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 <linux/device.h>
37 #include <linux/freezer.h>
38 #include <asm/uaccess.h>
41 #include <linux/atomic.h>
42 #include <linux/proc_fs.h>
43 #include <linux/seq_file.h>
44 #include <linux/compat.h>
45 #include <linux/rculist.h>
49 * There are three level of locking required by epoll :
53 * 3) ep->lock (spinlock)
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
63 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
67 * This mutex is acquired by ep_free() during the epoll file
68 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
78 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
87 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
90 * Events that require holding "epmutex" are very rare, while for
91 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
95 /* Epoll private bits inside the event mask */
96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EP_MAX_NESTS 4
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
103 #define EP_UNACTIVE_PTR ((void *) -1L)
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
107 struct epoll_filefd {
113 * Structure used to track possible nested calls, for too deep recursions
116 struct nested_call_node {
117 struct list_head llink;
123 * This structure is used as collector for nested calls, to check for
124 * maximum recursion dept and loop cycles.
126 struct nested_calls {
127 struct list_head tasks_call_list;
132 * Each file descriptor added to the eventpoll interface will
133 * have an entry of this type linked to the "rbr" RB tree.
134 * Avoid increasing the size of this struct, there can be many thousands
135 * of these on a server and we do not want this to take another cache line.
139 /* RB tree node links this structure to the eventpoll RB tree */
141 /* Used to free the struct epitem */
145 /* List header used to link this structure to the eventpoll ready list */
146 struct list_head rdllink;
149 * Works together "struct eventpoll"->ovflist in keeping the
150 * single linked chain of items.
154 /* The file descriptor information this item refers to */
155 struct epoll_filefd ffd;
157 /* Number of active wait queue attached to poll operations */
160 /* List containing poll wait queues */
161 struct list_head pwqlist;
163 /* The "container" of this item */
164 struct eventpoll *ep;
166 /* List header used to link this item to the "struct file" items list */
167 struct list_head fllink;
169 /* wakeup_source used when EPOLLWAKEUP is set */
170 struct wakeup_source __rcu *ws;
172 /* The structure that describe the interested events and the source fd */
173 struct epoll_event event;
177 * This structure is stored inside the "private_data" member of the file
178 * structure and represents the main data structure for the eventpoll
182 /* Protect the access to this structure */
186 * This mutex is used to ensure that files are not removed
187 * while epoll is using them. This is held during the event
188 * collection loop, the file cleanup path, the epoll file exit
189 * code and the ctl operations.
193 /* Wait queue used by sys_epoll_wait() */
194 wait_queue_head_t wq;
196 /* Wait queue used by file->poll() */
197 wait_queue_head_t poll_wait;
199 /* List of ready file descriptors */
200 struct list_head rdllist;
202 /* RB tree root used to store monitored fd structs */
206 * This is a single linked list that chains all the "struct epitem" that
207 * happened while transferring ready events to userspace w/out
210 struct epitem *ovflist;
212 /* wakeup_source used when ep_scan_ready_list is running */
213 struct wakeup_source *ws;
215 /* The user that created the eventpoll descriptor */
216 struct user_struct *user;
220 /* used to optimize loop detection check */
222 struct list_head visited_list_link;
225 /* Wait structure used by the poll hooks */
226 struct eppoll_entry {
227 /* List header used to link this structure to the "struct epitem" */
228 struct list_head llink;
230 /* The "base" pointer is set to the container "struct epitem" */
234 * Wait queue item that will be linked to the target file wait
239 /* The wait queue head that linked the "wait" wait queue item */
240 wait_queue_head_t *whead;
243 /* Wrapper struct used by poll queueing */
249 /* Used by the ep_send_events() function as callback private data */
250 struct ep_send_events_data {
252 struct epoll_event __user *events;
256 * Configuration options available inside /proc/sys/fs/epoll/
258 /* Maximum number of epoll watched descriptors, per user */
259 static long max_user_watches __read_mostly;
262 * This mutex is used to serialize ep_free() and eventpoll_release_file().
264 static DEFINE_MUTEX(epmutex);
266 /* Used to check for epoll file descriptor inclusion loops */
267 static struct nested_calls poll_loop_ncalls;
269 /* Used for safe wake up implementation */
270 static struct nested_calls poll_safewake_ncalls;
272 /* Used to call file's f_op->poll() under the nested calls boundaries */
273 static struct nested_calls poll_readywalk_ncalls;
275 /* Slab cache used to allocate "struct epitem" */
276 static struct kmem_cache *epi_cache __read_mostly;
278 /* Slab cache used to allocate "struct eppoll_entry" */
279 static struct kmem_cache *pwq_cache __read_mostly;
281 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
282 static LIST_HEAD(visited_list);
285 * List of files with newly added links, where we may need to limit the number
286 * of emanating paths. Protected by the epmutex.
288 static LIST_HEAD(tfile_check_list);
292 #include <linux/sysctl.h>
295 static long long_max = LONG_MAX;
297 struct ctl_table epoll_table[] = {
299 .procname = "max_user_watches",
300 .data = &max_user_watches,
301 .maxlen = sizeof(max_user_watches),
303 .proc_handler = proc_doulongvec_minmax,
309 #endif /* CONFIG_SYSCTL */
311 static const struct file_operations eventpoll_fops;
313 static inline int is_file_epoll(struct file *f)
315 return f->f_op == &eventpoll_fops;
318 /* Setup the structure that is used as key for the RB tree */
319 static inline void ep_set_ffd(struct epoll_filefd *ffd,
320 struct file *file, int fd)
326 /* Compare RB tree keys */
327 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
328 struct epoll_filefd *p2)
330 return (p1->file > p2->file ? +1:
331 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
334 /* Tells us if the item is currently linked */
335 static inline int ep_is_linked(struct list_head *p)
337 return !list_empty(p);
340 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
342 return container_of(p, struct eppoll_entry, wait);
345 /* Get the "struct epitem" from a wait queue pointer */
346 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
348 return container_of(p, struct eppoll_entry, wait)->base;
351 /* Get the "struct epitem" from an epoll queue wrapper */
352 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
354 return container_of(p, struct ep_pqueue, pt)->epi;
357 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
358 static inline int ep_op_has_event(int op)
360 return op != EPOLL_CTL_DEL;
363 /* Initialize the poll safe wake up structure */
364 static void ep_nested_calls_init(struct nested_calls *ncalls)
366 INIT_LIST_HEAD(&ncalls->tasks_call_list);
367 spin_lock_init(&ncalls->lock);
371 * ep_events_available - Checks if ready events might be available.
373 * @ep: Pointer to the eventpoll context.
375 * Returns: Returns a value different than zero if ready events are available,
378 static inline int ep_events_available(struct eventpoll *ep)
380 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
384 * ep_call_nested - Perform a bound (possibly) nested call, by checking
385 * that the recursion limit is not exceeded, and that
386 * the same nested call (by the meaning of same cookie) is
389 * @ncalls: Pointer to the nested_calls structure to be used for this call.
390 * @max_nests: Maximum number of allowed nesting calls.
391 * @nproc: Nested call core function pointer.
392 * @priv: Opaque data to be passed to the @nproc callback.
393 * @cookie: Cookie to be used to identify this nested call.
394 * @ctx: This instance context.
396 * Returns: Returns the code returned by the @nproc callback, or -1 if
397 * the maximum recursion limit has been exceeded.
399 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
400 int (*nproc)(void *, void *, int), void *priv,
401 void *cookie, void *ctx)
403 int error, call_nests = 0;
405 struct list_head *lsthead = &ncalls->tasks_call_list;
406 struct nested_call_node *tncur;
407 struct nested_call_node tnode;
409 spin_lock_irqsave(&ncalls->lock, flags);
412 * Try to see if the current task is already inside this wakeup call.
413 * We use a list here, since the population inside this set is always
416 list_for_each_entry(tncur, lsthead, llink) {
417 if (tncur->ctx == ctx &&
418 (tncur->cookie == cookie || ++call_nests > max_nests)) {
420 * Ops ... loop detected or maximum nest level reached.
421 * We abort this wake by breaking the cycle itself.
428 /* Add the current task and cookie to the list */
430 tnode.cookie = cookie;
431 list_add(&tnode.llink, lsthead);
433 spin_unlock_irqrestore(&ncalls->lock, flags);
435 /* Call the nested function */
436 error = (*nproc)(priv, cookie, call_nests);
438 /* Remove the current task from the list */
439 spin_lock_irqsave(&ncalls->lock, flags);
440 list_del(&tnode.llink);
442 spin_unlock_irqrestore(&ncalls->lock, flags);
448 * As described in commit 0ccf831cb lockdep: annotate epoll
449 * the use of wait queues used by epoll is done in a very controlled
450 * manner. Wake ups can nest inside each other, but are never done
451 * with the same locking. For example:
454 * efd1 = epoll_create();
455 * efd2 = epoll_create();
456 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
457 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
459 * When a packet arrives to the device underneath "dfd", the net code will
460 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
461 * callback wakeup entry on that queue, and the wake_up() performed by the
462 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
463 * (efd1) notices that it may have some event ready, so it needs to wake up
464 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
465 * that ends up in another wake_up(), after having checked about the
466 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
467 * avoid stack blasting.
469 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
470 * this special case of epoll.
472 #ifdef CONFIG_DEBUG_LOCK_ALLOC
473 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
474 unsigned long events, int subclass)
478 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
479 wake_up_locked_poll(wqueue, events);
480 spin_unlock_irqrestore(&wqueue->lock, flags);
483 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
484 unsigned long events, int subclass)
486 wake_up_poll(wqueue, events);
490 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
492 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
498 * Perform a safe wake up of the poll wait list. The problem is that
499 * with the new callback'd wake up system, it is possible that the
500 * poll callback is reentered from inside the call to wake_up() done
501 * on the poll wait queue head. The rule is that we cannot reenter the
502 * wake up code from the same task more than EP_MAX_NESTS times,
503 * and we cannot reenter the same wait queue head at all. This will
504 * enable to have a hierarchy of epoll file descriptor of no more than
507 static void ep_poll_safewake(wait_queue_head_t *wq)
509 int this_cpu = get_cpu();
511 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
512 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
517 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
519 wait_queue_head_t *whead;
522 /* If it is cleared by POLLFREE, it should be rcu-safe */
523 whead = rcu_dereference(pwq->whead);
525 remove_wait_queue(whead, &pwq->wait);
530 * This function unregisters poll callbacks from the associated file
531 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
534 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
536 struct list_head *lsthead = &epi->pwqlist;
537 struct eppoll_entry *pwq;
539 while (!list_empty(lsthead)) {
540 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
542 list_del(&pwq->llink);
543 ep_remove_wait_queue(pwq);
544 kmem_cache_free(pwq_cache, pwq);
548 /* call only when ep->mtx is held */
549 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
551 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
554 /* call only when ep->mtx is held */
555 static inline void ep_pm_stay_awake(struct epitem *epi)
557 struct wakeup_source *ws = ep_wakeup_source(epi);
563 static inline bool ep_has_wakeup_source(struct epitem *epi)
565 return rcu_access_pointer(epi->ws) ? true : false;
568 /* call when ep->mtx cannot be held (ep_poll_callback) */
569 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
571 struct wakeup_source *ws;
574 ws = rcu_dereference(epi->ws);
581 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
582 * the scan code, to call f_op->poll(). Also allows for
583 * O(NumReady) performance.
585 * @ep: Pointer to the epoll private data structure.
586 * @sproc: Pointer to the scan callback.
587 * @priv: Private opaque data passed to the @sproc callback.
588 * @depth: The current depth of recursive f_op->poll calls.
589 * @ep_locked: caller already holds ep->mtx
591 * Returns: The same integer error code returned by the @sproc callback.
593 static int ep_scan_ready_list(struct eventpoll *ep,
594 int (*sproc)(struct eventpoll *,
595 struct list_head *, void *),
596 void *priv, int depth, bool ep_locked)
598 int error, pwake = 0;
600 struct epitem *epi, *nepi;
604 * We need to lock this because we could be hit by
605 * eventpoll_release_file() and epoll_ctl().
609 mutex_lock_nested(&ep->mtx, depth);
612 * Steal the ready list, and re-init the original one to the
613 * empty list. Also, set ep->ovflist to NULL so that events
614 * happening while looping w/out locks, are not lost. We cannot
615 * have the poll callback to queue directly on ep->rdllist,
616 * because we want the "sproc" callback to be able to do it
619 spin_lock_irqsave(&ep->lock, flags);
620 list_splice_init(&ep->rdllist, &txlist);
622 spin_unlock_irqrestore(&ep->lock, flags);
625 * Now call the callback function.
627 error = (*sproc)(ep, &txlist, priv);
629 spin_lock_irqsave(&ep->lock, flags);
631 * During the time we spent inside the "sproc" callback, some
632 * other events might have been queued by the poll callback.
633 * We re-insert them inside the main ready-list here.
635 for (nepi = ep->ovflist; (epi = nepi) != NULL;
636 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
638 * We need to check if the item is already in the list.
639 * During the "sproc" callback execution time, items are
640 * queued into ->ovflist but the "txlist" might already
641 * contain them, and the list_splice() below takes care of them.
643 if (!ep_is_linked(&epi->rdllink)) {
644 list_add_tail(&epi->rdllink, &ep->rdllist);
645 ep_pm_stay_awake(epi);
649 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
650 * releasing the lock, events will be queued in the normal way inside
653 ep->ovflist = EP_UNACTIVE_PTR;
656 * Quickly re-inject items left on "txlist".
658 list_splice(&txlist, &ep->rdllist);
661 if (!list_empty(&ep->rdllist)) {
663 * Wake up (if active) both the eventpoll wait list and
664 * the ->poll() wait list (delayed after we release the lock).
666 if (waitqueue_active(&ep->wq))
667 wake_up_locked(&ep->wq);
668 if (waitqueue_active(&ep->poll_wait))
671 spin_unlock_irqrestore(&ep->lock, flags);
674 mutex_unlock(&ep->mtx);
676 /* We have to call this outside the lock */
678 ep_poll_safewake(&ep->poll_wait);
683 static void epi_rcu_free(struct rcu_head *head)
685 struct epitem *epi = container_of(head, struct epitem, rcu);
686 kmem_cache_free(epi_cache, epi);
690 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
691 * all the associated resources. Must be called with "mtx" held.
693 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
696 struct file *file = epi->ffd.file;
699 * Removes poll wait queue hooks. We _have_ to do this without holding
700 * the "ep->lock" otherwise a deadlock might occur. This because of the
701 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
702 * queue head lock when unregistering the wait queue. The wakeup callback
703 * will run by holding the wait queue head lock and will call our callback
704 * that will try to get "ep->lock".
706 ep_unregister_pollwait(ep, epi);
708 /* Remove the current item from the list of epoll hooks */
709 spin_lock(&file->f_lock);
710 list_del_rcu(&epi->fllink);
711 spin_unlock(&file->f_lock);
713 rb_erase(&epi->rbn, &ep->rbr);
715 spin_lock_irqsave(&ep->lock, flags);
716 if (ep_is_linked(&epi->rdllink))
717 list_del_init(&epi->rdllink);
718 spin_unlock_irqrestore(&ep->lock, flags);
720 wakeup_source_unregister(ep_wakeup_source(epi));
722 * At this point it is safe to free the eventpoll item. Use the union
723 * field epi->rcu, since we are trying to minimize the size of
724 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
725 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
726 * use of the rbn field.
728 call_rcu(&epi->rcu, epi_rcu_free);
730 atomic_long_dec(&ep->user->epoll_watches);
735 static void ep_free(struct eventpoll *ep)
740 /* We need to release all tasks waiting for these file */
741 if (waitqueue_active(&ep->poll_wait))
742 ep_poll_safewake(&ep->poll_wait);
745 * We need to lock this because we could be hit by
746 * eventpoll_release_file() while we're freeing the "struct eventpoll".
747 * We do not need to hold "ep->mtx" here because the epoll file
748 * is on the way to be removed and no one has references to it
749 * anymore. The only hit might come from eventpoll_release_file() but
750 * holding "epmutex" is sufficient here.
752 mutex_lock(&epmutex);
755 * Walks through the whole tree by unregistering poll callbacks.
757 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
758 epi = rb_entry(rbp, struct epitem, rbn);
760 ep_unregister_pollwait(ep, epi);
765 * Walks through the whole tree by freeing each "struct epitem". At this
766 * point we are sure no poll callbacks will be lingering around, and also by
767 * holding "epmutex" we can be sure that no file cleanup code will hit
768 * us during this operation. So we can avoid the lock on "ep->lock".
769 * We do not need to lock ep->mtx, either, we only do it to prevent
772 mutex_lock(&ep->mtx);
773 while ((rbp = rb_first(&ep->rbr)) != NULL) {
774 epi = rb_entry(rbp, struct epitem, rbn);
778 mutex_unlock(&ep->mtx);
780 mutex_unlock(&epmutex);
781 mutex_destroy(&ep->mtx);
783 wakeup_source_unregister(ep->ws);
787 static int ep_eventpoll_release(struct inode *inode, struct file *file)
789 struct eventpoll *ep = file->private_data;
797 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
799 pt->_key = epi->event.events;
801 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
804 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
807 struct epitem *epi, *tmp;
810 init_poll_funcptr(&pt, NULL);
812 list_for_each_entry_safe(epi, tmp, head, rdllink) {
813 if (ep_item_poll(epi, &pt))
814 return POLLIN | POLLRDNORM;
817 * Item has been dropped into the ready list by the poll
818 * callback, but it's not actually ready, as far as
819 * caller requested events goes. We can remove it here.
821 __pm_relax(ep_wakeup_source(epi));
822 list_del_init(&epi->rdllink);
829 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
832 struct readyevents_arg {
833 struct eventpoll *ep;
837 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
839 struct readyevents_arg *arg = priv;
841 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
842 call_nests + 1, arg->locked);
845 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
848 struct eventpoll *ep = file->private_data;
849 struct readyevents_arg arg;
852 * During ep_insert() we already hold the ep->mtx for the tfile.
853 * Prevent re-aquisition.
855 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
858 /* Insert inside our poll wait queue */
859 poll_wait(file, &ep->poll_wait, wait);
862 * Proceed to find out if wanted events are really available inside
863 * the ready list. This need to be done under ep_call_nested()
864 * supervision, since the call to f_op->poll() done on listed files
865 * could re-enter here.
867 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
868 ep_poll_readyevents_proc, &arg, ep, current);
870 return pollflags != -1 ? pollflags : 0;
873 #ifdef CONFIG_PROC_FS
874 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
876 struct eventpoll *ep = f->private_data;
879 mutex_lock(&ep->mtx);
880 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
881 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
883 seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
884 epi->ffd.fd, epi->event.events,
885 (long long)epi->event.data);
886 if (seq_has_overflowed(m))
889 mutex_unlock(&ep->mtx);
893 /* File callbacks that implement the eventpoll file behaviour */
894 static const struct file_operations eventpoll_fops = {
895 #ifdef CONFIG_PROC_FS
896 .show_fdinfo = ep_show_fdinfo,
898 .release = ep_eventpoll_release,
899 .poll = ep_eventpoll_poll,
900 .llseek = noop_llseek,
904 * This is called from eventpoll_release() to unlink files from the eventpoll
905 * interface. We need to have this facility to cleanup correctly files that are
906 * closed without being removed from the eventpoll interface.
908 void eventpoll_release_file(struct file *file)
910 struct eventpoll *ep;
911 struct epitem *epi, *next;
914 * We don't want to get "file->f_lock" because it is not
915 * necessary. It is not necessary because we're in the "struct file"
916 * cleanup path, and this means that no one is using this file anymore.
917 * So, for example, epoll_ctl() cannot hit here since if we reach this
918 * point, the file counter already went to zero and fget() would fail.
919 * The only hit might come from ep_free() but by holding the mutex
920 * will correctly serialize the operation. We do need to acquire
921 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
922 * from anywhere but ep_free().
924 * Besides, ep_remove() acquires the lock, so we can't hold it here.
926 mutex_lock(&epmutex);
927 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
929 mutex_lock_nested(&ep->mtx, 0);
931 mutex_unlock(&ep->mtx);
933 mutex_unlock(&epmutex);
936 static int ep_alloc(struct eventpoll **pep)
939 struct user_struct *user;
940 struct eventpoll *ep;
942 user = get_current_user();
944 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
948 spin_lock_init(&ep->lock);
949 mutex_init(&ep->mtx);
950 init_waitqueue_head(&ep->wq);
951 init_waitqueue_head(&ep->poll_wait);
952 INIT_LIST_HEAD(&ep->rdllist);
954 ep->ovflist = EP_UNACTIVE_PTR;
967 * Search the file inside the eventpoll tree. The RB tree operations
968 * are protected by the "mtx" mutex, and ep_find() must be called with
971 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
975 struct epitem *epi, *epir = NULL;
976 struct epoll_filefd ffd;
978 ep_set_ffd(&ffd, file, fd);
979 for (rbp = ep->rbr.rb_node; rbp; ) {
980 epi = rb_entry(rbp, struct epitem, rbn);
981 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
996 * This is the callback that is passed to the wait queue wakeup
997 * mechanism. It is called by the stored file descriptors when they
998 * have events to report.
1000 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1003 unsigned long flags;
1004 struct epitem *epi = ep_item_from_wait(wait);
1005 struct eventpoll *ep = epi->ep;
1007 if ((unsigned long)key & POLLFREE) {
1008 ep_pwq_from_wait(wait)->whead = NULL;
1010 * whead = NULL above can race with ep_remove_wait_queue()
1011 * which can do another remove_wait_queue() after us, so we
1012 * can't use __remove_wait_queue(). whead->lock is held by
1015 list_del_init(&wait->task_list);
1018 spin_lock_irqsave(&ep->lock, flags);
1021 * If the event mask does not contain any poll(2) event, we consider the
1022 * descriptor to be disabled. This condition is likely the effect of the
1023 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1024 * until the next EPOLL_CTL_MOD will be issued.
1026 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1030 * Check the events coming with the callback. At this stage, not
1031 * every device reports the events in the "key" parameter of the
1032 * callback. We need to be able to handle both cases here, hence the
1033 * test for "key" != NULL before the event match test.
1035 if (key && !((unsigned long) key & epi->event.events))
1039 * If we are transferring events to userspace, we can hold no locks
1040 * (because we're accessing user memory, and because of linux f_op->poll()
1041 * semantics). All the events that happen during that period of time are
1042 * chained in ep->ovflist and requeued later on.
1044 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1045 if (epi->next == EP_UNACTIVE_PTR) {
1046 epi->next = ep->ovflist;
1050 * Activate ep->ws since epi->ws may get
1051 * deactivated at any time.
1053 __pm_stay_awake(ep->ws);
1060 /* If this file is already in the ready list we exit soon */
1061 if (!ep_is_linked(&epi->rdllink)) {
1062 list_add_tail(&epi->rdllink, &ep->rdllist);
1063 ep_pm_stay_awake_rcu(epi);
1067 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1070 if (waitqueue_active(&ep->wq))
1071 wake_up_locked(&ep->wq);
1072 if (waitqueue_active(&ep->poll_wait))
1076 spin_unlock_irqrestore(&ep->lock, flags);
1078 /* We have to call this outside the lock */
1080 ep_poll_safewake(&ep->poll_wait);
1086 * This is the callback that is used to add our wait queue to the
1087 * target file wakeup lists.
1089 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1092 struct epitem *epi = ep_item_from_epqueue(pt);
1093 struct eppoll_entry *pwq;
1095 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1096 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1099 add_wait_queue(whead, &pwq->wait);
1100 list_add_tail(&pwq->llink, &epi->pwqlist);
1103 /* We have to signal that an error occurred */
1108 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1111 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1112 struct epitem *epic;
1116 epic = rb_entry(parent, struct epitem, rbn);
1117 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1119 p = &parent->rb_right;
1121 p = &parent->rb_left;
1123 rb_link_node(&epi->rbn, parent, p);
1124 rb_insert_color(&epi->rbn, &ep->rbr);
1129 #define PATH_ARR_SIZE 5
1131 * These are the number paths of length 1 to 5, that we are allowing to emanate
1132 * from a single file of interest. For example, we allow 1000 paths of length
1133 * 1, to emanate from each file of interest. This essentially represents the
1134 * potential wakeup paths, which need to be limited in order to avoid massive
1135 * uncontrolled wakeup storms. The common use case should be a single ep which
1136 * is connected to n file sources. In this case each file source has 1 path
1137 * of length 1. Thus, the numbers below should be more than sufficient. These
1138 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1139 * and delete can't add additional paths. Protected by the epmutex.
1141 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1142 static int path_count[PATH_ARR_SIZE];
1144 static int path_count_inc(int nests)
1146 /* Allow an arbitrary number of depth 1 paths */
1150 if (++path_count[nests] > path_limits[nests])
1155 static void path_count_init(void)
1159 for (i = 0; i < PATH_ARR_SIZE; i++)
1163 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1166 struct file *file = priv;
1167 struct file *child_file;
1170 /* CTL_DEL can remove links here, but that can't increase our count */
1172 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1173 child_file = epi->ep->file;
1174 if (is_file_epoll(child_file)) {
1175 if (list_empty(&child_file->f_ep_links)) {
1176 if (path_count_inc(call_nests)) {
1181 error = ep_call_nested(&poll_loop_ncalls,
1183 reverse_path_check_proc,
1184 child_file, child_file,
1190 printk(KERN_ERR "reverse_path_check_proc: "
1191 "file is not an ep!\n");
1199 * reverse_path_check - The tfile_check_list is list of file *, which have
1200 * links that are proposed to be newly added. We need to
1201 * make sure that those added links don't add too many
1202 * paths such that we will spend all our time waking up
1203 * eventpoll objects.
1205 * Returns: Returns zero if the proposed links don't create too many paths,
1208 static int reverse_path_check(void)
1211 struct file *current_file;
1213 /* let's call this for all tfiles */
1214 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1216 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1217 reverse_path_check_proc, current_file,
1218 current_file, current);
1225 static int ep_create_wakeup_source(struct epitem *epi)
1228 struct wakeup_source *ws;
1231 epi->ep->ws = wakeup_source_register("eventpoll");
1236 name = epi->ffd.file->f_path.dentry->d_name.name;
1237 ws = wakeup_source_register(name);
1241 rcu_assign_pointer(epi->ws, ws);
1246 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1247 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1249 struct wakeup_source *ws = ep_wakeup_source(epi);
1251 RCU_INIT_POINTER(epi->ws, NULL);
1254 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1255 * used internally by wakeup_source_remove, too (called by
1256 * wakeup_source_unregister), so we cannot use call_rcu
1259 wakeup_source_unregister(ws);
1263 * Must be called with "mtx" held.
1265 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1266 struct file *tfile, int fd, int full_check)
1268 int error, revents, pwake = 0;
1269 unsigned long flags;
1272 struct ep_pqueue epq;
1274 user_watches = atomic_long_read(&ep->user->epoll_watches);
1275 if (unlikely(user_watches >= max_user_watches))
1277 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1280 /* Item initialization follow here ... */
1281 INIT_LIST_HEAD(&epi->rdllink);
1282 INIT_LIST_HEAD(&epi->fllink);
1283 INIT_LIST_HEAD(&epi->pwqlist);
1285 ep_set_ffd(&epi->ffd, tfile, fd);
1286 epi->event = *event;
1288 epi->next = EP_UNACTIVE_PTR;
1289 if (epi->event.events & EPOLLWAKEUP) {
1290 error = ep_create_wakeup_source(epi);
1292 goto error_create_wakeup_source;
1294 RCU_INIT_POINTER(epi->ws, NULL);
1297 /* Initialize the poll table using the queue callback */
1299 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1302 * Attach the item to the poll hooks and get current event bits.
1303 * We can safely use the file* here because its usage count has
1304 * been increased by the caller of this function. Note that after
1305 * this operation completes, the poll callback can start hitting
1308 revents = ep_item_poll(epi, &epq.pt);
1311 * We have to check if something went wrong during the poll wait queue
1312 * install process. Namely an allocation for a wait queue failed due
1313 * high memory pressure.
1317 goto error_unregister;
1319 /* Add the current item to the list of active epoll hook for this file */
1320 spin_lock(&tfile->f_lock);
1321 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1322 spin_unlock(&tfile->f_lock);
1325 * Add the current item to the RB tree. All RB tree operations are
1326 * protected by "mtx", and ep_insert() is called with "mtx" held.
1328 ep_rbtree_insert(ep, epi);
1330 /* now check if we've created too many backpaths */
1332 if (full_check && reverse_path_check())
1333 goto error_remove_epi;
1335 /* We have to drop the new item inside our item list to keep track of it */
1336 spin_lock_irqsave(&ep->lock, flags);
1338 /* If the file is already "ready" we drop it inside the ready list */
1339 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1340 list_add_tail(&epi->rdllink, &ep->rdllist);
1341 ep_pm_stay_awake(epi);
1343 /* Notify waiting tasks that events are available */
1344 if (waitqueue_active(&ep->wq))
1345 wake_up_locked(&ep->wq);
1346 if (waitqueue_active(&ep->poll_wait))
1350 spin_unlock_irqrestore(&ep->lock, flags);
1352 atomic_long_inc(&ep->user->epoll_watches);
1354 /* We have to call this outside the lock */
1356 ep_poll_safewake(&ep->poll_wait);
1361 spin_lock(&tfile->f_lock);
1362 list_del_rcu(&epi->fllink);
1363 spin_unlock(&tfile->f_lock);
1365 rb_erase(&epi->rbn, &ep->rbr);
1368 ep_unregister_pollwait(ep, epi);
1371 * We need to do this because an event could have been arrived on some
1372 * allocated wait queue. Note that we don't care about the ep->ovflist
1373 * list, since that is used/cleaned only inside a section bound by "mtx".
1374 * And ep_insert() is called with "mtx" held.
1376 spin_lock_irqsave(&ep->lock, flags);
1377 if (ep_is_linked(&epi->rdllink))
1378 list_del_init(&epi->rdllink);
1379 spin_unlock_irqrestore(&ep->lock, flags);
1381 wakeup_source_unregister(ep_wakeup_source(epi));
1383 error_create_wakeup_source:
1384 kmem_cache_free(epi_cache, epi);
1390 * Modify the interest event mask by dropping an event if the new mask
1391 * has a match in the current file status. Must be called with "mtx" held.
1393 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1396 unsigned int revents;
1399 init_poll_funcptr(&pt, NULL);
1402 * Set the new event interest mask before calling f_op->poll();
1403 * otherwise we might miss an event that happens between the
1404 * f_op->poll() call and the new event set registering.
1406 epi->event.events = event->events; /* need barrier below */
1407 epi->event.data = event->data; /* protected by mtx */
1408 if (epi->event.events & EPOLLWAKEUP) {
1409 if (!ep_has_wakeup_source(epi))
1410 ep_create_wakeup_source(epi);
1411 } else if (ep_has_wakeup_source(epi)) {
1412 ep_destroy_wakeup_source(epi);
1416 * The following barrier has two effects:
1418 * 1) Flush epi changes above to other CPUs. This ensures
1419 * we do not miss events from ep_poll_callback if an
1420 * event occurs immediately after we call f_op->poll().
1421 * We need this because we did not take ep->lock while
1422 * changing epi above (but ep_poll_callback does take
1425 * 2) We also need to ensure we do not miss _past_ events
1426 * when calling f_op->poll(). This barrier also
1427 * pairs with the barrier in wq_has_sleeper (see
1428 * comments for wq_has_sleeper).
1430 * This barrier will now guarantee ep_poll_callback or f_op->poll
1431 * (or both) will notice the readiness of an item.
1436 * Get current event bits. We can safely use the file* here because
1437 * its usage count has been increased by the caller of this function.
1439 revents = ep_item_poll(epi, &pt);
1442 * If the item is "hot" and it is not registered inside the ready
1443 * list, push it inside.
1445 if (revents & event->events) {
1446 spin_lock_irq(&ep->lock);
1447 if (!ep_is_linked(&epi->rdllink)) {
1448 list_add_tail(&epi->rdllink, &ep->rdllist);
1449 ep_pm_stay_awake(epi);
1451 /* Notify waiting tasks that events are available */
1452 if (waitqueue_active(&ep->wq))
1453 wake_up_locked(&ep->wq);
1454 if (waitqueue_active(&ep->poll_wait))
1457 spin_unlock_irq(&ep->lock);
1460 /* We have to call this outside the lock */
1462 ep_poll_safewake(&ep->poll_wait);
1467 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1470 struct ep_send_events_data *esed = priv;
1472 unsigned int revents;
1474 struct epoll_event __user *uevent;
1475 struct wakeup_source *ws;
1478 init_poll_funcptr(&pt, NULL);
1481 * We can loop without lock because we are passed a task private list.
1482 * Items cannot vanish during the loop because ep_scan_ready_list() is
1483 * holding "mtx" during this call.
1485 for (eventcnt = 0, uevent = esed->events;
1486 !list_empty(head) && eventcnt < esed->maxevents;) {
1487 epi = list_first_entry(head, struct epitem, rdllink);
1490 * Activate ep->ws before deactivating epi->ws to prevent
1491 * triggering auto-suspend here (in case we reactive epi->ws
1494 * This could be rearranged to delay the deactivation of epi->ws
1495 * instead, but then epi->ws would temporarily be out of sync
1496 * with ep_is_linked().
1498 ws = ep_wakeup_source(epi);
1501 __pm_stay_awake(ep->ws);
1505 list_del_init(&epi->rdllink);
1507 revents = ep_item_poll(epi, &pt);
1510 * If the event mask intersect the caller-requested one,
1511 * deliver the event to userspace. Again, ep_scan_ready_list()
1512 * is holding "mtx", so no operations coming from userspace
1513 * can change the item.
1516 if (__put_user(revents, &uevent->events) ||
1517 __put_user(epi->event.data, &uevent->data)) {
1518 list_add(&epi->rdllink, head);
1519 ep_pm_stay_awake(epi);
1520 return eventcnt ? eventcnt : -EFAULT;
1524 if (epi->event.events & EPOLLONESHOT)
1525 epi->event.events &= EP_PRIVATE_BITS;
1526 else if (!(epi->event.events & EPOLLET)) {
1528 * If this file has been added with Level
1529 * Trigger mode, we need to insert back inside
1530 * the ready list, so that the next call to
1531 * epoll_wait() will check again the events
1532 * availability. At this point, no one can insert
1533 * into ep->rdllist besides us. The epoll_ctl()
1534 * callers are locked out by
1535 * ep_scan_ready_list() holding "mtx" and the
1536 * poll callback will queue them in ep->ovflist.
1538 list_add_tail(&epi->rdllink, &ep->rdllist);
1539 ep_pm_stay_awake(epi);
1547 static int ep_send_events(struct eventpoll *ep,
1548 struct epoll_event __user *events, int maxevents)
1550 struct ep_send_events_data esed;
1552 esed.maxevents = maxevents;
1553 esed.events = events;
1555 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1558 static inline struct timespec ep_set_mstimeout(long ms)
1560 struct timespec now, ts = {
1561 .tv_sec = ms / MSEC_PER_SEC,
1562 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1566 return timespec_add_safe(now, ts);
1570 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1573 * @ep: Pointer to the eventpoll context.
1574 * @events: Pointer to the userspace buffer where the ready events should be
1576 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1577 * @timeout: Maximum timeout for the ready events fetch operation, in
1578 * milliseconds. If the @timeout is zero, the function will not block,
1579 * while if the @timeout is less than zero, the function will block
1580 * until at least one event has been retrieved (or an error
1583 * Returns: Returns the number of ready events which have been fetched, or an
1584 * error code, in case of error.
1586 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1587 int maxevents, long timeout)
1589 int res = 0, eavail, timed_out = 0;
1590 unsigned long flags;
1593 ktime_t expires, *to = NULL;
1596 struct timespec end_time = ep_set_mstimeout(timeout);
1598 slack = select_estimate_accuracy(&end_time);
1600 *to = timespec_to_ktime(end_time);
1601 } else if (timeout == 0) {
1603 * Avoid the unnecessary trip to the wait queue loop, if the
1604 * caller specified a non blocking operation.
1607 spin_lock_irqsave(&ep->lock, flags);
1612 spin_lock_irqsave(&ep->lock, flags);
1614 if (!ep_events_available(ep)) {
1616 * We don't have any available event to return to the caller.
1617 * We need to sleep here, and we will be wake up by
1618 * ep_poll_callback() when events will become available.
1620 init_waitqueue_entry(&wait, current);
1621 __add_wait_queue_exclusive(&ep->wq, &wait);
1625 * We don't want to sleep if the ep_poll_callback() sends us
1626 * a wakeup in between. That's why we set the task state
1627 * to TASK_INTERRUPTIBLE before doing the checks.
1629 set_current_state(TASK_INTERRUPTIBLE);
1630 if (ep_events_available(ep) || timed_out)
1632 if (signal_pending(current)) {
1637 spin_unlock_irqrestore(&ep->lock, flags);
1638 if (!freezable_schedule_hrtimeout_range(to, slack,
1642 spin_lock_irqsave(&ep->lock, flags);
1645 __remove_wait_queue(&ep->wq, &wait);
1646 __set_current_state(TASK_RUNNING);
1649 /* Is it worth to try to dig for events ? */
1650 eavail = ep_events_available(ep);
1652 spin_unlock_irqrestore(&ep->lock, flags);
1655 * Try to transfer events to user space. In case we get 0 events and
1656 * there's still timeout left over, we go trying again in search of
1659 if (!res && eavail &&
1660 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1667 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1668 * API, to verify that adding an epoll file inside another
1669 * epoll structure, does not violate the constraints, in
1670 * terms of closed loops, or too deep chains (which can
1671 * result in excessive stack usage).
1673 * @priv: Pointer to the epoll file to be currently checked.
1674 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1675 * data structure pointer.
1676 * @call_nests: Current dept of the @ep_call_nested() call stack.
1678 * Returns: Returns zero if adding the epoll @file inside current epoll
1679 * structure @ep does not violate the constraints, or -1 otherwise.
1681 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1684 struct file *file = priv;
1685 struct eventpoll *ep = file->private_data;
1686 struct eventpoll *ep_tovisit;
1687 struct rb_node *rbp;
1690 mutex_lock_nested(&ep->mtx, call_nests + 1);
1692 list_add(&ep->visited_list_link, &visited_list);
1693 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1694 epi = rb_entry(rbp, struct epitem, rbn);
1695 if (unlikely(is_file_epoll(epi->ffd.file))) {
1696 ep_tovisit = epi->ffd.file->private_data;
1697 if (ep_tovisit->visited)
1699 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1700 ep_loop_check_proc, epi->ffd.file,
1701 ep_tovisit, current);
1706 * If we've reached a file that is not associated with
1707 * an ep, then we need to check if the newly added
1708 * links are going to add too many wakeup paths. We do
1709 * this by adding it to the tfile_check_list, if it's
1710 * not already there, and calling reverse_path_check()
1711 * during ep_insert().
1713 if (list_empty(&epi->ffd.file->f_tfile_llink))
1714 list_add(&epi->ffd.file->f_tfile_llink,
1718 mutex_unlock(&ep->mtx);
1724 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1725 * another epoll file (represented by @ep) does not create
1726 * closed loops or too deep chains.
1728 * @ep: Pointer to the epoll private data structure.
1729 * @file: Pointer to the epoll file to be checked.
1731 * Returns: Returns zero if adding the epoll @file inside current epoll
1732 * structure @ep does not violate the constraints, or -1 otherwise.
1734 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1737 struct eventpoll *ep_cur, *ep_next;
1739 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1740 ep_loop_check_proc, file, ep, current);
1741 /* clear visited list */
1742 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1743 visited_list_link) {
1744 ep_cur->visited = 0;
1745 list_del(&ep_cur->visited_list_link);
1750 static void clear_tfile_check_list(void)
1754 /* first clear the tfile_check_list */
1755 while (!list_empty(&tfile_check_list)) {
1756 file = list_first_entry(&tfile_check_list, struct file,
1758 list_del_init(&file->f_tfile_llink);
1760 INIT_LIST_HEAD(&tfile_check_list);
1764 * Open an eventpoll file descriptor.
1766 SYSCALL_DEFINE1(epoll_create1, int, flags)
1769 struct eventpoll *ep = NULL;
1772 /* Check the EPOLL_* constant for consistency. */
1773 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1775 if (flags & ~EPOLL_CLOEXEC)
1778 * Create the internal data structure ("struct eventpoll").
1780 error = ep_alloc(&ep);
1784 * Creates all the items needed to setup an eventpoll file. That is,
1785 * a file structure and a free file descriptor.
1787 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1792 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1793 O_RDWR | (flags & O_CLOEXEC));
1795 error = PTR_ERR(file);
1799 fd_install(fd, file);
1809 SYSCALL_DEFINE1(epoll_create, int, size)
1814 return sys_epoll_create1(0);
1818 * The following function implements the controller interface for
1819 * the eventpoll file that enables the insertion/removal/change of
1820 * file descriptors inside the interest set.
1822 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1823 struct epoll_event __user *, event)
1828 struct eventpoll *ep;
1830 struct epoll_event epds;
1831 struct eventpoll *tep = NULL;
1834 if (ep_op_has_event(op) &&
1835 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1843 /* Get the "struct file *" for the target file */
1848 /* The target file descriptor must support poll */
1850 if (!tf.file->f_op->poll)
1851 goto error_tgt_fput;
1853 /* Check if EPOLLWAKEUP is allowed */
1854 if (ep_op_has_event(op))
1855 ep_take_care_of_epollwakeup(&epds);
1858 * We have to check that the file structure underneath the file descriptor
1859 * the user passed to us _is_ an eventpoll file. And also we do not permit
1860 * adding an epoll file descriptor inside itself.
1863 if (f.file == tf.file || !is_file_epoll(f.file))
1864 goto error_tgt_fput;
1867 * At this point it is safe to assume that the "private_data" contains
1868 * our own data structure.
1870 ep = f.file->private_data;
1873 * When we insert an epoll file descriptor, inside another epoll file
1874 * descriptor, there is the change of creating closed loops, which are
1875 * better be handled here, than in more critical paths. While we are
1876 * checking for loops we also determine the list of files reachable
1877 * and hang them on the tfile_check_list, so we can check that we
1878 * haven't created too many possible wakeup paths.
1880 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1881 * the epoll file descriptor is attaching directly to a wakeup source,
1882 * unless the epoll file descriptor is nested. The purpose of taking the
1883 * 'epmutex' on add is to prevent complex toplogies such as loops and
1884 * deep wakeup paths from forming in parallel through multiple
1885 * EPOLL_CTL_ADD operations.
1887 mutex_lock_nested(&ep->mtx, 0);
1888 if (op == EPOLL_CTL_ADD) {
1889 if (!list_empty(&f.file->f_ep_links) ||
1890 is_file_epoll(tf.file)) {
1892 mutex_unlock(&ep->mtx);
1893 mutex_lock(&epmutex);
1894 if (is_file_epoll(tf.file)) {
1896 if (ep_loop_check(ep, tf.file) != 0) {
1897 clear_tfile_check_list();
1898 goto error_tgt_fput;
1901 list_add(&tf.file->f_tfile_llink,
1903 mutex_lock_nested(&ep->mtx, 0);
1904 if (is_file_epoll(tf.file)) {
1905 tep = tf.file->private_data;
1906 mutex_lock_nested(&tep->mtx, 1);
1912 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1913 * above, we can be sure to be able to use the item looked up by
1914 * ep_find() till we release the mutex.
1916 epi = ep_find(ep, tf.file, fd);
1922 epds.events |= POLLERR | POLLHUP;
1923 error = ep_insert(ep, &epds, tf.file, fd, full_check);
1927 clear_tfile_check_list();
1931 error = ep_remove(ep, epi);
1937 epds.events |= POLLERR | POLLHUP;
1938 error = ep_modify(ep, epi, &epds);
1944 mutex_unlock(&tep->mtx);
1945 mutex_unlock(&ep->mtx);
1949 mutex_unlock(&epmutex);
1960 * Implement the event wait interface for the eventpoll file. It is the kernel
1961 * part of the user space epoll_wait(2).
1963 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1964 int, maxevents, int, timeout)
1968 struct eventpoll *ep;
1970 /* The maximum number of event must be greater than zero */
1971 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1974 /* Verify that the area passed by the user is writeable */
1975 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1978 /* Get the "struct file *" for the eventpoll file */
1984 * We have to check that the file structure underneath the fd
1985 * the user passed to us _is_ an eventpoll file.
1988 if (!is_file_epoll(f.file))
1992 * At this point it is safe to assume that the "private_data" contains
1993 * our own data structure.
1995 ep = f.file->private_data;
1997 /* Time to fish for events ... */
1998 error = ep_poll(ep, events, maxevents, timeout);
2006 * Implement the event wait interface for the eventpoll file. It is the kernel
2007 * part of the user space epoll_pwait(2).
2009 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2010 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2014 sigset_t ksigmask, sigsaved;
2017 * If the caller wants a certain signal mask to be set during the wait,
2021 if (sigsetsize != sizeof(sigset_t))
2023 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2025 sigsaved = current->blocked;
2026 set_current_blocked(&ksigmask);
2029 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2032 * If we changed the signal mask, we need to restore the original one.
2033 * In case we've got a signal while waiting, we do not restore the
2034 * signal mask yet, and we allow do_signal() to deliver the signal on
2035 * the way back to userspace, before the signal mask is restored.
2038 if (error == -EINTR) {
2039 memcpy(¤t->saved_sigmask, &sigsaved,
2041 set_restore_sigmask();
2043 set_current_blocked(&sigsaved);
2049 #ifdef CONFIG_COMPAT
2050 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2051 struct epoll_event __user *, events,
2052 int, maxevents, int, timeout,
2053 const compat_sigset_t __user *, sigmask,
2054 compat_size_t, sigsetsize)
2057 compat_sigset_t csigmask;
2058 sigset_t ksigmask, sigsaved;
2061 * If the caller wants a certain signal mask to be set during the wait,
2065 if (sigsetsize != sizeof(compat_sigset_t))
2067 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2069 sigset_from_compat(&ksigmask, &csigmask);
2070 sigsaved = current->blocked;
2071 set_current_blocked(&ksigmask);
2074 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2077 * If we changed the signal mask, we need to restore the original one.
2078 * In case we've got a signal while waiting, we do not restore the
2079 * signal mask yet, and we allow do_signal() to deliver the signal on
2080 * the way back to userspace, before the signal mask is restored.
2083 if (err == -EINTR) {
2084 memcpy(¤t->saved_sigmask, &sigsaved,
2086 set_restore_sigmask();
2088 set_current_blocked(&sigsaved);
2095 static int __init eventpoll_init(void)
2101 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2103 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2105 BUG_ON(max_user_watches < 0);
2108 * Initialize the structure used to perform epoll file descriptor
2109 * inclusion loops checks.
2111 ep_nested_calls_init(&poll_loop_ncalls);
2113 /* Initialize the structure used to perform safe poll wait head wake ups */
2114 ep_nested_calls_init(&poll_safewake_ncalls);
2116 /* Initialize the structure used to perform file's f_op->poll() calls */
2117 ep_nested_calls_init(&poll_readywalk_ncalls);
2120 * We can have many thousands of epitems, so prevent this from
2121 * using an extra cache line on 64-bit (and smaller) CPUs
2123 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2125 /* Allocates slab cache used to allocate "struct epitem" items */
2126 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2127 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2129 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2130 pwq_cache = kmem_cache_create("eventpoll_pwq",
2131 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2135 fs_initcall(eventpoll_init);