2 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * Davide Libenzi <davidel@xmailserver.org>
14 #include <linux/init.h>
15 #include <linux/kernel.h>
16 #include <linux/sched/signal.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/device.h>
37 #include <linux/uaccess.h>
40 #include <linux/atomic.h>
41 #include <linux/proc_fs.h>
42 #include <linux/seq_file.h>
43 #include <linux/compat.h>
44 #include <linux/rculist.h>
45 #include <net/busy_poll.h>
49 * There are three level of locking required by epoll :
53 * 3) ep->lock (rwlock)
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a rwlock (ep->lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
63 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
67 * This mutex is acquired by ep_free() during the epoll file
68 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
78 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
87 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
90 * Events that require holding "epmutex" are very rare, while for
91 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
95 /* Epoll private bits inside the event mask */
96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
98 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
100 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
101 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
103 /* Maximum number of nesting allowed inside epoll sets */
104 #define EP_MAX_NESTS 4
106 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
108 #define EP_UNACTIVE_PTR ((void *) -1L)
110 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
112 struct epoll_filefd
{
118 * Structure used to track possible nested calls, for too deep recursions
121 struct nested_call_node
{
122 struct list_head llink
;
128 * This structure is used as collector for nested calls, to check for
129 * maximum recursion dept and loop cycles.
131 struct nested_calls
{
132 struct list_head tasks_call_list
;
137 * Each file descriptor added to the eventpoll interface will
138 * have an entry of this type linked to the "rbr" RB tree.
139 * Avoid increasing the size of this struct, there can be many thousands
140 * of these on a server and we do not want this to take another cache line.
144 /* RB tree node links this structure to the eventpoll RB tree */
146 /* Used to free the struct epitem */
150 /* List header used to link this structure to the eventpoll ready list */
151 struct list_head rdllink
;
154 * Works together "struct eventpoll"->ovflist in keeping the
155 * single linked chain of items.
159 /* The file descriptor information this item refers to */
160 struct epoll_filefd ffd
;
162 /* Number of active wait queue attached to poll operations */
165 /* List containing poll wait queues */
166 struct list_head pwqlist
;
168 /* The "container" of this item */
169 struct eventpoll
*ep
;
171 /* List header used to link this item to the "struct file" items list */
172 struct list_head fllink
;
174 /* wakeup_source used when EPOLLWAKEUP is set */
175 struct wakeup_source __rcu
*ws
;
177 /* The structure that describe the interested events and the source fd */
178 struct epoll_event event
;
182 * This structure is stored inside the "private_data" member of the file
183 * structure and represents the main data structure for the eventpoll
188 * This mutex is used to ensure that files are not removed
189 * while epoll is using them. This is held during the event
190 * collection loop, the file cleanup path, the epoll file exit
191 * code and the ctl operations.
195 /* Wait queue used by sys_epoll_wait() */
196 wait_queue_head_t wq
;
198 /* Wait queue used by file->poll() */
199 wait_queue_head_t poll_wait
;
201 /* List of ready file descriptors */
202 struct list_head rdllist
;
204 /* Lock which protects rdllist and ovflist */
207 /* RB tree root used to store monitored fd structs */
208 struct rb_root_cached rbr
;
211 * This is a single linked list that chains all the "struct epitem" that
212 * happened while transferring ready events to userspace w/out
215 struct epitem
*ovflist
;
217 /* wakeup_source used when ep_scan_ready_list is running */
218 struct wakeup_source
*ws
;
220 /* The user that created the eventpoll descriptor */
221 struct user_struct
*user
;
225 /* used to optimize loop detection check */
227 struct list_head visited_list_link
;
229 #ifdef CONFIG_NET_RX_BUSY_POLL
230 /* used to track busy poll napi_id */
231 unsigned int napi_id
;
235 /* Wait structure used by the poll hooks */
236 struct eppoll_entry
{
237 /* List header used to link this structure to the "struct epitem" */
238 struct list_head llink
;
240 /* The "base" pointer is set to the container "struct epitem" */
244 * Wait queue item that will be linked to the target file wait
247 wait_queue_entry_t wait
;
249 /* The wait queue head that linked the "wait" wait queue item */
250 wait_queue_head_t
*whead
;
253 /* Wrapper struct used by poll queueing */
259 /* Used by the ep_send_events() function as callback private data */
260 struct ep_send_events_data
{
262 struct epoll_event __user
*events
;
267 * Configuration options available inside /proc/sys/fs/epoll/
269 /* Maximum number of epoll watched descriptors, per user */
270 static long max_user_watches __read_mostly
;
273 * This mutex is used to serialize ep_free() and eventpoll_release_file().
275 static DEFINE_MUTEX(epmutex
);
277 /* Used to check for epoll file descriptor inclusion loops */
278 static struct nested_calls poll_loop_ncalls
;
280 /* Slab cache used to allocate "struct epitem" */
281 static struct kmem_cache
*epi_cache __read_mostly
;
283 /* Slab cache used to allocate "struct eppoll_entry" */
284 static struct kmem_cache
*pwq_cache __read_mostly
;
286 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
287 static LIST_HEAD(visited_list
);
290 * List of files with newly added links, where we may need to limit the number
291 * of emanating paths. Protected by the epmutex.
293 static LIST_HEAD(tfile_check_list
);
297 #include <linux/sysctl.h>
300 static long long_max
= LONG_MAX
;
302 struct ctl_table epoll_table
[] = {
304 .procname
= "max_user_watches",
305 .data
= &max_user_watches
,
306 .maxlen
= sizeof(max_user_watches
),
308 .proc_handler
= proc_doulongvec_minmax
,
314 #endif /* CONFIG_SYSCTL */
316 static const struct file_operations eventpoll_fops
;
318 static inline int is_file_epoll(struct file
*f
)
320 return f
->f_op
== &eventpoll_fops
;
323 /* Setup the structure that is used as key for the RB tree */
324 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
325 struct file
*file
, int fd
)
331 /* Compare RB tree keys */
332 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
333 struct epoll_filefd
*p2
)
335 return (p1
->file
> p2
->file
? +1:
336 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
339 /* Tells us if the item is currently linked */
340 static inline int ep_is_linked(struct epitem
*epi
)
342 return !list_empty(&epi
->rdllink
);
345 static inline struct eppoll_entry
*ep_pwq_from_wait(wait_queue_entry_t
*p
)
347 return container_of(p
, struct eppoll_entry
, wait
);
350 /* Get the "struct epitem" from a wait queue pointer */
351 static inline struct epitem
*ep_item_from_wait(wait_queue_entry_t
*p
)
353 return container_of(p
, struct eppoll_entry
, wait
)->base
;
356 /* Get the "struct epitem" from an epoll queue wrapper */
357 static inline struct epitem
*ep_item_from_epqueue(poll_table
*p
)
359 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
362 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
363 static inline int ep_op_has_event(int op
)
365 return op
!= EPOLL_CTL_DEL
;
368 /* Initialize the poll safe wake up structure */
369 static void ep_nested_calls_init(struct nested_calls
*ncalls
)
371 INIT_LIST_HEAD(&ncalls
->tasks_call_list
);
372 spin_lock_init(&ncalls
->lock
);
376 * ep_events_available - Checks if ready events might be available.
378 * @ep: Pointer to the eventpoll context.
380 * Returns: Returns a value different than zero if ready events are available,
383 static inline int ep_events_available(struct eventpoll
*ep
)
385 return !list_empty_careful(&ep
->rdllist
) ||
386 READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
;
389 #ifdef CONFIG_NET_RX_BUSY_POLL
390 static bool ep_busy_loop_end(void *p
, unsigned long start_time
)
392 struct eventpoll
*ep
= p
;
394 return ep_events_available(ep
) || busy_loop_timeout(start_time
);
398 * Busy poll if globally on and supporting sockets found && no events,
399 * busy loop will return if need_resched or ep_events_available.
401 * we must do our busy polling with irqs enabled
403 static void ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
405 unsigned int napi_id
= READ_ONCE(ep
->napi_id
);
407 if ((napi_id
>= MIN_NAPI_ID
) && net_busy_loop_on())
408 napi_busy_loop(napi_id
, nonblock
? NULL
: ep_busy_loop_end
, ep
);
411 static inline void ep_reset_busy_poll_napi_id(struct eventpoll
*ep
)
418 * Set epoll busy poll NAPI ID from sk.
420 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
422 struct eventpoll
*ep
;
423 unsigned int napi_id
;
428 if (!net_busy_loop_on())
431 sock
= sock_from_file(epi
->ffd
.file
, &err
);
439 napi_id
= READ_ONCE(sk
->sk_napi_id
);
442 /* Non-NAPI IDs can be rejected
444 * Nothing to do if we already have this ID
446 if (napi_id
< MIN_NAPI_ID
|| napi_id
== ep
->napi_id
)
449 /* record NAPI ID for use in next busy poll */
450 ep
->napi_id
= napi_id
;
455 static inline void ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
459 static inline void ep_reset_busy_poll_napi_id(struct eventpoll
*ep
)
463 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
467 #endif /* CONFIG_NET_RX_BUSY_POLL */
470 * ep_call_nested - Perform a bound (possibly) nested call, by checking
471 * that the recursion limit is not exceeded, and that
472 * the same nested call (by the meaning of same cookie) is
475 * @ncalls: Pointer to the nested_calls structure to be used for this call.
476 * @nproc: Nested call core function pointer.
477 * @priv: Opaque data to be passed to the @nproc callback.
478 * @cookie: Cookie to be used to identify this nested call.
479 * @ctx: This instance context.
481 * Returns: Returns the code returned by the @nproc callback, or -1 if
482 * the maximum recursion limit has been exceeded.
484 static int ep_call_nested(struct nested_calls
*ncalls
,
485 int (*nproc
)(void *, void *, int), void *priv
,
486 void *cookie
, void *ctx
)
488 int error
, call_nests
= 0;
490 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
491 struct nested_call_node
*tncur
;
492 struct nested_call_node tnode
;
494 spin_lock_irqsave(&ncalls
->lock
, flags
);
497 * Try to see if the current task is already inside this wakeup call.
498 * We use a list here, since the population inside this set is always
501 list_for_each_entry(tncur
, lsthead
, llink
) {
502 if (tncur
->ctx
== ctx
&&
503 (tncur
->cookie
== cookie
|| ++call_nests
> EP_MAX_NESTS
)) {
505 * Ops ... loop detected or maximum nest level reached.
506 * We abort this wake by breaking the cycle itself.
513 /* Add the current task and cookie to the list */
515 tnode
.cookie
= cookie
;
516 list_add(&tnode
.llink
, lsthead
);
518 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
520 /* Call the nested function */
521 error
= (*nproc
)(priv
, cookie
, call_nests
);
523 /* Remove the current task from the list */
524 spin_lock_irqsave(&ncalls
->lock
, flags
);
525 list_del(&tnode
.llink
);
527 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
533 * As described in commit 0ccf831cb lockdep: annotate epoll
534 * the use of wait queues used by epoll is done in a very controlled
535 * manner. Wake ups can nest inside each other, but are never done
536 * with the same locking. For example:
539 * efd1 = epoll_create();
540 * efd2 = epoll_create();
541 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
542 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
544 * When a packet arrives to the device underneath "dfd", the net code will
545 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
546 * callback wakeup entry on that queue, and the wake_up() performed by the
547 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
548 * (efd1) notices that it may have some event ready, so it needs to wake up
549 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
550 * that ends up in another wake_up(), after having checked about the
551 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
552 * avoid stack blasting.
554 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
555 * this special case of epoll.
557 #ifdef CONFIG_DEBUG_LOCK_ALLOC
559 static struct nested_calls poll_safewake_ncalls
;
561 static int ep_poll_wakeup_proc(void *priv
, void *cookie
, int call_nests
)
564 wait_queue_head_t
*wqueue
= (wait_queue_head_t
*)cookie
;
566 spin_lock_irqsave_nested(&wqueue
->lock
, flags
, call_nests
+ 1);
567 wake_up_locked_poll(wqueue
, EPOLLIN
);
568 spin_unlock_irqrestore(&wqueue
->lock
, flags
);
573 static void ep_poll_safewake(wait_queue_head_t
*wq
)
575 int this_cpu
= get_cpu();
577 ep_call_nested(&poll_safewake_ncalls
,
578 ep_poll_wakeup_proc
, NULL
, wq
, (void *) (long) this_cpu
);
585 static void ep_poll_safewake(wait_queue_head_t
*wq
)
587 wake_up_poll(wq
, EPOLLIN
);
592 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
594 wait_queue_head_t
*whead
;
598 * If it is cleared by POLLFREE, it should be rcu-safe.
599 * If we read NULL we need a barrier paired with
600 * smp_store_release() in ep_poll_callback(), otherwise
601 * we rely on whead->lock.
603 whead
= smp_load_acquire(&pwq
->whead
);
605 remove_wait_queue(whead
, &pwq
->wait
);
610 * This function unregisters poll callbacks from the associated file
611 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
614 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
616 struct list_head
*lsthead
= &epi
->pwqlist
;
617 struct eppoll_entry
*pwq
;
619 while (!list_empty(lsthead
)) {
620 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
622 list_del(&pwq
->llink
);
623 ep_remove_wait_queue(pwq
);
624 kmem_cache_free(pwq_cache
, pwq
);
628 /* call only when ep->mtx is held */
629 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
631 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
634 /* call only when ep->mtx is held */
635 static inline void ep_pm_stay_awake(struct epitem
*epi
)
637 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
643 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
645 return rcu_access_pointer(epi
->ws
) ? true : false;
648 /* call when ep->mtx cannot be held (ep_poll_callback) */
649 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
651 struct wakeup_source
*ws
;
654 ws
= rcu_dereference(epi
->ws
);
661 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
662 * the scan code, to call f_op->poll(). Also allows for
663 * O(NumReady) performance.
665 * @ep: Pointer to the epoll private data structure.
666 * @sproc: Pointer to the scan callback.
667 * @priv: Private opaque data passed to the @sproc callback.
668 * @depth: The current depth of recursive f_op->poll calls.
669 * @ep_locked: caller already holds ep->mtx
671 * Returns: The same integer error code returned by the @sproc callback.
673 static __poll_t
ep_scan_ready_list(struct eventpoll
*ep
,
674 __poll_t (*sproc
)(struct eventpoll
*,
675 struct list_head
*, void *),
676 void *priv
, int depth
, bool ep_locked
)
680 struct epitem
*epi
, *nepi
;
683 lockdep_assert_irqs_enabled();
686 * We need to lock this because we could be hit by
687 * eventpoll_release_file() and epoll_ctl().
691 mutex_lock_nested(&ep
->mtx
, depth
);
694 * Steal the ready list, and re-init the original one to the
695 * empty list. Also, set ep->ovflist to NULL so that events
696 * happening while looping w/out locks, are not lost. We cannot
697 * have the poll callback to queue directly on ep->rdllist,
698 * because we want the "sproc" callback to be able to do it
701 write_lock_irq(&ep
->lock
);
702 list_splice_init(&ep
->rdllist
, &txlist
);
703 WRITE_ONCE(ep
->ovflist
, NULL
);
704 write_unlock_irq(&ep
->lock
);
707 * Now call the callback function.
709 res
= (*sproc
)(ep
, &txlist
, priv
);
711 write_lock_irq(&ep
->lock
);
713 * During the time we spent inside the "sproc" callback, some
714 * other events might have been queued by the poll callback.
715 * We re-insert them inside the main ready-list here.
717 for (nepi
= READ_ONCE(ep
->ovflist
); (epi
= nepi
) != NULL
;
718 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
720 * We need to check if the item is already in the list.
721 * During the "sproc" callback execution time, items are
722 * queued into ->ovflist but the "txlist" might already
723 * contain them, and the list_splice() below takes care of them.
725 if (!ep_is_linked(epi
)) {
727 * ->ovflist is LIFO, so we have to reverse it in order
730 list_add(&epi
->rdllink
, &ep
->rdllist
);
731 ep_pm_stay_awake(epi
);
735 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
736 * releasing the lock, events will be queued in the normal way inside
739 WRITE_ONCE(ep
->ovflist
, EP_UNACTIVE_PTR
);
742 * Quickly re-inject items left on "txlist".
744 list_splice(&txlist
, &ep
->rdllist
);
747 if (!list_empty(&ep
->rdllist
)) {
749 * Wake up (if active) both the eventpoll wait list and
750 * the ->poll() wait list (delayed after we release the lock).
752 if (waitqueue_active(&ep
->wq
))
754 if (waitqueue_active(&ep
->poll_wait
))
757 write_unlock_irq(&ep
->lock
);
760 mutex_unlock(&ep
->mtx
);
762 /* We have to call this outside the lock */
764 ep_poll_safewake(&ep
->poll_wait
);
769 static void epi_rcu_free(struct rcu_head
*head
)
771 struct epitem
*epi
= container_of(head
, struct epitem
, rcu
);
772 kmem_cache_free(epi_cache
, epi
);
776 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
777 * all the associated resources. Must be called with "mtx" held.
779 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
781 struct file
*file
= epi
->ffd
.file
;
783 lockdep_assert_irqs_enabled();
786 * Removes poll wait queue hooks.
788 ep_unregister_pollwait(ep
, epi
);
790 /* Remove the current item from the list of epoll hooks */
791 spin_lock(&file
->f_lock
);
792 list_del_rcu(&epi
->fllink
);
793 spin_unlock(&file
->f_lock
);
795 rb_erase_cached(&epi
->rbn
, &ep
->rbr
);
797 write_lock_irq(&ep
->lock
);
798 if (ep_is_linked(epi
))
799 list_del_init(&epi
->rdllink
);
800 write_unlock_irq(&ep
->lock
);
802 wakeup_source_unregister(ep_wakeup_source(epi
));
804 * At this point it is safe to free the eventpoll item. Use the union
805 * field epi->rcu, since we are trying to minimize the size of
806 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
807 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
808 * use of the rbn field.
810 call_rcu(&epi
->rcu
, epi_rcu_free
);
812 atomic_long_dec(&ep
->user
->epoll_watches
);
817 static void ep_free(struct eventpoll
*ep
)
822 /* We need to release all tasks waiting for these file */
823 if (waitqueue_active(&ep
->poll_wait
))
824 ep_poll_safewake(&ep
->poll_wait
);
827 * We need to lock this because we could be hit by
828 * eventpoll_release_file() while we're freeing the "struct eventpoll".
829 * We do not need to hold "ep->mtx" here because the epoll file
830 * is on the way to be removed and no one has references to it
831 * anymore. The only hit might come from eventpoll_release_file() but
832 * holding "epmutex" is sufficient here.
834 mutex_lock(&epmutex
);
837 * Walks through the whole tree by unregistering poll callbacks.
839 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
840 epi
= rb_entry(rbp
, struct epitem
, rbn
);
842 ep_unregister_pollwait(ep
, epi
);
847 * Walks through the whole tree by freeing each "struct epitem". At this
848 * point we are sure no poll callbacks will be lingering around, and also by
849 * holding "epmutex" we can be sure that no file cleanup code will hit
850 * us during this operation. So we can avoid the lock on "ep->lock".
851 * We do not need to lock ep->mtx, either, we only do it to prevent
854 mutex_lock(&ep
->mtx
);
855 while ((rbp
= rb_first_cached(&ep
->rbr
)) != NULL
) {
856 epi
= rb_entry(rbp
, struct epitem
, rbn
);
860 mutex_unlock(&ep
->mtx
);
862 mutex_unlock(&epmutex
);
863 mutex_destroy(&ep
->mtx
);
865 wakeup_source_unregister(ep
->ws
);
869 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
871 struct eventpoll
*ep
= file
->private_data
;
879 static __poll_t
ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
881 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
885 * Differs from ep_eventpoll_poll() in that internal callers already have
886 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
887 * is correctly annotated.
889 static __poll_t
ep_item_poll(const struct epitem
*epi
, poll_table
*pt
,
892 struct eventpoll
*ep
;
895 pt
->_key
= epi
->event
.events
;
896 if (!is_file_epoll(epi
->ffd
.file
))
897 return vfs_poll(epi
->ffd
.file
, pt
) & epi
->event
.events
;
899 ep
= epi
->ffd
.file
->private_data
;
900 poll_wait(epi
->ffd
.file
, &ep
->poll_wait
, pt
);
901 locked
= pt
&& (pt
->_qproc
== ep_ptable_queue_proc
);
903 return ep_scan_ready_list(epi
->ffd
.file
->private_data
,
904 ep_read_events_proc
, &depth
, depth
,
905 locked
) & epi
->event
.events
;
908 static __poll_t
ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
911 struct epitem
*epi
, *tmp
;
913 int depth
= *(int *)priv
;
915 init_poll_funcptr(&pt
, NULL
);
918 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
919 if (ep_item_poll(epi
, &pt
, depth
)) {
920 return EPOLLIN
| EPOLLRDNORM
;
923 * Item has been dropped into the ready list by the poll
924 * callback, but it's not actually ready, as far as
925 * caller requested events goes. We can remove it here.
927 __pm_relax(ep_wakeup_source(epi
));
928 list_del_init(&epi
->rdllink
);
935 static __poll_t
ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
937 struct eventpoll
*ep
= file
->private_data
;
940 /* Insert inside our poll wait queue */
941 poll_wait(file
, &ep
->poll_wait
, wait
);
944 * Proceed to find out if wanted events are really available inside
947 return ep_scan_ready_list(ep
, ep_read_events_proc
,
948 &depth
, depth
, false);
951 #ifdef CONFIG_PROC_FS
952 static void ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
954 struct eventpoll
*ep
= f
->private_data
;
957 mutex_lock(&ep
->mtx
);
958 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
959 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
960 struct inode
*inode
= file_inode(epi
->ffd
.file
);
962 seq_printf(m
, "tfd: %8d events: %8x data: %16llx "
963 " pos:%lli ino:%lx sdev:%x\n",
964 epi
->ffd
.fd
, epi
->event
.events
,
965 (long long)epi
->event
.data
,
966 (long long)epi
->ffd
.file
->f_pos
,
967 inode
->i_ino
, inode
->i_sb
->s_dev
);
968 if (seq_has_overflowed(m
))
971 mutex_unlock(&ep
->mtx
);
975 /* File callbacks that implement the eventpoll file behaviour */
976 static const struct file_operations eventpoll_fops
= {
977 #ifdef CONFIG_PROC_FS
978 .show_fdinfo
= ep_show_fdinfo
,
980 .release
= ep_eventpoll_release
,
981 .poll
= ep_eventpoll_poll
,
982 .llseek
= noop_llseek
,
986 * This is called from eventpoll_release() to unlink files from the eventpoll
987 * interface. We need to have this facility to cleanup correctly files that are
988 * closed without being removed from the eventpoll interface.
990 void eventpoll_release_file(struct file
*file
)
992 struct eventpoll
*ep
;
993 struct epitem
*epi
, *next
;
996 * We don't want to get "file->f_lock" because it is not
997 * necessary. It is not necessary because we're in the "struct file"
998 * cleanup path, and this means that no one is using this file anymore.
999 * So, for example, epoll_ctl() cannot hit here since if we reach this
1000 * point, the file counter already went to zero and fget() would fail.
1001 * The only hit might come from ep_free() but by holding the mutex
1002 * will correctly serialize the operation. We do need to acquire
1003 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
1004 * from anywhere but ep_free().
1006 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1008 mutex_lock(&epmutex
);
1009 list_for_each_entry_safe(epi
, next
, &file
->f_ep_links
, fllink
) {
1011 mutex_lock_nested(&ep
->mtx
, 0);
1013 mutex_unlock(&ep
->mtx
);
1015 mutex_unlock(&epmutex
);
1018 static int ep_alloc(struct eventpoll
**pep
)
1021 struct user_struct
*user
;
1022 struct eventpoll
*ep
;
1024 user
= get_current_user();
1026 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
1030 mutex_init(&ep
->mtx
);
1031 rwlock_init(&ep
->lock
);
1032 init_waitqueue_head(&ep
->wq
);
1033 init_waitqueue_head(&ep
->poll_wait
);
1034 INIT_LIST_HEAD(&ep
->rdllist
);
1035 ep
->rbr
= RB_ROOT_CACHED
;
1036 ep
->ovflist
= EP_UNACTIVE_PTR
;
1049 * Search the file inside the eventpoll tree. The RB tree operations
1050 * are protected by the "mtx" mutex, and ep_find() must be called with
1053 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
1056 struct rb_node
*rbp
;
1057 struct epitem
*epi
, *epir
= NULL
;
1058 struct epoll_filefd ffd
;
1060 ep_set_ffd(&ffd
, file
, fd
);
1061 for (rbp
= ep
->rbr
.rb_root
.rb_node
; rbp
; ) {
1062 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1063 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
1065 rbp
= rbp
->rb_right
;
1077 #ifdef CONFIG_CHECKPOINT_RESTORE
1078 static struct epitem
*ep_find_tfd(struct eventpoll
*ep
, int tfd
, unsigned long toff
)
1080 struct rb_node
*rbp
;
1083 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1084 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1085 if (epi
->ffd
.fd
== tfd
) {
1097 struct file
*get_epoll_tfile_raw_ptr(struct file
*file
, int tfd
,
1100 struct file
*file_raw
;
1101 struct eventpoll
*ep
;
1104 if (!is_file_epoll(file
))
1105 return ERR_PTR(-EINVAL
);
1107 ep
= file
->private_data
;
1109 mutex_lock(&ep
->mtx
);
1110 epi
= ep_find_tfd(ep
, tfd
, toff
);
1112 file_raw
= epi
->ffd
.file
;
1114 file_raw
= ERR_PTR(-ENOENT
);
1115 mutex_unlock(&ep
->mtx
);
1119 #endif /* CONFIG_CHECKPOINT_RESTORE */
1122 * Adds a new entry to the tail of the list in a lockless way, i.e.
1123 * multiple CPUs are allowed to call this function concurrently.
1125 * Beware: it is necessary to prevent any other modifications of the
1126 * existing list until all changes are completed, in other words
1127 * concurrent list_add_tail_lockless() calls should be protected
1128 * with a read lock, where write lock acts as a barrier which
1129 * makes sure all list_add_tail_lockless() calls are fully
1132 * Also an element can be locklessly added to the list only in one
1133 * direction i.e. either to the tail either to the head, otherwise
1134 * concurrent access will corrupt the list.
1136 * Returns %false if element has been already added to the list, %true
1139 static inline bool list_add_tail_lockless(struct list_head
*new,
1140 struct list_head
*head
)
1142 struct list_head
*prev
;
1145 * This is simple 'new->next = head' operation, but cmpxchg()
1146 * is used in order to detect that same element has been just
1147 * added to the list from another CPU: the winner observes
1150 if (cmpxchg(&new->next
, new, head
) != new)
1154 * Initially ->next of a new element must be updated with the head
1155 * (we are inserting to the tail) and only then pointers are atomically
1156 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1157 * updated before pointers are actually swapped and pointers are
1158 * swapped before prev->next is updated.
1161 prev
= xchg(&head
->prev
, new);
1164 * It is safe to modify prev->next and new->prev, because a new element
1165 * is added only to the tail and new->next is updated before XCHG.
1175 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1176 * i.e. multiple CPUs are allowed to call this function concurrently.
1178 * Returns %false if epi element has been already chained, %true otherwise.
1180 static inline bool chain_epi_lockless(struct epitem
*epi
)
1182 struct eventpoll
*ep
= epi
->ep
;
1184 /* Check that the same epi has not been just chained from another CPU */
1185 if (cmpxchg(&epi
->next
, EP_UNACTIVE_PTR
, NULL
) != EP_UNACTIVE_PTR
)
1188 /* Atomically exchange tail */
1189 epi
->next
= xchg(&ep
->ovflist
, epi
);
1195 * This is the callback that is passed to the wait queue wakeup
1196 * mechanism. It is called by the stored file descriptors when they
1197 * have events to report.
1199 * This callback takes a read lock in order not to content with concurrent
1200 * events from another file descriptors, thus all modifications to ->rdllist
1201 * or ->ovflist are lockless. Read lock is paired with the write lock from
1202 * ep_scan_ready_list(), which stops all list modifications and guarantees
1203 * that lists state is seen correctly.
1205 * Another thing worth to mention is that ep_poll_callback() can be called
1206 * concurrently for the same @epi from different CPUs if poll table was inited
1207 * with several wait queues entries. Plural wakeup from different CPUs of a
1208 * single wait queue is serialized by wq.lock, but the case when multiple wait
1209 * queues are used should be detected accordingly. This is detected using
1210 * cmpxchg() operation.
1212 static int ep_poll_callback(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
1215 struct epitem
*epi
= ep_item_from_wait(wait
);
1216 struct eventpoll
*ep
= epi
->ep
;
1217 __poll_t pollflags
= key_to_poll(key
);
1218 unsigned long flags
;
1221 read_lock_irqsave(&ep
->lock
, flags
);
1223 ep_set_busy_poll_napi_id(epi
);
1226 * If the event mask does not contain any poll(2) event, we consider the
1227 * descriptor to be disabled. This condition is likely the effect of the
1228 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1229 * until the next EPOLL_CTL_MOD will be issued.
1231 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1235 * Check the events coming with the callback. At this stage, not
1236 * every device reports the events in the "key" parameter of the
1237 * callback. We need to be able to handle both cases here, hence the
1238 * test for "key" != NULL before the event match test.
1240 if (pollflags
&& !(pollflags
& epi
->event
.events
))
1244 * If we are transferring events to userspace, we can hold no locks
1245 * (because we're accessing user memory, and because of linux f_op->poll()
1246 * semantics). All the events that happen during that period of time are
1247 * chained in ep->ovflist and requeued later on.
1249 if (READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
) {
1250 if (epi
->next
== EP_UNACTIVE_PTR
&&
1251 chain_epi_lockless(epi
))
1252 ep_pm_stay_awake_rcu(epi
);
1256 /* If this file is already in the ready list we exit soon */
1257 if (!ep_is_linked(epi
) &&
1258 list_add_tail_lockless(&epi
->rdllink
, &ep
->rdllist
)) {
1259 ep_pm_stay_awake_rcu(epi
);
1263 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1266 if (waitqueue_active(&ep
->wq
)) {
1267 if ((epi
->event
.events
& EPOLLEXCLUSIVE
) &&
1268 !(pollflags
& POLLFREE
)) {
1269 switch (pollflags
& EPOLLINOUT_BITS
) {
1271 if (epi
->event
.events
& EPOLLIN
)
1275 if (epi
->event
.events
& EPOLLOUT
)
1285 if (waitqueue_active(&ep
->poll_wait
))
1289 read_unlock_irqrestore(&ep
->lock
, flags
);
1291 /* We have to call this outside the lock */
1293 ep_poll_safewake(&ep
->poll_wait
);
1295 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
))
1298 if (pollflags
& POLLFREE
) {
1300 * If we race with ep_remove_wait_queue() it can miss
1301 * ->whead = NULL and do another remove_wait_queue() after
1302 * us, so we can't use __remove_wait_queue().
1304 list_del_init(&wait
->entry
);
1306 * ->whead != NULL protects us from the race with ep_free()
1307 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1308 * held by the caller. Once we nullify it, nothing protects
1309 * ep/epi or even wait.
1311 smp_store_release(&ep_pwq_from_wait(wait
)->whead
, NULL
);
1318 * This is the callback that is used to add our wait queue to the
1319 * target file wakeup lists.
1321 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1324 struct epitem
*epi
= ep_item_from_epqueue(pt
);
1325 struct eppoll_entry
*pwq
;
1327 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
1328 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1331 if (epi
->event
.events
& EPOLLEXCLUSIVE
)
1332 add_wait_queue_exclusive(whead
, &pwq
->wait
);
1334 add_wait_queue(whead
, &pwq
->wait
);
1335 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
1338 /* We have to signal that an error occurred */
1343 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1346 struct rb_node
**p
= &ep
->rbr
.rb_root
.rb_node
, *parent
= NULL
;
1347 struct epitem
*epic
;
1348 bool leftmost
= true;
1352 epic
= rb_entry(parent
, struct epitem
, rbn
);
1353 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1355 p
= &parent
->rb_right
;
1358 p
= &parent
->rb_left
;
1360 rb_link_node(&epi
->rbn
, parent
, p
);
1361 rb_insert_color_cached(&epi
->rbn
, &ep
->rbr
, leftmost
);
1366 #define PATH_ARR_SIZE 5
1368 * These are the number paths of length 1 to 5, that we are allowing to emanate
1369 * from a single file of interest. For example, we allow 1000 paths of length
1370 * 1, to emanate from each file of interest. This essentially represents the
1371 * potential wakeup paths, which need to be limited in order to avoid massive
1372 * uncontrolled wakeup storms. The common use case should be a single ep which
1373 * is connected to n file sources. In this case each file source has 1 path
1374 * of length 1. Thus, the numbers below should be more than sufficient. These
1375 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1376 * and delete can't add additional paths. Protected by the epmutex.
1378 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1379 static int path_count
[PATH_ARR_SIZE
];
1381 static int path_count_inc(int nests
)
1383 /* Allow an arbitrary number of depth 1 paths */
1387 if (++path_count
[nests
] > path_limits
[nests
])
1392 static void path_count_init(void)
1396 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1400 static int reverse_path_check_proc(void *priv
, void *cookie
, int call_nests
)
1403 struct file
*file
= priv
;
1404 struct file
*child_file
;
1407 /* CTL_DEL can remove links here, but that can't increase our count */
1409 list_for_each_entry_rcu(epi
, &file
->f_ep_links
, fllink
) {
1410 child_file
= epi
->ep
->file
;
1411 if (is_file_epoll(child_file
)) {
1412 if (list_empty(&child_file
->f_ep_links
)) {
1413 if (path_count_inc(call_nests
)) {
1418 error
= ep_call_nested(&poll_loop_ncalls
,
1419 reverse_path_check_proc
,
1420 child_file
, child_file
,
1426 printk(KERN_ERR
"reverse_path_check_proc: "
1427 "file is not an ep!\n");
1435 * reverse_path_check - The tfile_check_list is list of file *, which have
1436 * links that are proposed to be newly added. We need to
1437 * make sure that those added links don't add too many
1438 * paths such that we will spend all our time waking up
1439 * eventpoll objects.
1441 * Returns: Returns zero if the proposed links don't create too many paths,
1444 static int reverse_path_check(void)
1447 struct file
*current_file
;
1449 /* let's call this for all tfiles */
1450 list_for_each_entry(current_file
, &tfile_check_list
, f_tfile_llink
) {
1452 error
= ep_call_nested(&poll_loop_ncalls
,
1453 reverse_path_check_proc
, current_file
,
1454 current_file
, current
);
1461 static int ep_create_wakeup_source(struct epitem
*epi
)
1464 struct wakeup_source
*ws
;
1467 epi
->ep
->ws
= wakeup_source_register("eventpoll");
1472 name
= epi
->ffd
.file
->f_path
.dentry
->d_name
.name
;
1473 ws
= wakeup_source_register(name
);
1477 rcu_assign_pointer(epi
->ws
, ws
);
1482 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1483 static noinline
void ep_destroy_wakeup_source(struct epitem
*epi
)
1485 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
1487 RCU_INIT_POINTER(epi
->ws
, NULL
);
1490 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1491 * used internally by wakeup_source_remove, too (called by
1492 * wakeup_source_unregister), so we cannot use call_rcu
1495 wakeup_source_unregister(ws
);
1499 * Must be called with "mtx" held.
1501 static int ep_insert(struct eventpoll
*ep
, const struct epoll_event
*event
,
1502 struct file
*tfile
, int fd
, int full_check
)
1504 int error
, pwake
= 0;
1508 struct ep_pqueue epq
;
1510 lockdep_assert_irqs_enabled();
1512 user_watches
= atomic_long_read(&ep
->user
->epoll_watches
);
1513 if (unlikely(user_watches
>= max_user_watches
))
1515 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
1518 /* Item initialization follow here ... */
1519 INIT_LIST_HEAD(&epi
->rdllink
);
1520 INIT_LIST_HEAD(&epi
->fllink
);
1521 INIT_LIST_HEAD(&epi
->pwqlist
);
1523 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
1524 epi
->event
= *event
;
1526 epi
->next
= EP_UNACTIVE_PTR
;
1527 if (epi
->event
.events
& EPOLLWAKEUP
) {
1528 error
= ep_create_wakeup_source(epi
);
1530 goto error_create_wakeup_source
;
1532 RCU_INIT_POINTER(epi
->ws
, NULL
);
1535 /* Initialize the poll table using the queue callback */
1537 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1540 * Attach the item to the poll hooks and get current event bits.
1541 * We can safely use the file* here because its usage count has
1542 * been increased by the caller of this function. Note that after
1543 * this operation completes, the poll callback can start hitting
1546 revents
= ep_item_poll(epi
, &epq
.pt
, 1);
1549 * We have to check if something went wrong during the poll wait queue
1550 * install process. Namely an allocation for a wait queue failed due
1551 * high memory pressure.
1555 goto error_unregister
;
1557 /* Add the current item to the list of active epoll hook for this file */
1558 spin_lock(&tfile
->f_lock
);
1559 list_add_tail_rcu(&epi
->fllink
, &tfile
->f_ep_links
);
1560 spin_unlock(&tfile
->f_lock
);
1563 * Add the current item to the RB tree. All RB tree operations are
1564 * protected by "mtx", and ep_insert() is called with "mtx" held.
1566 ep_rbtree_insert(ep
, epi
);
1568 /* now check if we've created too many backpaths */
1570 if (full_check
&& reverse_path_check())
1571 goto error_remove_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 atomic_long_inc(&ep
->user
->epoll_watches
);
1595 /* We have to call this outside the lock */
1597 ep_poll_safewake(&ep
->poll_wait
);
1602 spin_lock(&tfile
->f_lock
);
1603 list_del_rcu(&epi
->fllink
);
1604 spin_unlock(&tfile
->f_lock
);
1606 rb_erase_cached(&epi
->rbn
, &ep
->rbr
);
1609 ep_unregister_pollwait(ep
, epi
);
1612 * We need to do this because an event could have been arrived on some
1613 * allocated wait queue. Note that we don't care about the ep->ovflist
1614 * list, since that is used/cleaned only inside a section bound by "mtx".
1615 * And ep_insert() is called with "mtx" held.
1617 write_lock_irq(&ep
->lock
);
1618 if (ep_is_linked(epi
))
1619 list_del_init(&epi
->rdllink
);
1620 write_unlock_irq(&ep
->lock
);
1622 wakeup_source_unregister(ep_wakeup_source(epi
));
1624 error_create_wakeup_source
:
1625 kmem_cache_free(epi_cache
, epi
);
1631 * Modify the interest event mask by dropping an event if the new mask
1632 * has a match in the current file status. Must be called with "mtx" held.
1634 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
,
1635 const struct epoll_event
*event
)
1640 lockdep_assert_irqs_enabled();
1642 init_poll_funcptr(&pt
, NULL
);
1645 * Set the new event interest mask before calling f_op->poll();
1646 * otherwise we might miss an event that happens between the
1647 * f_op->poll() call and the new event set registering.
1649 epi
->event
.events
= event
->events
; /* need barrier below */
1650 epi
->event
.data
= event
->data
; /* protected by mtx */
1651 if (epi
->event
.events
& EPOLLWAKEUP
) {
1652 if (!ep_has_wakeup_source(epi
))
1653 ep_create_wakeup_source(epi
);
1654 } else if (ep_has_wakeup_source(epi
)) {
1655 ep_destroy_wakeup_source(epi
);
1659 * The following barrier has two effects:
1661 * 1) Flush epi changes above to other CPUs. This ensures
1662 * we do not miss events from ep_poll_callback if an
1663 * event occurs immediately after we call f_op->poll().
1664 * We need this because we did not take ep->lock while
1665 * changing epi above (but ep_poll_callback does take
1668 * 2) We also need to ensure we do not miss _past_ events
1669 * when calling f_op->poll(). This barrier also
1670 * pairs with the barrier in wq_has_sleeper (see
1671 * comments for wq_has_sleeper).
1673 * This barrier will now guarantee ep_poll_callback or f_op->poll
1674 * (or both) will notice the readiness of an item.
1679 * Get current event bits. We can safely use the file* here because
1680 * its usage count has been increased by the caller of this function.
1681 * If the item is "hot" and it is not registered inside the ready
1682 * list, push it inside.
1684 if (ep_item_poll(epi
, &pt
, 1)) {
1685 write_lock_irq(&ep
->lock
);
1686 if (!ep_is_linked(epi
)) {
1687 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1688 ep_pm_stay_awake(epi
);
1690 /* Notify waiting tasks that events are available */
1691 if (waitqueue_active(&ep
->wq
))
1693 if (waitqueue_active(&ep
->poll_wait
))
1696 write_unlock_irq(&ep
->lock
);
1699 /* We have to call this outside the lock */
1701 ep_poll_safewake(&ep
->poll_wait
);
1706 static __poll_t
ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1709 struct ep_send_events_data
*esed
= priv
;
1711 struct epitem
*epi
, *tmp
;
1712 struct epoll_event __user
*uevent
= esed
->events
;
1713 struct wakeup_source
*ws
;
1716 init_poll_funcptr(&pt
, NULL
);
1720 * We can loop without lock because we are passed a task private list.
1721 * Items cannot vanish during the loop because ep_scan_ready_list() is
1722 * holding "mtx" during this call.
1724 lockdep_assert_held(&ep
->mtx
);
1726 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
1727 if (esed
->res
>= esed
->maxevents
)
1731 * Activate ep->ws before deactivating epi->ws to prevent
1732 * triggering auto-suspend here (in case we reactive epi->ws
1735 * This could be rearranged to delay the deactivation of epi->ws
1736 * instead, but then epi->ws would temporarily be out of sync
1737 * with ep_is_linked().
1739 ws
= ep_wakeup_source(epi
);
1742 __pm_stay_awake(ep
->ws
);
1746 list_del_init(&epi
->rdllink
);
1749 * If the event mask intersect the caller-requested one,
1750 * deliver the event to userspace. Again, ep_scan_ready_list()
1751 * is holding ep->mtx, so no operations coming from userspace
1752 * can change the item.
1754 revents
= ep_item_poll(epi
, &pt
, 1);
1758 if (__put_user(revents
, &uevent
->events
) ||
1759 __put_user(epi
->event
.data
, &uevent
->data
)) {
1760 list_add(&epi
->rdllink
, head
);
1761 ep_pm_stay_awake(epi
);
1763 esed
->res
= -EFAULT
;
1768 if (epi
->event
.events
& EPOLLONESHOT
)
1769 epi
->event
.events
&= EP_PRIVATE_BITS
;
1770 else if (!(epi
->event
.events
& EPOLLET
)) {
1772 * If this file has been added with Level
1773 * Trigger mode, we need to insert back inside
1774 * the ready list, so that the next call to
1775 * epoll_wait() will check again the events
1776 * availability. At this point, no one can insert
1777 * into ep->rdllist besides us. The epoll_ctl()
1778 * callers are locked out by
1779 * ep_scan_ready_list() holding "mtx" and the
1780 * poll callback will queue them in ep->ovflist.
1782 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1783 ep_pm_stay_awake(epi
);
1790 static int ep_send_events(struct eventpoll
*ep
,
1791 struct epoll_event __user
*events
, int maxevents
)
1793 struct ep_send_events_data esed
;
1795 esed
.maxevents
= maxevents
;
1796 esed
.events
= events
;
1798 ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
, 0, false);
1802 static inline struct timespec64
ep_set_mstimeout(long ms
)
1804 struct timespec64 now
, ts
= {
1805 .tv_sec
= ms
/ MSEC_PER_SEC
,
1806 .tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
),
1809 ktime_get_ts64(&now
);
1810 return timespec64_add_safe(now
, ts
);
1814 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1817 * @ep: Pointer to the eventpoll context.
1818 * @events: Pointer to the userspace buffer where the ready events should be
1820 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1821 * @timeout: Maximum timeout for the ready events fetch operation, in
1822 * milliseconds. If the @timeout is zero, the function will not block,
1823 * while if the @timeout is less than zero, the function will block
1824 * until at least one event has been retrieved (or an error
1827 * Returns: Returns the number of ready events which have been fetched, or an
1828 * error code, in case of error.
1830 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1831 int maxevents
, long timeout
)
1833 int res
= 0, eavail
, timed_out
= 0;
1835 bool waiter
= false;
1836 wait_queue_entry_t wait
;
1837 ktime_t expires
, *to
= NULL
;
1839 lockdep_assert_irqs_enabled();
1842 struct timespec64 end_time
= ep_set_mstimeout(timeout
);
1844 slack
= select_estimate_accuracy(&end_time
);
1846 *to
= timespec64_to_ktime(end_time
);
1847 } else if (timeout
== 0) {
1849 * Avoid the unnecessary trip to the wait queue loop, if the
1850 * caller specified a non blocking operation. We still need
1851 * lock because we could race and not see an epi being added
1852 * to the ready list while in irq callback. Thus incorrectly
1853 * returning 0 back to userspace.
1857 write_lock_irq(&ep
->lock
);
1858 eavail
= ep_events_available(ep
);
1859 write_unlock_irq(&ep
->lock
);
1866 if (!ep_events_available(ep
))
1867 ep_busy_loop(ep
, timed_out
);
1869 eavail
= ep_events_available(ep
);
1874 * Busy poll timed out. Drop NAPI ID for now, we can add
1875 * it back in when we have moved a socket with a valid NAPI
1876 * ID onto the ready list.
1878 ep_reset_busy_poll_napi_id(ep
);
1881 * We don't have any available event to return to the caller. We need
1882 * to sleep here, and we will be woken by ep_poll_callback() when events
1887 init_waitqueue_entry(&wait
, current
);
1889 spin_lock_irq(&ep
->wq
.lock
);
1890 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1891 spin_unlock_irq(&ep
->wq
.lock
);
1896 * We don't want to sleep if the ep_poll_callback() sends us
1897 * a wakeup in between. That's why we set the task state
1898 * to TASK_INTERRUPTIBLE before doing the checks.
1900 set_current_state(TASK_INTERRUPTIBLE
);
1902 * Always short-circuit for fatal signals to allow
1903 * threads to make a timely exit without the chance of
1904 * finding more events available and fetching
1907 if (fatal_signal_pending(current
)) {
1912 eavail
= ep_events_available(ep
);
1915 if (signal_pending(current
)) {
1920 if (!schedule_hrtimeout_range(to
, slack
, HRTIMER_MODE_ABS
)) {
1926 __set_current_state(TASK_RUNNING
);
1930 * Try to transfer events to user space. In case we get 0 events and
1931 * there's still timeout left over, we go trying again in search of
1934 if (!res
&& eavail
&&
1935 !(res
= ep_send_events(ep
, events
, maxevents
)) && !timed_out
)
1939 spin_lock_irq(&ep
->wq
.lock
);
1940 __remove_wait_queue(&ep
->wq
, &wait
);
1941 spin_unlock_irq(&ep
->wq
.lock
);
1948 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1949 * API, to verify that adding an epoll file inside another
1950 * epoll structure, does not violate the constraints, in
1951 * terms of closed loops, or too deep chains (which can
1952 * result in excessive stack usage).
1954 * @priv: Pointer to the epoll file to be currently checked.
1955 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1956 * data structure pointer.
1957 * @call_nests: Current dept of the @ep_call_nested() call stack.
1959 * Returns: Returns zero if adding the epoll @file inside current epoll
1960 * structure @ep does not violate the constraints, or -1 otherwise.
1962 static int ep_loop_check_proc(void *priv
, void *cookie
, int call_nests
)
1965 struct file
*file
= priv
;
1966 struct eventpoll
*ep
= file
->private_data
;
1967 struct eventpoll
*ep_tovisit
;
1968 struct rb_node
*rbp
;
1971 mutex_lock_nested(&ep
->mtx
, call_nests
+ 1);
1973 list_add(&ep
->visited_list_link
, &visited_list
);
1974 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1975 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1976 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1977 ep_tovisit
= epi
->ffd
.file
->private_data
;
1978 if (ep_tovisit
->visited
)
1980 error
= ep_call_nested(&poll_loop_ncalls
,
1981 ep_loop_check_proc
, epi
->ffd
.file
,
1982 ep_tovisit
, current
);
1987 * If we've reached a file that is not associated with
1988 * an ep, then we need to check if the newly added
1989 * links are going to add too many wakeup paths. We do
1990 * this by adding it to the tfile_check_list, if it's
1991 * not already there, and calling reverse_path_check()
1992 * during ep_insert().
1994 if (list_empty(&epi
->ffd
.file
->f_tfile_llink
))
1995 list_add(&epi
->ffd
.file
->f_tfile_llink
,
1999 mutex_unlock(&ep
->mtx
);
2005 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2006 * another epoll file (represented by @ep) does not create
2007 * closed loops or too deep chains.
2009 * @ep: Pointer to the epoll private data structure.
2010 * @file: Pointer to the epoll file to be checked.
2012 * Returns: Returns zero if adding the epoll @file inside current epoll
2013 * structure @ep does not violate the constraints, or -1 otherwise.
2015 static int ep_loop_check(struct eventpoll
*ep
, struct file
*file
)
2018 struct eventpoll
*ep_cur
, *ep_next
;
2020 ret
= ep_call_nested(&poll_loop_ncalls
,
2021 ep_loop_check_proc
, file
, ep
, current
);
2022 /* clear visited list */
2023 list_for_each_entry_safe(ep_cur
, ep_next
, &visited_list
,
2024 visited_list_link
) {
2025 ep_cur
->visited
= 0;
2026 list_del(&ep_cur
->visited_list_link
);
2031 static void clear_tfile_check_list(void)
2035 /* first clear the tfile_check_list */
2036 while (!list_empty(&tfile_check_list
)) {
2037 file
= list_first_entry(&tfile_check_list
, struct file
,
2039 list_del_init(&file
->f_tfile_llink
);
2041 INIT_LIST_HEAD(&tfile_check_list
);
2045 * Open an eventpoll file descriptor.
2047 static int do_epoll_create(int flags
)
2050 struct eventpoll
*ep
= NULL
;
2053 /* Check the EPOLL_* constant for consistency. */
2054 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
2056 if (flags
& ~EPOLL_CLOEXEC
)
2059 * Create the internal data structure ("struct eventpoll").
2061 error
= ep_alloc(&ep
);
2065 * Creates all the items needed to setup an eventpoll file. That is,
2066 * a file structure and a free file descriptor.
2068 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
2073 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
2074 O_RDWR
| (flags
& O_CLOEXEC
));
2076 error
= PTR_ERR(file
);
2080 fd_install(fd
, file
);
2090 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
2092 return do_epoll_create(flags
);
2095 SYSCALL_DEFINE1(epoll_create
, int, size
)
2100 return do_epoll_create(0);
2104 * The following function implements the controller interface for
2105 * the eventpoll file that enables the insertion/removal/change of
2106 * file descriptors inside the interest set.
2108 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
2109 struct epoll_event __user
*, event
)
2114 struct eventpoll
*ep
;
2116 struct epoll_event epds
;
2117 struct eventpoll
*tep
= NULL
;
2120 if (ep_op_has_event(op
) &&
2121 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
2129 /* Get the "struct file *" for the target file */
2134 /* The target file descriptor must support poll */
2136 if (!file_can_poll(tf
.file
))
2137 goto error_tgt_fput
;
2139 /* Check if EPOLLWAKEUP is allowed */
2140 if (ep_op_has_event(op
))
2141 ep_take_care_of_epollwakeup(&epds
);
2144 * We have to check that the file structure underneath the file descriptor
2145 * the user passed to us _is_ an eventpoll file. And also we do not permit
2146 * adding an epoll file descriptor inside itself.
2149 if (f
.file
== tf
.file
|| !is_file_epoll(f
.file
))
2150 goto error_tgt_fput
;
2153 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2154 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2155 * Also, we do not currently supported nested exclusive wakeups.
2157 if (ep_op_has_event(op
) && (epds
.events
& EPOLLEXCLUSIVE
)) {
2158 if (op
== EPOLL_CTL_MOD
)
2159 goto error_tgt_fput
;
2160 if (op
== EPOLL_CTL_ADD
&& (is_file_epoll(tf
.file
) ||
2161 (epds
.events
& ~EPOLLEXCLUSIVE_OK_BITS
)))
2162 goto error_tgt_fput
;
2166 * At this point it is safe to assume that the "private_data" contains
2167 * our own data structure.
2169 ep
= f
.file
->private_data
;
2172 * When we insert an epoll file descriptor, inside another epoll file
2173 * descriptor, there is the change of creating closed loops, which are
2174 * better be handled here, than in more critical paths. While we are
2175 * checking for loops we also determine the list of files reachable
2176 * and hang them on the tfile_check_list, so we can check that we
2177 * haven't created too many possible wakeup paths.
2179 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2180 * the epoll file descriptor is attaching directly to a wakeup source,
2181 * unless the epoll file descriptor is nested. The purpose of taking the
2182 * 'epmutex' on add is to prevent complex toplogies such as loops and
2183 * deep wakeup paths from forming in parallel through multiple
2184 * EPOLL_CTL_ADD operations.
2186 mutex_lock_nested(&ep
->mtx
, 0);
2187 if (op
== EPOLL_CTL_ADD
) {
2188 if (!list_empty(&f
.file
->f_ep_links
) ||
2189 is_file_epoll(tf
.file
)) {
2191 mutex_unlock(&ep
->mtx
);
2192 mutex_lock(&epmutex
);
2193 if (is_file_epoll(tf
.file
)) {
2195 if (ep_loop_check(ep
, tf
.file
) != 0) {
2196 clear_tfile_check_list();
2197 goto error_tgt_fput
;
2200 list_add(&tf
.file
->f_tfile_llink
,
2202 mutex_lock_nested(&ep
->mtx
, 0);
2203 if (is_file_epoll(tf
.file
)) {
2204 tep
= tf
.file
->private_data
;
2205 mutex_lock_nested(&tep
->mtx
, 1);
2211 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2212 * above, we can be sure to be able to use the item looked up by
2213 * ep_find() till we release the mutex.
2215 epi
= ep_find(ep
, tf
.file
, fd
);
2221 epds
.events
|= EPOLLERR
| EPOLLHUP
;
2222 error
= ep_insert(ep
, &epds
, tf
.file
, fd
, full_check
);
2226 clear_tfile_check_list();
2230 error
= ep_remove(ep
, epi
);
2236 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
)) {
2237 epds
.events
|= EPOLLERR
| EPOLLHUP
;
2238 error
= ep_modify(ep
, epi
, &epds
);
2245 mutex_unlock(&tep
->mtx
);
2246 mutex_unlock(&ep
->mtx
);
2250 mutex_unlock(&epmutex
);
2261 * Implement the event wait interface for the eventpoll file. It is the kernel
2262 * part of the user space epoll_wait(2).
2264 static int do_epoll_wait(int epfd
, struct epoll_event __user
*events
,
2265 int maxevents
, int timeout
)
2269 struct eventpoll
*ep
;
2271 /* The maximum number of event must be greater than zero */
2272 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
2275 /* Verify that the area passed by the user is writeable */
2276 if (!access_ok(events
, maxevents
* sizeof(struct epoll_event
)))
2279 /* Get the "struct file *" for the eventpoll file */
2285 * We have to check that the file structure underneath the fd
2286 * the user passed to us _is_ an eventpoll file.
2289 if (!is_file_epoll(f
.file
))
2293 * At this point it is safe to assume that the "private_data" contains
2294 * our own data structure.
2296 ep
= f
.file
->private_data
;
2298 /* Time to fish for events ... */
2299 error
= ep_poll(ep
, events
, maxevents
, timeout
);
2306 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
2307 int, maxevents
, int, timeout
)
2309 return do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2313 * Implement the event wait interface for the eventpoll file. It is the kernel
2314 * part of the user space epoll_pwait(2).
2316 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
2317 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
2321 sigset_t ksigmask
, sigsaved
;
2324 * If the caller wants a certain signal mask to be set during the wait,
2327 error
= set_user_sigmask(sigmask
, &ksigmask
, &sigsaved
, sigsetsize
);
2331 error
= do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2333 restore_user_sigmask(sigmask
, &sigsaved
);
2338 #ifdef CONFIG_COMPAT
2339 COMPAT_SYSCALL_DEFINE6(epoll_pwait
, int, epfd
,
2340 struct epoll_event __user
*, events
,
2341 int, maxevents
, int, timeout
,
2342 const compat_sigset_t __user
*, sigmask
,
2343 compat_size_t
, sigsetsize
)
2346 sigset_t ksigmask
, sigsaved
;
2349 * If the caller wants a certain signal mask to be set during the wait,
2352 err
= set_compat_user_sigmask(sigmask
, &ksigmask
, &sigsaved
, sigsetsize
);
2356 err
= do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2358 restore_user_sigmask(sigmask
, &sigsaved
);
2364 static int __init
eventpoll_init(void)
2370 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2372 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2374 BUG_ON(max_user_watches
< 0);
2377 * Initialize the structure used to perform epoll file descriptor
2378 * inclusion loops checks.
2380 ep_nested_calls_init(&poll_loop_ncalls
);
2382 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2383 /* Initialize the structure used to perform safe poll wait head wake ups */
2384 ep_nested_calls_init(&poll_safewake_ncalls
);
2388 * We can have many thousands of epitems, so prevent this from
2389 * using an extra cache line on 64-bit (and smaller) CPUs
2391 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2393 /* Allocates slab cache used to allocate "struct epitem" items */
2394 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2395 0, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2397 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2398 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2399 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2403 fs_initcall(eventpoll_init
);