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 <linux/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 to epoll_ctl(EPOLL_CTL_DEL).
66 * It is also acquired when inserting an epoll fd onto another epoll
67 * fd. We do this so that we walk the epoll tree and ensure that this
68 * insertion does not create a cycle of epoll file descriptors, which
69 * could lead to deadlock. We need a global mutex to prevent two
70 * simultaneous inserts (A into B and B into A) from racing and
71 * constructing a cycle without either insert observing that it is
73 * It is necessary to acquire multiple "ep->mtx"es at once in the
74 * case when one epoll fd is added to another. In this case, we
75 * always acquire the locks in the order of nesting (i.e. after
76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77 * before e2->mtx). Since we disallow cycles of epoll file
78 * descriptors, this ensures that the mutexes are well-ordered. In
79 * order to communicate this nesting to lockdep, when walking a tree
80 * of epoll file descriptors, we use the current recursion depth as
82 * It is possible to drop the "ep->mtx" and to use the global
83 * mutex "epmutex" (together with "ep->lock") to have it working,
84 * but having "ep->mtx" will make the interface more scalable.
85 * Events that require holding "epmutex" are very rare, while for
86 * normal operations the epoll private "ep->mtx" will guarantee
87 * a better scalability.
90 /* Epoll private bits inside the event mask */
91 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
93 /* Maximum number of nesting allowed inside epoll sets */
94 #define EP_MAX_NESTS 4
96 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
98 #define EP_UNACTIVE_PTR ((void *) -1L)
100 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
102 struct epoll_filefd {
108 * Structure used to track possible nested calls, for too deep recursions
111 struct nested_call_node {
112 struct list_head llink;
118 * This structure is used as collector for nested calls, to check for
119 * maximum recursion dept and loop cycles.
121 struct nested_calls {
122 struct list_head tasks_call_list;
127 * Each file descriptor added to the eventpoll interface will
128 * have an entry of this type linked to the "rbr" RB tree.
131 /* RB tree node used to link this structure to the eventpoll RB tree */
134 /* List header used to link this structure to the eventpoll ready list */
135 struct list_head rdllink;
138 * Works together "struct eventpoll"->ovflist in keeping the
139 * single linked chain of items.
143 /* The file descriptor information this item refers to */
144 struct epoll_filefd ffd;
146 /* Number of active wait queue attached to poll operations */
149 /* List containing poll wait queues */
150 struct list_head pwqlist;
152 /* The "container" of this item */
153 struct eventpoll *ep;
155 /* List header used to link this item to the "struct file" items list */
156 struct list_head fllink;
158 /* The structure that describe the interested events and the source fd */
159 struct epoll_event event;
163 * This structure is stored inside the "private_data" member of the file
164 * structure and represents the main data structure for the eventpoll
168 /* Protect the access to this structure */
172 * This mutex is used to ensure that files are not removed
173 * while epoll is using them. This is held during the event
174 * collection loop, the file cleanup path, the epoll file exit
175 * code and the ctl operations.
179 /* Wait queue used by sys_epoll_wait() */
180 wait_queue_head_t wq;
182 /* Wait queue used by file->poll() */
183 wait_queue_head_t poll_wait;
185 /* List of ready file descriptors */
186 struct list_head rdllist;
188 /* RB tree root used to store monitored fd structs */
192 * This is a single linked list that chains all the "struct epitem" that
193 * happened while transferring ready events to userspace w/out
196 struct epitem *ovflist;
198 /* The user that created the eventpoll descriptor */
199 struct user_struct *user;
202 /* Wait structure used by the poll hooks */
203 struct eppoll_entry {
204 /* List header used to link this structure to the "struct epitem" */
205 struct list_head llink;
207 /* The "base" pointer is set to the container "struct epitem" */
211 * Wait queue item that will be linked to the target file wait
216 /* The wait queue head that linked the "wait" wait queue item */
217 wait_queue_head_t *whead;
220 /* Wrapper struct used by poll queueing */
226 /* Used by the ep_send_events() function as callback private data */
227 struct ep_send_events_data {
229 struct epoll_event __user *events;
233 * Configuration options available inside /proc/sys/fs/epoll/
235 /* Maximum number of epoll watched descriptors, per user */
236 static long max_user_watches __read_mostly;
239 * This mutex is used to serialize ep_free() and eventpoll_release_file().
241 static DEFINE_MUTEX(epmutex);
243 /* Used to check for epoll file descriptor inclusion loops */
244 static struct nested_calls poll_loop_ncalls;
246 /* Used for safe wake up implementation */
247 static struct nested_calls poll_safewake_ncalls;
249 /* Used to call file's f_op->poll() under the nested calls boundaries */
250 static struct nested_calls poll_readywalk_ncalls;
252 /* Slab cache used to allocate "struct epitem" */
253 static struct kmem_cache *epi_cache __read_mostly;
255 /* Slab cache used to allocate "struct eppoll_entry" */
256 static struct kmem_cache *pwq_cache __read_mostly;
260 #include <linux/sysctl.h>
263 static long long_max = LONG_MAX;
265 ctl_table epoll_table[] = {
267 .procname = "max_user_watches",
268 .data = &max_user_watches,
269 .maxlen = sizeof(max_user_watches),
271 .proc_handler = proc_doulongvec_minmax,
277 #endif /* CONFIG_SYSCTL */
280 /* Setup the structure that is used as key for the RB tree */
281 static inline void ep_set_ffd(struct epoll_filefd *ffd,
282 struct file *file, int fd)
288 /* Compare RB tree keys */
289 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
290 struct epoll_filefd *p2)
292 return (p1->file > p2->file ? +1:
293 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
296 /* Tells us if the item is currently linked */
297 static inline int ep_is_linked(struct list_head *p)
299 return !list_empty(p);
302 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
304 return container_of(p, struct eppoll_entry, wait);
307 /* Get the "struct epitem" from a wait queue pointer */
308 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
310 return container_of(p, struct eppoll_entry, wait)->base;
313 /* Get the "struct epitem" from an epoll queue wrapper */
314 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
316 return container_of(p, struct ep_pqueue, pt)->epi;
319 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
320 static inline int ep_op_has_event(int op)
322 return op != EPOLL_CTL_DEL;
325 /* Initialize the poll safe wake up structure */
326 static void ep_nested_calls_init(struct nested_calls *ncalls)
328 INIT_LIST_HEAD(&ncalls->tasks_call_list);
329 spin_lock_init(&ncalls->lock);
333 * ep_events_available - Checks if ready events might be available.
335 * @ep: Pointer to the eventpoll context.
337 * Returns: Returns a value different than zero if ready events are available,
340 static inline int ep_events_available(struct eventpoll *ep)
342 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
346 * ep_call_nested - Perform a bound (possibly) nested call, by checking
347 * that the recursion limit is not exceeded, and that
348 * the same nested call (by the meaning of same cookie) is
351 * @ncalls: Pointer to the nested_calls structure to be used for this call.
352 * @max_nests: Maximum number of allowed nesting calls.
353 * @nproc: Nested call core function pointer.
354 * @priv: Opaque data to be passed to the @nproc callback.
355 * @cookie: Cookie to be used to identify this nested call.
356 * @ctx: This instance context.
358 * Returns: Returns the code returned by the @nproc callback, or -1 if
359 * the maximum recursion limit has been exceeded.
361 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
362 int (*nproc)(void *, void *, int), void *priv,
363 void *cookie, void *ctx)
365 int error, call_nests = 0;
367 struct list_head *lsthead = &ncalls->tasks_call_list;
368 struct nested_call_node *tncur;
369 struct nested_call_node tnode;
371 spin_lock_irqsave(&ncalls->lock, flags);
374 * Try to see if the current task is already inside this wakeup call.
375 * We use a list here, since the population inside this set is always
378 list_for_each_entry(tncur, lsthead, llink) {
379 if (tncur->ctx == ctx &&
380 (tncur->cookie == cookie || ++call_nests > max_nests)) {
382 * Ops ... loop detected or maximum nest level reached.
383 * We abort this wake by breaking the cycle itself.
390 /* Add the current task and cookie to the list */
392 tnode.cookie = cookie;
393 list_add(&tnode.llink, lsthead);
395 spin_unlock_irqrestore(&ncalls->lock, flags);
397 /* Call the nested function */
398 error = (*nproc)(priv, cookie, call_nests);
400 /* Remove the current task from the list */
401 spin_lock_irqsave(&ncalls->lock, flags);
402 list_del(&tnode.llink);
404 spin_unlock_irqrestore(&ncalls->lock, flags);
409 #ifdef CONFIG_DEBUG_LOCK_ALLOC
410 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
411 unsigned long events, int subclass)
415 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
416 wake_up_locked_poll(wqueue, events);
417 spin_unlock_irqrestore(&wqueue->lock, flags);
420 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
421 unsigned long events, int subclass)
423 wake_up_poll(wqueue, events);
427 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
429 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
435 * Perform a safe wake up of the poll wait list. The problem is that
436 * with the new callback'd wake up system, it is possible that the
437 * poll callback is reentered from inside the call to wake_up() done
438 * on the poll wait queue head. The rule is that we cannot reenter the
439 * wake up code from the same task more than EP_MAX_NESTS times,
440 * and we cannot reenter the same wait queue head at all. This will
441 * enable to have a hierarchy of epoll file descriptor of no more than
444 static void ep_poll_safewake(wait_queue_head_t *wq)
446 int this_cpu = get_cpu();
448 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
449 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
454 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
456 wait_queue_head_t *whead;
459 /* If it is cleared by POLLFREE, it should be rcu-safe */
460 whead = rcu_dereference(pwq->whead);
462 remove_wait_queue(whead, &pwq->wait);
467 * This function unregisters poll callbacks from the associated file
468 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
471 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
473 struct list_head *lsthead = &epi->pwqlist;
474 struct eppoll_entry *pwq;
476 while (!list_empty(lsthead)) {
477 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
479 list_del(&pwq->llink);
480 ep_remove_wait_queue(pwq);
481 kmem_cache_free(pwq_cache, pwq);
486 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
487 * the scan code, to call f_op->poll(). Also allows for
488 * O(NumReady) performance.
490 * @ep: Pointer to the epoll private data structure.
491 * @sproc: Pointer to the scan callback.
492 * @priv: Private opaque data passed to the @sproc callback.
493 * @depth: The current depth of recursive f_op->poll calls.
495 * Returns: The same integer error code returned by the @sproc callback.
497 static int ep_scan_ready_list(struct eventpoll *ep,
498 int (*sproc)(struct eventpoll *,
499 struct list_head *, void *),
503 int error, pwake = 0;
505 struct epitem *epi, *nepi;
509 * We need to lock this because we could be hit by
510 * eventpoll_release_file() and epoll_ctl().
512 mutex_lock_nested(&ep->mtx, depth);
515 * Steal the ready list, and re-init the original one to the
516 * empty list. Also, set ep->ovflist to NULL so that events
517 * happening while looping w/out locks, are not lost. We cannot
518 * have the poll callback to queue directly on ep->rdllist,
519 * because we want the "sproc" callback to be able to do it
522 spin_lock_irqsave(&ep->lock, flags);
523 list_splice_init(&ep->rdllist, &txlist);
525 spin_unlock_irqrestore(&ep->lock, flags);
528 * Now call the callback function.
530 error = (*sproc)(ep, &txlist, priv);
532 spin_lock_irqsave(&ep->lock, flags);
534 * During the time we spent inside the "sproc" callback, some
535 * other events might have been queued by the poll callback.
536 * We re-insert them inside the main ready-list here.
538 for (nepi = ep->ovflist; (epi = nepi) != NULL;
539 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
541 * We need to check if the item is already in the list.
542 * During the "sproc" callback execution time, items are
543 * queued into ->ovflist but the "txlist" might already
544 * contain them, and the list_splice() below takes care of them.
546 if (!ep_is_linked(&epi->rdllink))
547 list_add_tail(&epi->rdllink, &ep->rdllist);
550 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
551 * releasing the lock, events will be queued in the normal way inside
554 ep->ovflist = EP_UNACTIVE_PTR;
557 * Quickly re-inject items left on "txlist".
559 list_splice(&txlist, &ep->rdllist);
561 if (!list_empty(&ep->rdllist)) {
563 * Wake up (if active) both the eventpoll wait list and
564 * the ->poll() wait list (delayed after we release the lock).
566 if (waitqueue_active(&ep->wq))
567 wake_up_locked(&ep->wq);
568 if (waitqueue_active(&ep->poll_wait))
571 spin_unlock_irqrestore(&ep->lock, flags);
573 mutex_unlock(&ep->mtx);
575 /* We have to call this outside the lock */
577 ep_poll_safewake(&ep->poll_wait);
583 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
584 * all the associated resources. Must be called with "mtx" held.
586 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
589 struct file *file = epi->ffd.file;
592 * Removes poll wait queue hooks. We _have_ to do this without holding
593 * the "ep->lock" otherwise a deadlock might occur. This because of the
594 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
595 * queue head lock when unregistering the wait queue. The wakeup callback
596 * will run by holding the wait queue head lock and will call our callback
597 * that will try to get "ep->lock".
599 ep_unregister_pollwait(ep, epi);
601 /* Remove the current item from the list of epoll hooks */
602 spin_lock(&file->f_lock);
603 if (ep_is_linked(&epi->fllink))
604 list_del_init(&epi->fllink);
605 spin_unlock(&file->f_lock);
607 rb_erase(&epi->rbn, &ep->rbr);
609 spin_lock_irqsave(&ep->lock, flags);
610 if (ep_is_linked(&epi->rdllink))
611 list_del_init(&epi->rdllink);
612 spin_unlock_irqrestore(&ep->lock, flags);
614 /* At this point it is safe to free the eventpoll item */
615 kmem_cache_free(epi_cache, epi);
617 atomic_long_dec(&ep->user->epoll_watches);
622 static void ep_free(struct eventpoll *ep)
627 /* We need to release all tasks waiting for these file */
628 if (waitqueue_active(&ep->poll_wait))
629 ep_poll_safewake(&ep->poll_wait);
632 * We need to lock this because we could be hit by
633 * eventpoll_release_file() while we're freeing the "struct eventpoll".
634 * We do not need to hold "ep->mtx" here because the epoll file
635 * is on the way to be removed and no one has references to it
636 * anymore. The only hit might come from eventpoll_release_file() but
637 * holding "epmutex" is sufficient here.
639 mutex_lock(&epmutex);
642 * Walks through the whole tree by unregistering poll callbacks.
644 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
645 epi = rb_entry(rbp, struct epitem, rbn);
647 ep_unregister_pollwait(ep, epi);
651 * Walks through the whole tree by freeing each "struct epitem". At this
652 * point we are sure no poll callbacks will be lingering around, and also by
653 * holding "epmutex" we can be sure that no file cleanup code will hit
654 * us during this operation. So we can avoid the lock on "ep->lock".
656 while ((rbp = rb_first(&ep->rbr)) != NULL) {
657 epi = rb_entry(rbp, struct epitem, rbn);
661 mutex_unlock(&epmutex);
662 mutex_destroy(&ep->mtx);
667 static int ep_eventpoll_release(struct inode *inode, struct file *file)
669 struct eventpoll *ep = file->private_data;
677 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
680 struct epitem *epi, *tmp;
682 list_for_each_entry_safe(epi, tmp, head, rdllink) {
683 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
685 return POLLIN | POLLRDNORM;
688 * Item has been dropped into the ready list by the poll
689 * callback, but it's not actually ready, as far as
690 * caller requested events goes. We can remove it here.
692 list_del_init(&epi->rdllink);
699 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
701 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
704 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
707 struct eventpoll *ep = file->private_data;
709 /* Insert inside our poll wait queue */
710 poll_wait(file, &ep->poll_wait, wait);
713 * Proceed to find out if wanted events are really available inside
714 * the ready list. This need to be done under ep_call_nested()
715 * supervision, since the call to f_op->poll() done on listed files
716 * could re-enter here.
718 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
719 ep_poll_readyevents_proc, ep, ep, current);
721 return pollflags != -1 ? pollflags : 0;
724 /* File callbacks that implement the eventpoll file behaviour */
725 static const struct file_operations eventpoll_fops = {
726 .release = ep_eventpoll_release,
727 .poll = ep_eventpoll_poll,
728 .llseek = noop_llseek,
731 /* Fast test to see if the file is an eventpoll file */
732 static inline int is_file_epoll(struct file *f)
734 return f->f_op == &eventpoll_fops;
738 * This is called from eventpoll_release() to unlink files from the eventpoll
739 * interface. We need to have this facility to cleanup correctly files that are
740 * closed without being removed from the eventpoll interface.
742 void eventpoll_release_file(struct file *file)
744 struct list_head *lsthead = &file->f_ep_links;
745 struct eventpoll *ep;
749 * We don't want to get "file->f_lock" because it is not
750 * necessary. It is not necessary because we're in the "struct file"
751 * cleanup path, and this means that no one is using this file anymore.
752 * So, for example, epoll_ctl() cannot hit here since if we reach this
753 * point, the file counter already went to zero and fget() would fail.
754 * The only hit might come from ep_free() but by holding the mutex
755 * will correctly serialize the operation. We do need to acquire
756 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
757 * from anywhere but ep_free().
759 * Besides, ep_remove() acquires the lock, so we can't hold it here.
761 mutex_lock(&epmutex);
763 while (!list_empty(lsthead)) {
764 epi = list_first_entry(lsthead, struct epitem, fllink);
767 list_del_init(&epi->fllink);
768 mutex_lock_nested(&ep->mtx, 0);
770 mutex_unlock(&ep->mtx);
773 mutex_unlock(&epmutex);
776 static int ep_alloc(struct eventpoll **pep)
779 struct user_struct *user;
780 struct eventpoll *ep;
782 user = get_current_user();
784 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
788 spin_lock_init(&ep->lock);
789 mutex_init(&ep->mtx);
790 init_waitqueue_head(&ep->wq);
791 init_waitqueue_head(&ep->poll_wait);
792 INIT_LIST_HEAD(&ep->rdllist);
794 ep->ovflist = EP_UNACTIVE_PTR;
807 * Search the file inside the eventpoll tree. The RB tree operations
808 * are protected by the "mtx" mutex, and ep_find() must be called with
811 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
815 struct epitem *epi, *epir = NULL;
816 struct epoll_filefd ffd;
818 ep_set_ffd(&ffd, file, fd);
819 for (rbp = ep->rbr.rb_node; rbp; ) {
820 epi = rb_entry(rbp, struct epitem, rbn);
821 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
836 * This is the callback that is passed to the wait queue wakeup
837 * mechanism. It is called by the stored file descriptors when they
838 * have events to report.
840 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
844 struct epitem *epi = ep_item_from_wait(wait);
845 struct eventpoll *ep = epi->ep;
847 if ((unsigned long)key & POLLFREE) {
848 ep_pwq_from_wait(wait)->whead = NULL;
850 * whead = NULL above can race with ep_remove_wait_queue()
851 * which can do another remove_wait_queue() after us, so we
852 * can't use __remove_wait_queue(). whead->lock is held by
855 list_del_init(&wait->task_list);
858 spin_lock_irqsave(&ep->lock, flags);
861 * If the event mask does not contain any poll(2) event, we consider the
862 * descriptor to be disabled. This condition is likely the effect of the
863 * EPOLLONESHOT bit that disables the descriptor when an event is received,
864 * until the next EPOLL_CTL_MOD will be issued.
866 if (!(epi->event.events & ~EP_PRIVATE_BITS))
870 * Check the events coming with the callback. At this stage, not
871 * every device reports the events in the "key" parameter of the
872 * callback. We need to be able to handle both cases here, hence the
873 * test for "key" != NULL before the event match test.
875 if (key && !((unsigned long) key & epi->event.events))
879 * If we are transferring events to userspace, we can hold no locks
880 * (because we're accessing user memory, and because of linux f_op->poll()
881 * semantics). All the events that happen during that period of time are
882 * chained in ep->ovflist and requeued later on.
884 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
885 if (epi->next == EP_UNACTIVE_PTR) {
886 epi->next = ep->ovflist;
892 /* If this file is already in the ready list we exit soon */
893 if (!ep_is_linked(&epi->rdllink))
894 list_add_tail(&epi->rdllink, &ep->rdllist);
897 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
900 if (waitqueue_active(&ep->wq))
901 wake_up_locked(&ep->wq);
902 if (waitqueue_active(&ep->poll_wait))
906 spin_unlock_irqrestore(&ep->lock, flags);
908 /* We have to call this outside the lock */
910 ep_poll_safewake(&ep->poll_wait);
916 * This is the callback that is used to add our wait queue to the
917 * target file wakeup lists.
919 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
922 struct epitem *epi = ep_item_from_epqueue(pt);
923 struct eppoll_entry *pwq;
925 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
926 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
929 add_wait_queue(whead, &pwq->wait);
930 list_add_tail(&pwq->llink, &epi->pwqlist);
933 /* We have to signal that an error occurred */
938 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
941 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
946 epic = rb_entry(parent, struct epitem, rbn);
947 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
949 p = &parent->rb_right;
951 p = &parent->rb_left;
953 rb_link_node(&epi->rbn, parent, p);
954 rb_insert_color(&epi->rbn, &ep->rbr);
958 * Must be called with "mtx" held.
960 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
961 struct file *tfile, int fd)
963 int error, revents, pwake = 0;
967 struct ep_pqueue epq;
969 user_watches = atomic_long_read(&ep->user->epoll_watches);
970 if (unlikely(user_watches >= max_user_watches))
972 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
975 /* Item initialization follow here ... */
976 INIT_LIST_HEAD(&epi->rdllink);
977 INIT_LIST_HEAD(&epi->fllink);
978 INIT_LIST_HEAD(&epi->pwqlist);
980 ep_set_ffd(&epi->ffd, tfile, fd);
983 epi->next = EP_UNACTIVE_PTR;
985 /* Initialize the poll table using the queue callback */
987 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
990 * Attach the item to the poll hooks and get current event bits.
991 * We can safely use the file* here because its usage count has
992 * been increased by the caller of this function. Note that after
993 * this operation completes, the poll callback can start hitting
996 revents = tfile->f_op->poll(tfile, &epq.pt);
999 * We have to check if something went wrong during the poll wait queue
1000 * install process. Namely an allocation for a wait queue failed due
1001 * high memory pressure.
1005 goto error_unregister;
1007 /* Add the current item to the list of active epoll hook for this file */
1008 spin_lock(&tfile->f_lock);
1009 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1010 spin_unlock(&tfile->f_lock);
1013 * Add the current item to the RB tree. All RB tree operations are
1014 * protected by "mtx", and ep_insert() is called with "mtx" held.
1016 ep_rbtree_insert(ep, epi);
1018 /* We have to drop the new item inside our item list to keep track of it */
1019 spin_lock_irqsave(&ep->lock, flags);
1021 /* If the file is already "ready" we drop it inside the ready list */
1022 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1023 list_add_tail(&epi->rdllink, &ep->rdllist);
1025 /* Notify waiting tasks that events are available */
1026 if (waitqueue_active(&ep->wq))
1027 wake_up_locked(&ep->wq);
1028 if (waitqueue_active(&ep->poll_wait))
1032 spin_unlock_irqrestore(&ep->lock, flags);
1034 atomic_long_inc(&ep->user->epoll_watches);
1036 /* We have to call this outside the lock */
1038 ep_poll_safewake(&ep->poll_wait);
1043 ep_unregister_pollwait(ep, epi);
1046 * We need to do this because an event could have been arrived on some
1047 * allocated wait queue. Note that we don't care about the ep->ovflist
1048 * list, since that is used/cleaned only inside a section bound by "mtx".
1049 * And ep_insert() is called with "mtx" held.
1051 spin_lock_irqsave(&ep->lock, flags);
1052 if (ep_is_linked(&epi->rdllink))
1053 list_del_init(&epi->rdllink);
1054 spin_unlock_irqrestore(&ep->lock, flags);
1056 kmem_cache_free(epi_cache, epi);
1062 * Modify the interest event mask by dropping an event if the new mask
1063 * has a match in the current file status. Must be called with "mtx" held.
1065 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1068 unsigned int revents;
1071 * Set the new event interest mask before calling f_op->poll();
1072 * otherwise we might miss an event that happens between the
1073 * f_op->poll() call and the new event set registering.
1075 epi->event.events = event->events;
1076 epi->event.data = event->data; /* protected by mtx */
1079 * Get current event bits. We can safely use the file* here because
1080 * its usage count has been increased by the caller of this function.
1082 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1085 * If the item is "hot" and it is not registered inside the ready
1086 * list, push it inside.
1088 if (revents & event->events) {
1089 spin_lock_irq(&ep->lock);
1090 if (!ep_is_linked(&epi->rdllink)) {
1091 list_add_tail(&epi->rdllink, &ep->rdllist);
1093 /* Notify waiting tasks that events are available */
1094 if (waitqueue_active(&ep->wq))
1095 wake_up_locked(&ep->wq);
1096 if (waitqueue_active(&ep->poll_wait))
1099 spin_unlock_irq(&ep->lock);
1102 /* We have to call this outside the lock */
1104 ep_poll_safewake(&ep->poll_wait);
1109 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1112 struct ep_send_events_data *esed = priv;
1114 unsigned int revents;
1116 struct epoll_event __user *uevent;
1119 * We can loop without lock because we are passed a task private list.
1120 * Items cannot vanish during the loop because ep_scan_ready_list() is
1121 * holding "mtx" during this call.
1123 for (eventcnt = 0, uevent = esed->events;
1124 !list_empty(head) && eventcnt < esed->maxevents;) {
1125 epi = list_first_entry(head, struct epitem, rdllink);
1127 list_del_init(&epi->rdllink);
1129 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1133 * If the event mask intersect the caller-requested one,
1134 * deliver the event to userspace. Again, ep_scan_ready_list()
1135 * is holding "mtx", so no operations coming from userspace
1136 * can change the item.
1139 if (__put_user(revents, &uevent->events) ||
1140 __put_user(epi->event.data, &uevent->data)) {
1141 list_add(&epi->rdllink, head);
1142 return eventcnt ? eventcnt : -EFAULT;
1146 if (epi->event.events & EPOLLONESHOT)
1147 epi->event.events &= EP_PRIVATE_BITS;
1148 else if (!(epi->event.events & EPOLLET)) {
1150 * If this file has been added with Level
1151 * Trigger mode, we need to insert back inside
1152 * the ready list, so that the next call to
1153 * epoll_wait() will check again the events
1154 * availability. At this point, no one can insert
1155 * into ep->rdllist besides us. The epoll_ctl()
1156 * callers are locked out by
1157 * ep_scan_ready_list() holding "mtx" and the
1158 * poll callback will queue them in ep->ovflist.
1160 list_add_tail(&epi->rdllink, &ep->rdllist);
1168 static int ep_send_events(struct eventpoll *ep,
1169 struct epoll_event __user *events, int maxevents)
1171 struct ep_send_events_data esed;
1173 esed.maxevents = maxevents;
1174 esed.events = events;
1176 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
1179 static inline struct timespec ep_set_mstimeout(long ms)
1181 struct timespec now, ts = {
1182 .tv_sec = ms / MSEC_PER_SEC,
1183 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1187 return timespec_add_safe(now, ts);
1191 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1194 * @ep: Pointer to the eventpoll context.
1195 * @events: Pointer to the userspace buffer where the ready events should be
1197 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1198 * @timeout: Maximum timeout for the ready events fetch operation, in
1199 * milliseconds. If the @timeout is zero, the function will not block,
1200 * while if the @timeout is less than zero, the function will block
1201 * until at least one event has been retrieved (or an error
1204 * Returns: Returns the number of ready events which have been fetched, or an
1205 * error code, in case of error.
1207 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1208 int maxevents, long timeout)
1210 int res = 0, eavail, timed_out = 0;
1211 unsigned long flags;
1214 ktime_t expires, *to = NULL;
1217 struct timespec end_time = ep_set_mstimeout(timeout);
1219 slack = select_estimate_accuracy(&end_time);
1221 *to = timespec_to_ktime(end_time);
1222 } else if (timeout == 0) {
1224 * Avoid the unnecessary trip to the wait queue loop, if the
1225 * caller specified a non blocking operation.
1228 spin_lock_irqsave(&ep->lock, flags);
1233 spin_lock_irqsave(&ep->lock, flags);
1235 if (!ep_events_available(ep)) {
1237 * We don't have any available event to return to the caller.
1238 * We need to sleep here, and we will be wake up by
1239 * ep_poll_callback() when events will become available.
1241 init_waitqueue_entry(&wait, current);
1242 __add_wait_queue_exclusive(&ep->wq, &wait);
1246 * We don't want to sleep if the ep_poll_callback() sends us
1247 * a wakeup in between. That's why we set the task state
1248 * to TASK_INTERRUPTIBLE before doing the checks.
1250 set_current_state(TASK_INTERRUPTIBLE);
1251 if (ep_events_available(ep) || timed_out)
1253 if (signal_pending(current)) {
1258 spin_unlock_irqrestore(&ep->lock, flags);
1259 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1262 spin_lock_irqsave(&ep->lock, flags);
1264 __remove_wait_queue(&ep->wq, &wait);
1266 set_current_state(TASK_RUNNING);
1269 /* Is it worth to try to dig for events ? */
1270 eavail = ep_events_available(ep);
1272 spin_unlock_irqrestore(&ep->lock, flags);
1275 * Try to transfer events to user space. In case we get 0 events and
1276 * there's still timeout left over, we go trying again in search of
1279 if (!res && eavail &&
1280 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1287 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1288 * API, to verify that adding an epoll file inside another
1289 * epoll structure, does not violate the constraints, in
1290 * terms of closed loops, or too deep chains (which can
1291 * result in excessive stack usage).
1293 * @priv: Pointer to the epoll file to be currently checked.
1294 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1295 * data structure pointer.
1296 * @call_nests: Current dept of the @ep_call_nested() call stack.
1298 * Returns: Returns zero if adding the epoll @file inside current epoll
1299 * structure @ep does not violate the constraints, or -1 otherwise.
1301 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1304 struct file *file = priv;
1305 struct eventpoll *ep = file->private_data;
1306 struct rb_node *rbp;
1309 mutex_lock_nested(&ep->mtx, call_nests + 1);
1310 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1311 epi = rb_entry(rbp, struct epitem, rbn);
1312 if (unlikely(is_file_epoll(epi->ffd.file))) {
1313 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1314 ep_loop_check_proc, epi->ffd.file,
1315 epi->ffd.file->private_data, current);
1320 mutex_unlock(&ep->mtx);
1326 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1327 * another epoll file (represented by @ep) does not create
1328 * closed loops or too deep chains.
1330 * @ep: Pointer to the epoll private data structure.
1331 * @file: Pointer to the epoll file to be checked.
1333 * Returns: Returns zero if adding the epoll @file inside current epoll
1334 * structure @ep does not violate the constraints, or -1 otherwise.
1336 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1338 return ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1339 ep_loop_check_proc, file, ep, current);
1343 * Open an eventpoll file descriptor.
1345 SYSCALL_DEFINE1(epoll_create1, int, flags)
1348 struct eventpoll *ep = NULL;
1350 /* Check the EPOLL_* constant for consistency. */
1351 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1353 if (flags & ~EPOLL_CLOEXEC)
1356 * Create the internal data structure ("struct eventpoll").
1358 error = ep_alloc(&ep);
1362 * Creates all the items needed to setup an eventpoll file. That is,
1363 * a file structure and a free file descriptor.
1365 error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1366 O_RDWR | (flags & O_CLOEXEC));
1373 SYSCALL_DEFINE1(epoll_create, int, size)
1378 return sys_epoll_create1(0);
1382 * The following function implements the controller interface for
1383 * the eventpoll file that enables the insertion/removal/change of
1384 * file descriptors inside the interest set.
1386 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1387 struct epoll_event __user *, event)
1390 int did_lock_epmutex = 0;
1391 struct file *file, *tfile;
1392 struct eventpoll *ep;
1394 struct epoll_event epds;
1397 if (ep_op_has_event(op) &&
1398 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1401 /* Get the "struct file *" for the eventpoll file */
1407 /* Get the "struct file *" for the target file */
1412 /* The target file descriptor must support poll */
1414 if (!tfile->f_op || !tfile->f_op->poll)
1415 goto error_tgt_fput;
1418 * We have to check that the file structure underneath the file descriptor
1419 * the user passed to us _is_ an eventpoll file. And also we do not permit
1420 * adding an epoll file descriptor inside itself.
1423 if (file == tfile || !is_file_epoll(file))
1424 goto error_tgt_fput;
1427 * At this point it is safe to assume that the "private_data" contains
1428 * our own data structure.
1430 ep = file->private_data;
1433 * When we insert an epoll file descriptor, inside another epoll file
1434 * descriptor, there is the change of creating closed loops, which are
1435 * better be handled here, than in more critical paths.
1437 * We hold epmutex across the loop check and the insert in this case, in
1438 * order to prevent two separate inserts from racing and each doing the
1439 * insert "at the same time" such that ep_loop_check passes on both
1440 * before either one does the insert, thereby creating a cycle.
1442 if (unlikely(is_file_epoll(tfile) && op == EPOLL_CTL_ADD)) {
1443 mutex_lock(&epmutex);
1444 did_lock_epmutex = 1;
1446 if (ep_loop_check(ep, tfile) != 0)
1447 goto error_tgt_fput;
1451 mutex_lock_nested(&ep->mtx, 0);
1454 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1455 * above, we can be sure to be able to use the item looked up by
1456 * ep_find() till we release the mutex.
1458 epi = ep_find(ep, tfile, fd);
1464 epds.events |= POLLERR | POLLHUP;
1465 error = ep_insert(ep, &epds, tfile, fd);
1471 error = ep_remove(ep, epi);
1477 epds.events |= POLLERR | POLLHUP;
1478 error = ep_modify(ep, epi, &epds);
1483 mutex_unlock(&ep->mtx);
1486 if (unlikely(did_lock_epmutex))
1487 mutex_unlock(&epmutex);
1498 * Implement the event wait interface for the eventpoll file. It is the kernel
1499 * part of the user space epoll_wait(2).
1501 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1502 int, maxevents, int, timeout)
1506 struct eventpoll *ep;
1508 /* The maximum number of event must be greater than zero */
1509 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1512 /* Verify that the area passed by the user is writeable */
1513 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1518 /* Get the "struct file *" for the eventpoll file */
1525 * We have to check that the file structure underneath the fd
1526 * the user passed to us _is_ an eventpoll file.
1529 if (!is_file_epoll(file))
1533 * At this point it is safe to assume that the "private_data" contains
1534 * our own data structure.
1536 ep = file->private_data;
1538 /* Time to fish for events ... */
1539 error = ep_poll(ep, events, maxevents, timeout);
1548 #ifdef HAVE_SET_RESTORE_SIGMASK
1551 * Implement the event wait interface for the eventpoll file. It is the kernel
1552 * part of the user space epoll_pwait(2).
1554 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1555 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1559 sigset_t ksigmask, sigsaved;
1562 * If the caller wants a certain signal mask to be set during the wait,
1566 if (sigsetsize != sizeof(sigset_t))
1568 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1570 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1571 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1574 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1577 * If we changed the signal mask, we need to restore the original one.
1578 * In case we've got a signal while waiting, we do not restore the
1579 * signal mask yet, and we allow do_signal() to deliver the signal on
1580 * the way back to userspace, before the signal mask is restored.
1583 if (error == -EINTR) {
1584 memcpy(¤t->saved_sigmask, &sigsaved,
1586 set_restore_sigmask();
1588 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1594 #endif /* HAVE_SET_RESTORE_SIGMASK */
1596 static int __init eventpoll_init(void)
1602 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1604 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1606 BUG_ON(max_user_watches < 0);
1609 * Initialize the structure used to perform epoll file descriptor
1610 * inclusion loops checks.
1612 ep_nested_calls_init(&poll_loop_ncalls);
1614 /* Initialize the structure used to perform safe poll wait head wake ups */
1615 ep_nested_calls_init(&poll_safewake_ncalls);
1617 /* Initialize the structure used to perform file's f_op->poll() calls */
1618 ep_nested_calls_init(&poll_readywalk_ncalls);
1620 /* Allocates slab cache used to allocate "struct epitem" items */
1621 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1622 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1624 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1625 pwq_cache = kmem_cache_create("eventpoll_pwq",
1626 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1630 fs_initcall(eventpoll_init);