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 :
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 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * This mutex is acquired by ep_free() during the epoll file
63 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
66 * It is also acquired when inserting an epoll fd onto another epoll
67 * fd. We do this so that we walk the epoll tree and ensure that this
68 * insertion does not create a cycle of epoll file descriptors, which
69 * could lead to deadlock. We need a global mutex to prevent two
70 * simultaneous inserts (A into B and B into A) from racing and
71 * constructing a cycle without either insert observing that it is
73 * It is necessary to acquire multiple "ep->mtx"es at once in the
74 * case when one epoll fd is added to another. In this case, we
75 * always acquire the locks in the order of nesting (i.e. after
76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77 * before e2->mtx). Since we disallow cycles of epoll file
78 * descriptors, this ensures that the mutexes are well-ordered. In
79 * order to communicate this nesting to lockdep, when walking a tree
80 * of epoll file descriptors, we use the current recursion depth as
82 * It is possible to drop the "ep->mtx" and to use the global
83 * mutex "epmutex" (together with "ep->lock") to have it working,
84 * but having "ep->mtx" will make the interface more scalable.
85 * Events that require holding "epmutex" are very rare, while for
86 * normal operations the epoll private "ep->mtx" will guarantee
87 * a better scalability.
90 /* Epoll private bits inside the event mask */
91 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
93 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
95 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
96 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EP_MAX_NESTS 4
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
103 #define EP_UNACTIVE_PTR ((void *) -1L)
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
107 struct epoll_filefd
{
112 /* Wait structure used by the poll hooks */
113 struct eppoll_entry
{
114 /* List header used to link this structure to the "struct epitem" */
115 struct eppoll_entry
*next
;
117 /* The "base" pointer is set to the container "struct epitem" */
121 * Wait queue item that will be linked to the target file wait
124 wait_queue_entry_t wait
;
126 /* The wait queue head that linked the "wait" wait queue item */
127 wait_queue_head_t
*whead
;
131 * Each file descriptor added to the eventpoll interface will
132 * have an entry of this type linked to the "rbr" RB tree.
133 * Avoid increasing the size of this struct, there can be many thousands
134 * of these on a server and we do not want this to take another cache line.
138 /* RB tree node links this structure to the eventpoll RB tree */
140 /* Used to free the struct epitem */
144 /* List header used to link this structure to the eventpoll ready list */
145 struct list_head rdllink
;
148 * Works together "struct eventpoll"->ovflist in keeping the
149 * single linked chain of items.
153 /* The file descriptor information this item refers to */
154 struct epoll_filefd ffd
;
156 /* List containing poll wait queues */
157 struct eppoll_entry
*pwqlist
;
159 /* The "container" of this item */
160 struct eventpoll
*ep
;
162 /* List header used to link this item to the "struct file" items list */
163 struct hlist_node fllink
;
165 /* wakeup_source used when EPOLLWAKEUP is set */
166 struct wakeup_source __rcu
*ws
;
168 /* The structure that describe the interested events and the source fd */
169 struct epoll_event event
;
173 * This structure is stored inside the "private_data" member of the file
174 * structure and represents the main data structure for the eventpoll
179 * This mutex is used to ensure that files are not removed
180 * while epoll is using them. This is held during the event
181 * collection loop, the file cleanup path, the epoll file exit
182 * code and the ctl operations.
186 /* Wait queue used by sys_epoll_wait() */
187 wait_queue_head_t wq
;
189 /* Wait queue used by file->poll() */
190 wait_queue_head_t poll_wait
;
192 /* List of ready file descriptors */
193 struct list_head rdllist
;
195 /* Lock which protects rdllist and ovflist */
198 /* RB tree root used to store monitored fd structs */
199 struct rb_root_cached rbr
;
202 * This is a single linked list that chains all the "struct epitem" that
203 * happened while transferring ready events to userspace w/out
206 struct epitem
*ovflist
;
208 /* wakeup_source used when ep_scan_ready_list is running */
209 struct wakeup_source
*ws
;
211 /* The user that created the eventpoll descriptor */
212 struct user_struct
*user
;
216 /* used to optimize loop detection check */
218 struct hlist_head refs
;
220 #ifdef CONFIG_NET_RX_BUSY_POLL
221 /* used to track busy poll napi_id */
222 unsigned int napi_id
;
225 #ifdef CONFIG_DEBUG_LOCK_ALLOC
226 /* tracks wakeup nests for lockdep validation */
231 /* Wrapper struct used by poll queueing */
238 * Configuration options available inside /proc/sys/fs/epoll/
240 /* Maximum number of epoll watched descriptors, per user */
241 static long max_user_watches __read_mostly
;
244 * This mutex is used to serialize ep_free() and eventpoll_release_file().
246 static DEFINE_MUTEX(epmutex
);
248 static u64 loop_check_gen
= 0;
250 /* Used to check for epoll file descriptor inclusion loops */
251 static struct eventpoll
*inserting_into
;
253 /* Slab cache used to allocate "struct epitem" */
254 static struct kmem_cache
*epi_cache __read_mostly
;
256 /* Slab cache used to allocate "struct eppoll_entry" */
257 static struct kmem_cache
*pwq_cache __read_mostly
;
260 * List of files with newly added links, where we may need to limit the number
261 * of emanating paths. Protected by the epmutex.
263 struct epitems_head
{
264 struct hlist_head epitems
;
265 struct epitems_head
*next
;
267 static struct epitems_head
*tfile_check_list
= EP_UNACTIVE_PTR
;
269 static struct kmem_cache
*ephead_cache __read_mostly
;
271 static inline void free_ephead(struct epitems_head
*head
)
274 kmem_cache_free(ephead_cache
, head
);
277 static void list_file(struct file
*file
)
279 struct epitems_head
*head
;
281 head
= container_of(file
->f_ep
, struct epitems_head
, epitems
);
283 head
->next
= tfile_check_list
;
284 tfile_check_list
= head
;
288 static void unlist_file(struct epitems_head
*head
)
290 struct epitems_head
*to_free
= head
;
291 struct hlist_node
*p
= rcu_dereference(hlist_first_rcu(&head
->epitems
));
293 struct epitem
*epi
= container_of(p
, struct epitem
, fllink
);
294 spin_lock(&epi
->ffd
.file
->f_lock
);
295 if (!hlist_empty(&head
->epitems
))
298 spin_unlock(&epi
->ffd
.file
->f_lock
);
300 free_ephead(to_free
);
305 #include <linux/sysctl.h>
307 static long long_zero
;
308 static long long_max
= LONG_MAX
;
310 static struct ctl_table epoll_table
[] = {
312 .procname
= "max_user_watches",
313 .data
= &max_user_watches
,
314 .maxlen
= sizeof(max_user_watches
),
316 .proc_handler
= proc_doulongvec_minmax
,
317 .extra1
= &long_zero
,
323 static void __init
epoll_sysctls_init(void)
325 register_sysctl("fs/epoll", epoll_table
);
328 #define epoll_sysctls_init() do { } while (0)
329 #endif /* CONFIG_SYSCTL */
331 static const struct file_operations eventpoll_fops
;
333 static inline int is_file_epoll(struct file
*f
)
335 return f
->f_op
== &eventpoll_fops
;
338 /* Setup the structure that is used as key for the RB tree */
339 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
340 struct file
*file
, int fd
)
346 /* Compare RB tree keys */
347 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
348 struct epoll_filefd
*p2
)
350 return (p1
->file
> p2
->file
? +1:
351 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
354 /* Tells us if the item is currently linked */
355 static inline int ep_is_linked(struct epitem
*epi
)
357 return !list_empty(&epi
->rdllink
);
360 static inline struct eppoll_entry
*ep_pwq_from_wait(wait_queue_entry_t
*p
)
362 return container_of(p
, struct eppoll_entry
, wait
);
365 /* Get the "struct epitem" from a wait queue pointer */
366 static inline struct epitem
*ep_item_from_wait(wait_queue_entry_t
*p
)
368 return container_of(p
, struct eppoll_entry
, wait
)->base
;
372 * ep_events_available - Checks if ready events might be available.
374 * @ep: Pointer to the eventpoll context.
376 * Return: a value different than %zero if ready events are available,
377 * or %zero otherwise.
379 static inline int ep_events_available(struct eventpoll
*ep
)
381 return !list_empty_careful(&ep
->rdllist
) ||
382 READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
;
385 #ifdef CONFIG_NET_RX_BUSY_POLL
386 static bool ep_busy_loop_end(void *p
, unsigned long start_time
)
388 struct eventpoll
*ep
= p
;
390 return ep_events_available(ep
) || busy_loop_timeout(start_time
);
394 * Busy poll if globally on and supporting sockets found && no events,
395 * busy loop will return if need_resched or ep_events_available.
397 * we must do our busy polling with irqs enabled
399 static bool ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
401 unsigned int napi_id
= READ_ONCE(ep
->napi_id
);
403 if ((napi_id
>= MIN_NAPI_ID
) && net_busy_loop_on()) {
404 napi_busy_loop(napi_id
, nonblock
? NULL
: ep_busy_loop_end
, ep
, false,
406 if (ep_events_available(ep
))
409 * Busy poll timed out. Drop NAPI ID for now, we can add
410 * it back in when we have moved a socket with a valid NAPI
411 * ID onto the ready list.
420 * Set epoll busy poll NAPI ID from sk.
422 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
424 struct eventpoll
*ep
;
425 unsigned int napi_id
;
429 if (!net_busy_loop_on())
432 sock
= sock_from_file(epi
->ffd
.file
);
440 napi_id
= READ_ONCE(sk
->sk_napi_id
);
443 /* Non-NAPI IDs can be rejected
445 * Nothing to do if we already have this ID
447 if (napi_id
< MIN_NAPI_ID
|| napi_id
== ep
->napi_id
)
450 /* record NAPI ID for use in next busy poll */
451 ep
->napi_id
= napi_id
;
456 static inline bool ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
461 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
465 #endif /* CONFIG_NET_RX_BUSY_POLL */
468 * As described in commit 0ccf831cb lockdep: annotate epoll
469 * the use of wait queues used by epoll is done in a very controlled
470 * manner. Wake ups can nest inside each other, but are never done
471 * with the same locking. For example:
474 * efd1 = epoll_create();
475 * efd2 = epoll_create();
476 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
477 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
479 * When a packet arrives to the device underneath "dfd", the net code will
480 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
481 * callback wakeup entry on that queue, and the wake_up() performed by the
482 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
483 * (efd1) notices that it may have some event ready, so it needs to wake up
484 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
485 * that ends up in another wake_up(), after having checked about the
486 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
487 * avoid stack blasting.
489 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
490 * this special case of epoll.
492 #ifdef CONFIG_DEBUG_LOCK_ALLOC
494 static void ep_poll_safewake(struct eventpoll
*ep
, struct epitem
*epi
,
497 struct eventpoll
*ep_src
;
502 * To set the subclass or nesting level for spin_lock_irqsave_nested()
503 * it might be natural to create a per-cpu nest count. However, since
504 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
505 * schedule() in the -rt kernel, the per-cpu variable are no longer
506 * protected. Thus, we are introducing a per eventpoll nest field.
507 * If we are not being call from ep_poll_callback(), epi is NULL and
508 * we are at the first level of nesting, 0. Otherwise, we are being
509 * called from ep_poll_callback() and if a previous wakeup source is
510 * not an epoll file itself, we are at depth 1 since the wakeup source
511 * is depth 0. If the wakeup source is a previous epoll file in the
512 * wakeup chain then we use its nests value and record ours as
513 * nests + 1. The previous epoll file nests value is stable since its
514 * already holding its own poll_wait.lock.
517 if ((is_file_epoll(epi
->ffd
.file
))) {
518 ep_src
= epi
->ffd
.file
->private_data
;
519 nests
= ep_src
->nests
;
524 spin_lock_irqsave_nested(&ep
->poll_wait
.lock
, flags
, nests
);
525 ep
->nests
= nests
+ 1;
526 wake_up_locked_poll(&ep
->poll_wait
, EPOLLIN
| pollflags
);
528 spin_unlock_irqrestore(&ep
->poll_wait
.lock
, flags
);
533 static void ep_poll_safewake(struct eventpoll
*ep
, struct epitem
*epi
,
536 wake_up_poll(&ep
->poll_wait
, EPOLLIN
| pollflags
);
541 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
543 wait_queue_head_t
*whead
;
547 * If it is cleared by POLLFREE, it should be rcu-safe.
548 * If we read NULL we need a barrier paired with
549 * smp_store_release() in ep_poll_callback(), otherwise
550 * we rely on whead->lock.
552 whead
= smp_load_acquire(&pwq
->whead
);
554 remove_wait_queue(whead
, &pwq
->wait
);
559 * This function unregisters poll callbacks from the associated file
560 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
563 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
565 struct eppoll_entry
**p
= &epi
->pwqlist
;
566 struct eppoll_entry
*pwq
;
568 while ((pwq
= *p
) != NULL
) {
570 ep_remove_wait_queue(pwq
);
571 kmem_cache_free(pwq_cache
, pwq
);
575 /* call only when ep->mtx is held */
576 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
578 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
581 /* call only when ep->mtx is held */
582 static inline void ep_pm_stay_awake(struct epitem
*epi
)
584 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
590 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
592 return rcu_access_pointer(epi
->ws
) ? true : false;
595 /* call when ep->mtx cannot be held (ep_poll_callback) */
596 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
598 struct wakeup_source
*ws
;
601 ws
= rcu_dereference(epi
->ws
);
609 * ep->mutex needs to be held because we could be hit by
610 * eventpoll_release_file() and epoll_ctl().
612 static void ep_start_scan(struct eventpoll
*ep
, struct list_head
*txlist
)
615 * Steal the ready list, and re-init the original one to the
616 * empty list. Also, set ep->ovflist to NULL so that events
617 * happening while looping w/out locks, are not lost. We cannot
618 * have the poll callback to queue directly on ep->rdllist,
619 * because we want the "sproc" callback to be able to do it
622 lockdep_assert_irqs_enabled();
623 write_lock_irq(&ep
->lock
);
624 list_splice_init(&ep
->rdllist
, txlist
);
625 WRITE_ONCE(ep
->ovflist
, NULL
);
626 write_unlock_irq(&ep
->lock
);
629 static void ep_done_scan(struct eventpoll
*ep
,
630 struct list_head
*txlist
)
632 struct epitem
*epi
, *nepi
;
634 write_lock_irq(&ep
->lock
);
636 * During the time we spent inside the "sproc" callback, some
637 * other events might have been queued by the poll callback.
638 * We re-insert them inside the main ready-list here.
640 for (nepi
= READ_ONCE(ep
->ovflist
); (epi
= nepi
) != NULL
;
641 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
643 * We need to check if the item is already in the list.
644 * During the "sproc" callback execution time, items are
645 * queued into ->ovflist but the "txlist" might already
646 * contain them, and the list_splice() below takes care of them.
648 if (!ep_is_linked(epi
)) {
650 * ->ovflist is LIFO, so we have to reverse it in order
653 list_add(&epi
->rdllink
, &ep
->rdllist
);
654 ep_pm_stay_awake(epi
);
658 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
659 * releasing the lock, events will be queued in the normal way inside
662 WRITE_ONCE(ep
->ovflist
, EP_UNACTIVE_PTR
);
665 * Quickly re-inject items left on "txlist".
667 list_splice(txlist
, &ep
->rdllist
);
670 if (!list_empty(&ep
->rdllist
)) {
671 if (waitqueue_active(&ep
->wq
))
675 write_unlock_irq(&ep
->lock
);
678 static void epi_rcu_free(struct rcu_head
*head
)
680 struct epitem
*epi
= container_of(head
, struct epitem
, rcu
);
681 kmem_cache_free(epi_cache
, epi
);
685 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
686 * all the associated resources. Must be called with "mtx" held.
688 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
690 struct file
*file
= epi
->ffd
.file
;
691 struct epitems_head
*to_free
;
692 struct hlist_head
*head
;
694 lockdep_assert_irqs_enabled();
697 * Removes poll wait queue hooks.
699 ep_unregister_pollwait(ep
, epi
);
701 /* Remove the current item from the list of epoll hooks */
702 spin_lock(&file
->f_lock
);
705 if (head
->first
== &epi
->fllink
&& !epi
->fllink
.next
) {
707 if (!is_file_epoll(file
)) {
708 struct epitems_head
*v
;
709 v
= container_of(head
, struct epitems_head
, epitems
);
710 if (!smp_load_acquire(&v
->next
))
714 hlist_del_rcu(&epi
->fllink
);
715 spin_unlock(&file
->f_lock
);
716 free_ephead(to_free
);
718 rb_erase_cached(&epi
->rbn
, &ep
->rbr
);
720 write_lock_irq(&ep
->lock
);
721 if (ep_is_linked(epi
))
722 list_del_init(&epi
->rdllink
);
723 write_unlock_irq(&ep
->lock
);
725 wakeup_source_unregister(ep_wakeup_source(epi
));
727 * At this point it is safe to free the eventpoll item. Use the union
728 * field epi->rcu, since we are trying to minimize the size of
729 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
730 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
731 * use of the rbn field.
733 call_rcu(&epi
->rcu
, epi_rcu_free
);
735 percpu_counter_dec(&ep
->user
->epoll_watches
);
740 static void ep_free(struct eventpoll
*ep
)
745 /* We need to release all tasks waiting for these file */
746 if (waitqueue_active(&ep
->poll_wait
))
747 ep_poll_safewake(ep
, NULL
, 0);
750 * We need to lock this because we could be hit by
751 * eventpoll_release_file() while we're freeing the "struct eventpoll".
752 * We do not need to hold "ep->mtx" here because the epoll file
753 * is on the way to be removed and no one has references to it
754 * anymore. The only hit might come from eventpoll_release_file() but
755 * holding "epmutex" is sufficient here.
757 mutex_lock(&epmutex
);
760 * Walks through the whole tree by unregistering poll callbacks.
762 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
763 epi
= rb_entry(rbp
, struct epitem
, rbn
);
765 ep_unregister_pollwait(ep
, epi
);
770 * Walks through the whole tree by freeing each "struct epitem". At this
771 * point we are sure no poll callbacks will be lingering around, and also by
772 * holding "epmutex" we can be sure that no file cleanup code will hit
773 * us during this operation. So we can avoid the lock on "ep->lock".
774 * We do not need to lock ep->mtx, either, we only do it to prevent
777 mutex_lock(&ep
->mtx
);
778 while ((rbp
= rb_first_cached(&ep
->rbr
)) != NULL
) {
779 epi
= rb_entry(rbp
, struct epitem
, rbn
);
783 mutex_unlock(&ep
->mtx
);
785 mutex_unlock(&epmutex
);
786 mutex_destroy(&ep
->mtx
);
788 wakeup_source_unregister(ep
->ws
);
792 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
794 struct eventpoll
*ep
= file
->private_data
;
802 static __poll_t
ep_item_poll(const struct epitem
*epi
, poll_table
*pt
, int depth
);
804 static __poll_t
__ep_eventpoll_poll(struct file
*file
, poll_table
*wait
, int depth
)
806 struct eventpoll
*ep
= file
->private_data
;
808 struct epitem
*epi
, *tmp
;
812 init_poll_funcptr(&pt
, NULL
);
814 /* Insert inside our poll wait queue */
815 poll_wait(file
, &ep
->poll_wait
, wait
);
818 * Proceed to find out if wanted events are really available inside
821 mutex_lock_nested(&ep
->mtx
, depth
);
822 ep_start_scan(ep
, &txlist
);
823 list_for_each_entry_safe(epi
, tmp
, &txlist
, rdllink
) {
824 if (ep_item_poll(epi
, &pt
, depth
+ 1)) {
825 res
= EPOLLIN
| EPOLLRDNORM
;
829 * Item has been dropped into the ready list by the poll
830 * callback, but it's not actually ready, as far as
831 * caller requested events goes. We can remove it here.
833 __pm_relax(ep_wakeup_source(epi
));
834 list_del_init(&epi
->rdllink
);
837 ep_done_scan(ep
, &txlist
);
838 mutex_unlock(&ep
->mtx
);
843 * Differs from ep_eventpoll_poll() in that internal callers already have
844 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
845 * is correctly annotated.
847 static __poll_t
ep_item_poll(const struct epitem
*epi
, poll_table
*pt
,
850 struct file
*file
= epi
->ffd
.file
;
853 pt
->_key
= epi
->event
.events
;
854 if (!is_file_epoll(file
))
855 res
= vfs_poll(file
, pt
);
857 res
= __ep_eventpoll_poll(file
, pt
, depth
);
858 return res
& epi
->event
.events
;
861 static __poll_t
ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
863 return __ep_eventpoll_poll(file
, wait
, 0);
866 #ifdef CONFIG_PROC_FS
867 static void ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
869 struct eventpoll
*ep
= f
->private_data
;
872 mutex_lock(&ep
->mtx
);
873 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
874 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
875 struct inode
*inode
= file_inode(epi
->ffd
.file
);
877 seq_printf(m
, "tfd: %8d events: %8x data: %16llx "
878 " pos:%lli ino:%lx sdev:%x\n",
879 epi
->ffd
.fd
, epi
->event
.events
,
880 (long long)epi
->event
.data
,
881 (long long)epi
->ffd
.file
->f_pos
,
882 inode
->i_ino
, inode
->i_sb
->s_dev
);
883 if (seq_has_overflowed(m
))
886 mutex_unlock(&ep
->mtx
);
890 /* File callbacks that implement the eventpoll file behaviour */
891 static const struct file_operations eventpoll_fops
= {
892 #ifdef CONFIG_PROC_FS
893 .show_fdinfo
= ep_show_fdinfo
,
895 .release
= ep_eventpoll_release
,
896 .poll
= ep_eventpoll_poll
,
897 .llseek
= noop_llseek
,
901 * This is called from eventpoll_release() to unlink files from the eventpoll
902 * interface. We need to have this facility to cleanup correctly files that are
903 * closed without being removed from the eventpoll interface.
905 void eventpoll_release_file(struct file
*file
)
907 struct eventpoll
*ep
;
909 struct hlist_node
*next
;
912 * We don't want to get "file->f_lock" because it is not
913 * necessary. It is not necessary because we're in the "struct file"
914 * cleanup path, and this means that no one is using this file anymore.
915 * So, for example, epoll_ctl() cannot hit here since if we reach this
916 * point, the file counter already went to zero and fget() would fail.
917 * The only hit might come from ep_free() but by holding the mutex
918 * will correctly serialize the operation. We do need to acquire
919 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
920 * from anywhere but ep_free().
922 * Besides, ep_remove() acquires the lock, so we can't hold it here.
924 mutex_lock(&epmutex
);
925 if (unlikely(!file
->f_ep
)) {
926 mutex_unlock(&epmutex
);
929 hlist_for_each_entry_safe(epi
, next
, file
->f_ep
, fllink
) {
931 mutex_lock_nested(&ep
->mtx
, 0);
933 mutex_unlock(&ep
->mtx
);
935 mutex_unlock(&epmutex
);
938 static int ep_alloc(struct eventpoll
**pep
)
941 struct user_struct
*user
;
942 struct eventpoll
*ep
;
944 user
= get_current_user();
946 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
950 mutex_init(&ep
->mtx
);
951 rwlock_init(&ep
->lock
);
952 init_waitqueue_head(&ep
->wq
);
953 init_waitqueue_head(&ep
->poll_wait
);
954 INIT_LIST_HEAD(&ep
->rdllist
);
955 ep
->rbr
= RB_ROOT_CACHED
;
956 ep
->ovflist
= EP_UNACTIVE_PTR
;
969 * Search the file inside the eventpoll tree. The RB tree operations
970 * are protected by the "mtx" mutex, and ep_find() must be called with
973 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
977 struct epitem
*epi
, *epir
= NULL
;
978 struct epoll_filefd ffd
;
980 ep_set_ffd(&ffd
, file
, fd
);
981 for (rbp
= ep
->rbr
.rb_root
.rb_node
; rbp
; ) {
982 epi
= rb_entry(rbp
, struct epitem
, rbn
);
983 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
998 static struct epitem
*ep_find_tfd(struct eventpoll
*ep
, int tfd
, unsigned long toff
)
1000 struct rb_node
*rbp
;
1003 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1004 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1005 if (epi
->ffd
.fd
== tfd
) {
1017 struct file
*get_epoll_tfile_raw_ptr(struct file
*file
, int tfd
,
1020 struct file
*file_raw
;
1021 struct eventpoll
*ep
;
1024 if (!is_file_epoll(file
))
1025 return ERR_PTR(-EINVAL
);
1027 ep
= file
->private_data
;
1029 mutex_lock(&ep
->mtx
);
1030 epi
= ep_find_tfd(ep
, tfd
, toff
);
1032 file_raw
= epi
->ffd
.file
;
1034 file_raw
= ERR_PTR(-ENOENT
);
1035 mutex_unlock(&ep
->mtx
);
1039 #endif /* CONFIG_KCMP */
1042 * Adds a new entry to the tail of the list in a lockless way, i.e.
1043 * multiple CPUs are allowed to call this function concurrently.
1045 * Beware: it is necessary to prevent any other modifications of the
1046 * existing list until all changes are completed, in other words
1047 * concurrent list_add_tail_lockless() calls should be protected
1048 * with a read lock, where write lock acts as a barrier which
1049 * makes sure all list_add_tail_lockless() calls are fully
1052 * Also an element can be locklessly added to the list only in one
1053 * direction i.e. either to the tail or to the head, otherwise
1054 * concurrent access will corrupt the list.
1056 * Return: %false if element has been already added to the list, %true
1059 static inline bool list_add_tail_lockless(struct list_head
*new,
1060 struct list_head
*head
)
1062 struct list_head
*prev
;
1065 * This is simple 'new->next = head' operation, but cmpxchg()
1066 * is used in order to detect that same element has been just
1067 * added to the list from another CPU: the winner observes
1070 if (!try_cmpxchg(&new->next
, &new, head
))
1074 * Initially ->next of a new element must be updated with the head
1075 * (we are inserting to the tail) and only then pointers are atomically
1076 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1077 * updated before pointers are actually swapped and pointers are
1078 * swapped before prev->next is updated.
1081 prev
= xchg(&head
->prev
, new);
1084 * It is safe to modify prev->next and new->prev, because a new element
1085 * is added only to the tail and new->next is updated before XCHG.
1095 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1096 * i.e. multiple CPUs are allowed to call this function concurrently.
1098 * Return: %false if epi element has been already chained, %true otherwise.
1100 static inline bool chain_epi_lockless(struct epitem
*epi
)
1102 struct eventpoll
*ep
= epi
->ep
;
1104 /* Fast preliminary check */
1105 if (epi
->next
!= EP_UNACTIVE_PTR
)
1108 /* Check that the same epi has not been just chained from another CPU */
1109 if (cmpxchg(&epi
->next
, EP_UNACTIVE_PTR
, NULL
) != EP_UNACTIVE_PTR
)
1112 /* Atomically exchange tail */
1113 epi
->next
= xchg(&ep
->ovflist
, epi
);
1119 * This is the callback that is passed to the wait queue wakeup
1120 * mechanism. It is called by the stored file descriptors when they
1121 * have events to report.
1123 * This callback takes a read lock in order not to contend with concurrent
1124 * events from another file descriptor, thus all modifications to ->rdllist
1125 * or ->ovflist are lockless. Read lock is paired with the write lock from
1126 * ep_scan_ready_list(), which stops all list modifications and guarantees
1127 * that lists state is seen correctly.
1129 * Another thing worth to mention is that ep_poll_callback() can be called
1130 * concurrently for the same @epi from different CPUs if poll table was inited
1131 * with several wait queues entries. Plural wakeup from different CPUs of a
1132 * single wait queue is serialized by wq.lock, but the case when multiple wait
1133 * queues are used should be detected accordingly. This is detected using
1134 * cmpxchg() operation.
1136 static int ep_poll_callback(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
1139 struct epitem
*epi
= ep_item_from_wait(wait
);
1140 struct eventpoll
*ep
= epi
->ep
;
1141 __poll_t pollflags
= key_to_poll(key
);
1142 unsigned long flags
;
1145 read_lock_irqsave(&ep
->lock
, flags
);
1147 ep_set_busy_poll_napi_id(epi
);
1150 * If the event mask does not contain any poll(2) event, we consider the
1151 * descriptor to be disabled. This condition is likely the effect of the
1152 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1153 * until the next EPOLL_CTL_MOD will be issued.
1155 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1159 * Check the events coming with the callback. At this stage, not
1160 * every device reports the events in the "key" parameter of the
1161 * callback. We need to be able to handle both cases here, hence the
1162 * test for "key" != NULL before the event match test.
1164 if (pollflags
&& !(pollflags
& epi
->event
.events
))
1168 * If we are transferring events to userspace, we can hold no locks
1169 * (because we're accessing user memory, and because of linux f_op->poll()
1170 * semantics). All the events that happen during that period of time are
1171 * chained in ep->ovflist and requeued later on.
1173 if (READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
) {
1174 if (chain_epi_lockless(epi
))
1175 ep_pm_stay_awake_rcu(epi
);
1176 } else if (!ep_is_linked(epi
)) {
1177 /* In the usual case, add event to ready list. */
1178 if (list_add_tail_lockless(&epi
->rdllink
, &ep
->rdllist
))
1179 ep_pm_stay_awake_rcu(epi
);
1183 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1186 if (waitqueue_active(&ep
->wq
)) {
1187 if ((epi
->event
.events
& EPOLLEXCLUSIVE
) &&
1188 !(pollflags
& POLLFREE
)) {
1189 switch (pollflags
& EPOLLINOUT_BITS
) {
1191 if (epi
->event
.events
& EPOLLIN
)
1195 if (epi
->event
.events
& EPOLLOUT
)
1205 if (waitqueue_active(&ep
->poll_wait
))
1209 read_unlock_irqrestore(&ep
->lock
, flags
);
1211 /* We have to call this outside the lock */
1213 ep_poll_safewake(ep
, epi
, pollflags
& EPOLL_URING_WAKE
);
1215 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
))
1218 if (pollflags
& POLLFREE
) {
1220 * If we race with ep_remove_wait_queue() it can miss
1221 * ->whead = NULL and do another remove_wait_queue() after
1222 * us, so we can't use __remove_wait_queue().
1224 list_del_init(&wait
->entry
);
1226 * ->whead != NULL protects us from the race with ep_free()
1227 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1228 * held by the caller. Once we nullify it, nothing protects
1229 * ep/epi or even wait.
1231 smp_store_release(&ep_pwq_from_wait(wait
)->whead
, NULL
);
1238 * This is the callback that is used to add our wait queue to the
1239 * target file wakeup lists.
1241 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1244 struct ep_pqueue
*epq
= container_of(pt
, struct ep_pqueue
, pt
);
1245 struct epitem
*epi
= epq
->epi
;
1246 struct eppoll_entry
*pwq
;
1248 if (unlikely(!epi
)) // an earlier allocation has failed
1251 pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
);
1252 if (unlikely(!pwq
)) {
1257 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1260 if (epi
->event
.events
& EPOLLEXCLUSIVE
)
1261 add_wait_queue_exclusive(whead
, &pwq
->wait
);
1263 add_wait_queue(whead
, &pwq
->wait
);
1264 pwq
->next
= epi
->pwqlist
;
1268 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1271 struct rb_node
**p
= &ep
->rbr
.rb_root
.rb_node
, *parent
= NULL
;
1272 struct epitem
*epic
;
1273 bool leftmost
= true;
1277 epic
= rb_entry(parent
, struct epitem
, rbn
);
1278 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1280 p
= &parent
->rb_right
;
1283 p
= &parent
->rb_left
;
1285 rb_link_node(&epi
->rbn
, parent
, p
);
1286 rb_insert_color_cached(&epi
->rbn
, &ep
->rbr
, leftmost
);
1291 #define PATH_ARR_SIZE 5
1293 * These are the number paths of length 1 to 5, that we are allowing to emanate
1294 * from a single file of interest. For example, we allow 1000 paths of length
1295 * 1, to emanate from each file of interest. This essentially represents the
1296 * potential wakeup paths, which need to be limited in order to avoid massive
1297 * uncontrolled wakeup storms. The common use case should be a single ep which
1298 * is connected to n file sources. In this case each file source has 1 path
1299 * of length 1. Thus, the numbers below should be more than sufficient. These
1300 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1301 * and delete can't add additional paths. Protected by the epmutex.
1303 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1304 static int path_count
[PATH_ARR_SIZE
];
1306 static int path_count_inc(int nests
)
1308 /* Allow an arbitrary number of depth 1 paths */
1312 if (++path_count
[nests
] > path_limits
[nests
])
1317 static void path_count_init(void)
1321 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1325 static int reverse_path_check_proc(struct hlist_head
*refs
, int depth
)
1330 if (depth
> EP_MAX_NESTS
) /* too deep nesting */
1333 /* CTL_DEL can remove links here, but that can't increase our count */
1334 hlist_for_each_entry_rcu(epi
, refs
, fllink
) {
1335 struct hlist_head
*refs
= &epi
->ep
->refs
;
1336 if (hlist_empty(refs
))
1337 error
= path_count_inc(depth
);
1339 error
= reverse_path_check_proc(refs
, depth
+ 1);
1347 * reverse_path_check - The tfile_check_list is list of epitem_head, which have
1348 * links that are proposed to be newly added. We need to
1349 * make sure that those added links don't add too many
1350 * paths such that we will spend all our time waking up
1351 * eventpoll objects.
1353 * Return: %zero if the proposed links don't create too many paths,
1356 static int reverse_path_check(void)
1358 struct epitems_head
*p
;
1360 for (p
= tfile_check_list
; p
!= EP_UNACTIVE_PTR
; p
= p
->next
) {
1364 error
= reverse_path_check_proc(&p
->epitems
, 0);
1372 static int ep_create_wakeup_source(struct epitem
*epi
)
1374 struct name_snapshot n
;
1375 struct wakeup_source
*ws
;
1378 epi
->ep
->ws
= wakeup_source_register(NULL
, "eventpoll");
1383 take_dentry_name_snapshot(&n
, epi
->ffd
.file
->f_path
.dentry
);
1384 ws
= wakeup_source_register(NULL
, n
.name
.name
);
1385 release_dentry_name_snapshot(&n
);
1389 rcu_assign_pointer(epi
->ws
, ws
);
1394 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1395 static noinline
void ep_destroy_wakeup_source(struct epitem
*epi
)
1397 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
1399 RCU_INIT_POINTER(epi
->ws
, NULL
);
1402 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1403 * used internally by wakeup_source_remove, too (called by
1404 * wakeup_source_unregister), so we cannot use call_rcu
1407 wakeup_source_unregister(ws
);
1410 static int attach_epitem(struct file
*file
, struct epitem
*epi
)
1412 struct epitems_head
*to_free
= NULL
;
1413 struct hlist_head
*head
= NULL
;
1414 struct eventpoll
*ep
= NULL
;
1416 if (is_file_epoll(file
))
1417 ep
= file
->private_data
;
1421 } else if (!READ_ONCE(file
->f_ep
)) {
1423 to_free
= kmem_cache_zalloc(ephead_cache
, GFP_KERNEL
);
1426 head
= &to_free
->epitems
;
1428 spin_lock(&file
->f_lock
);
1430 if (unlikely(!head
)) {
1431 spin_unlock(&file
->f_lock
);
1437 hlist_add_head_rcu(&epi
->fllink
, file
->f_ep
);
1438 spin_unlock(&file
->f_lock
);
1439 free_ephead(to_free
);
1444 * Must be called with "mtx" held.
1446 static int ep_insert(struct eventpoll
*ep
, const struct epoll_event
*event
,
1447 struct file
*tfile
, int fd
, int full_check
)
1449 int error
, pwake
= 0;
1452 struct ep_pqueue epq
;
1453 struct eventpoll
*tep
= NULL
;
1455 if (is_file_epoll(tfile
))
1456 tep
= tfile
->private_data
;
1458 lockdep_assert_irqs_enabled();
1460 if (unlikely(percpu_counter_compare(&ep
->user
->epoll_watches
,
1461 max_user_watches
) >= 0))
1463 percpu_counter_inc(&ep
->user
->epoll_watches
);
1465 if (!(epi
= kmem_cache_zalloc(epi_cache
, GFP_KERNEL
))) {
1466 percpu_counter_dec(&ep
->user
->epoll_watches
);
1470 /* Item initialization follow here ... */
1471 INIT_LIST_HEAD(&epi
->rdllink
);
1473 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
1474 epi
->event
= *event
;
1475 epi
->next
= EP_UNACTIVE_PTR
;
1478 mutex_lock_nested(&tep
->mtx
, 1);
1479 /* Add the current item to the list of active epoll hook for this file */
1480 if (unlikely(attach_epitem(tfile
, epi
) < 0)) {
1482 mutex_unlock(&tep
->mtx
);
1483 kmem_cache_free(epi_cache
, epi
);
1484 percpu_counter_dec(&ep
->user
->epoll_watches
);
1488 if (full_check
&& !tep
)
1492 * Add the current item to the RB tree. All RB tree operations are
1493 * protected by "mtx", and ep_insert() is called with "mtx" held.
1495 ep_rbtree_insert(ep
, epi
);
1497 mutex_unlock(&tep
->mtx
);
1499 /* now check if we've created too many backpaths */
1500 if (unlikely(full_check
&& reverse_path_check())) {
1505 if (epi
->event
.events
& EPOLLWAKEUP
) {
1506 error
= ep_create_wakeup_source(epi
);
1513 /* Initialize the poll table using the queue callback */
1515 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1518 * Attach the item to the poll hooks and get current event bits.
1519 * We can safely use the file* here because its usage count has
1520 * been increased by the caller of this function. Note that after
1521 * this operation completes, the poll callback can start hitting
1524 revents
= ep_item_poll(epi
, &epq
.pt
, 1);
1527 * We have to check if something went wrong during the poll wait queue
1528 * install process. Namely an allocation for a wait queue failed due
1529 * high memory pressure.
1531 if (unlikely(!epq
.epi
)) {
1536 /* We have to drop the new item inside our item list to keep track of it */
1537 write_lock_irq(&ep
->lock
);
1539 /* record NAPI ID of new item if present */
1540 ep_set_busy_poll_napi_id(epi
);
1542 /* If the file is already "ready" we drop it inside the ready list */
1543 if (revents
&& !ep_is_linked(epi
)) {
1544 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1545 ep_pm_stay_awake(epi
);
1547 /* Notify waiting tasks that events are available */
1548 if (waitqueue_active(&ep
->wq
))
1550 if (waitqueue_active(&ep
->poll_wait
))
1554 write_unlock_irq(&ep
->lock
);
1556 /* We have to call this outside the lock */
1558 ep_poll_safewake(ep
, NULL
, 0);
1564 * Modify the interest event mask by dropping an event if the new mask
1565 * has a match in the current file status. Must be called with "mtx" held.
1567 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
,
1568 const struct epoll_event
*event
)
1573 lockdep_assert_irqs_enabled();
1575 init_poll_funcptr(&pt
, NULL
);
1578 * Set the new event interest mask before calling f_op->poll();
1579 * otherwise we might miss an event that happens between the
1580 * f_op->poll() call and the new event set registering.
1582 epi
->event
.events
= event
->events
; /* need barrier below */
1583 epi
->event
.data
= event
->data
; /* protected by mtx */
1584 if (epi
->event
.events
& EPOLLWAKEUP
) {
1585 if (!ep_has_wakeup_source(epi
))
1586 ep_create_wakeup_source(epi
);
1587 } else if (ep_has_wakeup_source(epi
)) {
1588 ep_destroy_wakeup_source(epi
);
1592 * The following barrier has two effects:
1594 * 1) Flush epi changes above to other CPUs. This ensures
1595 * we do not miss events from ep_poll_callback if an
1596 * event occurs immediately after we call f_op->poll().
1597 * We need this because we did not take ep->lock while
1598 * changing epi above (but ep_poll_callback does take
1601 * 2) We also need to ensure we do not miss _past_ events
1602 * when calling f_op->poll(). This barrier also
1603 * pairs with the barrier in wq_has_sleeper (see
1604 * comments for wq_has_sleeper).
1606 * This barrier will now guarantee ep_poll_callback or f_op->poll
1607 * (or both) will notice the readiness of an item.
1612 * Get current event bits. We can safely use the file* here because
1613 * its usage count has been increased by the caller of this function.
1614 * If the item is "hot" and it is not registered inside the ready
1615 * list, push it inside.
1617 if (ep_item_poll(epi
, &pt
, 1)) {
1618 write_lock_irq(&ep
->lock
);
1619 if (!ep_is_linked(epi
)) {
1620 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1621 ep_pm_stay_awake(epi
);
1623 /* Notify waiting tasks that events are available */
1624 if (waitqueue_active(&ep
->wq
))
1626 if (waitqueue_active(&ep
->poll_wait
))
1629 write_unlock_irq(&ep
->lock
);
1632 /* We have to call this outside the lock */
1634 ep_poll_safewake(ep
, NULL
, 0);
1639 static int ep_send_events(struct eventpoll
*ep
,
1640 struct epoll_event __user
*events
, int maxevents
)
1642 struct epitem
*epi
, *tmp
;
1648 * Always short-circuit for fatal signals to allow threads to make a
1649 * timely exit without the chance of finding more events available and
1650 * fetching repeatedly.
1652 if (fatal_signal_pending(current
))
1655 init_poll_funcptr(&pt
, NULL
);
1657 mutex_lock(&ep
->mtx
);
1658 ep_start_scan(ep
, &txlist
);
1661 * We can loop without lock because we are passed a task private list.
1662 * Items cannot vanish during the loop we are holding ep->mtx.
1664 list_for_each_entry_safe(epi
, tmp
, &txlist
, rdllink
) {
1665 struct wakeup_source
*ws
;
1668 if (res
>= maxevents
)
1672 * Activate ep->ws before deactivating epi->ws to prevent
1673 * triggering auto-suspend here (in case we reactive epi->ws
1676 * This could be rearranged to delay the deactivation of epi->ws
1677 * instead, but then epi->ws would temporarily be out of sync
1678 * with ep_is_linked().
1680 ws
= ep_wakeup_source(epi
);
1683 __pm_stay_awake(ep
->ws
);
1687 list_del_init(&epi
->rdllink
);
1690 * If the event mask intersect the caller-requested one,
1691 * deliver the event to userspace. Again, we are holding ep->mtx,
1692 * so no operations coming from userspace can change the item.
1694 revents
= ep_item_poll(epi
, &pt
, 1);
1698 events
= epoll_put_uevent(revents
, epi
->event
.data
, events
);
1700 list_add(&epi
->rdllink
, &txlist
);
1701 ep_pm_stay_awake(epi
);
1707 if (epi
->event
.events
& EPOLLONESHOT
)
1708 epi
->event
.events
&= EP_PRIVATE_BITS
;
1709 else if (!(epi
->event
.events
& EPOLLET
)) {
1711 * If this file has been added with Level
1712 * Trigger mode, we need to insert back inside
1713 * the ready list, so that the next call to
1714 * epoll_wait() will check again the events
1715 * availability. At this point, no one can insert
1716 * into ep->rdllist besides us. The epoll_ctl()
1717 * callers are locked out by
1718 * ep_scan_ready_list() holding "mtx" and the
1719 * poll callback will queue them in ep->ovflist.
1721 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1722 ep_pm_stay_awake(epi
);
1725 ep_done_scan(ep
, &txlist
);
1726 mutex_unlock(&ep
->mtx
);
1731 static struct timespec64
*ep_timeout_to_timespec(struct timespec64
*to
, long ms
)
1733 struct timespec64 now
;
1744 to
->tv_sec
= ms
/ MSEC_PER_SEC
;
1745 to
->tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
);
1747 ktime_get_ts64(&now
);
1748 *to
= timespec64_add_safe(now
, *to
);
1753 * autoremove_wake_function, but remove even on failure to wake up, because we
1754 * know that default_wake_function/ttwu will only fail if the thread is already
1755 * woken, and in that case the ep_poll loop will remove the entry anyways, not
1758 static int ep_autoremove_wake_function(struct wait_queue_entry
*wq_entry
,
1759 unsigned int mode
, int sync
, void *key
)
1761 int ret
= default_wake_function(wq_entry
, mode
, sync
, key
);
1763 list_del_init(&wq_entry
->entry
);
1768 * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
1771 * @ep: Pointer to the eventpoll context.
1772 * @events: Pointer to the userspace buffer where the ready events should be
1774 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1775 * @timeout: Maximum timeout for the ready events fetch operation, in
1776 * timespec. If the timeout is zero, the function will not block,
1777 * while if the @timeout ptr is NULL, the function will block
1778 * until at least one event has been retrieved (or an error
1781 * Return: the number of ready events which have been fetched, or an
1782 * error code, in case of error.
1784 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1785 int maxevents
, struct timespec64
*timeout
)
1787 int res
, eavail
, timed_out
= 0;
1789 wait_queue_entry_t wait
;
1790 ktime_t expires
, *to
= NULL
;
1792 lockdep_assert_irqs_enabled();
1794 if (timeout
&& (timeout
->tv_sec
| timeout
->tv_nsec
)) {
1795 slack
= select_estimate_accuracy(timeout
);
1797 *to
= timespec64_to_ktime(*timeout
);
1798 } else if (timeout
) {
1800 * Avoid the unnecessary trip to the wait queue loop, if the
1801 * caller specified a non blocking operation.
1807 * This call is racy: We may or may not see events that are being added
1808 * to the ready list under the lock (e.g., in IRQ callbacks). For cases
1809 * with a non-zero timeout, this thread will check the ready list under
1810 * lock and will add to the wait queue. For cases with a zero
1811 * timeout, the user by definition should not care and will have to
1814 eavail
= ep_events_available(ep
);
1819 * Try to transfer events to user space. In case we get
1820 * 0 events and there's still timeout left over, we go
1821 * trying again in search of more luck.
1823 res
= ep_send_events(ep
, events
, maxevents
);
1831 eavail
= ep_busy_loop(ep
, timed_out
);
1835 if (signal_pending(current
))
1839 * Internally init_wait() uses autoremove_wake_function(),
1840 * thus wait entry is removed from the wait queue on each
1841 * wakeup. Why it is important? In case of several waiters
1842 * each new wakeup will hit the next waiter, giving it the
1843 * chance to harvest new event. Otherwise wakeup can be
1844 * lost. This is also good performance-wise, because on
1845 * normal wakeup path no need to call __remove_wait_queue()
1846 * explicitly, thus ep->lock is not taken, which halts the
1849 * In fact, we now use an even more aggressive function that
1850 * unconditionally removes, because we don't reuse the wait
1851 * entry between loop iterations. This lets us also avoid the
1852 * performance issue if a process is killed, causing all of its
1853 * threads to wake up without being removed normally.
1856 wait
.func
= ep_autoremove_wake_function
;
1858 write_lock_irq(&ep
->lock
);
1860 * Barrierless variant, waitqueue_active() is called under
1861 * the same lock on wakeup ep_poll_callback() side, so it
1862 * is safe to avoid an explicit barrier.
1864 __set_current_state(TASK_INTERRUPTIBLE
);
1867 * Do the final check under the lock. ep_scan_ready_list()
1868 * plays with two lists (->rdllist and ->ovflist) and there
1869 * is always a race when both lists are empty for short
1870 * period of time although events are pending, so lock is
1873 eavail
= ep_events_available(ep
);
1875 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1877 write_unlock_irq(&ep
->lock
);
1880 timed_out
= !schedule_hrtimeout_range(to
, slack
,
1882 __set_current_state(TASK_RUNNING
);
1885 * We were woken up, thus go and try to harvest some events.
1886 * If timed out and still on the wait queue, recheck eavail
1887 * carefully under lock, below.
1891 if (!list_empty_careful(&wait
.entry
)) {
1892 write_lock_irq(&ep
->lock
);
1894 * If the thread timed out and is not on the wait queue,
1895 * it means that the thread was woken up after its
1896 * timeout expired before it could reacquire the lock.
1897 * Thus, when wait.entry is empty, it needs to harvest
1901 eavail
= list_empty(&wait
.entry
);
1902 __remove_wait_queue(&ep
->wq
, &wait
);
1903 write_unlock_irq(&ep
->lock
);
1909 * ep_loop_check_proc - verify that adding an epoll file inside another
1910 * epoll structure does not violate the constraints, in
1911 * terms of closed loops, or too deep chains (which can
1912 * result in excessive stack usage).
1914 * @ep: the &struct eventpoll to be currently checked.
1915 * @depth: Current depth of the path being checked.
1917 * Return: %zero if adding the epoll @file inside current epoll
1918 * structure @ep does not violate the constraints, or %-1 otherwise.
1920 static int ep_loop_check_proc(struct eventpoll
*ep
, int depth
)
1923 struct rb_node
*rbp
;
1926 mutex_lock_nested(&ep
->mtx
, depth
+ 1);
1927 ep
->gen
= loop_check_gen
;
1928 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1929 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1930 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1931 struct eventpoll
*ep_tovisit
;
1932 ep_tovisit
= epi
->ffd
.file
->private_data
;
1933 if (ep_tovisit
->gen
== loop_check_gen
)
1935 if (ep_tovisit
== inserting_into
|| depth
> EP_MAX_NESTS
)
1938 error
= ep_loop_check_proc(ep_tovisit
, depth
+ 1);
1943 * If we've reached a file that is not associated with
1944 * an ep, then we need to check if the newly added
1945 * links are going to add too many wakeup paths. We do
1946 * this by adding it to the tfile_check_list, if it's
1947 * not already there, and calling reverse_path_check()
1948 * during ep_insert().
1950 list_file(epi
->ffd
.file
);
1953 mutex_unlock(&ep
->mtx
);
1959 * ep_loop_check - Performs a check to verify that adding an epoll file (@to)
1960 * into another epoll file (represented by @ep) does not create
1961 * closed loops or too deep chains.
1963 * @ep: Pointer to the epoll we are inserting into.
1964 * @to: Pointer to the epoll to be inserted.
1966 * Return: %zero if adding the epoll @to inside the epoll @from
1967 * does not violate the constraints, or %-1 otherwise.
1969 static int ep_loop_check(struct eventpoll
*ep
, struct eventpoll
*to
)
1971 inserting_into
= ep
;
1972 return ep_loop_check_proc(to
, 0);
1975 static void clear_tfile_check_list(void)
1978 while (tfile_check_list
!= EP_UNACTIVE_PTR
) {
1979 struct epitems_head
*head
= tfile_check_list
;
1980 tfile_check_list
= head
->next
;
1987 * Open an eventpoll file descriptor.
1989 static int do_epoll_create(int flags
)
1992 struct eventpoll
*ep
= NULL
;
1995 /* Check the EPOLL_* constant for consistency. */
1996 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
1998 if (flags
& ~EPOLL_CLOEXEC
)
2001 * Create the internal data structure ("struct eventpoll").
2003 error
= ep_alloc(&ep
);
2007 * Creates all the items needed to setup an eventpoll file. That is,
2008 * a file structure and a free file descriptor.
2010 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
2015 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
2016 O_RDWR
| (flags
& O_CLOEXEC
));
2018 error
= PTR_ERR(file
);
2022 fd_install(fd
, file
);
2032 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
2034 return do_epoll_create(flags
);
2037 SYSCALL_DEFINE1(epoll_create
, int, size
)
2042 return do_epoll_create(0);
2045 static inline int epoll_mutex_lock(struct mutex
*mutex
, int depth
,
2049 mutex_lock_nested(mutex
, depth
);
2052 if (mutex_trylock(mutex
))
2057 int do_epoll_ctl(int epfd
, int op
, int fd
, struct epoll_event
*epds
,
2063 struct eventpoll
*ep
;
2065 struct eventpoll
*tep
= NULL
;
2072 /* Get the "struct file *" for the target file */
2077 /* The target file descriptor must support poll */
2079 if (!file_can_poll(tf
.file
))
2080 goto error_tgt_fput
;
2082 /* Check if EPOLLWAKEUP is allowed */
2083 if (ep_op_has_event(op
))
2084 ep_take_care_of_epollwakeup(epds
);
2087 * We have to check that the file structure underneath the file descriptor
2088 * the user passed to us _is_ an eventpoll file. And also we do not permit
2089 * adding an epoll file descriptor inside itself.
2092 if (f
.file
== tf
.file
|| !is_file_epoll(f
.file
))
2093 goto error_tgt_fput
;
2096 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2097 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2098 * Also, we do not currently supported nested exclusive wakeups.
2100 if (ep_op_has_event(op
) && (epds
->events
& EPOLLEXCLUSIVE
)) {
2101 if (op
== EPOLL_CTL_MOD
)
2102 goto error_tgt_fput
;
2103 if (op
== EPOLL_CTL_ADD
&& (is_file_epoll(tf
.file
) ||
2104 (epds
->events
& ~EPOLLEXCLUSIVE_OK_BITS
)))
2105 goto error_tgt_fput
;
2109 * At this point it is safe to assume that the "private_data" contains
2110 * our own data structure.
2112 ep
= f
.file
->private_data
;
2115 * When we insert an epoll file descriptor inside another epoll file
2116 * descriptor, there is the chance of creating closed loops, which are
2117 * better be handled here, than in more critical paths. While we are
2118 * checking for loops we also determine the list of files reachable
2119 * and hang them on the tfile_check_list, so we can check that we
2120 * haven't created too many possible wakeup paths.
2122 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2123 * the epoll file descriptor is attaching directly to a wakeup source,
2124 * unless the epoll file descriptor is nested. The purpose of taking the
2125 * 'epmutex' on add is to prevent complex toplogies such as loops and
2126 * deep wakeup paths from forming in parallel through multiple
2127 * EPOLL_CTL_ADD operations.
2129 error
= epoll_mutex_lock(&ep
->mtx
, 0, nonblock
);
2131 goto error_tgt_fput
;
2132 if (op
== EPOLL_CTL_ADD
) {
2133 if (READ_ONCE(f
.file
->f_ep
) || ep
->gen
== loop_check_gen
||
2134 is_file_epoll(tf
.file
)) {
2135 mutex_unlock(&ep
->mtx
);
2136 error
= epoll_mutex_lock(&epmutex
, 0, nonblock
);
2138 goto error_tgt_fput
;
2141 if (is_file_epoll(tf
.file
)) {
2142 tep
= tf
.file
->private_data
;
2144 if (ep_loop_check(ep
, tep
) != 0)
2145 goto error_tgt_fput
;
2147 error
= epoll_mutex_lock(&ep
->mtx
, 0, nonblock
);
2149 goto error_tgt_fput
;
2154 * Try to lookup the file inside our RB tree. Since we grabbed "mtx"
2155 * above, we can be sure to be able to use the item looked up by
2156 * ep_find() till we release the mutex.
2158 epi
= ep_find(ep
, tf
.file
, fd
);
2164 epds
->events
|= EPOLLERR
| EPOLLHUP
;
2165 error
= ep_insert(ep
, epds
, tf
.file
, fd
, full_check
);
2171 error
= ep_remove(ep
, epi
);
2177 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
)) {
2178 epds
->events
|= EPOLLERR
| EPOLLHUP
;
2179 error
= ep_modify(ep
, epi
, epds
);
2185 mutex_unlock(&ep
->mtx
);
2189 clear_tfile_check_list();
2191 mutex_unlock(&epmutex
);
2203 * The following function implements the controller interface for
2204 * the eventpoll file that enables the insertion/removal/change of
2205 * file descriptors inside the interest set.
2207 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
2208 struct epoll_event __user
*, event
)
2210 struct epoll_event epds
;
2212 if (ep_op_has_event(op
) &&
2213 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
2216 return do_epoll_ctl(epfd
, op
, fd
, &epds
, false);
2220 * Implement the event wait interface for the eventpoll file. It is the kernel
2221 * part of the user space epoll_wait(2).
2223 static int do_epoll_wait(int epfd
, struct epoll_event __user
*events
,
2224 int maxevents
, struct timespec64
*to
)
2228 struct eventpoll
*ep
;
2230 /* The maximum number of event must be greater than zero */
2231 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
2234 /* Verify that the area passed by the user is writeable */
2235 if (!access_ok(events
, maxevents
* sizeof(struct epoll_event
)))
2238 /* Get the "struct file *" for the eventpoll file */
2244 * We have to check that the file structure underneath the fd
2245 * the user passed to us _is_ an eventpoll file.
2248 if (!is_file_epoll(f
.file
))
2252 * At this point it is safe to assume that the "private_data" contains
2253 * our own data structure.
2255 ep
= f
.file
->private_data
;
2257 /* Time to fish for events ... */
2258 error
= ep_poll(ep
, events
, maxevents
, to
);
2265 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
2266 int, maxevents
, int, timeout
)
2268 struct timespec64 to
;
2270 return do_epoll_wait(epfd
, events
, maxevents
,
2271 ep_timeout_to_timespec(&to
, timeout
));
2275 * Implement the event wait interface for the eventpoll file. It is the kernel
2276 * part of the user space epoll_pwait(2).
2278 static int do_epoll_pwait(int epfd
, struct epoll_event __user
*events
,
2279 int maxevents
, struct timespec64
*to
,
2280 const sigset_t __user
*sigmask
, size_t sigsetsize
)
2285 * If the caller wants a certain signal mask to be set during the wait,
2288 error
= set_user_sigmask(sigmask
, sigsetsize
);
2292 error
= do_epoll_wait(epfd
, events
, maxevents
, to
);
2294 restore_saved_sigmask_unless(error
== -EINTR
);
2299 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
2300 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
2303 struct timespec64 to
;
2305 return do_epoll_pwait(epfd
, events
, maxevents
,
2306 ep_timeout_to_timespec(&to
, timeout
),
2307 sigmask
, sigsetsize
);
2310 SYSCALL_DEFINE6(epoll_pwait2
, int, epfd
, struct epoll_event __user
*, events
,
2311 int, maxevents
, const struct __kernel_timespec __user
*, timeout
,
2312 const sigset_t __user
*, sigmask
, size_t, sigsetsize
)
2314 struct timespec64 ts
, *to
= NULL
;
2317 if (get_timespec64(&ts
, timeout
))
2320 if (poll_select_set_timeout(to
, ts
.tv_sec
, ts
.tv_nsec
))
2324 return do_epoll_pwait(epfd
, events
, maxevents
, to
,
2325 sigmask
, sigsetsize
);
2328 #ifdef CONFIG_COMPAT
2329 static int do_compat_epoll_pwait(int epfd
, struct epoll_event __user
*events
,
2330 int maxevents
, struct timespec64
*timeout
,
2331 const compat_sigset_t __user
*sigmask
,
2332 compat_size_t sigsetsize
)
2337 * If the caller wants a certain signal mask to be set during the wait,
2340 err
= set_compat_user_sigmask(sigmask
, sigsetsize
);
2344 err
= do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2346 restore_saved_sigmask_unless(err
== -EINTR
);
2351 COMPAT_SYSCALL_DEFINE6(epoll_pwait
, int, epfd
,
2352 struct epoll_event __user
*, events
,
2353 int, maxevents
, int, timeout
,
2354 const compat_sigset_t __user
*, sigmask
,
2355 compat_size_t
, sigsetsize
)
2357 struct timespec64 to
;
2359 return do_compat_epoll_pwait(epfd
, events
, maxevents
,
2360 ep_timeout_to_timespec(&to
, timeout
),
2361 sigmask
, sigsetsize
);
2364 COMPAT_SYSCALL_DEFINE6(epoll_pwait2
, int, epfd
,
2365 struct epoll_event __user
*, events
,
2367 const struct __kernel_timespec __user
*, timeout
,
2368 const compat_sigset_t __user
*, sigmask
,
2369 compat_size_t
, sigsetsize
)
2371 struct timespec64 ts
, *to
= NULL
;
2374 if (get_timespec64(&ts
, timeout
))
2377 if (poll_select_set_timeout(to
, ts
.tv_sec
, ts
.tv_nsec
))
2381 return do_compat_epoll_pwait(epfd
, events
, maxevents
, to
,
2382 sigmask
, sigsetsize
);
2387 static int __init
eventpoll_init(void)
2393 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2395 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2397 BUG_ON(max_user_watches
< 0);
2400 * We can have many thousands of epitems, so prevent this from
2401 * using an extra cache line on 64-bit (and smaller) CPUs
2403 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2405 /* Allocates slab cache used to allocate "struct epitem" items */
2406 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2407 0, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2409 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2410 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2411 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2412 epoll_sysctls_init();
2414 ephead_cache
= kmem_cache_create("ep_head",
2415 sizeof(struct epitems_head
), 0, SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2419 fs_initcall(eventpoll_init
);