2 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * Davide Libenzi <davidel@xmailserver.org>
14 #include <linux/init.h>
15 #include <linux/kernel.h>
16 #include <linux/sched.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/device.h>
37 #include <asm/uaccess.h>
40 #include <linux/atomic.h>
41 #include <linux/proc_fs.h>
42 #include <linux/seq_file.h>
43 #include <linux/compat.h>
47 * There are three level of locking required by epoll :
51 * 3) ep->lock (spinlock)
53 * The acquire order is the one listed above, from 1 to 3.
54 * We need a spinlock (ep->lock) because we manipulate objects
55 * from inside the poll callback, that might be triggered from
56 * a wake_up() that in turn might be called from IRQ context.
57 * So we can't sleep inside the poll callback and hence we need
58 * a spinlock. During the event transfer loop (from kernel to
59 * user space) we could end up sleeping due a copy_to_user(), so
60 * we need a lock that will allow us to sleep. This lock is a
61 * mutex (ep->mtx). It is acquired during the event transfer loop,
62 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
63 * Then we also need a global mutex to serialize eventpoll_release_file()
65 * This mutex is acquired by ep_free() during the epoll file
66 * cleanup path and it is also acquired by eventpoll_release_file()
67 * if a file has been pushed inside an epoll set and it is then
68 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
69 * It is also acquired when inserting an epoll fd onto another epoll
70 * fd. We do this so that we walk the epoll tree and ensure that this
71 * insertion does not create a cycle of epoll file descriptors, which
72 * could lead to deadlock. We need a global mutex to prevent two
73 * simultaneous inserts (A into B and B into A) from racing and
74 * constructing a cycle without either insert observing that it is
76 * It is necessary to acquire multiple "ep->mtx"es at once in the
77 * case when one epoll fd is added to another. In this case, we
78 * always acquire the locks in the order of nesting (i.e. after
79 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
80 * before e2->mtx). Since we disallow cycles of epoll file
81 * descriptors, this ensures that the mutexes are well-ordered. In
82 * order to communicate this nesting to lockdep, when walking a tree
83 * of epoll file descriptors, we use the current recursion depth as
85 * It is possible to drop the "ep->mtx" and to use the global
86 * mutex "epmutex" (together with "ep->lock") to have it working,
87 * but having "ep->mtx" will make the interface more scalable.
88 * Events that require holding "epmutex" are very rare, while for
89 * normal operations the epoll private "ep->mtx" will guarantee
90 * a better scalability.
93 /* Epoll private bits inside the event mask */
94 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
96 /* Maximum number of nesting allowed inside epoll sets */
97 #define EP_MAX_NESTS 4
99 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
101 #define EP_UNACTIVE_PTR ((void *) -1L)
103 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
105 struct epoll_filefd
{
111 * Structure used to track possible nested calls, for too deep recursions
114 struct nested_call_node
{
115 struct list_head llink
;
121 * This structure is used as collector for nested calls, to check for
122 * maximum recursion dept and loop cycles.
124 struct nested_calls
{
125 struct list_head tasks_call_list
;
130 * Each file descriptor added to the eventpoll interface will
131 * have an entry of this type linked to the "rbr" RB tree.
132 * Avoid increasing the size of this struct, there can be many thousands
133 * of these on a server and we do not want this to take another cache line.
136 /* RB tree node used to link this structure to the eventpoll RB tree */
139 /* List header used to link this structure to the eventpoll ready list */
140 struct list_head rdllink
;
143 * Works together "struct eventpoll"->ovflist in keeping the
144 * single linked chain of items.
148 /* The file descriptor information this item refers to */
149 struct epoll_filefd ffd
;
151 /* Number of active wait queue attached to poll operations */
154 /* List containing poll wait queues */
155 struct list_head pwqlist
;
157 /* The "container" of this item */
158 struct eventpoll
*ep
;
160 /* List header used to link this item to the "struct file" items list */
161 struct list_head fllink
;
163 /* wakeup_source used when EPOLLWAKEUP is set */
164 struct wakeup_source __rcu
*ws
;
166 /* The structure that describe the interested events and the source fd */
167 struct epoll_event event
;
171 * This structure is stored inside the "private_data" member of the file
172 * structure and represents the main data structure for the eventpoll
176 /* Protect the access to this structure */
180 * This mutex is used to ensure that files are not removed
181 * while epoll is using them. This is held during the event
182 * collection loop, the file cleanup path, the epoll file exit
183 * code and the ctl operations.
187 /* Wait queue used by sys_epoll_wait() */
188 wait_queue_head_t wq
;
190 /* Wait queue used by file->poll() */
191 wait_queue_head_t poll_wait
;
193 /* List of ready file descriptors */
194 struct list_head rdllist
;
196 /* RB tree root used to store monitored fd structs */
200 * This is a single linked list that chains all the "struct epitem" that
201 * happened while transferring ready events to userspace w/out
204 struct epitem
*ovflist
;
206 /* wakeup_source used when ep_scan_ready_list is running */
207 struct wakeup_source
*ws
;
209 /* The user that created the eventpoll descriptor */
210 struct user_struct
*user
;
214 /* used to optimize loop detection check */
216 struct list_head visited_list_link
;
219 /* Wait structure used by the poll hooks */
220 struct eppoll_entry
{
221 /* List header used to link this structure to the "struct epitem" */
222 struct list_head llink
;
224 /* The "base" pointer is set to the container "struct epitem" */
228 * Wait queue item that will be linked to the target file wait
233 /* The wait queue head that linked the "wait" wait queue item */
234 wait_queue_head_t
*whead
;
237 /* Wrapper struct used by poll queueing */
243 /* Used by the ep_send_events() function as callback private data */
244 struct ep_send_events_data
{
246 struct epoll_event __user
*events
;
250 * Configuration options available inside /proc/sys/fs/epoll/
252 /* Maximum number of epoll watched descriptors, per user */
253 static long max_user_watches __read_mostly
;
256 * This mutex is used to serialize ep_free() and eventpoll_release_file().
258 static DEFINE_MUTEX(epmutex
);
260 /* Used to check for epoll file descriptor inclusion loops */
261 static struct nested_calls poll_loop_ncalls
;
263 /* Used for safe wake up implementation */
264 static struct nested_calls poll_safewake_ncalls
;
266 /* Used to call file's f_op->poll() under the nested calls boundaries */
267 static struct nested_calls poll_readywalk_ncalls
;
269 /* Slab cache used to allocate "struct epitem" */
270 static struct kmem_cache
*epi_cache __read_mostly
;
272 /* Slab cache used to allocate "struct eppoll_entry" */
273 static struct kmem_cache
*pwq_cache __read_mostly
;
275 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
276 static LIST_HEAD(visited_list
);
279 * List of files with newly added links, where we may need to limit the number
280 * of emanating paths. Protected by the epmutex.
282 static LIST_HEAD(tfile_check_list
);
286 #include <linux/sysctl.h>
289 static long long_max
= LONG_MAX
;
291 ctl_table epoll_table
[] = {
293 .procname
= "max_user_watches",
294 .data
= &max_user_watches
,
295 .maxlen
= sizeof(max_user_watches
),
297 .proc_handler
= proc_doulongvec_minmax
,
303 #endif /* CONFIG_SYSCTL */
305 static const struct file_operations eventpoll_fops
;
307 static inline int is_file_epoll(struct file
*f
)
309 return f
->f_op
== &eventpoll_fops
;
312 /* Setup the structure that is used as key for the RB tree */
313 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
314 struct file
*file
, int fd
)
320 /* Compare RB tree keys */
321 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
322 struct epoll_filefd
*p2
)
324 return (p1
->file
> p2
->file
? +1:
325 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
328 /* Tells us if the item is currently linked */
329 static inline int ep_is_linked(struct list_head
*p
)
331 return !list_empty(p
);
334 static inline struct eppoll_entry
*ep_pwq_from_wait(wait_queue_t
*p
)
336 return container_of(p
, struct eppoll_entry
, wait
);
339 /* Get the "struct epitem" from a wait queue pointer */
340 static inline struct epitem
*ep_item_from_wait(wait_queue_t
*p
)
342 return container_of(p
, struct eppoll_entry
, wait
)->base
;
345 /* Get the "struct epitem" from an epoll queue wrapper */
346 static inline struct epitem
*ep_item_from_epqueue(poll_table
*p
)
348 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
351 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
352 static inline int ep_op_has_event(int op
)
354 return op
!= EPOLL_CTL_DEL
;
357 /* Initialize the poll safe wake up structure */
358 static void ep_nested_calls_init(struct nested_calls
*ncalls
)
360 INIT_LIST_HEAD(&ncalls
->tasks_call_list
);
361 spin_lock_init(&ncalls
->lock
);
365 * ep_events_available - Checks if ready events might be available.
367 * @ep: Pointer to the eventpoll context.
369 * Returns: Returns a value different than zero if ready events are available,
372 static inline int ep_events_available(struct eventpoll
*ep
)
374 return !list_empty(&ep
->rdllist
) || ep
->ovflist
!= EP_UNACTIVE_PTR
;
378 * ep_call_nested - Perform a bound (possibly) nested call, by checking
379 * that the recursion limit is not exceeded, and that
380 * the same nested call (by the meaning of same cookie) is
383 * @ncalls: Pointer to the nested_calls structure to be used for this call.
384 * @max_nests: Maximum number of allowed nesting calls.
385 * @nproc: Nested call core function pointer.
386 * @priv: Opaque data to be passed to the @nproc callback.
387 * @cookie: Cookie to be used to identify this nested call.
388 * @ctx: This instance context.
390 * Returns: Returns the code returned by the @nproc callback, or -1 if
391 * the maximum recursion limit has been exceeded.
393 static int ep_call_nested(struct nested_calls
*ncalls
, int max_nests
,
394 int (*nproc
)(void *, void *, int), void *priv
,
395 void *cookie
, void *ctx
)
397 int error
, call_nests
= 0;
399 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
400 struct nested_call_node
*tncur
;
401 struct nested_call_node tnode
;
403 spin_lock_irqsave(&ncalls
->lock
, flags
);
406 * Try to see if the current task is already inside this wakeup call.
407 * We use a list here, since the population inside this set is always
410 list_for_each_entry(tncur
, lsthead
, llink
) {
411 if (tncur
->ctx
== ctx
&&
412 (tncur
->cookie
== cookie
|| ++call_nests
> max_nests
)) {
414 * Ops ... loop detected or maximum nest level reached.
415 * We abort this wake by breaking the cycle itself.
422 /* Add the current task and cookie to the list */
424 tnode
.cookie
= cookie
;
425 list_add(&tnode
.llink
, lsthead
);
427 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
429 /* Call the nested function */
430 error
= (*nproc
)(priv
, cookie
, call_nests
);
432 /* Remove the current task from the list */
433 spin_lock_irqsave(&ncalls
->lock
, flags
);
434 list_del(&tnode
.llink
);
436 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
442 * As described in commit 0ccf831cb lockdep: annotate epoll
443 * the use of wait queues used by epoll is done in a very controlled
444 * manner. Wake ups can nest inside each other, but are never done
445 * with the same locking. For example:
448 * efd1 = epoll_create();
449 * efd2 = epoll_create();
450 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
451 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
453 * When a packet arrives to the device underneath "dfd", the net code will
454 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
455 * callback wakeup entry on that queue, and the wake_up() performed by the
456 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
457 * (efd1) notices that it may have some event ready, so it needs to wake up
458 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
459 * that ends up in another wake_up(), after having checked about the
460 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
461 * avoid stack blasting.
463 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
464 * this special case of epoll.
466 #ifdef CONFIG_DEBUG_LOCK_ALLOC
467 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
468 unsigned long events
, int subclass
)
472 spin_lock_irqsave_nested(&wqueue
->lock
, flags
, subclass
);
473 wake_up_locked_poll(wqueue
, events
);
474 spin_unlock_irqrestore(&wqueue
->lock
, flags
);
477 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
478 unsigned long events
, int subclass
)
480 wake_up_poll(wqueue
, events
);
484 static int ep_poll_wakeup_proc(void *priv
, void *cookie
, int call_nests
)
486 ep_wake_up_nested((wait_queue_head_t
*) cookie
, POLLIN
,
492 * Perform a safe wake up of the poll wait list. The problem is that
493 * with the new callback'd wake up system, it is possible that the
494 * poll callback is reentered from inside the call to wake_up() done
495 * on the poll wait queue head. The rule is that we cannot reenter the
496 * wake up code from the same task more than EP_MAX_NESTS times,
497 * and we cannot reenter the same wait queue head at all. This will
498 * enable to have a hierarchy of epoll file descriptor of no more than
501 static void ep_poll_safewake(wait_queue_head_t
*wq
)
503 int this_cpu
= get_cpu();
505 ep_call_nested(&poll_safewake_ncalls
, EP_MAX_NESTS
,
506 ep_poll_wakeup_proc
, NULL
, wq
, (void *) (long) this_cpu
);
511 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
513 wait_queue_head_t
*whead
;
516 /* If it is cleared by POLLFREE, it should be rcu-safe */
517 whead
= rcu_dereference(pwq
->whead
);
519 remove_wait_queue(whead
, &pwq
->wait
);
524 * This function unregisters poll callbacks from the associated file
525 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
528 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
530 struct list_head
*lsthead
= &epi
->pwqlist
;
531 struct eppoll_entry
*pwq
;
533 while (!list_empty(lsthead
)) {
534 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
536 list_del(&pwq
->llink
);
537 ep_remove_wait_queue(pwq
);
538 kmem_cache_free(pwq_cache
, pwq
);
542 /* call only when ep->mtx is held */
543 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
545 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
548 /* call only when ep->mtx is held */
549 static inline void ep_pm_stay_awake(struct epitem
*epi
)
551 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
557 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
559 return rcu_access_pointer(epi
->ws
) ? true : false;
562 /* call when ep->mtx cannot be held (ep_poll_callback) */
563 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
565 struct wakeup_source
*ws
;
568 ws
= rcu_dereference(epi
->ws
);
575 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
576 * the scan code, to call f_op->poll(). Also allows for
577 * O(NumReady) performance.
579 * @ep: Pointer to the epoll private data structure.
580 * @sproc: Pointer to the scan callback.
581 * @priv: Private opaque data passed to the @sproc callback.
582 * @depth: The current depth of recursive f_op->poll calls.
584 * Returns: The same integer error code returned by the @sproc callback.
586 static int ep_scan_ready_list(struct eventpoll
*ep
,
587 int (*sproc
)(struct eventpoll
*,
588 struct list_head
*, void *),
592 int error
, pwake
= 0;
594 struct epitem
*epi
, *nepi
;
598 * We need to lock this because we could be hit by
599 * eventpoll_release_file() and epoll_ctl().
601 mutex_lock_nested(&ep
->mtx
, depth
);
604 * Steal the ready list, and re-init the original one to the
605 * empty list. Also, set ep->ovflist to NULL so that events
606 * happening while looping w/out locks, are not lost. We cannot
607 * have the poll callback to queue directly on ep->rdllist,
608 * because we want the "sproc" callback to be able to do it
611 spin_lock_irqsave(&ep
->lock
, flags
);
612 list_splice_init(&ep
->rdllist
, &txlist
);
614 spin_unlock_irqrestore(&ep
->lock
, flags
);
617 * Now call the callback function.
619 error
= (*sproc
)(ep
, &txlist
, priv
);
621 spin_lock_irqsave(&ep
->lock
, flags
);
623 * During the time we spent inside the "sproc" callback, some
624 * other events might have been queued by the poll callback.
625 * We re-insert them inside the main ready-list here.
627 for (nepi
= ep
->ovflist
; (epi
= nepi
) != NULL
;
628 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
630 * We need to check if the item is already in the list.
631 * During the "sproc" callback execution time, items are
632 * queued into ->ovflist but the "txlist" might already
633 * contain them, and the list_splice() below takes care of them.
635 if (!ep_is_linked(&epi
->rdllink
)) {
636 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
637 ep_pm_stay_awake(epi
);
641 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
642 * releasing the lock, events will be queued in the normal way inside
645 ep
->ovflist
= EP_UNACTIVE_PTR
;
648 * Quickly re-inject items left on "txlist".
650 list_splice(&txlist
, &ep
->rdllist
);
653 if (!list_empty(&ep
->rdllist
)) {
655 * Wake up (if active) both the eventpoll wait list and
656 * the ->poll() wait list (delayed after we release the lock).
658 if (waitqueue_active(&ep
->wq
))
659 wake_up_locked(&ep
->wq
);
660 if (waitqueue_active(&ep
->poll_wait
))
663 spin_unlock_irqrestore(&ep
->lock
, flags
);
665 mutex_unlock(&ep
->mtx
);
667 /* We have to call this outside the lock */
669 ep_poll_safewake(&ep
->poll_wait
);
675 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
676 * all the associated resources. Must be called with "mtx" held.
678 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
681 struct file
*file
= epi
->ffd
.file
;
684 * Removes poll wait queue hooks. We _have_ to do this without holding
685 * the "ep->lock" otherwise a deadlock might occur. This because of the
686 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
687 * queue head lock when unregistering the wait queue. The wakeup callback
688 * will run by holding the wait queue head lock and will call our callback
689 * that will try to get "ep->lock".
691 ep_unregister_pollwait(ep
, epi
);
693 /* Remove the current item from the list of epoll hooks */
694 spin_lock(&file
->f_lock
);
695 if (ep_is_linked(&epi
->fllink
))
696 list_del_init(&epi
->fllink
);
697 spin_unlock(&file
->f_lock
);
699 rb_erase(&epi
->rbn
, &ep
->rbr
);
701 spin_lock_irqsave(&ep
->lock
, flags
);
702 if (ep_is_linked(&epi
->rdllink
))
703 list_del_init(&epi
->rdllink
);
704 spin_unlock_irqrestore(&ep
->lock
, flags
);
706 wakeup_source_unregister(ep_wakeup_source(epi
));
708 /* At this point it is safe to free the eventpoll item */
709 kmem_cache_free(epi_cache
, epi
);
711 atomic_long_dec(&ep
->user
->epoll_watches
);
716 static void ep_free(struct eventpoll
*ep
)
721 /* We need to release all tasks waiting for these file */
722 if (waitqueue_active(&ep
->poll_wait
))
723 ep_poll_safewake(&ep
->poll_wait
);
726 * We need to lock this because we could be hit by
727 * eventpoll_release_file() while we're freeing the "struct eventpoll".
728 * We do not need to hold "ep->mtx" here because the epoll file
729 * is on the way to be removed and no one has references to it
730 * anymore. The only hit might come from eventpoll_release_file() but
731 * holding "epmutex" is sufficient here.
733 mutex_lock(&epmutex
);
736 * Walks through the whole tree by unregistering poll callbacks.
738 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
739 epi
= rb_entry(rbp
, struct epitem
, rbn
);
741 ep_unregister_pollwait(ep
, epi
);
746 * Walks through the whole tree by freeing each "struct epitem". At this
747 * point we are sure no poll callbacks will be lingering around, and also by
748 * holding "epmutex" we can be sure that no file cleanup code will hit
749 * us during this operation. So we can avoid the lock on "ep->lock".
750 * We do not need to lock ep->mtx, either, we only do it to prevent
753 mutex_lock(&ep
->mtx
);
754 while ((rbp
= rb_first(&ep
->rbr
)) != NULL
) {
755 epi
= rb_entry(rbp
, struct epitem
, rbn
);
759 mutex_unlock(&ep
->mtx
);
761 mutex_unlock(&epmutex
);
762 mutex_destroy(&ep
->mtx
);
764 wakeup_source_unregister(ep
->ws
);
768 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
770 struct eventpoll
*ep
= file
->private_data
;
778 static inline unsigned int ep_item_poll(struct epitem
*epi
, poll_table
*pt
)
780 pt
->_key
= epi
->event
.events
;
782 return epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, pt
) & epi
->event
.events
;
785 static int ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
788 struct epitem
*epi
, *tmp
;
791 init_poll_funcptr(&pt
, NULL
);
793 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
794 if (ep_item_poll(epi
, &pt
))
795 return POLLIN
| POLLRDNORM
;
798 * Item has been dropped into the ready list by the poll
799 * callback, but it's not actually ready, as far as
800 * caller requested events goes. We can remove it here.
802 __pm_relax(ep_wakeup_source(epi
));
803 list_del_init(&epi
->rdllink
);
810 static int ep_poll_readyevents_proc(void *priv
, void *cookie
, int call_nests
)
812 return ep_scan_ready_list(priv
, ep_read_events_proc
, NULL
, call_nests
+ 1);
815 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
818 struct eventpoll
*ep
= file
->private_data
;
820 /* Insert inside our poll wait queue */
821 poll_wait(file
, &ep
->poll_wait
, wait
);
824 * Proceed to find out if wanted events are really available inside
825 * the ready list. This need to be done under ep_call_nested()
826 * supervision, since the call to f_op->poll() done on listed files
827 * could re-enter here.
829 pollflags
= ep_call_nested(&poll_readywalk_ncalls
, EP_MAX_NESTS
,
830 ep_poll_readyevents_proc
, ep
, ep
, current
);
832 return pollflags
!= -1 ? pollflags
: 0;
835 #ifdef CONFIG_PROC_FS
836 static int ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
838 struct eventpoll
*ep
= f
->private_data
;
842 mutex_lock(&ep
->mtx
);
843 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
844 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
846 ret
= seq_printf(m
, "tfd: %8d events: %8x data: %16llx\n",
847 epi
->ffd
.fd
, epi
->event
.events
,
848 (long long)epi
->event
.data
);
852 mutex_unlock(&ep
->mtx
);
858 /* File callbacks that implement the eventpoll file behaviour */
859 static const struct file_operations eventpoll_fops
= {
860 #ifdef CONFIG_PROC_FS
861 .show_fdinfo
= ep_show_fdinfo
,
863 .release
= ep_eventpoll_release
,
864 .poll
= ep_eventpoll_poll
,
865 .llseek
= noop_llseek
,
869 * This is called from eventpoll_release() to unlink files from the eventpoll
870 * interface. We need to have this facility to cleanup correctly files that are
871 * closed without being removed from the eventpoll interface.
873 void eventpoll_release_file(struct file
*file
)
875 struct list_head
*lsthead
= &file
->f_ep_links
;
876 struct eventpoll
*ep
;
880 * We don't want to get "file->f_lock" because it is not
881 * necessary. It is not necessary because we're in the "struct file"
882 * cleanup path, and this means that no one is using this file anymore.
883 * So, for example, epoll_ctl() cannot hit here since if we reach this
884 * point, the file counter already went to zero and fget() would fail.
885 * The only hit might come from ep_free() but by holding the mutex
886 * will correctly serialize the operation. We do need to acquire
887 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
888 * from anywhere but ep_free().
890 * Besides, ep_remove() acquires the lock, so we can't hold it here.
892 mutex_lock(&epmutex
);
894 while (!list_empty(lsthead
)) {
895 epi
= list_first_entry(lsthead
, struct epitem
, fllink
);
898 list_del_init(&epi
->fllink
);
899 mutex_lock_nested(&ep
->mtx
, 0);
901 mutex_unlock(&ep
->mtx
);
904 mutex_unlock(&epmutex
);
907 static int ep_alloc(struct eventpoll
**pep
)
910 struct user_struct
*user
;
911 struct eventpoll
*ep
;
913 user
= get_current_user();
915 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
919 spin_lock_init(&ep
->lock
);
920 mutex_init(&ep
->mtx
);
921 init_waitqueue_head(&ep
->wq
);
922 init_waitqueue_head(&ep
->poll_wait
);
923 INIT_LIST_HEAD(&ep
->rdllist
);
925 ep
->ovflist
= EP_UNACTIVE_PTR
;
938 * Search the file inside the eventpoll tree. The RB tree operations
939 * are protected by the "mtx" mutex, and ep_find() must be called with
942 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
946 struct epitem
*epi
, *epir
= NULL
;
947 struct epoll_filefd ffd
;
949 ep_set_ffd(&ffd
, file
, fd
);
950 for (rbp
= ep
->rbr
.rb_node
; rbp
; ) {
951 epi
= rb_entry(rbp
, struct epitem
, rbn
);
952 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
967 * This is the callback that is passed to the wait queue wakeup
968 * mechanism. It is called by the stored file descriptors when they
969 * have events to report.
971 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
975 struct epitem
*epi
= ep_item_from_wait(wait
);
976 struct eventpoll
*ep
= epi
->ep
;
978 if ((unsigned long)key
& POLLFREE
) {
979 ep_pwq_from_wait(wait
)->whead
= NULL
;
981 * whead = NULL above can race with ep_remove_wait_queue()
982 * which can do another remove_wait_queue() after us, so we
983 * can't use __remove_wait_queue(). whead->lock is held by
986 list_del_init(&wait
->task_list
);
989 spin_lock_irqsave(&ep
->lock
, flags
);
992 * If the event mask does not contain any poll(2) event, we consider the
993 * descriptor to be disabled. This condition is likely the effect of the
994 * EPOLLONESHOT bit that disables the descriptor when an event is received,
995 * until the next EPOLL_CTL_MOD will be issued.
997 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1001 * Check the events coming with the callback. At this stage, not
1002 * every device reports the events in the "key" parameter of the
1003 * callback. We need to be able to handle both cases here, hence the
1004 * test for "key" != NULL before the event match test.
1006 if (key
&& !((unsigned long) key
& epi
->event
.events
))
1010 * If we are transferring events to userspace, we can hold no locks
1011 * (because we're accessing user memory, and because of linux f_op->poll()
1012 * semantics). All the events that happen during that period of time are
1013 * chained in ep->ovflist and requeued later on.
1015 if (unlikely(ep
->ovflist
!= EP_UNACTIVE_PTR
)) {
1016 if (epi
->next
== EP_UNACTIVE_PTR
) {
1017 epi
->next
= ep
->ovflist
;
1021 * Activate ep->ws since epi->ws may get
1022 * deactivated at any time.
1024 __pm_stay_awake(ep
->ws
);
1031 /* If this file is already in the ready list we exit soon */
1032 if (!ep_is_linked(&epi
->rdllink
)) {
1033 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1034 ep_pm_stay_awake_rcu(epi
);
1038 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1041 if (waitqueue_active(&ep
->wq
))
1042 wake_up_locked(&ep
->wq
);
1043 if (waitqueue_active(&ep
->poll_wait
))
1047 spin_unlock_irqrestore(&ep
->lock
, flags
);
1049 /* We have to call this outside the lock */
1051 ep_poll_safewake(&ep
->poll_wait
);
1057 * This is the callback that is used to add our wait queue to the
1058 * target file wakeup lists.
1060 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1063 struct epitem
*epi
= ep_item_from_epqueue(pt
);
1064 struct eppoll_entry
*pwq
;
1066 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
1067 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1070 add_wait_queue(whead
, &pwq
->wait
);
1071 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
1074 /* We have to signal that an error occurred */
1079 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1082 struct rb_node
**p
= &ep
->rbr
.rb_node
, *parent
= NULL
;
1083 struct epitem
*epic
;
1087 epic
= rb_entry(parent
, struct epitem
, rbn
);
1088 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1090 p
= &parent
->rb_right
;
1092 p
= &parent
->rb_left
;
1094 rb_link_node(&epi
->rbn
, parent
, p
);
1095 rb_insert_color(&epi
->rbn
, &ep
->rbr
);
1100 #define PATH_ARR_SIZE 5
1102 * These are the number paths of length 1 to 5, that we are allowing to emanate
1103 * from a single file of interest. For example, we allow 1000 paths of length
1104 * 1, to emanate from each file of interest. This essentially represents the
1105 * potential wakeup paths, which need to be limited in order to avoid massive
1106 * uncontrolled wakeup storms. The common use case should be a single ep which
1107 * is connected to n file sources. In this case each file source has 1 path
1108 * of length 1. Thus, the numbers below should be more than sufficient. These
1109 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1110 * and delete can't add additional paths. Protected by the epmutex.
1112 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1113 static int path_count
[PATH_ARR_SIZE
];
1115 static int path_count_inc(int nests
)
1117 /* Allow an arbitrary number of depth 1 paths */
1121 if (++path_count
[nests
] > path_limits
[nests
])
1126 static void path_count_init(void)
1130 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1134 static int reverse_path_check_proc(void *priv
, void *cookie
, int call_nests
)
1137 struct file
*file
= priv
;
1138 struct file
*child_file
;
1141 list_for_each_entry(epi
, &file
->f_ep_links
, fllink
) {
1142 child_file
= epi
->ep
->file
;
1143 if (is_file_epoll(child_file
)) {
1144 if (list_empty(&child_file
->f_ep_links
)) {
1145 if (path_count_inc(call_nests
)) {
1150 error
= ep_call_nested(&poll_loop_ncalls
,
1152 reverse_path_check_proc
,
1153 child_file
, child_file
,
1159 printk(KERN_ERR
"reverse_path_check_proc: "
1160 "file is not an ep!\n");
1167 * reverse_path_check - The tfile_check_list is list of file *, which have
1168 * links that are proposed to be newly added. We need to
1169 * make sure that those added links don't add too many
1170 * paths such that we will spend all our time waking up
1171 * eventpoll objects.
1173 * Returns: Returns zero if the proposed links don't create too many paths,
1176 static int reverse_path_check(void)
1179 struct file
*current_file
;
1181 /* let's call this for all tfiles */
1182 list_for_each_entry(current_file
, &tfile_check_list
, f_tfile_llink
) {
1184 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1185 reverse_path_check_proc
, current_file
,
1186 current_file
, current
);
1193 static int ep_create_wakeup_source(struct epitem
*epi
)
1196 struct wakeup_source
*ws
;
1199 epi
->ep
->ws
= wakeup_source_register("eventpoll");
1204 name
= epi
->ffd
.file
->f_path
.dentry
->d_name
.name
;
1205 ws
= wakeup_source_register(name
);
1209 rcu_assign_pointer(epi
->ws
, ws
);
1214 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1215 static noinline
void ep_destroy_wakeup_source(struct epitem
*epi
)
1217 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
1219 RCU_INIT_POINTER(epi
->ws
, NULL
);
1222 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1223 * used internally by wakeup_source_remove, too (called by
1224 * wakeup_source_unregister), so we cannot use call_rcu
1227 wakeup_source_unregister(ws
);
1231 * Must be called with "mtx" held.
1233 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
1234 struct file
*tfile
, int fd
)
1236 int error
, revents
, pwake
= 0;
1237 unsigned long flags
;
1240 struct ep_pqueue epq
;
1242 user_watches
= atomic_long_read(&ep
->user
->epoll_watches
);
1243 if (unlikely(user_watches
>= max_user_watches
))
1245 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
1248 /* Item initialization follow here ... */
1249 INIT_LIST_HEAD(&epi
->rdllink
);
1250 INIT_LIST_HEAD(&epi
->fllink
);
1251 INIT_LIST_HEAD(&epi
->pwqlist
);
1253 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
1254 epi
->event
= *event
;
1256 epi
->next
= EP_UNACTIVE_PTR
;
1257 if (epi
->event
.events
& EPOLLWAKEUP
) {
1258 error
= ep_create_wakeup_source(epi
);
1260 goto error_create_wakeup_source
;
1262 RCU_INIT_POINTER(epi
->ws
, NULL
);
1265 /* Initialize the poll table using the queue callback */
1267 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1270 * Attach the item to the poll hooks and get current event bits.
1271 * We can safely use the file* here because its usage count has
1272 * been increased by the caller of this function. Note that after
1273 * this operation completes, the poll callback can start hitting
1276 revents
= ep_item_poll(epi
, &epq
.pt
);
1279 * We have to check if something went wrong during the poll wait queue
1280 * install process. Namely an allocation for a wait queue failed due
1281 * high memory pressure.
1285 goto error_unregister
;
1287 /* Add the current item to the list of active epoll hook for this file */
1288 spin_lock(&tfile
->f_lock
);
1289 list_add_tail(&epi
->fllink
, &tfile
->f_ep_links
);
1290 spin_unlock(&tfile
->f_lock
);
1293 * Add the current item to the RB tree. All RB tree operations are
1294 * protected by "mtx", and ep_insert() is called with "mtx" held.
1296 ep_rbtree_insert(ep
, epi
);
1298 /* now check if we've created too many backpaths */
1300 if (reverse_path_check())
1301 goto error_remove_epi
;
1303 /* We have to drop the new item inside our item list to keep track of it */
1304 spin_lock_irqsave(&ep
->lock
, flags
);
1306 /* If the file is already "ready" we drop it inside the ready list */
1307 if ((revents
& event
->events
) && !ep_is_linked(&epi
->rdllink
)) {
1308 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1309 ep_pm_stay_awake(epi
);
1311 /* Notify waiting tasks that events are available */
1312 if (waitqueue_active(&ep
->wq
))
1313 wake_up_locked(&ep
->wq
);
1314 if (waitqueue_active(&ep
->poll_wait
))
1318 spin_unlock_irqrestore(&ep
->lock
, flags
);
1320 atomic_long_inc(&ep
->user
->epoll_watches
);
1322 /* We have to call this outside the lock */
1324 ep_poll_safewake(&ep
->poll_wait
);
1329 spin_lock(&tfile
->f_lock
);
1330 if (ep_is_linked(&epi
->fllink
))
1331 list_del_init(&epi
->fllink
);
1332 spin_unlock(&tfile
->f_lock
);
1334 rb_erase(&epi
->rbn
, &ep
->rbr
);
1337 ep_unregister_pollwait(ep
, epi
);
1340 * We need to do this because an event could have been arrived on some
1341 * allocated wait queue. Note that we don't care about the ep->ovflist
1342 * list, since that is used/cleaned only inside a section bound by "mtx".
1343 * And ep_insert() is called with "mtx" held.
1345 spin_lock_irqsave(&ep
->lock
, flags
);
1346 if (ep_is_linked(&epi
->rdllink
))
1347 list_del_init(&epi
->rdllink
);
1348 spin_unlock_irqrestore(&ep
->lock
, flags
);
1350 wakeup_source_unregister(ep_wakeup_source(epi
));
1352 error_create_wakeup_source
:
1353 kmem_cache_free(epi_cache
, epi
);
1359 * Modify the interest event mask by dropping an event if the new mask
1360 * has a match in the current file status. Must be called with "mtx" held.
1362 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
, struct epoll_event
*event
)
1365 unsigned int revents
;
1368 init_poll_funcptr(&pt
, NULL
);
1371 * Set the new event interest mask before calling f_op->poll();
1372 * otherwise we might miss an event that happens between the
1373 * f_op->poll() call and the new event set registering.
1375 epi
->event
.events
= event
->events
; /* need barrier below */
1376 epi
->event
.data
= event
->data
; /* protected by mtx */
1377 if (epi
->event
.events
& EPOLLWAKEUP
) {
1378 if (!ep_has_wakeup_source(epi
))
1379 ep_create_wakeup_source(epi
);
1380 } else if (ep_has_wakeup_source(epi
)) {
1381 ep_destroy_wakeup_source(epi
);
1385 * The following barrier has two effects:
1387 * 1) Flush epi changes above to other CPUs. This ensures
1388 * we do not miss events from ep_poll_callback if an
1389 * event occurs immediately after we call f_op->poll().
1390 * We need this because we did not take ep->lock while
1391 * changing epi above (but ep_poll_callback does take
1394 * 2) We also need to ensure we do not miss _past_ events
1395 * when calling f_op->poll(). This barrier also
1396 * pairs with the barrier in wq_has_sleeper (see
1397 * comments for wq_has_sleeper).
1399 * This barrier will now guarantee ep_poll_callback or f_op->poll
1400 * (or both) will notice the readiness of an item.
1405 * Get current event bits. We can safely use the file* here because
1406 * its usage count has been increased by the caller of this function.
1408 revents
= ep_item_poll(epi
, &pt
);
1411 * If the item is "hot" and it is not registered inside the ready
1412 * list, push it inside.
1414 if (revents
& event
->events
) {
1415 spin_lock_irq(&ep
->lock
);
1416 if (!ep_is_linked(&epi
->rdllink
)) {
1417 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1418 ep_pm_stay_awake(epi
);
1420 /* Notify waiting tasks that events are available */
1421 if (waitqueue_active(&ep
->wq
))
1422 wake_up_locked(&ep
->wq
);
1423 if (waitqueue_active(&ep
->poll_wait
))
1426 spin_unlock_irq(&ep
->lock
);
1429 /* We have to call this outside the lock */
1431 ep_poll_safewake(&ep
->poll_wait
);
1436 static int ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1439 struct ep_send_events_data
*esed
= priv
;
1441 unsigned int revents
;
1443 struct epoll_event __user
*uevent
;
1444 struct wakeup_source
*ws
;
1447 init_poll_funcptr(&pt
, NULL
);
1450 * We can loop without lock because we are passed a task private list.
1451 * Items cannot vanish during the loop because ep_scan_ready_list() is
1452 * holding "mtx" during this call.
1454 for (eventcnt
= 0, uevent
= esed
->events
;
1455 !list_empty(head
) && eventcnt
< esed
->maxevents
;) {
1456 epi
= list_first_entry(head
, struct epitem
, rdllink
);
1459 * Activate ep->ws before deactivating epi->ws to prevent
1460 * triggering auto-suspend here (in case we reactive epi->ws
1463 * This could be rearranged to delay the deactivation of epi->ws
1464 * instead, but then epi->ws would temporarily be out of sync
1465 * with ep_is_linked().
1467 ws
= ep_wakeup_source(epi
);
1470 __pm_stay_awake(ep
->ws
);
1474 list_del_init(&epi
->rdllink
);
1476 revents
= ep_item_poll(epi
, &pt
);
1479 * If the event mask intersect the caller-requested one,
1480 * deliver the event to userspace. Again, ep_scan_ready_list()
1481 * is holding "mtx", so no operations coming from userspace
1482 * can change the item.
1485 if (__put_user(revents
, &uevent
->events
) ||
1486 __put_user(epi
->event
.data
, &uevent
->data
)) {
1487 list_add(&epi
->rdllink
, head
);
1488 ep_pm_stay_awake(epi
);
1489 return eventcnt
? eventcnt
: -EFAULT
;
1493 if (epi
->event
.events
& EPOLLONESHOT
)
1494 epi
->event
.events
&= EP_PRIVATE_BITS
;
1495 else if (!(epi
->event
.events
& EPOLLET
)) {
1497 * If this file has been added with Level
1498 * Trigger mode, we need to insert back inside
1499 * the ready list, so that the next call to
1500 * epoll_wait() will check again the events
1501 * availability. At this point, no one can insert
1502 * into ep->rdllist besides us. The epoll_ctl()
1503 * callers are locked out by
1504 * ep_scan_ready_list() holding "mtx" and the
1505 * poll callback will queue them in ep->ovflist.
1507 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1508 ep_pm_stay_awake(epi
);
1516 static int ep_send_events(struct eventpoll
*ep
,
1517 struct epoll_event __user
*events
, int maxevents
)
1519 struct ep_send_events_data esed
;
1521 esed
.maxevents
= maxevents
;
1522 esed
.events
= events
;
1524 return ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
, 0);
1527 static inline struct timespec
ep_set_mstimeout(long ms
)
1529 struct timespec now
, ts
= {
1530 .tv_sec
= ms
/ MSEC_PER_SEC
,
1531 .tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
),
1535 return timespec_add_safe(now
, ts
);
1539 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1542 * @ep: Pointer to the eventpoll context.
1543 * @events: Pointer to the userspace buffer where the ready events should be
1545 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1546 * @timeout: Maximum timeout for the ready events fetch operation, in
1547 * milliseconds. If the @timeout is zero, the function will not block,
1548 * while if the @timeout is less than zero, the function will block
1549 * until at least one event has been retrieved (or an error
1552 * Returns: Returns the number of ready events which have been fetched, or an
1553 * error code, in case of error.
1555 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1556 int maxevents
, long timeout
)
1558 int res
= 0, eavail
, timed_out
= 0;
1559 unsigned long flags
;
1562 ktime_t expires
, *to
= NULL
;
1565 struct timespec end_time
= ep_set_mstimeout(timeout
);
1567 slack
= select_estimate_accuracy(&end_time
);
1569 *to
= timespec_to_ktime(end_time
);
1570 } else if (timeout
== 0) {
1572 * Avoid the unnecessary trip to the wait queue loop, if the
1573 * caller specified a non blocking operation.
1576 spin_lock_irqsave(&ep
->lock
, flags
);
1581 spin_lock_irqsave(&ep
->lock
, flags
);
1583 if (!ep_events_available(ep
)) {
1585 * We don't have any available event to return to the caller.
1586 * We need to sleep here, and we will be wake up by
1587 * ep_poll_callback() when events will become available.
1589 init_waitqueue_entry(&wait
, current
);
1590 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1594 * We don't want to sleep if the ep_poll_callback() sends us
1595 * a wakeup in between. That's why we set the task state
1596 * to TASK_INTERRUPTIBLE before doing the checks.
1598 set_current_state(TASK_INTERRUPTIBLE
);
1599 if (ep_events_available(ep
) || timed_out
)
1601 if (signal_pending(current
)) {
1606 spin_unlock_irqrestore(&ep
->lock
, flags
);
1607 if (!schedule_hrtimeout_range(to
, slack
, HRTIMER_MODE_ABS
))
1610 spin_lock_irqsave(&ep
->lock
, flags
);
1612 __remove_wait_queue(&ep
->wq
, &wait
);
1614 set_current_state(TASK_RUNNING
);
1617 /* Is it worth to try to dig for events ? */
1618 eavail
= ep_events_available(ep
);
1620 spin_unlock_irqrestore(&ep
->lock
, flags
);
1623 * Try to transfer events to user space. In case we get 0 events and
1624 * there's still timeout left over, we go trying again in search of
1627 if (!res
&& eavail
&&
1628 !(res
= ep_send_events(ep
, events
, maxevents
)) && !timed_out
)
1635 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1636 * API, to verify that adding an epoll file inside another
1637 * epoll structure, does not violate the constraints, in
1638 * terms of closed loops, or too deep chains (which can
1639 * result in excessive stack usage).
1641 * @priv: Pointer to the epoll file to be currently checked.
1642 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1643 * data structure pointer.
1644 * @call_nests: Current dept of the @ep_call_nested() call stack.
1646 * Returns: Returns zero if adding the epoll @file inside current epoll
1647 * structure @ep does not violate the constraints, or -1 otherwise.
1649 static int ep_loop_check_proc(void *priv
, void *cookie
, int call_nests
)
1652 struct file
*file
= priv
;
1653 struct eventpoll
*ep
= file
->private_data
;
1654 struct eventpoll
*ep_tovisit
;
1655 struct rb_node
*rbp
;
1658 mutex_lock_nested(&ep
->mtx
, call_nests
+ 1);
1660 list_add(&ep
->visited_list_link
, &visited_list
);
1661 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1662 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1663 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1664 ep_tovisit
= epi
->ffd
.file
->private_data
;
1665 if (ep_tovisit
->visited
)
1667 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1668 ep_loop_check_proc
, epi
->ffd
.file
,
1669 ep_tovisit
, current
);
1674 * If we've reached a file that is not associated with
1675 * an ep, then we need to check if the newly added
1676 * links are going to add too many wakeup paths. We do
1677 * this by adding it to the tfile_check_list, if it's
1678 * not already there, and calling reverse_path_check()
1679 * during ep_insert().
1681 if (list_empty(&epi
->ffd
.file
->f_tfile_llink
))
1682 list_add(&epi
->ffd
.file
->f_tfile_llink
,
1686 mutex_unlock(&ep
->mtx
);
1692 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1693 * another epoll file (represented by @ep) does not create
1694 * closed loops or too deep chains.
1696 * @ep: Pointer to the epoll private data structure.
1697 * @file: Pointer to the epoll file to be checked.
1699 * Returns: Returns zero if adding the epoll @file inside current epoll
1700 * structure @ep does not violate the constraints, or -1 otherwise.
1702 static int ep_loop_check(struct eventpoll
*ep
, struct file
*file
)
1705 struct eventpoll
*ep_cur
, *ep_next
;
1707 ret
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1708 ep_loop_check_proc
, file
, ep
, current
);
1709 /* clear visited list */
1710 list_for_each_entry_safe(ep_cur
, ep_next
, &visited_list
,
1711 visited_list_link
) {
1712 ep_cur
->visited
= 0;
1713 list_del(&ep_cur
->visited_list_link
);
1718 static void clear_tfile_check_list(void)
1722 /* first clear the tfile_check_list */
1723 while (!list_empty(&tfile_check_list
)) {
1724 file
= list_first_entry(&tfile_check_list
, struct file
,
1726 list_del_init(&file
->f_tfile_llink
);
1728 INIT_LIST_HEAD(&tfile_check_list
);
1732 * Open an eventpoll file descriptor.
1734 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
1737 struct eventpoll
*ep
= NULL
;
1740 /* Check the EPOLL_* constant for consistency. */
1741 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
1743 if (flags
& ~EPOLL_CLOEXEC
)
1746 * Create the internal data structure ("struct eventpoll").
1748 error
= ep_alloc(&ep
);
1752 * Creates all the items needed to setup an eventpoll file. That is,
1753 * a file structure and a free file descriptor.
1755 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
1760 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
1761 O_RDWR
| (flags
& O_CLOEXEC
));
1763 error
= PTR_ERR(file
);
1767 fd_install(fd
, file
);
1777 SYSCALL_DEFINE1(epoll_create
, int, size
)
1782 return sys_epoll_create1(0);
1786 * The following function implements the controller interface for
1787 * the eventpoll file that enables the insertion/removal/change of
1788 * file descriptors inside the interest set.
1790 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
1791 struct epoll_event __user
*, event
)
1794 int did_lock_epmutex
= 0;
1796 struct eventpoll
*ep
;
1798 struct epoll_event epds
;
1801 if (ep_op_has_event(op
) &&
1802 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
1810 /* Get the "struct file *" for the target file */
1815 /* The target file descriptor must support poll */
1817 if (!tf
.file
->f_op
|| !tf
.file
->f_op
->poll
)
1818 goto error_tgt_fput
;
1820 /* Check if EPOLLWAKEUP is allowed */
1821 if ((epds
.events
& EPOLLWAKEUP
) && !capable(CAP_BLOCK_SUSPEND
))
1822 epds
.events
&= ~EPOLLWAKEUP
;
1825 * We have to check that the file structure underneath the file descriptor
1826 * the user passed to us _is_ an eventpoll file. And also we do not permit
1827 * adding an epoll file descriptor inside itself.
1830 if (f
.file
== tf
.file
|| !is_file_epoll(f
.file
))
1831 goto error_tgt_fput
;
1834 * At this point it is safe to assume that the "private_data" contains
1835 * our own data structure.
1837 ep
= f
.file
->private_data
;
1840 * When we insert an epoll file descriptor, inside another epoll file
1841 * descriptor, there is the change of creating closed loops, which are
1842 * better be handled here, than in more critical paths. While we are
1843 * checking for loops we also determine the list of files reachable
1844 * and hang them on the tfile_check_list, so we can check that we
1845 * haven't created too many possible wakeup paths.
1847 * We need to hold the epmutex across both ep_insert and ep_remove
1848 * b/c we want to make sure we are looking at a coherent view of
1851 if (op
== EPOLL_CTL_ADD
|| op
== EPOLL_CTL_DEL
) {
1852 mutex_lock(&epmutex
);
1853 did_lock_epmutex
= 1;
1855 if (op
== EPOLL_CTL_ADD
) {
1856 if (is_file_epoll(tf
.file
)) {
1858 if (ep_loop_check(ep
, tf
.file
) != 0) {
1859 clear_tfile_check_list();
1860 goto error_tgt_fput
;
1863 list_add(&tf
.file
->f_tfile_llink
, &tfile_check_list
);
1866 mutex_lock_nested(&ep
->mtx
, 0);
1869 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1870 * above, we can be sure to be able to use the item looked up by
1871 * ep_find() till we release the mutex.
1873 epi
= ep_find(ep
, tf
.file
, fd
);
1879 epds
.events
|= POLLERR
| POLLHUP
;
1880 error
= ep_insert(ep
, &epds
, tf
.file
, fd
);
1883 clear_tfile_check_list();
1887 error
= ep_remove(ep
, epi
);
1893 epds
.events
|= POLLERR
| POLLHUP
;
1894 error
= ep_modify(ep
, epi
, &epds
);
1899 mutex_unlock(&ep
->mtx
);
1902 if (did_lock_epmutex
)
1903 mutex_unlock(&epmutex
);
1914 * Implement the event wait interface for the eventpoll file. It is the kernel
1915 * part of the user space epoll_wait(2).
1917 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
1918 int, maxevents
, int, timeout
)
1922 struct eventpoll
*ep
;
1924 /* The maximum number of event must be greater than zero */
1925 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
1928 /* Verify that the area passed by the user is writeable */
1929 if (!access_ok(VERIFY_WRITE
, events
, maxevents
* sizeof(struct epoll_event
)))
1932 /* Get the "struct file *" for the eventpoll file */
1938 * We have to check that the file structure underneath the fd
1939 * the user passed to us _is_ an eventpoll file.
1942 if (!is_file_epoll(f
.file
))
1946 * At this point it is safe to assume that the "private_data" contains
1947 * our own data structure.
1949 ep
= f
.file
->private_data
;
1951 /* Time to fish for events ... */
1952 error
= ep_poll(ep
, events
, maxevents
, timeout
);
1960 * Implement the event wait interface for the eventpoll file. It is the kernel
1961 * part of the user space epoll_pwait(2).
1963 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
1964 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
1968 sigset_t ksigmask
, sigsaved
;
1971 * If the caller wants a certain signal mask to be set during the wait,
1975 if (sigsetsize
!= sizeof(sigset_t
))
1977 if (copy_from_user(&ksigmask
, sigmask
, sizeof(ksigmask
)))
1979 sigsaved
= current
->blocked
;
1980 set_current_blocked(&ksigmask
);
1983 error
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
1986 * If we changed the signal mask, we need to restore the original one.
1987 * In case we've got a signal while waiting, we do not restore the
1988 * signal mask yet, and we allow do_signal() to deliver the signal on
1989 * the way back to userspace, before the signal mask is restored.
1992 if (error
== -EINTR
) {
1993 memcpy(¤t
->saved_sigmask
, &sigsaved
,
1995 set_restore_sigmask();
1997 set_current_blocked(&sigsaved
);
2003 #ifdef CONFIG_COMPAT
2004 COMPAT_SYSCALL_DEFINE6(epoll_pwait
, int, epfd
,
2005 struct epoll_event __user
*, events
,
2006 int, maxevents
, int, timeout
,
2007 const compat_sigset_t __user
*, sigmask
,
2008 compat_size_t
, sigsetsize
)
2011 compat_sigset_t csigmask
;
2012 sigset_t ksigmask
, sigsaved
;
2015 * If the caller wants a certain signal mask to be set during the wait,
2019 if (sigsetsize
!= sizeof(compat_sigset_t
))
2021 if (copy_from_user(&csigmask
, sigmask
, sizeof(csigmask
)))
2023 sigset_from_compat(&ksigmask
, &csigmask
);
2024 sigsaved
= current
->blocked
;
2025 set_current_blocked(&ksigmask
);
2028 err
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
2031 * If we changed the signal mask, we need to restore the original one.
2032 * In case we've got a signal while waiting, we do not restore the
2033 * signal mask yet, and we allow do_signal() to deliver the signal on
2034 * the way back to userspace, before the signal mask is restored.
2037 if (err
== -EINTR
) {
2038 memcpy(¤t
->saved_sigmask
, &sigsaved
,
2040 set_restore_sigmask();
2042 set_current_blocked(&sigsaved
);
2049 static int __init
eventpoll_init(void)
2055 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2057 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2059 BUG_ON(max_user_watches
< 0);
2062 * Initialize the structure used to perform epoll file descriptor
2063 * inclusion loops checks.
2065 ep_nested_calls_init(&poll_loop_ncalls
);
2067 /* Initialize the structure used to perform safe poll wait head wake ups */
2068 ep_nested_calls_init(&poll_safewake_ncalls
);
2070 /* Initialize the structure used to perform file's f_op->poll() calls */
2071 ep_nested_calls_init(&poll_readywalk_ncalls
);
2074 * We can have many thousands of epitems, so prevent this from
2075 * using an extra cache line on 64-bit (and smaller) CPUs
2077 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2079 /* Allocates slab cache used to allocate "struct epitem" items */
2080 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2081 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
2083 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2084 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2085 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
, NULL
);
2089 fs_initcall(eventpoll_init
);