1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * fs/eventpoll.c (Efficient event retrieval implementation)
4 * Copyright (C) 2001,...,2009 Davide Libenzi
6 * Davide Libenzi <davidel@xmailserver.org>
9 #include <linux/init.h>
10 #include <linux/kernel.h>
11 #include <linux/sched/signal.h>
13 #include <linux/file.h>
14 #include <linux/signal.h>
15 #include <linux/errno.h>
17 #include <linux/slab.h>
18 #include <linux/poll.h>
19 #include <linux/string.h>
20 #include <linux/list.h>
21 #include <linux/hash.h>
22 #include <linux/spinlock.h>
23 #include <linux/syscalls.h>
24 #include <linux/rbtree.h>
25 #include <linux/wait.h>
26 #include <linux/eventpoll.h>
27 #include <linux/mount.h>
28 #include <linux/bitops.h>
29 #include <linux/mutex.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/device.h>
32 #include <linux/uaccess.h>
35 #include <linux/atomic.h>
36 #include <linux/proc_fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/compat.h>
39 #include <linux/rculist.h>
40 #include <net/busy_poll.h>
44 * There are three level of locking required by epoll :
46 * 1) epnested_mutex (mutex)
48 * 3) ep->lock (rwlock)
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a rwlock (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 * The epnested_mutex is acquired when inserting an epoll fd onto another
61 * epoll fd. We do this so that we walk the epoll tree and ensure that this
62 * insertion does not create a cycle of epoll file descriptors, which
63 * could lead to deadlock. We need a global mutex to prevent two
64 * simultaneous inserts (A into B and B into A) from racing and
65 * constructing a cycle without either insert observing that it is
67 * It is necessary to acquire multiple "ep->mtx"es at once in the
68 * case when one epoll fd is added to another. In this case, we
69 * always acquire the locks in the order of nesting (i.e. after
70 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
71 * before e2->mtx). Since we disallow cycles of epoll file
72 * descriptors, this ensures that the mutexes are well-ordered. In
73 * order to communicate this nesting to lockdep, when walking a tree
74 * of epoll file descriptors, we use the current recursion depth as
76 * It is possible to drop the "ep->mtx" and to use the global
77 * mutex "epnested_mutex" (together with "ep->lock") to have it working,
78 * but having "ep->mtx" will make the interface more scalable.
79 * Events that require holding "epnested_mutex" are very rare, while for
80 * normal operations the epoll private "ep->mtx" will guarantee
81 * a better scalability.
84 /* Epoll private bits inside the event mask */
85 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
87 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
89 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
90 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
92 /* Maximum number of nesting allowed inside epoll sets */
93 #define EP_MAX_NESTS 4
95 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
97 #define EP_UNACTIVE_PTR ((void *) -1L)
99 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
101 struct epoll_filefd {
106 /* Wait structure used by the poll hooks */
107 struct eppoll_entry {
108 /* List header used to link this structure to the "struct epitem" */
109 struct eppoll_entry *next;
111 /* The "base" pointer is set to the container "struct epitem" */
115 * Wait queue item that will be linked to the target file wait
118 wait_queue_entry_t wait;
120 /* The wait queue head that linked the "wait" wait queue item */
121 wait_queue_head_t *whead;
125 * Each file descriptor added to the eventpoll interface will
126 * have an entry of this type linked to the "rbr" RB tree.
127 * Avoid increasing the size of this struct, there can be many thousands
128 * of these on a server and we do not want this to take another cache line.
132 /* RB tree node links this structure to the eventpoll RB tree */
134 /* Used to free the struct epitem */
138 /* List header used to link this structure to the eventpoll ready list */
139 struct list_head rdllink;
142 * Works together "struct eventpoll"->ovflist in keeping the
143 * single linked chain of items.
147 /* The file descriptor information this item refers to */
148 struct epoll_filefd ffd;
151 * Protected by file->f_lock, true for to-be-released epitem already
152 * removed from the "struct file" items list; together with
153 * eventpoll->refcount orchestrates "struct eventpoll" disposal
157 /* List containing poll wait queues */
158 struct eppoll_entry *pwqlist;
160 /* The "container" of this item */
161 struct eventpoll *ep;
163 /* List header used to link this item to the "struct file" items list */
164 struct hlist_node fllink;
166 /* wakeup_source used when EPOLLWAKEUP is set */
167 struct wakeup_source __rcu *ws;
169 /* The structure that describe the interested events and the source fd */
170 struct epoll_event event;
174 * This structure is stored inside the "private_data" member of the file
175 * structure and represents the main data structure for the eventpoll
180 * This mutex is used to ensure that files are not removed
181 * while epoll is using them. This is held during the event
182 * collection loop, the file cleanup path, the epoll file exit
183 * code and the ctl operations.
187 /* Wait queue used by sys_epoll_wait() */
188 wait_queue_head_t wq;
190 /* Wait queue used by file->poll() */
191 wait_queue_head_t poll_wait;
193 /* List of ready file descriptors */
194 struct list_head rdllist;
196 /* Lock which protects rdllist and ovflist */
199 /* RB tree root used to store monitored fd structs */
200 struct rb_root_cached rbr;
203 * This is a single linked list that chains all the "struct epitem" that
204 * happened while transferring ready events to userspace w/out
207 struct epitem *ovflist;
209 /* wakeup_source used when ep_send_events or __ep_eventpoll_poll is running */
210 struct wakeup_source *ws;
212 /* The user that created the eventpoll descriptor */
213 struct user_struct *user;
217 /* used to optimize loop detection check */
219 struct hlist_head refs;
222 * usage count, used together with epitem->dying to
223 * orchestrate the disposal of this struct
227 #ifdef CONFIG_NET_RX_BUSY_POLL
228 /* used to track busy poll napi_id */
229 unsigned int napi_id;
232 #ifdef CONFIG_DEBUG_LOCK_ALLOC
233 /* tracks wakeup nests for lockdep validation */
238 /* Wrapper struct used by poll queueing */
245 * Configuration options available inside /proc/sys/fs/epoll/
247 /* Maximum number of epoll watched descriptors, per user */
248 static long max_user_watches __read_mostly;
250 /* Used for cycles detection */
251 static DEFINE_MUTEX(epnested_mutex);
253 static u64 loop_check_gen = 0;
255 /* Used to check for epoll file descriptor inclusion loops */
256 static struct eventpoll *inserting_into;
258 /* Slab cache used to allocate "struct epitem" */
259 static struct kmem_cache *epi_cache __ro_after_init;
261 /* Slab cache used to allocate "struct eppoll_entry" */
262 static struct kmem_cache *pwq_cache __ro_after_init;
265 * List of files with newly added links, where we may need to limit the number
266 * of emanating paths. Protected by the epnested_mutex.
268 struct epitems_head {
269 struct hlist_head epitems;
270 struct epitems_head *next;
272 static struct epitems_head *tfile_check_list = EP_UNACTIVE_PTR;
274 static struct kmem_cache *ephead_cache __ro_after_init;
276 static inline void free_ephead(struct epitems_head *head)
279 kmem_cache_free(ephead_cache, head);
282 static void list_file(struct file *file)
284 struct epitems_head *head;
286 head = container_of(file->f_ep, struct epitems_head, epitems);
288 head->next = tfile_check_list;
289 tfile_check_list = head;
293 static void unlist_file(struct epitems_head *head)
295 struct epitems_head *to_free = head;
296 struct hlist_node *p = rcu_dereference(hlist_first_rcu(&head->epitems));
298 struct epitem *epi= container_of(p, struct epitem, fllink);
299 spin_lock(&epi->ffd.file->f_lock);
300 if (!hlist_empty(&head->epitems))
303 spin_unlock(&epi->ffd.file->f_lock);
305 free_ephead(to_free);
310 #include <linux/sysctl.h>
312 static long long_zero;
313 static long long_max = LONG_MAX;
315 static struct ctl_table epoll_table[] = {
317 .procname = "max_user_watches",
318 .data = &max_user_watches,
319 .maxlen = sizeof(max_user_watches),
321 .proc_handler = proc_doulongvec_minmax,
322 .extra1 = &long_zero,
327 static void __init epoll_sysctls_init(void)
329 register_sysctl("fs/epoll", epoll_table);
332 #define epoll_sysctls_init() do { } while (0)
333 #endif /* CONFIG_SYSCTL */
335 static const struct file_operations eventpoll_fops;
337 static inline int is_file_epoll(struct file *f)
339 return f->f_op == &eventpoll_fops;
342 /* Setup the structure that is used as key for the RB tree */
343 static inline void ep_set_ffd(struct epoll_filefd *ffd,
344 struct file *file, int fd)
350 /* Compare RB tree keys */
351 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
352 struct epoll_filefd *p2)
354 return (p1->file > p2->file ? +1:
355 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
358 /* Tells us if the item is currently linked */
359 static inline int ep_is_linked(struct epitem *epi)
361 return !list_empty(&epi->rdllink);
364 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
366 return container_of(p, struct eppoll_entry, wait);
369 /* Get the "struct epitem" from a wait queue pointer */
370 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
372 return container_of(p, struct eppoll_entry, wait)->base;
376 * ep_events_available - Checks if ready events might be available.
378 * @ep: Pointer to the eventpoll context.
380 * Return: a value different than %zero if ready events are available,
381 * or %zero otherwise.
383 static inline int ep_events_available(struct eventpoll *ep)
385 return !list_empty_careful(&ep->rdllist) ||
386 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
389 #ifdef CONFIG_NET_RX_BUSY_POLL
390 static bool ep_busy_loop_end(void *p, unsigned long start_time)
392 struct eventpoll *ep = p;
394 return ep_events_available(ep) || busy_loop_timeout(start_time);
398 * Busy poll if globally on and supporting sockets found && no events,
399 * busy loop will return if need_resched or ep_events_available.
401 * we must do our busy polling with irqs enabled
403 static bool ep_busy_loop(struct eventpoll *ep, int nonblock)
405 unsigned int napi_id = READ_ONCE(ep->napi_id);
407 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on()) {
408 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep, false,
410 if (ep_events_available(ep))
413 * Busy poll timed out. Drop NAPI ID for now, we can add
414 * it back in when we have moved a socket with a valid NAPI
415 * ID onto the ready list.
424 * Set epoll busy poll NAPI ID from sk.
426 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
428 struct eventpoll *ep;
429 unsigned int napi_id;
433 if (!net_busy_loop_on())
436 sock = sock_from_file(epi->ffd.file);
444 napi_id = READ_ONCE(sk->sk_napi_id);
447 /* Non-NAPI IDs can be rejected
449 * Nothing to do if we already have this ID
451 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
454 /* record NAPI ID for use in next busy poll */
455 ep->napi_id = napi_id;
460 static inline bool ep_busy_loop(struct eventpoll *ep, int nonblock)
465 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
469 #endif /* CONFIG_NET_RX_BUSY_POLL */
472 * As described in commit 0ccf831cb lockdep: annotate epoll
473 * the use of wait queues used by epoll is done in a very controlled
474 * manner. Wake ups can nest inside each other, but are never done
475 * with the same locking. For example:
478 * efd1 = epoll_create();
479 * efd2 = epoll_create();
480 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
481 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
483 * When a packet arrives to the device underneath "dfd", the net code will
484 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
485 * callback wakeup entry on that queue, and the wake_up() performed by the
486 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
487 * (efd1) notices that it may have some event ready, so it needs to wake up
488 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
489 * that ends up in another wake_up(), after having checked about the
490 * recursion constraints. That are, no more than EP_MAX_NESTS, to avoid
493 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
494 * this special case of epoll.
496 #ifdef CONFIG_DEBUG_LOCK_ALLOC
498 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
501 struct eventpoll *ep_src;
506 * To set the subclass or nesting level for spin_lock_irqsave_nested()
507 * it might be natural to create a per-cpu nest count. However, since
508 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
509 * schedule() in the -rt kernel, the per-cpu variable are no longer
510 * protected. Thus, we are introducing a per eventpoll nest field.
511 * If we are not being call from ep_poll_callback(), epi is NULL and
512 * we are at the first level of nesting, 0. Otherwise, we are being
513 * called from ep_poll_callback() and if a previous wakeup source is
514 * not an epoll file itself, we are at depth 1 since the wakeup source
515 * is depth 0. If the wakeup source is a previous epoll file in the
516 * wakeup chain then we use its nests value and record ours as
517 * nests + 1. The previous epoll file nests value is stable since its
518 * already holding its own poll_wait.lock.
521 if ((is_file_epoll(epi->ffd.file))) {
522 ep_src = epi->ffd.file->private_data;
523 nests = ep_src->nests;
528 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
529 ep->nests = nests + 1;
530 wake_up_locked_poll(&ep->poll_wait, EPOLLIN | pollflags);
532 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
537 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
540 wake_up_poll(&ep->poll_wait, EPOLLIN | pollflags);
545 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
547 wait_queue_head_t *whead;
551 * If it is cleared by POLLFREE, it should be rcu-safe.
552 * If we read NULL we need a barrier paired with
553 * smp_store_release() in ep_poll_callback(), otherwise
554 * we rely on whead->lock.
556 whead = smp_load_acquire(&pwq->whead);
558 remove_wait_queue(whead, &pwq->wait);
563 * This function unregisters poll callbacks from the associated file
564 * descriptor. Must be called with "mtx" held.
566 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
568 struct eppoll_entry **p = &epi->pwqlist;
569 struct eppoll_entry *pwq;
571 while ((pwq = *p) != NULL) {
573 ep_remove_wait_queue(pwq);
574 kmem_cache_free(pwq_cache, pwq);
578 /* call only when ep->mtx is held */
579 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
581 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
584 /* call only when ep->mtx is held */
585 static inline void ep_pm_stay_awake(struct epitem *epi)
587 struct wakeup_source *ws = ep_wakeup_source(epi);
593 static inline bool ep_has_wakeup_source(struct epitem *epi)
595 return rcu_access_pointer(epi->ws) ? true : false;
598 /* call when ep->mtx cannot be held (ep_poll_callback) */
599 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
601 struct wakeup_source *ws;
604 ws = rcu_dereference(epi->ws);
612 * ep->mutex needs to be held because we could be hit by
613 * eventpoll_release_file() and epoll_ctl().
615 static void ep_start_scan(struct eventpoll *ep, struct list_head *txlist)
618 * Steal the ready list, and re-init the original one to the
619 * empty list. Also, set ep->ovflist to NULL so that events
620 * happening while looping w/out locks, are not lost. We cannot
621 * have the poll callback to queue directly on ep->rdllist,
622 * because we want the "sproc" callback to be able to do it
625 lockdep_assert_irqs_enabled();
626 write_lock_irq(&ep->lock);
627 list_splice_init(&ep->rdllist, txlist);
628 WRITE_ONCE(ep->ovflist, NULL);
629 write_unlock_irq(&ep->lock);
632 static void ep_done_scan(struct eventpoll *ep,
633 struct list_head *txlist)
635 struct epitem *epi, *nepi;
637 write_lock_irq(&ep->lock);
639 * During the time we spent inside the "sproc" callback, some
640 * other events might have been queued by the poll callback.
641 * We re-insert them inside the main ready-list here.
643 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
644 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
646 * We need to check if the item is already in the list.
647 * During the "sproc" callback execution time, items are
648 * queued into ->ovflist but the "txlist" might already
649 * contain them, and the list_splice() below takes care of them.
651 if (!ep_is_linked(epi)) {
653 * ->ovflist is LIFO, so we have to reverse it in order
656 list_add(&epi->rdllink, &ep->rdllist);
657 ep_pm_stay_awake(epi);
661 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
662 * releasing the lock, events will be queued in the normal way inside
665 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
668 * Quickly re-inject items left on "txlist".
670 list_splice(txlist, &ep->rdllist);
673 if (!list_empty(&ep->rdllist)) {
674 if (waitqueue_active(&ep->wq))
678 write_unlock_irq(&ep->lock);
681 static void ep_get(struct eventpoll *ep)
683 refcount_inc(&ep->refcount);
687 * Returns true if the event poll can be disposed
689 static bool ep_refcount_dec_and_test(struct eventpoll *ep)
691 if (!refcount_dec_and_test(&ep->refcount))
694 WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep->rbr.rb_root));
698 static void ep_free(struct eventpoll *ep)
700 mutex_destroy(&ep->mtx);
702 wakeup_source_unregister(ep->ws);
707 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
708 * all the associated resources. Must be called with "mtx" held.
709 * If the dying flag is set, do the removal only if force is true.
710 * This prevents ep_clear_and_put() from dropping all the ep references
711 * while running concurrently with eventpoll_release_file().
712 * Returns true if the eventpoll can be disposed.
714 static bool __ep_remove(struct eventpoll *ep, struct epitem *epi, bool force)
716 struct file *file = epi->ffd.file;
717 struct epitems_head *to_free;
718 struct hlist_head *head;
720 lockdep_assert_irqs_enabled();
723 * Removes poll wait queue hooks.
725 ep_unregister_pollwait(ep, epi);
727 /* Remove the current item from the list of epoll hooks */
728 spin_lock(&file->f_lock);
729 if (epi->dying && !force) {
730 spin_unlock(&file->f_lock);
736 if (head->first == &epi->fllink && !epi->fllink.next) {
738 if (!is_file_epoll(file)) {
739 struct epitems_head *v;
740 v = container_of(head, struct epitems_head, epitems);
741 if (!smp_load_acquire(&v->next))
745 hlist_del_rcu(&epi->fllink);
746 spin_unlock(&file->f_lock);
747 free_ephead(to_free);
749 rb_erase_cached(&epi->rbn, &ep->rbr);
751 write_lock_irq(&ep->lock);
752 if (ep_is_linked(epi))
753 list_del_init(&epi->rdllink);
754 write_unlock_irq(&ep->lock);
756 wakeup_source_unregister(ep_wakeup_source(epi));
758 * At this point it is safe to free the eventpoll item. Use the union
759 * field epi->rcu, since we are trying to minimize the size of
760 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
761 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
762 * use of the rbn field.
766 percpu_counter_dec(&ep->user->epoll_watches);
767 return ep_refcount_dec_and_test(ep);
771 * ep_remove variant for callers owing an additional reference to the ep
773 static void ep_remove_safe(struct eventpoll *ep, struct epitem *epi)
775 WARN_ON_ONCE(__ep_remove(ep, epi, false));
778 static void ep_clear_and_put(struct eventpoll *ep)
780 struct rb_node *rbp, *next;
784 /* We need to release all tasks waiting for these file */
785 if (waitqueue_active(&ep->poll_wait))
786 ep_poll_safewake(ep, NULL, 0);
788 mutex_lock(&ep->mtx);
791 * Walks through the whole tree by unregistering poll callbacks.
793 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
794 epi = rb_entry(rbp, struct epitem, rbn);
796 ep_unregister_pollwait(ep, epi);
801 * Walks through the whole tree and try to free each "struct epitem".
802 * Note that ep_remove_safe() will not remove the epitem in case of a
803 * racing eventpoll_release_file(); the latter will do the removal.
804 * At this point we are sure no poll callbacks will be lingering around.
805 * Since we still own a reference to the eventpoll struct, the loop can't
808 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = next) {
810 epi = rb_entry(rbp, struct epitem, rbn);
811 ep_remove_safe(ep, epi);
815 dispose = ep_refcount_dec_and_test(ep);
816 mutex_unlock(&ep->mtx);
822 static int ep_eventpoll_release(struct inode *inode, struct file *file)
824 struct eventpoll *ep = file->private_data;
827 ep_clear_and_put(ep);
832 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt, int depth);
834 static __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int depth)
836 struct eventpoll *ep = file->private_data;
838 struct epitem *epi, *tmp;
842 init_poll_funcptr(&pt, NULL);
844 /* Insert inside our poll wait queue */
845 poll_wait(file, &ep->poll_wait, wait);
848 * Proceed to find out if wanted events are really available inside
851 mutex_lock_nested(&ep->mtx, depth);
852 ep_start_scan(ep, &txlist);
853 list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {
854 if (ep_item_poll(epi, &pt, depth + 1)) {
855 res = EPOLLIN | EPOLLRDNORM;
859 * Item has been dropped into the ready list by the poll
860 * callback, but it's not actually ready, as far as
861 * caller requested events goes. We can remove it here.
863 __pm_relax(ep_wakeup_source(epi));
864 list_del_init(&epi->rdllink);
867 ep_done_scan(ep, &txlist);
868 mutex_unlock(&ep->mtx);
873 * Differs from ep_eventpoll_poll() in that internal callers already have
874 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
875 * is correctly annotated.
877 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
880 struct file *file = epi->ffd.file;
883 pt->_key = epi->event.events;
884 if (!is_file_epoll(file))
885 res = vfs_poll(file, pt);
887 res = __ep_eventpoll_poll(file, pt, depth);
888 return res & epi->event.events;
891 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
893 return __ep_eventpoll_poll(file, wait, 0);
896 #ifdef CONFIG_PROC_FS
897 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
899 struct eventpoll *ep = f->private_data;
902 mutex_lock(&ep->mtx);
903 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
904 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
905 struct inode *inode = file_inode(epi->ffd.file);
907 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
908 " pos:%lli ino:%lx sdev:%x\n",
909 epi->ffd.fd, epi->event.events,
910 (long long)epi->event.data,
911 (long long)epi->ffd.file->f_pos,
912 inode->i_ino, inode->i_sb->s_dev);
913 if (seq_has_overflowed(m))
916 mutex_unlock(&ep->mtx);
920 /* File callbacks that implement the eventpoll file behaviour */
921 static const struct file_operations eventpoll_fops = {
922 #ifdef CONFIG_PROC_FS
923 .show_fdinfo = ep_show_fdinfo,
925 .release = ep_eventpoll_release,
926 .poll = ep_eventpoll_poll,
927 .llseek = noop_llseek,
931 * This is called from eventpoll_release() to unlink files from the eventpoll
932 * interface. We need to have this facility to cleanup correctly files that are
933 * closed without being removed from the eventpoll interface.
935 void eventpoll_release_file(struct file *file)
937 struct eventpoll *ep;
942 * Use the 'dying' flag to prevent a concurrent ep_clear_and_put() from
943 * touching the epitems list before eventpoll_release_file() can access
947 spin_lock(&file->f_lock);
948 if (file->f_ep && file->f_ep->first) {
949 epi = hlist_entry(file->f_ep->first, struct epitem, fllink);
951 spin_unlock(&file->f_lock);
954 * ep access is safe as we still own a reference to the ep
958 mutex_lock(&ep->mtx);
959 dispose = __ep_remove(ep, epi, true);
960 mutex_unlock(&ep->mtx);
966 spin_unlock(&file->f_lock);
969 static int ep_alloc(struct eventpoll **pep)
971 struct eventpoll *ep;
973 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
977 mutex_init(&ep->mtx);
978 rwlock_init(&ep->lock);
979 init_waitqueue_head(&ep->wq);
980 init_waitqueue_head(&ep->poll_wait);
981 INIT_LIST_HEAD(&ep->rdllist);
982 ep->rbr = RB_ROOT_CACHED;
983 ep->ovflist = EP_UNACTIVE_PTR;
984 ep->user = get_current_user();
985 refcount_set(&ep->refcount, 1);
993 * Search the file inside the eventpoll tree. The RB tree operations
994 * are protected by the "mtx" mutex, and ep_find() must be called with
997 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1000 struct rb_node *rbp;
1001 struct epitem *epi, *epir = NULL;
1002 struct epoll_filefd ffd;
1004 ep_set_ffd(&ffd, file, fd);
1005 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1006 epi = rb_entry(rbp, struct epitem, rbn);
1007 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1009 rbp = rbp->rb_right;
1022 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1024 struct rb_node *rbp;
1027 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1028 epi = rb_entry(rbp, struct epitem, rbn);
1029 if (epi->ffd.fd == tfd) {
1041 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1044 struct file *file_raw;
1045 struct eventpoll *ep;
1048 if (!is_file_epoll(file))
1049 return ERR_PTR(-EINVAL);
1051 ep = file->private_data;
1053 mutex_lock(&ep->mtx);
1054 epi = ep_find_tfd(ep, tfd, toff);
1056 file_raw = epi->ffd.file;
1058 file_raw = ERR_PTR(-ENOENT);
1059 mutex_unlock(&ep->mtx);
1063 #endif /* CONFIG_KCMP */
1066 * Adds a new entry to the tail of the list in a lockless way, i.e.
1067 * multiple CPUs are allowed to call this function concurrently.
1069 * Beware: it is necessary to prevent any other modifications of the
1070 * existing list until all changes are completed, in other words
1071 * concurrent list_add_tail_lockless() calls should be protected
1072 * with a read lock, where write lock acts as a barrier which
1073 * makes sure all list_add_tail_lockless() calls are fully
1076 * Also an element can be locklessly added to the list only in one
1077 * direction i.e. either to the tail or to the head, otherwise
1078 * concurrent access will corrupt the list.
1080 * Return: %false if element has been already added to the list, %true
1083 static inline bool list_add_tail_lockless(struct list_head *new,
1084 struct list_head *head)
1086 struct list_head *prev;
1089 * This is simple 'new->next = head' operation, but cmpxchg()
1090 * is used in order to detect that same element has been just
1091 * added to the list from another CPU: the winner observes
1094 if (!try_cmpxchg(&new->next, &new, head))
1098 * Initially ->next of a new element must be updated with the head
1099 * (we are inserting to the tail) and only then pointers are atomically
1100 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1101 * updated before pointers are actually swapped and pointers are
1102 * swapped before prev->next is updated.
1105 prev = xchg(&head->prev, new);
1108 * It is safe to modify prev->next and new->prev, because a new element
1109 * is added only to the tail and new->next is updated before XCHG.
1119 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1120 * i.e. multiple CPUs are allowed to call this function concurrently.
1122 * Return: %false if epi element has been already chained, %true otherwise.
1124 static inline bool chain_epi_lockless(struct epitem *epi)
1126 struct eventpoll *ep = epi->ep;
1128 /* Fast preliminary check */
1129 if (epi->next != EP_UNACTIVE_PTR)
1132 /* Check that the same epi has not been just chained from another CPU */
1133 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1136 /* Atomically exchange tail */
1137 epi->next = xchg(&ep->ovflist, epi);
1143 * This is the callback that is passed to the wait queue wakeup
1144 * mechanism. It is called by the stored file descriptors when they
1145 * have events to report.
1147 * This callback takes a read lock in order not to contend with concurrent
1148 * events from another file descriptor, thus all modifications to ->rdllist
1149 * or ->ovflist are lockless. Read lock is paired with the write lock from
1150 * ep_start/done_scan(), which stops all list modifications and guarantees
1151 * that lists state is seen correctly.
1153 * Another thing worth to mention is that ep_poll_callback() can be called
1154 * concurrently for the same @epi from different CPUs if poll table was inited
1155 * with several wait queues entries. Plural wakeup from different CPUs of a
1156 * single wait queue is serialized by wq.lock, but the case when multiple wait
1157 * queues are used should be detected accordingly. This is detected using
1158 * cmpxchg() operation.
1160 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1163 struct epitem *epi = ep_item_from_wait(wait);
1164 struct eventpoll *ep = epi->ep;
1165 __poll_t pollflags = key_to_poll(key);
1166 unsigned long flags;
1169 read_lock_irqsave(&ep->lock, flags);
1171 ep_set_busy_poll_napi_id(epi);
1174 * If the event mask does not contain any poll(2) event, we consider the
1175 * descriptor to be disabled. This condition is likely the effect of the
1176 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1177 * until the next EPOLL_CTL_MOD will be issued.
1179 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1183 * Check the events coming with the callback. At this stage, not
1184 * every device reports the events in the "key" parameter of the
1185 * callback. We need to be able to handle both cases here, hence the
1186 * test for "key" != NULL before the event match test.
1188 if (pollflags && !(pollflags & epi->event.events))
1192 * If we are transferring events to userspace, we can hold no locks
1193 * (because we're accessing user memory, and because of linux f_op->poll()
1194 * semantics). All the events that happen during that period of time are
1195 * chained in ep->ovflist and requeued later on.
1197 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1198 if (chain_epi_lockless(epi))
1199 ep_pm_stay_awake_rcu(epi);
1200 } else if (!ep_is_linked(epi)) {
1201 /* In the usual case, add event to ready list. */
1202 if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1203 ep_pm_stay_awake_rcu(epi);
1207 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1210 if (waitqueue_active(&ep->wq)) {
1211 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1212 !(pollflags & POLLFREE)) {
1213 switch (pollflags & EPOLLINOUT_BITS) {
1215 if (epi->event.events & EPOLLIN)
1219 if (epi->event.events & EPOLLOUT)
1229 if (waitqueue_active(&ep->poll_wait))
1233 read_unlock_irqrestore(&ep->lock, flags);
1235 /* We have to call this outside the lock */
1237 ep_poll_safewake(ep, epi, pollflags & EPOLL_URING_WAKE);
1239 if (!(epi->event.events & EPOLLEXCLUSIVE))
1242 if (pollflags & POLLFREE) {
1244 * If we race with ep_remove_wait_queue() it can miss
1245 * ->whead = NULL and do another remove_wait_queue() after
1246 * us, so we can't use __remove_wait_queue().
1248 list_del_init(&wait->entry);
1250 * ->whead != NULL protects us from the race with
1251 * ep_clear_and_put() or ep_remove(), ep_remove_wait_queue()
1252 * takes whead->lock held by the caller. Once we nullify it,
1253 * nothing protects ep/epi or even wait.
1255 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1262 * This is the callback that is used to add our wait queue to the
1263 * target file wakeup lists.
1265 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1268 struct ep_pqueue *epq = container_of(pt, struct ep_pqueue, pt);
1269 struct epitem *epi = epq->epi;
1270 struct eppoll_entry *pwq;
1272 if (unlikely(!epi)) // an earlier allocation has failed
1275 pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL);
1276 if (unlikely(!pwq)) {
1281 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1284 if (epi->event.events & EPOLLEXCLUSIVE)
1285 add_wait_queue_exclusive(whead, &pwq->wait);
1287 add_wait_queue(whead, &pwq->wait);
1288 pwq->next = epi->pwqlist;
1292 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1295 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1296 struct epitem *epic;
1297 bool leftmost = true;
1301 epic = rb_entry(parent, struct epitem, rbn);
1302 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1304 p = &parent->rb_right;
1307 p = &parent->rb_left;
1309 rb_link_node(&epi->rbn, parent, p);
1310 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1315 #define PATH_ARR_SIZE 5
1317 * These are the number paths of length 1 to 5, that we are allowing to emanate
1318 * from a single file of interest. For example, we allow 1000 paths of length
1319 * 1, to emanate from each file of interest. This essentially represents the
1320 * potential wakeup paths, which need to be limited in order to avoid massive
1321 * uncontrolled wakeup storms. The common use case should be a single ep which
1322 * is connected to n file sources. In this case each file source has 1 path
1323 * of length 1. Thus, the numbers below should be more than sufficient. These
1324 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1325 * and delete can't add additional paths. Protected by the epnested_mutex.
1327 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1328 static int path_count[PATH_ARR_SIZE];
1330 static int path_count_inc(int nests)
1332 /* Allow an arbitrary number of depth 1 paths */
1336 if (++path_count[nests] > path_limits[nests])
1341 static void path_count_init(void)
1345 for (i = 0; i < PATH_ARR_SIZE; i++)
1349 static int reverse_path_check_proc(struct hlist_head *refs, int depth)
1354 if (depth > EP_MAX_NESTS) /* too deep nesting */
1357 /* CTL_DEL can remove links here, but that can't increase our count */
1358 hlist_for_each_entry_rcu(epi, refs, fllink) {
1359 struct hlist_head *refs = &epi->ep->refs;
1360 if (hlist_empty(refs))
1361 error = path_count_inc(depth);
1363 error = reverse_path_check_proc(refs, depth + 1);
1371 * reverse_path_check - The tfile_check_list is list of epitem_head, which have
1372 * links that are proposed to be newly added. We need to
1373 * make sure that those added links don't add too many
1374 * paths such that we will spend all our time waking up
1375 * eventpoll objects.
1377 * Return: %zero if the proposed links don't create too many paths,
1380 static int reverse_path_check(void)
1382 struct epitems_head *p;
1384 for (p = tfile_check_list; p != EP_UNACTIVE_PTR; p = p->next) {
1388 error = reverse_path_check_proc(&p->epitems, 0);
1396 static int ep_create_wakeup_source(struct epitem *epi)
1398 struct name_snapshot n;
1399 struct wakeup_source *ws;
1402 epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1407 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1408 ws = wakeup_source_register(NULL, n.name.name);
1409 release_dentry_name_snapshot(&n);
1413 rcu_assign_pointer(epi->ws, ws);
1418 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1419 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1421 struct wakeup_source *ws = ep_wakeup_source(epi);
1423 RCU_INIT_POINTER(epi->ws, NULL);
1426 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1427 * used internally by wakeup_source_remove, too (called by
1428 * wakeup_source_unregister), so we cannot use call_rcu
1431 wakeup_source_unregister(ws);
1434 static int attach_epitem(struct file *file, struct epitem *epi)
1436 struct epitems_head *to_free = NULL;
1437 struct hlist_head *head = NULL;
1438 struct eventpoll *ep = NULL;
1440 if (is_file_epoll(file))
1441 ep = file->private_data;
1445 } else if (!READ_ONCE(file->f_ep)) {
1447 to_free = kmem_cache_zalloc(ephead_cache, GFP_KERNEL);
1450 head = &to_free->epitems;
1452 spin_lock(&file->f_lock);
1454 if (unlikely(!head)) {
1455 spin_unlock(&file->f_lock);
1461 hlist_add_head_rcu(&epi->fllink, file->f_ep);
1462 spin_unlock(&file->f_lock);
1463 free_ephead(to_free);
1468 * Must be called with "mtx" held.
1470 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1471 struct file *tfile, int fd, int full_check)
1473 int error, pwake = 0;
1476 struct ep_pqueue epq;
1477 struct eventpoll *tep = NULL;
1479 if (is_file_epoll(tfile))
1480 tep = tfile->private_data;
1482 lockdep_assert_irqs_enabled();
1484 if (unlikely(percpu_counter_compare(&ep->user->epoll_watches,
1485 max_user_watches) >= 0))
1487 percpu_counter_inc(&ep->user->epoll_watches);
1489 if (!(epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL))) {
1490 percpu_counter_dec(&ep->user->epoll_watches);
1494 /* Item initialization follow here ... */
1495 INIT_LIST_HEAD(&epi->rdllink);
1497 ep_set_ffd(&epi->ffd, tfile, fd);
1498 epi->event = *event;
1499 epi->next = EP_UNACTIVE_PTR;
1502 mutex_lock_nested(&tep->mtx, 1);
1503 /* Add the current item to the list of active epoll hook for this file */
1504 if (unlikely(attach_epitem(tfile, epi) < 0)) {
1506 mutex_unlock(&tep->mtx);
1507 kmem_cache_free(epi_cache, epi);
1508 percpu_counter_dec(&ep->user->epoll_watches);
1512 if (full_check && !tep)
1516 * Add the current item to the RB tree. All RB tree operations are
1517 * protected by "mtx", and ep_insert() is called with "mtx" held.
1519 ep_rbtree_insert(ep, epi);
1521 mutex_unlock(&tep->mtx);
1524 * ep_remove_safe() calls in the later error paths can't lead to
1525 * ep_free() as the ep file itself still holds an ep reference.
1529 /* now check if we've created too many backpaths */
1530 if (unlikely(full_check && reverse_path_check())) {
1531 ep_remove_safe(ep, epi);
1535 if (epi->event.events & EPOLLWAKEUP) {
1536 error = ep_create_wakeup_source(epi);
1538 ep_remove_safe(ep, epi);
1543 /* Initialize the poll table using the queue callback */
1545 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1548 * Attach the item to the poll hooks and get current event bits.
1549 * We can safely use the file* here because its usage count has
1550 * been increased by the caller of this function. Note that after
1551 * this operation completes, the poll callback can start hitting
1554 revents = ep_item_poll(epi, &epq.pt, 1);
1557 * We have to check if something went wrong during the poll wait queue
1558 * install process. Namely an allocation for a wait queue failed due
1559 * high memory pressure.
1561 if (unlikely(!epq.epi)) {
1562 ep_remove_safe(ep, epi);
1566 /* We have to drop the new item inside our item list to keep track of it */
1567 write_lock_irq(&ep->lock);
1569 /* record NAPI ID of new item if present */
1570 ep_set_busy_poll_napi_id(epi);
1572 /* If the file is already "ready" we drop it inside the ready list */
1573 if (revents && !ep_is_linked(epi)) {
1574 list_add_tail(&epi->rdllink, &ep->rdllist);
1575 ep_pm_stay_awake(epi);
1577 /* Notify waiting tasks that events are available */
1578 if (waitqueue_active(&ep->wq))
1580 if (waitqueue_active(&ep->poll_wait))
1584 write_unlock_irq(&ep->lock);
1586 /* We have to call this outside the lock */
1588 ep_poll_safewake(ep, NULL, 0);
1594 * Modify the interest event mask by dropping an event if the new mask
1595 * has a match in the current file status. Must be called with "mtx" held.
1597 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1598 const struct epoll_event *event)
1603 lockdep_assert_irqs_enabled();
1605 init_poll_funcptr(&pt, NULL);
1608 * Set the new event interest mask before calling f_op->poll();
1609 * otherwise we might miss an event that happens between the
1610 * f_op->poll() call and the new event set registering.
1612 epi->event.events = event->events; /* need barrier below */
1613 epi->event.data = event->data; /* protected by mtx */
1614 if (epi->event.events & EPOLLWAKEUP) {
1615 if (!ep_has_wakeup_source(epi))
1616 ep_create_wakeup_source(epi);
1617 } else if (ep_has_wakeup_source(epi)) {
1618 ep_destroy_wakeup_source(epi);
1622 * The following barrier has two effects:
1624 * 1) Flush epi changes above to other CPUs. This ensures
1625 * we do not miss events from ep_poll_callback if an
1626 * event occurs immediately after we call f_op->poll().
1627 * We need this because we did not take ep->lock while
1628 * changing epi above (but ep_poll_callback does take
1631 * 2) We also need to ensure we do not miss _past_ events
1632 * when calling f_op->poll(). This barrier also
1633 * pairs with the barrier in wq_has_sleeper (see
1634 * comments for wq_has_sleeper).
1636 * This barrier will now guarantee ep_poll_callback or f_op->poll
1637 * (or both) will notice the readiness of an item.
1642 * Get current event bits. We can safely use the file* here because
1643 * its usage count has been increased by the caller of this function.
1644 * If the item is "hot" and it is not registered inside the ready
1645 * list, push it inside.
1647 if (ep_item_poll(epi, &pt, 1)) {
1648 write_lock_irq(&ep->lock);
1649 if (!ep_is_linked(epi)) {
1650 list_add_tail(&epi->rdllink, &ep->rdllist);
1651 ep_pm_stay_awake(epi);
1653 /* Notify waiting tasks that events are available */
1654 if (waitqueue_active(&ep->wq))
1656 if (waitqueue_active(&ep->poll_wait))
1659 write_unlock_irq(&ep->lock);
1662 /* We have to call this outside the lock */
1664 ep_poll_safewake(ep, NULL, 0);
1669 static int ep_send_events(struct eventpoll *ep,
1670 struct epoll_event __user *events, int maxevents)
1672 struct epitem *epi, *tmp;
1678 * Always short-circuit for fatal signals to allow threads to make a
1679 * timely exit without the chance of finding more events available and
1680 * fetching repeatedly.
1682 if (fatal_signal_pending(current))
1685 init_poll_funcptr(&pt, NULL);
1687 mutex_lock(&ep->mtx);
1688 ep_start_scan(ep, &txlist);
1691 * We can loop without lock because we are passed a task private list.
1692 * Items cannot vanish during the loop we are holding ep->mtx.
1694 list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {
1695 struct wakeup_source *ws;
1698 if (res >= maxevents)
1702 * Activate ep->ws before deactivating epi->ws to prevent
1703 * triggering auto-suspend here (in case we reactive epi->ws
1706 * This could be rearranged to delay the deactivation of epi->ws
1707 * instead, but then epi->ws would temporarily be out of sync
1708 * with ep_is_linked().
1710 ws = ep_wakeup_source(epi);
1713 __pm_stay_awake(ep->ws);
1717 list_del_init(&epi->rdllink);
1720 * If the event mask intersect the caller-requested one,
1721 * deliver the event to userspace. Again, we are holding ep->mtx,
1722 * so no operations coming from userspace can change the item.
1724 revents = ep_item_poll(epi, &pt, 1);
1728 events = epoll_put_uevent(revents, epi->event.data, events);
1730 list_add(&epi->rdllink, &txlist);
1731 ep_pm_stay_awake(epi);
1737 if (epi->event.events & EPOLLONESHOT)
1738 epi->event.events &= EP_PRIVATE_BITS;
1739 else if (!(epi->event.events & EPOLLET)) {
1741 * If this file has been added with Level
1742 * Trigger mode, we need to insert back inside
1743 * the ready list, so that the next call to
1744 * epoll_wait() will check again the events
1745 * availability. At this point, no one can insert
1746 * into ep->rdllist besides us. The epoll_ctl()
1747 * callers are locked out by
1748 * ep_send_events() holding "mtx" and the
1749 * poll callback will queue them in ep->ovflist.
1751 list_add_tail(&epi->rdllink, &ep->rdllist);
1752 ep_pm_stay_awake(epi);
1755 ep_done_scan(ep, &txlist);
1756 mutex_unlock(&ep->mtx);
1761 static struct timespec64 *ep_timeout_to_timespec(struct timespec64 *to, long ms)
1763 struct timespec64 now;
1774 to->tv_sec = ms / MSEC_PER_SEC;
1775 to->tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC);
1777 ktime_get_ts64(&now);
1778 *to = timespec64_add_safe(now, *to);
1783 * autoremove_wake_function, but remove even on failure to wake up, because we
1784 * know that default_wake_function/ttwu will only fail if the thread is already
1785 * woken, and in that case the ep_poll loop will remove the entry anyways, not
1788 static int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,
1789 unsigned int mode, int sync, void *key)
1791 int ret = default_wake_function(wq_entry, mode, sync, key);
1794 * Pairs with list_empty_careful in ep_poll, and ensures future loop
1795 * iterations see the cause of this wakeup.
1797 list_del_init_careful(&wq_entry->entry);
1802 * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
1805 * @ep: Pointer to the eventpoll context.
1806 * @events: Pointer to the userspace buffer where the ready events should be
1808 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1809 * @timeout: Maximum timeout for the ready events fetch operation, in
1810 * timespec. If the timeout is zero, the function will not block,
1811 * while if the @timeout ptr is NULL, the function will block
1812 * until at least one event has been retrieved (or an error
1815 * Return: the number of ready events which have been fetched, or an
1816 * error code, in case of error.
1818 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1819 int maxevents, struct timespec64 *timeout)
1821 int res, eavail, timed_out = 0;
1823 wait_queue_entry_t wait;
1824 ktime_t expires, *to = NULL;
1826 lockdep_assert_irqs_enabled();
1828 if (timeout && (timeout->tv_sec | timeout->tv_nsec)) {
1829 slack = select_estimate_accuracy(timeout);
1831 *to = timespec64_to_ktime(*timeout);
1832 } else if (timeout) {
1834 * Avoid the unnecessary trip to the wait queue loop, if the
1835 * caller specified a non blocking operation.
1841 * This call is racy: We may or may not see events that are being added
1842 * to the ready list under the lock (e.g., in IRQ callbacks). For cases
1843 * with a non-zero timeout, this thread will check the ready list under
1844 * lock and will add to the wait queue. For cases with a zero
1845 * timeout, the user by definition should not care and will have to
1848 eavail = ep_events_available(ep);
1853 * Try to transfer events to user space. In case we get
1854 * 0 events and there's still timeout left over, we go
1855 * trying again in search of more luck.
1857 res = ep_send_events(ep, events, maxevents);
1865 eavail = ep_busy_loop(ep, timed_out);
1869 if (signal_pending(current))
1873 * Internally init_wait() uses autoremove_wake_function(),
1874 * thus wait entry is removed from the wait queue on each
1875 * wakeup. Why it is important? In case of several waiters
1876 * each new wakeup will hit the next waiter, giving it the
1877 * chance to harvest new event. Otherwise wakeup can be
1878 * lost. This is also good performance-wise, because on
1879 * normal wakeup path no need to call __remove_wait_queue()
1880 * explicitly, thus ep->lock is not taken, which halts the
1883 * In fact, we now use an even more aggressive function that
1884 * unconditionally removes, because we don't reuse the wait
1885 * entry between loop iterations. This lets us also avoid the
1886 * performance issue if a process is killed, causing all of its
1887 * threads to wake up without being removed normally.
1890 wait.func = ep_autoremove_wake_function;
1892 write_lock_irq(&ep->lock);
1894 * Barrierless variant, waitqueue_active() is called under
1895 * the same lock on wakeup ep_poll_callback() side, so it
1896 * is safe to avoid an explicit barrier.
1898 __set_current_state(TASK_INTERRUPTIBLE);
1901 * Do the final check under the lock. ep_start/done_scan()
1902 * plays with two lists (->rdllist and ->ovflist) and there
1903 * is always a race when both lists are empty for short
1904 * period of time although events are pending, so lock is
1907 eavail = ep_events_available(ep);
1909 __add_wait_queue_exclusive(&ep->wq, &wait);
1911 write_unlock_irq(&ep->lock);
1914 timed_out = !schedule_hrtimeout_range(to, slack,
1916 __set_current_state(TASK_RUNNING);
1919 * We were woken up, thus go and try to harvest some events.
1920 * If timed out and still on the wait queue, recheck eavail
1921 * carefully under lock, below.
1925 if (!list_empty_careful(&wait.entry)) {
1926 write_lock_irq(&ep->lock);
1928 * If the thread timed out and is not on the wait queue,
1929 * it means that the thread was woken up after its
1930 * timeout expired before it could reacquire the lock.
1931 * Thus, when wait.entry is empty, it needs to harvest
1935 eavail = list_empty(&wait.entry);
1936 __remove_wait_queue(&ep->wq, &wait);
1937 write_unlock_irq(&ep->lock);
1943 * ep_loop_check_proc - verify that adding an epoll file inside another
1944 * epoll structure does not violate the constraints, in
1945 * terms of closed loops, or too deep chains (which can
1946 * result in excessive stack usage).
1948 * @ep: the &struct eventpoll to be currently checked.
1949 * @depth: Current depth of the path being checked.
1951 * Return: %zero if adding the epoll @file inside current epoll
1952 * structure @ep does not violate the constraints, or %-1 otherwise.
1954 static int ep_loop_check_proc(struct eventpoll *ep, int depth)
1957 struct rb_node *rbp;
1960 mutex_lock_nested(&ep->mtx, depth + 1);
1961 ep->gen = loop_check_gen;
1962 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1963 epi = rb_entry(rbp, struct epitem, rbn);
1964 if (unlikely(is_file_epoll(epi->ffd.file))) {
1965 struct eventpoll *ep_tovisit;
1966 ep_tovisit = epi->ffd.file->private_data;
1967 if (ep_tovisit->gen == loop_check_gen)
1969 if (ep_tovisit == inserting_into || depth > EP_MAX_NESTS)
1972 error = ep_loop_check_proc(ep_tovisit, depth + 1);
1977 * If we've reached a file that is not associated with
1978 * an ep, then we need to check if the newly added
1979 * links are going to add too many wakeup paths. We do
1980 * this by adding it to the tfile_check_list, if it's
1981 * not already there, and calling reverse_path_check()
1982 * during ep_insert().
1984 list_file(epi->ffd.file);
1987 mutex_unlock(&ep->mtx);
1993 * ep_loop_check - Performs a check to verify that adding an epoll file (@to)
1994 * into another epoll file (represented by @ep) does not create
1995 * closed loops or too deep chains.
1997 * @ep: Pointer to the epoll we are inserting into.
1998 * @to: Pointer to the epoll to be inserted.
2000 * Return: %zero if adding the epoll @to inside the epoll @from
2001 * does not violate the constraints, or %-1 otherwise.
2003 static int ep_loop_check(struct eventpoll *ep, struct eventpoll *to)
2005 inserting_into = ep;
2006 return ep_loop_check_proc(to, 0);
2009 static void clear_tfile_check_list(void)
2012 while (tfile_check_list != EP_UNACTIVE_PTR) {
2013 struct epitems_head *head = tfile_check_list;
2014 tfile_check_list = head->next;
2021 * Open an eventpoll file descriptor.
2023 static int do_epoll_create(int flags)
2026 struct eventpoll *ep = NULL;
2029 /* Check the EPOLL_* constant for consistency. */
2030 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2032 if (flags & ~EPOLL_CLOEXEC)
2035 * Create the internal data structure ("struct eventpoll").
2037 error = ep_alloc(&ep);
2041 * Creates all the items needed to setup an eventpoll file. That is,
2042 * a file structure and a free file descriptor.
2044 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2049 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2050 O_RDWR | (flags & O_CLOEXEC));
2052 error = PTR_ERR(file);
2056 fd_install(fd, file);
2062 ep_clear_and_put(ep);
2066 SYSCALL_DEFINE1(epoll_create1, int, flags)
2068 return do_epoll_create(flags);
2071 SYSCALL_DEFINE1(epoll_create, int, size)
2076 return do_epoll_create(0);
2079 #ifdef CONFIG_PM_SLEEP
2080 static inline void ep_take_care_of_epollwakeup(struct epoll_event *epev)
2082 if ((epev->events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
2083 epev->events &= ~EPOLLWAKEUP;
2086 static inline void ep_take_care_of_epollwakeup(struct epoll_event *epev)
2088 epev->events &= ~EPOLLWAKEUP;
2092 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2096 mutex_lock_nested(mutex, depth);
2099 if (mutex_trylock(mutex))
2104 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2110 struct eventpoll *ep;
2112 struct eventpoll *tep = NULL;
2119 /* Get the "struct file *" for the target file */
2124 /* The target file descriptor must support poll */
2126 if (!file_can_poll(tf.file))
2127 goto error_tgt_fput;
2129 /* Check if EPOLLWAKEUP is allowed */
2130 if (ep_op_has_event(op))
2131 ep_take_care_of_epollwakeup(epds);
2134 * We have to check that the file structure underneath the file descriptor
2135 * the user passed to us _is_ an eventpoll file. And also we do not permit
2136 * adding an epoll file descriptor inside itself.
2139 if (f.file == tf.file || !is_file_epoll(f.file))
2140 goto error_tgt_fput;
2143 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2144 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2145 * Also, we do not currently supported nested exclusive wakeups.
2147 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2148 if (op == EPOLL_CTL_MOD)
2149 goto error_tgt_fput;
2150 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2151 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2152 goto error_tgt_fput;
2156 * At this point it is safe to assume that the "private_data" contains
2157 * our own data structure.
2159 ep = f.file->private_data;
2162 * When we insert an epoll file descriptor inside another epoll file
2163 * descriptor, there is the chance of creating closed loops, which are
2164 * better be handled here, than in more critical paths. While we are
2165 * checking for loops we also determine the list of files reachable
2166 * and hang them on the tfile_check_list, so we can check that we
2167 * haven't created too many possible wakeup paths.
2169 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2170 * the epoll file descriptor is attaching directly to a wakeup source,
2171 * unless the epoll file descriptor is nested. The purpose of taking the
2172 * 'epnested_mutex' on add is to prevent complex toplogies such as loops and
2173 * deep wakeup paths from forming in parallel through multiple
2174 * EPOLL_CTL_ADD operations.
2176 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2178 goto error_tgt_fput;
2179 if (op == EPOLL_CTL_ADD) {
2180 if (READ_ONCE(f.file->f_ep) || ep->gen == loop_check_gen ||
2181 is_file_epoll(tf.file)) {
2182 mutex_unlock(&ep->mtx);
2183 error = epoll_mutex_lock(&epnested_mutex, 0, nonblock);
2185 goto error_tgt_fput;
2188 if (is_file_epoll(tf.file)) {
2189 tep = tf.file->private_data;
2191 if (ep_loop_check(ep, tep) != 0)
2192 goto error_tgt_fput;
2194 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2196 goto error_tgt_fput;
2201 * Try to lookup the file inside our RB tree. Since we grabbed "mtx"
2202 * above, we can be sure to be able to use the item looked up by
2203 * ep_find() till we release the mutex.
2205 epi = ep_find(ep, tf.file, fd);
2211 epds->events |= EPOLLERR | EPOLLHUP;
2212 error = ep_insert(ep, epds, tf.file, fd, full_check);
2219 * The eventpoll itself is still alive: the refcount
2220 * can't go to zero here.
2222 ep_remove_safe(ep, epi);
2230 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2231 epds->events |= EPOLLERR | EPOLLHUP;
2232 error = ep_modify(ep, epi, epds);
2238 mutex_unlock(&ep->mtx);
2242 clear_tfile_check_list();
2244 mutex_unlock(&epnested_mutex);
2256 * The following function implements the controller interface for
2257 * the eventpoll file that enables the insertion/removal/change of
2258 * file descriptors inside the interest set.
2260 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2261 struct epoll_event __user *, event)
2263 struct epoll_event epds;
2265 if (ep_op_has_event(op) &&
2266 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2269 return do_epoll_ctl(epfd, op, fd, &epds, false);
2273 * Implement the event wait interface for the eventpoll file. It is the kernel
2274 * part of the user space epoll_wait(2).
2276 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2277 int maxevents, struct timespec64 *to)
2281 struct eventpoll *ep;
2283 /* The maximum number of event must be greater than zero */
2284 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2287 /* Verify that the area passed by the user is writeable */
2288 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2291 /* Get the "struct file *" for the eventpoll file */
2297 * We have to check that the file structure underneath the fd
2298 * the user passed to us _is_ an eventpoll file.
2301 if (!is_file_epoll(f.file))
2305 * At this point it is safe to assume that the "private_data" contains
2306 * our own data structure.
2308 ep = f.file->private_data;
2310 /* Time to fish for events ... */
2311 error = ep_poll(ep, events, maxevents, to);
2318 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2319 int, maxevents, int, timeout)
2321 struct timespec64 to;
2323 return do_epoll_wait(epfd, events, maxevents,
2324 ep_timeout_to_timespec(&to, timeout));
2328 * Implement the event wait interface for the eventpoll file. It is the kernel
2329 * part of the user space epoll_pwait(2).
2331 static int do_epoll_pwait(int epfd, struct epoll_event __user *events,
2332 int maxevents, struct timespec64 *to,
2333 const sigset_t __user *sigmask, size_t sigsetsize)
2338 * If the caller wants a certain signal mask to be set during the wait,
2341 error = set_user_sigmask(sigmask, sigsetsize);
2345 error = do_epoll_wait(epfd, events, maxevents, to);
2347 restore_saved_sigmask_unless(error == -EINTR);
2352 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2353 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2356 struct timespec64 to;
2358 return do_epoll_pwait(epfd, events, maxevents,
2359 ep_timeout_to_timespec(&to, timeout),
2360 sigmask, sigsetsize);
2363 SYSCALL_DEFINE6(epoll_pwait2, int, epfd, struct epoll_event __user *, events,
2364 int, maxevents, const struct __kernel_timespec __user *, timeout,
2365 const sigset_t __user *, sigmask, size_t, sigsetsize)
2367 struct timespec64 ts, *to = NULL;
2370 if (get_timespec64(&ts, timeout))
2373 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
2377 return do_epoll_pwait(epfd, events, maxevents, to,
2378 sigmask, sigsetsize);
2381 #ifdef CONFIG_COMPAT
2382 static int do_compat_epoll_pwait(int epfd, struct epoll_event __user *events,
2383 int maxevents, struct timespec64 *timeout,
2384 const compat_sigset_t __user *sigmask,
2385 compat_size_t sigsetsize)
2390 * If the caller wants a certain signal mask to be set during the wait,
2393 err = set_compat_user_sigmask(sigmask, sigsetsize);
2397 err = do_epoll_wait(epfd, events, maxevents, timeout);
2399 restore_saved_sigmask_unless(err == -EINTR);
2404 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2405 struct epoll_event __user *, events,
2406 int, maxevents, int, timeout,
2407 const compat_sigset_t __user *, sigmask,
2408 compat_size_t, sigsetsize)
2410 struct timespec64 to;
2412 return do_compat_epoll_pwait(epfd, events, maxevents,
2413 ep_timeout_to_timespec(&to, timeout),
2414 sigmask, sigsetsize);
2417 COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epfd,
2418 struct epoll_event __user *, events,
2420 const struct __kernel_timespec __user *, timeout,
2421 const compat_sigset_t __user *, sigmask,
2422 compat_size_t, sigsetsize)
2424 struct timespec64 ts, *to = NULL;
2427 if (get_timespec64(&ts, timeout))
2430 if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
2434 return do_compat_epoll_pwait(epfd, events, maxevents, to,
2435 sigmask, sigsetsize);
2440 static int __init eventpoll_init(void)
2446 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2448 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2450 BUG_ON(max_user_watches < 0);
2453 * We can have many thousands of epitems, so prevent this from
2454 * using an extra cache line on 64-bit (and smaller) CPUs
2456 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2458 /* Allocates slab cache used to allocate "struct epitem" items */
2459 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2460 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2462 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2463 pwq_cache = kmem_cache_create("eventpoll_pwq",
2464 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2465 epoll_sysctls_init();
2467 ephead_cache = kmem_cache_create("ep_head",
2468 sizeof(struct epitems_head), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2472 fs_initcall(eventpoll_init);