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_scan_ready_list 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
,
328 static void __init
epoll_sysctls_init(void)
330 register_sysctl("fs/epoll", epoll_table
);
333 #define epoll_sysctls_init() do { } while (0)
334 #endif /* CONFIG_SYSCTL */
336 static const struct file_operations eventpoll_fops
;
338 static inline int is_file_epoll(struct file
*f
)
340 return f
->f_op
== &eventpoll_fops
;
343 /* Setup the structure that is used as key for the RB tree */
344 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
345 struct file
*file
, int fd
)
351 /* Compare RB tree keys */
352 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
353 struct epoll_filefd
*p2
)
355 return (p1
->file
> p2
->file
? +1:
356 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
359 /* Tells us if the item is currently linked */
360 static inline int ep_is_linked(struct epitem
*epi
)
362 return !list_empty(&epi
->rdllink
);
365 static inline struct eppoll_entry
*ep_pwq_from_wait(wait_queue_entry_t
*p
)
367 return container_of(p
, struct eppoll_entry
, wait
);
370 /* Get the "struct epitem" from a wait queue pointer */
371 static inline struct epitem
*ep_item_from_wait(wait_queue_entry_t
*p
)
373 return container_of(p
, struct eppoll_entry
, wait
)->base
;
377 * ep_events_available - Checks if ready events might be available.
379 * @ep: Pointer to the eventpoll context.
381 * Return: a value different than %zero if ready events are available,
382 * or %zero otherwise.
384 static inline int ep_events_available(struct eventpoll
*ep
)
386 return !list_empty_careful(&ep
->rdllist
) ||
387 READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
;
390 #ifdef CONFIG_NET_RX_BUSY_POLL
391 static bool ep_busy_loop_end(void *p
, unsigned long start_time
)
393 struct eventpoll
*ep
= p
;
395 return ep_events_available(ep
) || busy_loop_timeout(start_time
);
399 * Busy poll if globally on and supporting sockets found && no events,
400 * busy loop will return if need_resched or ep_events_available.
402 * we must do our busy polling with irqs enabled
404 static bool ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
406 unsigned int napi_id
= READ_ONCE(ep
->napi_id
);
408 if ((napi_id
>= MIN_NAPI_ID
) && net_busy_loop_on()) {
409 napi_busy_loop(napi_id
, nonblock
? NULL
: ep_busy_loop_end
, ep
, false,
411 if (ep_events_available(ep
))
414 * Busy poll timed out. Drop NAPI ID for now, we can add
415 * it back in when we have moved a socket with a valid NAPI
416 * ID onto the ready list.
425 * Set epoll busy poll NAPI ID from sk.
427 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
429 struct eventpoll
*ep
;
430 unsigned int napi_id
;
434 if (!net_busy_loop_on())
437 sock
= sock_from_file(epi
->ffd
.file
);
445 napi_id
= READ_ONCE(sk
->sk_napi_id
);
448 /* Non-NAPI IDs can be rejected
450 * Nothing to do if we already have this ID
452 if (napi_id
< MIN_NAPI_ID
|| napi_id
== ep
->napi_id
)
455 /* record NAPI ID for use in next busy poll */
456 ep
->napi_id
= napi_id
;
461 static inline bool ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
466 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
470 #endif /* CONFIG_NET_RX_BUSY_POLL */
473 * As described in commit 0ccf831cb lockdep: annotate epoll
474 * the use of wait queues used by epoll is done in a very controlled
475 * manner. Wake ups can nest inside each other, but are never done
476 * with the same locking. For example:
479 * efd1 = epoll_create();
480 * efd2 = epoll_create();
481 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
482 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
484 * When a packet arrives to the device underneath "dfd", the net code will
485 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
486 * callback wakeup entry on that queue, and the wake_up() performed by the
487 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
488 * (efd1) notices that it may have some event ready, so it needs to wake up
489 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
490 * that ends up in another wake_up(), after having checked about the
491 * recursion constraints. That are, no more than EP_MAX_NESTS, to avoid
494 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
495 * this special case of epoll.
497 #ifdef CONFIG_DEBUG_LOCK_ALLOC
499 static void ep_poll_safewake(struct eventpoll
*ep
, struct epitem
*epi
,
502 struct eventpoll
*ep_src
;
507 * To set the subclass or nesting level for spin_lock_irqsave_nested()
508 * it might be natural to create a per-cpu nest count. However, since
509 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
510 * schedule() in the -rt kernel, the per-cpu variable are no longer
511 * protected. Thus, we are introducing a per eventpoll nest field.
512 * If we are not being call from ep_poll_callback(), epi is NULL and
513 * we are at the first level of nesting, 0. Otherwise, we are being
514 * called from ep_poll_callback() and if a previous wakeup source is
515 * not an epoll file itself, we are at depth 1 since the wakeup source
516 * is depth 0. If the wakeup source is a previous epoll file in the
517 * wakeup chain then we use its nests value and record ours as
518 * nests + 1. The previous epoll file nests value is stable since its
519 * already holding its own poll_wait.lock.
522 if ((is_file_epoll(epi
->ffd
.file
))) {
523 ep_src
= epi
->ffd
.file
->private_data
;
524 nests
= ep_src
->nests
;
529 spin_lock_irqsave_nested(&ep
->poll_wait
.lock
, flags
, nests
);
530 ep
->nests
= nests
+ 1;
531 wake_up_locked_poll(&ep
->poll_wait
, EPOLLIN
| pollflags
);
533 spin_unlock_irqrestore(&ep
->poll_wait
.lock
, flags
);
538 static void ep_poll_safewake(struct eventpoll
*ep
, struct epitem
*epi
,
541 wake_up_poll(&ep
->poll_wait
, EPOLLIN
| pollflags
);
546 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
548 wait_queue_head_t
*whead
;
552 * If it is cleared by POLLFREE, it should be rcu-safe.
553 * If we read NULL we need a barrier paired with
554 * smp_store_release() in ep_poll_callback(), otherwise
555 * we rely on whead->lock.
557 whead
= smp_load_acquire(&pwq
->whead
);
559 remove_wait_queue(whead
, &pwq
->wait
);
564 * This function unregisters poll callbacks from the associated file
565 * descriptor. Must be called with "mtx" held.
567 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
569 struct eppoll_entry
**p
= &epi
->pwqlist
;
570 struct eppoll_entry
*pwq
;
572 while ((pwq
= *p
) != NULL
) {
574 ep_remove_wait_queue(pwq
);
575 kmem_cache_free(pwq_cache
, pwq
);
579 /* call only when ep->mtx is held */
580 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
582 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
585 /* call only when ep->mtx is held */
586 static inline void ep_pm_stay_awake(struct epitem
*epi
)
588 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
594 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
596 return rcu_access_pointer(epi
->ws
) ? true : false;
599 /* call when ep->mtx cannot be held (ep_poll_callback) */
600 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
602 struct wakeup_source
*ws
;
605 ws
= rcu_dereference(epi
->ws
);
613 * ep->mutex needs to be held because we could be hit by
614 * eventpoll_release_file() and epoll_ctl().
616 static void ep_start_scan(struct eventpoll
*ep
, struct list_head
*txlist
)
619 * Steal the ready list, and re-init the original one to the
620 * empty list. Also, set ep->ovflist to NULL so that events
621 * happening while looping w/out locks, are not lost. We cannot
622 * have the poll callback to queue directly on ep->rdllist,
623 * because we want the "sproc" callback to be able to do it
626 lockdep_assert_irqs_enabled();
627 write_lock_irq(&ep
->lock
);
628 list_splice_init(&ep
->rdllist
, txlist
);
629 WRITE_ONCE(ep
->ovflist
, NULL
);
630 write_unlock_irq(&ep
->lock
);
633 static void ep_done_scan(struct eventpoll
*ep
,
634 struct list_head
*txlist
)
636 struct epitem
*epi
, *nepi
;
638 write_lock_irq(&ep
->lock
);
640 * During the time we spent inside the "sproc" callback, some
641 * other events might have been queued by the poll callback.
642 * We re-insert them inside the main ready-list here.
644 for (nepi
= READ_ONCE(ep
->ovflist
); (epi
= nepi
) != NULL
;
645 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
647 * We need to check if the item is already in the list.
648 * During the "sproc" callback execution time, items are
649 * queued into ->ovflist but the "txlist" might already
650 * contain them, and the list_splice() below takes care of them.
652 if (!ep_is_linked(epi
)) {
654 * ->ovflist is LIFO, so we have to reverse it in order
657 list_add(&epi
->rdllink
, &ep
->rdllist
);
658 ep_pm_stay_awake(epi
);
662 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
663 * releasing the lock, events will be queued in the normal way inside
666 WRITE_ONCE(ep
->ovflist
, EP_UNACTIVE_PTR
);
669 * Quickly re-inject items left on "txlist".
671 list_splice(txlist
, &ep
->rdllist
);
674 if (!list_empty(&ep
->rdllist
)) {
675 if (waitqueue_active(&ep
->wq
))
679 write_unlock_irq(&ep
->lock
);
682 static void epi_rcu_free(struct rcu_head
*head
)
684 struct epitem
*epi
= container_of(head
, struct epitem
, rcu
);
685 kmem_cache_free(epi_cache
, epi
);
688 static void ep_get(struct eventpoll
*ep
)
690 refcount_inc(&ep
->refcount
);
694 * Returns true if the event poll can be disposed
696 static bool ep_refcount_dec_and_test(struct eventpoll
*ep
)
698 if (!refcount_dec_and_test(&ep
->refcount
))
701 WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep
->rbr
.rb_root
));
705 static void ep_free(struct eventpoll
*ep
)
707 mutex_destroy(&ep
->mtx
);
709 wakeup_source_unregister(ep
->ws
);
714 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
715 * all the associated resources. Must be called with "mtx" held.
716 * If the dying flag is set, do the removal only if force is true.
717 * This prevents ep_clear_and_put() from dropping all the ep references
718 * while running concurrently with eventpoll_release_file().
719 * Returns true if the eventpoll can be disposed.
721 static bool __ep_remove(struct eventpoll
*ep
, struct epitem
*epi
, bool force
)
723 struct file
*file
= epi
->ffd
.file
;
724 struct epitems_head
*to_free
;
725 struct hlist_head
*head
;
727 lockdep_assert_irqs_enabled();
730 * Removes poll wait queue hooks.
732 ep_unregister_pollwait(ep
, epi
);
734 /* Remove the current item from the list of epoll hooks */
735 spin_lock(&file
->f_lock
);
736 if (epi
->dying
&& !force
) {
737 spin_unlock(&file
->f_lock
);
743 if (head
->first
== &epi
->fllink
&& !epi
->fllink
.next
) {
745 if (!is_file_epoll(file
)) {
746 struct epitems_head
*v
;
747 v
= container_of(head
, struct epitems_head
, epitems
);
748 if (!smp_load_acquire(&v
->next
))
752 hlist_del_rcu(&epi
->fllink
);
753 spin_unlock(&file
->f_lock
);
754 free_ephead(to_free
);
756 rb_erase_cached(&epi
->rbn
, &ep
->rbr
);
758 write_lock_irq(&ep
->lock
);
759 if (ep_is_linked(epi
))
760 list_del_init(&epi
->rdllink
);
761 write_unlock_irq(&ep
->lock
);
763 wakeup_source_unregister(ep_wakeup_source(epi
));
765 * At this point it is safe to free the eventpoll item. Use the union
766 * field epi->rcu, since we are trying to minimize the size of
767 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
768 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
769 * use of the rbn field.
771 call_rcu(&epi
->rcu
, epi_rcu_free
);
773 percpu_counter_dec(&ep
->user
->epoll_watches
);
774 return ep_refcount_dec_and_test(ep
);
778 * ep_remove variant for callers owing an additional reference to the ep
780 static void ep_remove_safe(struct eventpoll
*ep
, struct epitem
*epi
)
782 WARN_ON_ONCE(__ep_remove(ep
, epi
, false));
785 static void ep_clear_and_put(struct eventpoll
*ep
)
787 struct rb_node
*rbp
, *next
;
791 /* We need to release all tasks waiting for these file */
792 if (waitqueue_active(&ep
->poll_wait
))
793 ep_poll_safewake(ep
, NULL
, 0);
795 mutex_lock(&ep
->mtx
);
798 * Walks through the whole tree by unregistering poll callbacks.
800 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
801 epi
= rb_entry(rbp
, struct epitem
, rbn
);
803 ep_unregister_pollwait(ep
, epi
);
808 * Walks through the whole tree and try to free each "struct epitem".
809 * Note that ep_remove_safe() will not remove the epitem in case of a
810 * racing eventpoll_release_file(); the latter will do the removal.
811 * At this point we are sure no poll callbacks will be lingering around.
812 * Since we still own a reference to the eventpoll struct, the loop can't
815 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= next
) {
817 epi
= rb_entry(rbp
, struct epitem
, rbn
);
818 ep_remove_safe(ep
, epi
);
822 dispose
= ep_refcount_dec_and_test(ep
);
823 mutex_unlock(&ep
->mtx
);
829 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
831 struct eventpoll
*ep
= file
->private_data
;
834 ep_clear_and_put(ep
);
839 static __poll_t
ep_item_poll(const struct epitem
*epi
, poll_table
*pt
, int depth
);
841 static __poll_t
__ep_eventpoll_poll(struct file
*file
, poll_table
*wait
, int depth
)
843 struct eventpoll
*ep
= file
->private_data
;
845 struct epitem
*epi
, *tmp
;
849 init_poll_funcptr(&pt
, NULL
);
851 /* Insert inside our poll wait queue */
852 poll_wait(file
, &ep
->poll_wait
, wait
);
855 * Proceed to find out if wanted events are really available inside
858 mutex_lock_nested(&ep
->mtx
, depth
);
859 ep_start_scan(ep
, &txlist
);
860 list_for_each_entry_safe(epi
, tmp
, &txlist
, rdllink
) {
861 if (ep_item_poll(epi
, &pt
, depth
+ 1)) {
862 res
= EPOLLIN
| EPOLLRDNORM
;
866 * Item has been dropped into the ready list by the poll
867 * callback, but it's not actually ready, as far as
868 * caller requested events goes. We can remove it here.
870 __pm_relax(ep_wakeup_source(epi
));
871 list_del_init(&epi
->rdllink
);
874 ep_done_scan(ep
, &txlist
);
875 mutex_unlock(&ep
->mtx
);
880 * Differs from ep_eventpoll_poll() in that internal callers already have
881 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
882 * is correctly annotated.
884 static __poll_t
ep_item_poll(const struct epitem
*epi
, poll_table
*pt
,
887 struct file
*file
= epi
->ffd
.file
;
890 pt
->_key
= epi
->event
.events
;
891 if (!is_file_epoll(file
))
892 res
= vfs_poll(file
, pt
);
894 res
= __ep_eventpoll_poll(file
, pt
, depth
);
895 return res
& epi
->event
.events
;
898 static __poll_t
ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
900 return __ep_eventpoll_poll(file
, wait
, 0);
903 #ifdef CONFIG_PROC_FS
904 static void ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
906 struct eventpoll
*ep
= f
->private_data
;
909 mutex_lock(&ep
->mtx
);
910 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
911 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
912 struct inode
*inode
= file_inode(epi
->ffd
.file
);
914 seq_printf(m
, "tfd: %8d events: %8x data: %16llx "
915 " pos:%lli ino:%lx sdev:%x\n",
916 epi
->ffd
.fd
, epi
->event
.events
,
917 (long long)epi
->event
.data
,
918 (long long)epi
->ffd
.file
->f_pos
,
919 inode
->i_ino
, inode
->i_sb
->s_dev
);
920 if (seq_has_overflowed(m
))
923 mutex_unlock(&ep
->mtx
);
927 /* File callbacks that implement the eventpoll file behaviour */
928 static const struct file_operations eventpoll_fops
= {
929 #ifdef CONFIG_PROC_FS
930 .show_fdinfo
= ep_show_fdinfo
,
932 .release
= ep_eventpoll_release
,
933 .poll
= ep_eventpoll_poll
,
934 .llseek
= noop_llseek
,
938 * This is called from eventpoll_release() to unlink files from the eventpoll
939 * interface. We need to have this facility to cleanup correctly files that are
940 * closed without being removed from the eventpoll interface.
942 void eventpoll_release_file(struct file
*file
)
944 struct eventpoll
*ep
;
949 * Use the 'dying' flag to prevent a concurrent ep_clear_and_put() from
950 * touching the epitems list before eventpoll_release_file() can access
954 spin_lock(&file
->f_lock
);
955 if (file
->f_ep
&& file
->f_ep
->first
) {
956 epi
= hlist_entry(file
->f_ep
->first
, struct epitem
, fllink
);
958 spin_unlock(&file
->f_lock
);
961 * ep access is safe as we still own a reference to the ep
965 mutex_lock(&ep
->mtx
);
966 dispose
= __ep_remove(ep
, epi
, true);
967 mutex_unlock(&ep
->mtx
);
973 spin_unlock(&file
->f_lock
);
976 static int ep_alloc(struct eventpoll
**pep
)
978 struct eventpoll
*ep
;
980 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
984 mutex_init(&ep
->mtx
);
985 rwlock_init(&ep
->lock
);
986 init_waitqueue_head(&ep
->wq
);
987 init_waitqueue_head(&ep
->poll_wait
);
988 INIT_LIST_HEAD(&ep
->rdllist
);
989 ep
->rbr
= RB_ROOT_CACHED
;
990 ep
->ovflist
= EP_UNACTIVE_PTR
;
991 ep
->user
= get_current_user();
992 refcount_set(&ep
->refcount
, 1);
1000 * Search the file inside the eventpoll tree. The RB tree operations
1001 * are protected by the "mtx" mutex, and ep_find() must be called with
1004 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
1007 struct rb_node
*rbp
;
1008 struct epitem
*epi
, *epir
= NULL
;
1009 struct epoll_filefd ffd
;
1011 ep_set_ffd(&ffd
, file
, fd
);
1012 for (rbp
= ep
->rbr
.rb_root
.rb_node
; rbp
; ) {
1013 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1014 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
1016 rbp
= rbp
->rb_right
;
1029 static struct epitem
*ep_find_tfd(struct eventpoll
*ep
, int tfd
, unsigned long toff
)
1031 struct rb_node
*rbp
;
1034 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1035 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1036 if (epi
->ffd
.fd
== tfd
) {
1048 struct file
*get_epoll_tfile_raw_ptr(struct file
*file
, int tfd
,
1051 struct file
*file_raw
;
1052 struct eventpoll
*ep
;
1055 if (!is_file_epoll(file
))
1056 return ERR_PTR(-EINVAL
);
1058 ep
= file
->private_data
;
1060 mutex_lock(&ep
->mtx
);
1061 epi
= ep_find_tfd(ep
, tfd
, toff
);
1063 file_raw
= epi
->ffd
.file
;
1065 file_raw
= ERR_PTR(-ENOENT
);
1066 mutex_unlock(&ep
->mtx
);
1070 #endif /* CONFIG_KCMP */
1073 * Adds a new entry to the tail of the list in a lockless way, i.e.
1074 * multiple CPUs are allowed to call this function concurrently.
1076 * Beware: it is necessary to prevent any other modifications of the
1077 * existing list until all changes are completed, in other words
1078 * concurrent list_add_tail_lockless() calls should be protected
1079 * with a read lock, where write lock acts as a barrier which
1080 * makes sure all list_add_tail_lockless() calls are fully
1083 * Also an element can be locklessly added to the list only in one
1084 * direction i.e. either to the tail or to the head, otherwise
1085 * concurrent access will corrupt the list.
1087 * Return: %false if element has been already added to the list, %true
1090 static inline bool list_add_tail_lockless(struct list_head
*new,
1091 struct list_head
*head
)
1093 struct list_head
*prev
;
1096 * This is simple 'new->next = head' operation, but cmpxchg()
1097 * is used in order to detect that same element has been just
1098 * added to the list from another CPU: the winner observes
1101 if (!try_cmpxchg(&new->next
, &new, head
))
1105 * Initially ->next of a new element must be updated with the head
1106 * (we are inserting to the tail) and only then pointers are atomically
1107 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1108 * updated before pointers are actually swapped and pointers are
1109 * swapped before prev->next is updated.
1112 prev
= xchg(&head
->prev
, new);
1115 * It is safe to modify prev->next and new->prev, because a new element
1116 * is added only to the tail and new->next is updated before XCHG.
1126 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1127 * i.e. multiple CPUs are allowed to call this function concurrently.
1129 * Return: %false if epi element has been already chained, %true otherwise.
1131 static inline bool chain_epi_lockless(struct epitem
*epi
)
1133 struct eventpoll
*ep
= epi
->ep
;
1135 /* Fast preliminary check */
1136 if (epi
->next
!= EP_UNACTIVE_PTR
)
1139 /* Check that the same epi has not been just chained from another CPU */
1140 if (cmpxchg(&epi
->next
, EP_UNACTIVE_PTR
, NULL
) != EP_UNACTIVE_PTR
)
1143 /* Atomically exchange tail */
1144 epi
->next
= xchg(&ep
->ovflist
, epi
);
1150 * This is the callback that is passed to the wait queue wakeup
1151 * mechanism. It is called by the stored file descriptors when they
1152 * have events to report.
1154 * This callback takes a read lock in order not to contend with concurrent
1155 * events from another file descriptor, thus all modifications to ->rdllist
1156 * or ->ovflist are lockless. Read lock is paired with the write lock from
1157 * ep_scan_ready_list(), which stops all list modifications and guarantees
1158 * that lists state is seen correctly.
1160 * Another thing worth to mention is that ep_poll_callback() can be called
1161 * concurrently for the same @epi from different CPUs if poll table was inited
1162 * with several wait queues entries. Plural wakeup from different CPUs of a
1163 * single wait queue is serialized by wq.lock, but the case when multiple wait
1164 * queues are used should be detected accordingly. This is detected using
1165 * cmpxchg() operation.
1167 static int ep_poll_callback(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
1170 struct epitem
*epi
= ep_item_from_wait(wait
);
1171 struct eventpoll
*ep
= epi
->ep
;
1172 __poll_t pollflags
= key_to_poll(key
);
1173 unsigned long flags
;
1176 read_lock_irqsave(&ep
->lock
, flags
);
1178 ep_set_busy_poll_napi_id(epi
);
1181 * If the event mask does not contain any poll(2) event, we consider the
1182 * descriptor to be disabled. This condition is likely the effect of the
1183 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1184 * until the next EPOLL_CTL_MOD will be issued.
1186 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1190 * Check the events coming with the callback. At this stage, not
1191 * every device reports the events in the "key" parameter of the
1192 * callback. We need to be able to handle both cases here, hence the
1193 * test for "key" != NULL before the event match test.
1195 if (pollflags
&& !(pollflags
& epi
->event
.events
))
1199 * If we are transferring events to userspace, we can hold no locks
1200 * (because we're accessing user memory, and because of linux f_op->poll()
1201 * semantics). All the events that happen during that period of time are
1202 * chained in ep->ovflist and requeued later on.
1204 if (READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
) {
1205 if (chain_epi_lockless(epi
))
1206 ep_pm_stay_awake_rcu(epi
);
1207 } else if (!ep_is_linked(epi
)) {
1208 /* In the usual case, add event to ready list. */
1209 if (list_add_tail_lockless(&epi
->rdllink
, &ep
->rdllist
))
1210 ep_pm_stay_awake_rcu(epi
);
1214 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1217 if (waitqueue_active(&ep
->wq
)) {
1218 if ((epi
->event
.events
& EPOLLEXCLUSIVE
) &&
1219 !(pollflags
& POLLFREE
)) {
1220 switch (pollflags
& EPOLLINOUT_BITS
) {
1222 if (epi
->event
.events
& EPOLLIN
)
1226 if (epi
->event
.events
& EPOLLOUT
)
1236 if (waitqueue_active(&ep
->poll_wait
))
1240 read_unlock_irqrestore(&ep
->lock
, flags
);
1242 /* We have to call this outside the lock */
1244 ep_poll_safewake(ep
, epi
, pollflags
& EPOLL_URING_WAKE
);
1246 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
))
1249 if (pollflags
& POLLFREE
) {
1251 * If we race with ep_remove_wait_queue() it can miss
1252 * ->whead = NULL and do another remove_wait_queue() after
1253 * us, so we can't use __remove_wait_queue().
1255 list_del_init(&wait
->entry
);
1257 * ->whead != NULL protects us from the race with
1258 * ep_clear_and_put() or ep_remove(), ep_remove_wait_queue()
1259 * takes whead->lock held by the caller. Once we nullify it,
1260 * nothing protects ep/epi or even wait.
1262 smp_store_release(&ep_pwq_from_wait(wait
)->whead
, NULL
);
1269 * This is the callback that is used to add our wait queue to the
1270 * target file wakeup lists.
1272 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1275 struct ep_pqueue
*epq
= container_of(pt
, struct ep_pqueue
, pt
);
1276 struct epitem
*epi
= epq
->epi
;
1277 struct eppoll_entry
*pwq
;
1279 if (unlikely(!epi
)) // an earlier allocation has failed
1282 pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
);
1283 if (unlikely(!pwq
)) {
1288 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1291 if (epi
->event
.events
& EPOLLEXCLUSIVE
)
1292 add_wait_queue_exclusive(whead
, &pwq
->wait
);
1294 add_wait_queue(whead
, &pwq
->wait
);
1295 pwq
->next
= epi
->pwqlist
;
1299 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1302 struct rb_node
**p
= &ep
->rbr
.rb_root
.rb_node
, *parent
= NULL
;
1303 struct epitem
*epic
;
1304 bool leftmost
= true;
1308 epic
= rb_entry(parent
, struct epitem
, rbn
);
1309 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1311 p
= &parent
->rb_right
;
1314 p
= &parent
->rb_left
;
1316 rb_link_node(&epi
->rbn
, parent
, p
);
1317 rb_insert_color_cached(&epi
->rbn
, &ep
->rbr
, leftmost
);
1322 #define PATH_ARR_SIZE 5
1324 * These are the number paths of length 1 to 5, that we are allowing to emanate
1325 * from a single file of interest. For example, we allow 1000 paths of length
1326 * 1, to emanate from each file of interest. This essentially represents the
1327 * potential wakeup paths, which need to be limited in order to avoid massive
1328 * uncontrolled wakeup storms. The common use case should be a single ep which
1329 * is connected to n file sources. In this case each file source has 1 path
1330 * of length 1. Thus, the numbers below should be more than sufficient. These
1331 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1332 * and delete can't add additional paths. Protected by the epnested_mutex.
1334 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1335 static int path_count
[PATH_ARR_SIZE
];
1337 static int path_count_inc(int nests
)
1339 /* Allow an arbitrary number of depth 1 paths */
1343 if (++path_count
[nests
] > path_limits
[nests
])
1348 static void path_count_init(void)
1352 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1356 static int reverse_path_check_proc(struct hlist_head
*refs
, int depth
)
1361 if (depth
> EP_MAX_NESTS
) /* too deep nesting */
1364 /* CTL_DEL can remove links here, but that can't increase our count */
1365 hlist_for_each_entry_rcu(epi
, refs
, fllink
) {
1366 struct hlist_head
*refs
= &epi
->ep
->refs
;
1367 if (hlist_empty(refs
))
1368 error
= path_count_inc(depth
);
1370 error
= reverse_path_check_proc(refs
, depth
+ 1);
1378 * reverse_path_check - The tfile_check_list is list of epitem_head, which have
1379 * links that are proposed to be newly added. We need to
1380 * make sure that those added links don't add too many
1381 * paths such that we will spend all our time waking up
1382 * eventpoll objects.
1384 * Return: %zero if the proposed links don't create too many paths,
1387 static int reverse_path_check(void)
1389 struct epitems_head
*p
;
1391 for (p
= tfile_check_list
; p
!= EP_UNACTIVE_PTR
; p
= p
->next
) {
1395 error
= reverse_path_check_proc(&p
->epitems
, 0);
1403 static int ep_create_wakeup_source(struct epitem
*epi
)
1405 struct name_snapshot n
;
1406 struct wakeup_source
*ws
;
1409 epi
->ep
->ws
= wakeup_source_register(NULL
, "eventpoll");
1414 take_dentry_name_snapshot(&n
, epi
->ffd
.file
->f_path
.dentry
);
1415 ws
= wakeup_source_register(NULL
, n
.name
.name
);
1416 release_dentry_name_snapshot(&n
);
1420 rcu_assign_pointer(epi
->ws
, ws
);
1425 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1426 static noinline
void ep_destroy_wakeup_source(struct epitem
*epi
)
1428 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
1430 RCU_INIT_POINTER(epi
->ws
, NULL
);
1433 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1434 * used internally by wakeup_source_remove, too (called by
1435 * wakeup_source_unregister), so we cannot use call_rcu
1438 wakeup_source_unregister(ws
);
1441 static int attach_epitem(struct file
*file
, struct epitem
*epi
)
1443 struct epitems_head
*to_free
= NULL
;
1444 struct hlist_head
*head
= NULL
;
1445 struct eventpoll
*ep
= NULL
;
1447 if (is_file_epoll(file
))
1448 ep
= file
->private_data
;
1452 } else if (!READ_ONCE(file
->f_ep
)) {
1454 to_free
= kmem_cache_zalloc(ephead_cache
, GFP_KERNEL
);
1457 head
= &to_free
->epitems
;
1459 spin_lock(&file
->f_lock
);
1461 if (unlikely(!head
)) {
1462 spin_unlock(&file
->f_lock
);
1468 hlist_add_head_rcu(&epi
->fllink
, file
->f_ep
);
1469 spin_unlock(&file
->f_lock
);
1470 free_ephead(to_free
);
1475 * Must be called with "mtx" held.
1477 static int ep_insert(struct eventpoll
*ep
, const struct epoll_event
*event
,
1478 struct file
*tfile
, int fd
, int full_check
)
1480 int error
, pwake
= 0;
1483 struct ep_pqueue epq
;
1484 struct eventpoll
*tep
= NULL
;
1486 if (is_file_epoll(tfile
))
1487 tep
= tfile
->private_data
;
1489 lockdep_assert_irqs_enabled();
1491 if (unlikely(percpu_counter_compare(&ep
->user
->epoll_watches
,
1492 max_user_watches
) >= 0))
1494 percpu_counter_inc(&ep
->user
->epoll_watches
);
1496 if (!(epi
= kmem_cache_zalloc(epi_cache
, GFP_KERNEL
))) {
1497 percpu_counter_dec(&ep
->user
->epoll_watches
);
1501 /* Item initialization follow here ... */
1502 INIT_LIST_HEAD(&epi
->rdllink
);
1504 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
1505 epi
->event
= *event
;
1506 epi
->next
= EP_UNACTIVE_PTR
;
1509 mutex_lock_nested(&tep
->mtx
, 1);
1510 /* Add the current item to the list of active epoll hook for this file */
1511 if (unlikely(attach_epitem(tfile
, epi
) < 0)) {
1513 mutex_unlock(&tep
->mtx
);
1514 kmem_cache_free(epi_cache
, epi
);
1515 percpu_counter_dec(&ep
->user
->epoll_watches
);
1519 if (full_check
&& !tep
)
1523 * Add the current item to the RB tree. All RB tree operations are
1524 * protected by "mtx", and ep_insert() is called with "mtx" held.
1526 ep_rbtree_insert(ep
, epi
);
1528 mutex_unlock(&tep
->mtx
);
1531 * ep_remove_safe() calls in the later error paths can't lead to
1532 * ep_free() as the ep file itself still holds an ep reference.
1536 /* now check if we've created too many backpaths */
1537 if (unlikely(full_check
&& reverse_path_check())) {
1538 ep_remove_safe(ep
, epi
);
1542 if (epi
->event
.events
& EPOLLWAKEUP
) {
1543 error
= ep_create_wakeup_source(epi
);
1545 ep_remove_safe(ep
, epi
);
1550 /* Initialize the poll table using the queue callback */
1552 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1555 * Attach the item to the poll hooks and get current event bits.
1556 * We can safely use the file* here because its usage count has
1557 * been increased by the caller of this function. Note that after
1558 * this operation completes, the poll callback can start hitting
1561 revents
= ep_item_poll(epi
, &epq
.pt
, 1);
1564 * We have to check if something went wrong during the poll wait queue
1565 * install process. Namely an allocation for a wait queue failed due
1566 * high memory pressure.
1568 if (unlikely(!epq
.epi
)) {
1569 ep_remove_safe(ep
, epi
);
1573 /* We have to drop the new item inside our item list to keep track of it */
1574 write_lock_irq(&ep
->lock
);
1576 /* record NAPI ID of new item if present */
1577 ep_set_busy_poll_napi_id(epi
);
1579 /* If the file is already "ready" we drop it inside the ready list */
1580 if (revents
&& !ep_is_linked(epi
)) {
1581 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1582 ep_pm_stay_awake(epi
);
1584 /* Notify waiting tasks that events are available */
1585 if (waitqueue_active(&ep
->wq
))
1587 if (waitqueue_active(&ep
->poll_wait
))
1591 write_unlock_irq(&ep
->lock
);
1593 /* We have to call this outside the lock */
1595 ep_poll_safewake(ep
, NULL
, 0);
1601 * Modify the interest event mask by dropping an event if the new mask
1602 * has a match in the current file status. Must be called with "mtx" held.
1604 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
,
1605 const struct epoll_event
*event
)
1610 lockdep_assert_irqs_enabled();
1612 init_poll_funcptr(&pt
, NULL
);
1615 * Set the new event interest mask before calling f_op->poll();
1616 * otherwise we might miss an event that happens between the
1617 * f_op->poll() call and the new event set registering.
1619 epi
->event
.events
= event
->events
; /* need barrier below */
1620 epi
->event
.data
= event
->data
; /* protected by mtx */
1621 if (epi
->event
.events
& EPOLLWAKEUP
) {
1622 if (!ep_has_wakeup_source(epi
))
1623 ep_create_wakeup_source(epi
);
1624 } else if (ep_has_wakeup_source(epi
)) {
1625 ep_destroy_wakeup_source(epi
);
1629 * The following barrier has two effects:
1631 * 1) Flush epi changes above to other CPUs. This ensures
1632 * we do not miss events from ep_poll_callback if an
1633 * event occurs immediately after we call f_op->poll().
1634 * We need this because we did not take ep->lock while
1635 * changing epi above (but ep_poll_callback does take
1638 * 2) We also need to ensure we do not miss _past_ events
1639 * when calling f_op->poll(). This barrier also
1640 * pairs with the barrier in wq_has_sleeper (see
1641 * comments for wq_has_sleeper).
1643 * This barrier will now guarantee ep_poll_callback or f_op->poll
1644 * (or both) will notice the readiness of an item.
1649 * Get current event bits. We can safely use the file* here because
1650 * its usage count has been increased by the caller of this function.
1651 * If the item is "hot" and it is not registered inside the ready
1652 * list, push it inside.
1654 if (ep_item_poll(epi
, &pt
, 1)) {
1655 write_lock_irq(&ep
->lock
);
1656 if (!ep_is_linked(epi
)) {
1657 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1658 ep_pm_stay_awake(epi
);
1660 /* Notify waiting tasks that events are available */
1661 if (waitqueue_active(&ep
->wq
))
1663 if (waitqueue_active(&ep
->poll_wait
))
1666 write_unlock_irq(&ep
->lock
);
1669 /* We have to call this outside the lock */
1671 ep_poll_safewake(ep
, NULL
, 0);
1676 static int ep_send_events(struct eventpoll
*ep
,
1677 struct epoll_event __user
*events
, int maxevents
)
1679 struct epitem
*epi
, *tmp
;
1685 * Always short-circuit for fatal signals to allow threads to make a
1686 * timely exit without the chance of finding more events available and
1687 * fetching repeatedly.
1689 if (fatal_signal_pending(current
))
1692 init_poll_funcptr(&pt
, NULL
);
1694 mutex_lock(&ep
->mtx
);
1695 ep_start_scan(ep
, &txlist
);
1698 * We can loop without lock because we are passed a task private list.
1699 * Items cannot vanish during the loop we are holding ep->mtx.
1701 list_for_each_entry_safe(epi
, tmp
, &txlist
, rdllink
) {
1702 struct wakeup_source
*ws
;
1705 if (res
>= maxevents
)
1709 * Activate ep->ws before deactivating epi->ws to prevent
1710 * triggering auto-suspend here (in case we reactive epi->ws
1713 * This could be rearranged to delay the deactivation of epi->ws
1714 * instead, but then epi->ws would temporarily be out of sync
1715 * with ep_is_linked().
1717 ws
= ep_wakeup_source(epi
);
1720 __pm_stay_awake(ep
->ws
);
1724 list_del_init(&epi
->rdllink
);
1727 * If the event mask intersect the caller-requested one,
1728 * deliver the event to userspace. Again, we are holding ep->mtx,
1729 * so no operations coming from userspace can change the item.
1731 revents
= ep_item_poll(epi
, &pt
, 1);
1735 events
= epoll_put_uevent(revents
, epi
->event
.data
, events
);
1737 list_add(&epi
->rdllink
, &txlist
);
1738 ep_pm_stay_awake(epi
);
1744 if (epi
->event
.events
& EPOLLONESHOT
)
1745 epi
->event
.events
&= EP_PRIVATE_BITS
;
1746 else if (!(epi
->event
.events
& EPOLLET
)) {
1748 * If this file has been added with Level
1749 * Trigger mode, we need to insert back inside
1750 * the ready list, so that the next call to
1751 * epoll_wait() will check again the events
1752 * availability. At this point, no one can insert
1753 * into ep->rdllist besides us. The epoll_ctl()
1754 * callers are locked out by
1755 * ep_scan_ready_list() holding "mtx" and the
1756 * poll callback will queue them in ep->ovflist.
1758 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1759 ep_pm_stay_awake(epi
);
1762 ep_done_scan(ep
, &txlist
);
1763 mutex_unlock(&ep
->mtx
);
1768 static struct timespec64
*ep_timeout_to_timespec(struct timespec64
*to
, long ms
)
1770 struct timespec64 now
;
1781 to
->tv_sec
= ms
/ MSEC_PER_SEC
;
1782 to
->tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
);
1784 ktime_get_ts64(&now
);
1785 *to
= timespec64_add_safe(now
, *to
);
1790 * autoremove_wake_function, but remove even on failure to wake up, because we
1791 * know that default_wake_function/ttwu will only fail if the thread is already
1792 * woken, and in that case the ep_poll loop will remove the entry anyways, not
1795 static int ep_autoremove_wake_function(struct wait_queue_entry
*wq_entry
,
1796 unsigned int mode
, int sync
, void *key
)
1798 int ret
= default_wake_function(wq_entry
, mode
, sync
, key
);
1801 * Pairs with list_empty_careful in ep_poll, and ensures future loop
1802 * iterations see the cause of this wakeup.
1804 list_del_init_careful(&wq_entry
->entry
);
1809 * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
1812 * @ep: Pointer to the eventpoll context.
1813 * @events: Pointer to the userspace buffer where the ready events should be
1815 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1816 * @timeout: Maximum timeout for the ready events fetch operation, in
1817 * timespec. If the timeout is zero, the function will not block,
1818 * while if the @timeout ptr is NULL, the function will block
1819 * until at least one event has been retrieved (or an error
1822 * Return: the number of ready events which have been fetched, or an
1823 * error code, in case of error.
1825 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1826 int maxevents
, struct timespec64
*timeout
)
1828 int res
, eavail
, timed_out
= 0;
1830 wait_queue_entry_t wait
;
1831 ktime_t expires
, *to
= NULL
;
1833 lockdep_assert_irqs_enabled();
1835 if (timeout
&& (timeout
->tv_sec
| timeout
->tv_nsec
)) {
1836 slack
= select_estimate_accuracy(timeout
);
1838 *to
= timespec64_to_ktime(*timeout
);
1839 } else if (timeout
) {
1841 * Avoid the unnecessary trip to the wait queue loop, if the
1842 * caller specified a non blocking operation.
1848 * This call is racy: We may or may not see events that are being added
1849 * to the ready list under the lock (e.g., in IRQ callbacks). For cases
1850 * with a non-zero timeout, this thread will check the ready list under
1851 * lock and will add to the wait queue. For cases with a zero
1852 * timeout, the user by definition should not care and will have to
1855 eavail
= ep_events_available(ep
);
1860 * Try to transfer events to user space. In case we get
1861 * 0 events and there's still timeout left over, we go
1862 * trying again in search of more luck.
1864 res
= ep_send_events(ep
, events
, maxevents
);
1872 eavail
= ep_busy_loop(ep
, timed_out
);
1876 if (signal_pending(current
))
1880 * Internally init_wait() uses autoremove_wake_function(),
1881 * thus wait entry is removed from the wait queue on each
1882 * wakeup. Why it is important? In case of several waiters
1883 * each new wakeup will hit the next waiter, giving it the
1884 * chance to harvest new event. Otherwise wakeup can be
1885 * lost. This is also good performance-wise, because on
1886 * normal wakeup path no need to call __remove_wait_queue()
1887 * explicitly, thus ep->lock is not taken, which halts the
1890 * In fact, we now use an even more aggressive function that
1891 * unconditionally removes, because we don't reuse the wait
1892 * entry between loop iterations. This lets us also avoid the
1893 * performance issue if a process is killed, causing all of its
1894 * threads to wake up without being removed normally.
1897 wait
.func
= ep_autoremove_wake_function
;
1899 write_lock_irq(&ep
->lock
);
1901 * Barrierless variant, waitqueue_active() is called under
1902 * the same lock on wakeup ep_poll_callback() side, so it
1903 * is safe to avoid an explicit barrier.
1905 __set_current_state(TASK_INTERRUPTIBLE
);
1908 * Do the final check under the lock. ep_scan_ready_list()
1909 * plays with two lists (->rdllist and ->ovflist) and there
1910 * is always a race when both lists are empty for short
1911 * period of time although events are pending, so lock is
1914 eavail
= ep_events_available(ep
);
1916 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1918 write_unlock_irq(&ep
->lock
);
1921 timed_out
= !schedule_hrtimeout_range(to
, slack
,
1923 __set_current_state(TASK_RUNNING
);
1926 * We were woken up, thus go and try to harvest some events.
1927 * If timed out and still on the wait queue, recheck eavail
1928 * carefully under lock, below.
1932 if (!list_empty_careful(&wait
.entry
)) {
1933 write_lock_irq(&ep
->lock
);
1935 * If the thread timed out and is not on the wait queue,
1936 * it means that the thread was woken up after its
1937 * timeout expired before it could reacquire the lock.
1938 * Thus, when wait.entry is empty, it needs to harvest
1942 eavail
= list_empty(&wait
.entry
);
1943 __remove_wait_queue(&ep
->wq
, &wait
);
1944 write_unlock_irq(&ep
->lock
);
1950 * ep_loop_check_proc - verify that adding an epoll file inside another
1951 * epoll structure does not violate the constraints, in
1952 * terms of closed loops, or too deep chains (which can
1953 * result in excessive stack usage).
1955 * @ep: the &struct eventpoll to be currently checked.
1956 * @depth: Current depth of the path being checked.
1958 * Return: %zero if adding the epoll @file inside current epoll
1959 * structure @ep does not violate the constraints, or %-1 otherwise.
1961 static int ep_loop_check_proc(struct eventpoll
*ep
, int depth
)
1964 struct rb_node
*rbp
;
1967 mutex_lock_nested(&ep
->mtx
, depth
+ 1);
1968 ep
->gen
= loop_check_gen
;
1969 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1970 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1971 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1972 struct eventpoll
*ep_tovisit
;
1973 ep_tovisit
= epi
->ffd
.file
->private_data
;
1974 if (ep_tovisit
->gen
== loop_check_gen
)
1976 if (ep_tovisit
== inserting_into
|| depth
> EP_MAX_NESTS
)
1979 error
= ep_loop_check_proc(ep_tovisit
, depth
+ 1);
1984 * If we've reached a file that is not associated with
1985 * an ep, then we need to check if the newly added
1986 * links are going to add too many wakeup paths. We do
1987 * this by adding it to the tfile_check_list, if it's
1988 * not already there, and calling reverse_path_check()
1989 * during ep_insert().
1991 list_file(epi
->ffd
.file
);
1994 mutex_unlock(&ep
->mtx
);
2000 * ep_loop_check - Performs a check to verify that adding an epoll file (@to)
2001 * into another epoll file (represented by @ep) does not create
2002 * closed loops or too deep chains.
2004 * @ep: Pointer to the epoll we are inserting into.
2005 * @to: Pointer to the epoll to be inserted.
2007 * Return: %zero if adding the epoll @to inside the epoll @from
2008 * does not violate the constraints, or %-1 otherwise.
2010 static int ep_loop_check(struct eventpoll
*ep
, struct eventpoll
*to
)
2012 inserting_into
= ep
;
2013 return ep_loop_check_proc(to
, 0);
2016 static void clear_tfile_check_list(void)
2019 while (tfile_check_list
!= EP_UNACTIVE_PTR
) {
2020 struct epitems_head
*head
= tfile_check_list
;
2021 tfile_check_list
= head
->next
;
2028 * Open an eventpoll file descriptor.
2030 static int do_epoll_create(int flags
)
2033 struct eventpoll
*ep
= NULL
;
2036 /* Check the EPOLL_* constant for consistency. */
2037 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
2039 if (flags
& ~EPOLL_CLOEXEC
)
2042 * Create the internal data structure ("struct eventpoll").
2044 error
= ep_alloc(&ep
);
2048 * Creates all the items needed to setup an eventpoll file. That is,
2049 * a file structure and a free file descriptor.
2051 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
2056 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
2057 O_RDWR
| (flags
& O_CLOEXEC
));
2059 error
= PTR_ERR(file
);
2063 fd_install(fd
, file
);
2069 ep_clear_and_put(ep
);
2073 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
2075 return do_epoll_create(flags
);
2078 SYSCALL_DEFINE1(epoll_create
, int, size
)
2083 return do_epoll_create(0);
2086 #ifdef CONFIG_PM_SLEEP
2087 static inline void ep_take_care_of_epollwakeup(struct epoll_event
*epev
)
2089 if ((epev
->events
& EPOLLWAKEUP
) && !capable(CAP_BLOCK_SUSPEND
))
2090 epev
->events
&= ~EPOLLWAKEUP
;
2093 static inline void ep_take_care_of_epollwakeup(struct epoll_event
*epev
)
2095 epev
->events
&= ~EPOLLWAKEUP
;
2099 static inline int epoll_mutex_lock(struct mutex
*mutex
, int depth
,
2103 mutex_lock_nested(mutex
, depth
);
2106 if (mutex_trylock(mutex
))
2111 int do_epoll_ctl(int epfd
, int op
, int fd
, struct epoll_event
*epds
,
2117 struct eventpoll
*ep
;
2119 struct eventpoll
*tep
= NULL
;
2126 /* Get the "struct file *" for the target file */
2131 /* The target file descriptor must support poll */
2133 if (!file_can_poll(tf
.file
))
2134 goto error_tgt_fput
;
2136 /* Check if EPOLLWAKEUP is allowed */
2137 if (ep_op_has_event(op
))
2138 ep_take_care_of_epollwakeup(epds
);
2141 * We have to check that the file structure underneath the file descriptor
2142 * the user passed to us _is_ an eventpoll file. And also we do not permit
2143 * adding an epoll file descriptor inside itself.
2146 if (f
.file
== tf
.file
|| !is_file_epoll(f
.file
))
2147 goto error_tgt_fput
;
2150 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2151 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2152 * Also, we do not currently supported nested exclusive wakeups.
2154 if (ep_op_has_event(op
) && (epds
->events
& EPOLLEXCLUSIVE
)) {
2155 if (op
== EPOLL_CTL_MOD
)
2156 goto error_tgt_fput
;
2157 if (op
== EPOLL_CTL_ADD
&& (is_file_epoll(tf
.file
) ||
2158 (epds
->events
& ~EPOLLEXCLUSIVE_OK_BITS
)))
2159 goto error_tgt_fput
;
2163 * At this point it is safe to assume that the "private_data" contains
2164 * our own data structure.
2166 ep
= f
.file
->private_data
;
2169 * When we insert an epoll file descriptor inside another epoll file
2170 * descriptor, there is the chance of creating closed loops, which are
2171 * better be handled here, than in more critical paths. While we are
2172 * checking for loops we also determine the list of files reachable
2173 * and hang them on the tfile_check_list, so we can check that we
2174 * haven't created too many possible wakeup paths.
2176 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2177 * the epoll file descriptor is attaching directly to a wakeup source,
2178 * unless the epoll file descriptor is nested. The purpose of taking the
2179 * 'epnested_mutex' on add is to prevent complex toplogies such as loops and
2180 * deep wakeup paths from forming in parallel through multiple
2181 * EPOLL_CTL_ADD operations.
2183 error
= epoll_mutex_lock(&ep
->mtx
, 0, nonblock
);
2185 goto error_tgt_fput
;
2186 if (op
== EPOLL_CTL_ADD
) {
2187 if (READ_ONCE(f
.file
->f_ep
) || ep
->gen
== loop_check_gen
||
2188 is_file_epoll(tf
.file
)) {
2189 mutex_unlock(&ep
->mtx
);
2190 error
= epoll_mutex_lock(&epnested_mutex
, 0, nonblock
);
2192 goto error_tgt_fput
;
2195 if (is_file_epoll(tf
.file
)) {
2196 tep
= tf
.file
->private_data
;
2198 if (ep_loop_check(ep
, tep
) != 0)
2199 goto error_tgt_fput
;
2201 error
= epoll_mutex_lock(&ep
->mtx
, 0, nonblock
);
2203 goto error_tgt_fput
;
2208 * Try to lookup the file inside our RB tree. Since we grabbed "mtx"
2209 * above, we can be sure to be able to use the item looked up by
2210 * ep_find() till we release the mutex.
2212 epi
= ep_find(ep
, tf
.file
, fd
);
2218 epds
->events
|= EPOLLERR
| EPOLLHUP
;
2219 error
= ep_insert(ep
, epds
, tf
.file
, fd
, full_check
);
2226 * The eventpoll itself is still alive: the refcount
2227 * can't go to zero here.
2229 ep_remove_safe(ep
, epi
);
2237 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
)) {
2238 epds
->events
|= EPOLLERR
| EPOLLHUP
;
2239 error
= ep_modify(ep
, epi
, epds
);
2245 mutex_unlock(&ep
->mtx
);
2249 clear_tfile_check_list();
2251 mutex_unlock(&epnested_mutex
);
2263 * The following function implements the controller interface for
2264 * the eventpoll file that enables the insertion/removal/change of
2265 * file descriptors inside the interest set.
2267 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
2268 struct epoll_event __user
*, event
)
2270 struct epoll_event epds
;
2272 if (ep_op_has_event(op
) &&
2273 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
2276 return do_epoll_ctl(epfd
, op
, fd
, &epds
, false);
2280 * Implement the event wait interface for the eventpoll file. It is the kernel
2281 * part of the user space epoll_wait(2).
2283 static int do_epoll_wait(int epfd
, struct epoll_event __user
*events
,
2284 int maxevents
, struct timespec64
*to
)
2288 struct eventpoll
*ep
;
2290 /* The maximum number of event must be greater than zero */
2291 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
2294 /* Verify that the area passed by the user is writeable */
2295 if (!access_ok(events
, maxevents
* sizeof(struct epoll_event
)))
2298 /* Get the "struct file *" for the eventpoll file */
2304 * We have to check that the file structure underneath the fd
2305 * the user passed to us _is_ an eventpoll file.
2308 if (!is_file_epoll(f
.file
))
2312 * At this point it is safe to assume that the "private_data" contains
2313 * our own data structure.
2315 ep
= f
.file
->private_data
;
2317 /* Time to fish for events ... */
2318 error
= ep_poll(ep
, events
, maxevents
, to
);
2325 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
2326 int, maxevents
, int, timeout
)
2328 struct timespec64 to
;
2330 return do_epoll_wait(epfd
, events
, maxevents
,
2331 ep_timeout_to_timespec(&to
, timeout
));
2335 * Implement the event wait interface for the eventpoll file. It is the kernel
2336 * part of the user space epoll_pwait(2).
2338 static int do_epoll_pwait(int epfd
, struct epoll_event __user
*events
,
2339 int maxevents
, struct timespec64
*to
,
2340 const sigset_t __user
*sigmask
, size_t sigsetsize
)
2345 * If the caller wants a certain signal mask to be set during the wait,
2348 error
= set_user_sigmask(sigmask
, sigsetsize
);
2352 error
= do_epoll_wait(epfd
, events
, maxevents
, to
);
2354 restore_saved_sigmask_unless(error
== -EINTR
);
2359 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
2360 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
2363 struct timespec64 to
;
2365 return do_epoll_pwait(epfd
, events
, maxevents
,
2366 ep_timeout_to_timespec(&to
, timeout
),
2367 sigmask
, sigsetsize
);
2370 SYSCALL_DEFINE6(epoll_pwait2
, int, epfd
, struct epoll_event __user
*, events
,
2371 int, maxevents
, const struct __kernel_timespec __user
*, timeout
,
2372 const sigset_t __user
*, sigmask
, size_t, sigsetsize
)
2374 struct timespec64 ts
, *to
= NULL
;
2377 if (get_timespec64(&ts
, timeout
))
2380 if (poll_select_set_timeout(to
, ts
.tv_sec
, ts
.tv_nsec
))
2384 return do_epoll_pwait(epfd
, events
, maxevents
, to
,
2385 sigmask
, sigsetsize
);
2388 #ifdef CONFIG_COMPAT
2389 static int do_compat_epoll_pwait(int epfd
, struct epoll_event __user
*events
,
2390 int maxevents
, struct timespec64
*timeout
,
2391 const compat_sigset_t __user
*sigmask
,
2392 compat_size_t sigsetsize
)
2397 * If the caller wants a certain signal mask to be set during the wait,
2400 err
= set_compat_user_sigmask(sigmask
, sigsetsize
);
2404 err
= do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2406 restore_saved_sigmask_unless(err
== -EINTR
);
2411 COMPAT_SYSCALL_DEFINE6(epoll_pwait
, int, epfd
,
2412 struct epoll_event __user
*, events
,
2413 int, maxevents
, int, timeout
,
2414 const compat_sigset_t __user
*, sigmask
,
2415 compat_size_t
, sigsetsize
)
2417 struct timespec64 to
;
2419 return do_compat_epoll_pwait(epfd
, events
, maxevents
,
2420 ep_timeout_to_timespec(&to
, timeout
),
2421 sigmask
, sigsetsize
);
2424 COMPAT_SYSCALL_DEFINE6(epoll_pwait2
, int, epfd
,
2425 struct epoll_event __user
*, events
,
2427 const struct __kernel_timespec __user
*, timeout
,
2428 const compat_sigset_t __user
*, sigmask
,
2429 compat_size_t
, sigsetsize
)
2431 struct timespec64 ts
, *to
= NULL
;
2434 if (get_timespec64(&ts
, timeout
))
2437 if (poll_select_set_timeout(to
, ts
.tv_sec
, ts
.tv_nsec
))
2441 return do_compat_epoll_pwait(epfd
, events
, maxevents
, to
,
2442 sigmask
, sigsetsize
);
2447 static int __init
eventpoll_init(void)
2453 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2455 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2457 BUG_ON(max_user_watches
< 0);
2460 * We can have many thousands of epitems, so prevent this from
2461 * using an extra cache line on 64-bit (and smaller) CPUs
2463 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2465 /* Allocates slab cache used to allocate "struct epitem" items */
2466 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2467 0, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2469 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2470 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2471 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2472 epoll_sysctls_init();
2474 ephead_cache
= kmem_cache_create("ep_head",
2475 sizeof(struct epitems_head
), 0, SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2479 fs_initcall(eventpoll_init
);