to fish for events ... */
1338 error = ep_poll(ep, events, maxevents, timeout);
1347 #ifdef HAVE_SET_RESTORE_SIGMASK
1350 * Implement the event wait interface for the eventpoll file. It is the kernel
1351 * part of the user space epoll_pwait(2).
1353 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1354 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1358 sigset_t ksigmask, sigsaved;
1361 * If the caller wants a certain signal mask to be set during the wait,
1365 if (sigsetsize != sizeof(sigset_t))
1367 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1369 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1370 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1373 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1376 * If we changed the signal mask, we need to restore the original one.
1377 * In case we've got a signal while waiting, we do not restore the
1378 * signal mask yet, and we allow do_signal() to deliver the signal on
1379 * the way back to userspace, before the signal mask is restored.
1382 if (error == -EINTR) {
1383 memcpy(¤t->saved_sigmask, &sigsaved,
1385 set_restore_sigmask();
1387 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1393 #endif /* HAVE_SET_RESTORE_SIGMASK */
1395 static int __init eventpoll_init(void)
1401 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1403 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1406 /* Initialize the structure used to perform safe poll wait head wake ups */
1407 ep_nested_calls_init(&poll_safewake_ncalls);
1409 /* Initialize the structure used to perform file's f_op->poll() calls */
1410 ep_nested_calls_init(&poll_readywalk_ncalls);
1412 /* Allocates slab cache used to allocate "struct epitem" items */
1413 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1414 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1416 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1417 pwq_cache = kmem_cache_create("eventpoll_pwq",
1418 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1422 fs_initcall(eventpoll_init);