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>
46 * There are three level of locking required by epoll :
50 * 3) ep->lock (spinlock)
52 * The acquire order is the one listed above, from 1 to 3.
53 * We need a spinlock (ep->lock) because we manipulate objects
54 * from inside the poll callback, that might be triggered from
55 * a wake_up() that in turn might be called from IRQ context.
56 * So we can't sleep inside the poll callback and hence we need
57 * a spinlock. During the event transfer loop (from kernel to
58 * user space) we could end up sleeping due a copy_to_user(), so
59 * we need a lock that will allow us to sleep. This lock is a
60 * mutex (ep->mtx). It is acquired during the event transfer loop,
61 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
62 * Then we also need a global mutex to serialize eventpoll_release_file()
64 * This mutex is acquired by ep_free() during the epoll file
65 * cleanup path and it is also acquired by eventpoll_release_file()
66 * if a file has been pushed inside an epoll set and it is then
67 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
68 * It is also acquired when inserting an epoll fd onto another epoll
69 * fd. We do this so that we walk the epoll tree and ensure that this
70 * insertion does not create a cycle of epoll file descriptors, which
71 * could lead to deadlock. We need a global mutex to prevent two
72 * simultaneous inserts (A into B and B into A) from racing and
73 * constructing a cycle without either insert observing that it is
75 * It is necessary to acquire multiple "ep->mtx"es at once in the
76 * case when one epoll fd is added to another. In this case, we
77 * always acquire the locks in the order of nesting (i.e. after
78 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
79 * before e2->mtx). Since we disallow cycles of epoll file
80 * descriptors, this ensures that the mutexes are well-ordered. In
81 * order to communicate this nesting to lockdep, when walking a tree
82 * of epoll file descriptors, we use the current recursion depth as
84 * It is possible to drop the "ep->mtx" and to use the global
85 * mutex "epmutex" (together with "ep->lock") to have it working,
86 * but having "ep->mtx" will make the interface more scalable.
87 * Events that require holding "epmutex" are very rare, while for
88 * normal operations the epoll private "ep->mtx" will guarantee
89 * a better scalability.
92 /* Epoll private bits inside the event mask */
93 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
95 /* Maximum number of nesting allowed inside epoll sets */
96 #define EP_MAX_NESTS 4
98 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
100 #define EP_UNACTIVE_PTR ((void *) -1L)
102 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
104 struct epoll_filefd
{
110 * Structure used to track possible nested calls, for too deep recursions
113 struct nested_call_node
{
114 struct list_head llink
;
120 * This structure is used as collector for nested calls, to check for
121 * maximum recursion dept and loop cycles.
123 struct nested_calls
{
124 struct list_head tasks_call_list
;
129 * Each file descriptor added to the eventpoll interface will
130 * have an entry of this type linked to the "rbr" RB tree.
131 * Avoid increasing the size of this struct, there can be many thousands
132 * of these on a server and we do not want this to take another cache line.
135 /* RB tree node used to link this structure to the eventpoll RB tree */
138 /* List header used to link this structure to the eventpoll ready list */
139 struct list_head rdllink
;
142 * Works together "struct eventpoll"->ovflist in keeping the
143 * single linked chain of items.
147 /* The file descriptor information this item refers to */
148 struct epoll_filefd ffd
;
150 /* Number of active wait queue attached to poll operations */
153 /* List containing poll wait queues */
154 struct list_head pwqlist
;
156 /* The "container" of this item */
157 struct eventpoll
*ep
;
159 /* List header used to link this item to the "struct file" items list */
160 struct list_head fllink
;
162 /* wakeup_source used when EPOLLWAKEUP is set */
163 struct wakeup_source __rcu
*ws
;
165 /* The structure that describe the interested events and the source fd */
166 struct epoll_event event
;
170 * This structure is stored inside the "private_data" member of the file
171 * structure and represents the main data structure for the eventpoll
175 /* Protect the access to this structure */
179 * This mutex is used to ensure that files are not removed
180 * while epoll is using them. This is held during the event
181 * collection loop, the file cleanup path, the epoll file exit
182 * code and the ctl operations.
186 /* Wait queue used by sys_epoll_wait() */
187 wait_queue_head_t wq
;
189 /* Wait queue used by file->poll() */
190 wait_queue_head_t poll_wait
;
192 /* List of ready file descriptors */
193 struct list_head rdllist
;
195 /* RB tree root used to store monitored fd structs */
199 * This is a single linked list that chains all the "struct epitem" that
200 * happened while transferring ready events to userspace w/out
203 struct epitem
*ovflist
;
205 /* wakeup_source used when ep_scan_ready_list is running */
206 struct wakeup_source
*ws
;
208 /* The user that created the eventpoll descriptor */
209 struct user_struct
*user
;
213 /* used to optimize loop detection check */
215 struct list_head visited_list_link
;
218 /* Wait structure used by the poll hooks */
219 struct eppoll_entry
{
220 /* List header used to link this structure to the "struct epitem" */
221 struct list_head llink
;
223 /* The "base" pointer is set to the container "struct epitem" */
227 * Wait queue item that will be linked to the target file wait
232 /* The wait queue head that linked the "wait" wait queue item */
233 wait_queue_head_t
*whead
;
236 /* Wrapper struct used by poll queueing */
242 /* Used by the ep_send_events() function as callback private data */
243 struct ep_send_events_data
{
245 struct epoll_event __user
*events
;
249 * Configuration options available inside /proc/sys/fs/epoll/
251 /* Maximum number of epoll watched descriptors, per user */
252 static long max_user_watches __read_mostly
;
255 * This mutex is used to serialize ep_free() and eventpoll_release_file().
257 static DEFINE_MUTEX(epmutex
);
259 /* Used to check for epoll file descriptor inclusion loops */
260 static struct nested_calls poll_loop_ncalls
;
262 /* Used for safe wake up implementation */
263 static struct nested_calls poll_safewake_ncalls
;
265 /* Used to call file's f_op->poll() under the nested calls boundaries */
266 static struct nested_calls poll_readywalk_ncalls
;
268 /* Slab cache used to allocate "struct epitem" */
269 static struct kmem_cache
*epi_cache __read_mostly
;
271 /* Slab cache used to allocate "struct eppoll_entry" */
272 static struct kmem_cache
*pwq_cache __read_mostly
;
274 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
275 static LIST_HEAD(visited_list
);
278 * List of files with newly added links, where we may need to limit the number
279 * of emanating paths. Protected by the epmutex.
281 static LIST_HEAD(tfile_check_list
);
285 #include <linux/sysctl.h>
288 static long long_max
= LONG_MAX
;
290 ctl_table epoll_table
[] = {
292 .procname
= "max_user_watches",
293 .data
= &max_user_watches
,
294 .maxlen
= sizeof(max_user_watches
),
296 .proc_handler
= proc_doulongvec_minmax
,
302 #endif /* CONFIG_SYSCTL */
304 static const struct file_operations eventpoll_fops
;
306 static inline int is_file_epoll(struct file
*f
)
308 return f
->f_op
== &eventpoll_fops
;
311 /* Setup the structure that is used as key for the RB tree */
312 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
313 struct file
*file
, int fd
)
319 /* Compare RB tree keys */
320 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
321 struct epoll_filefd
*p2
)
323 return (p1
->file
> p2
->file
? +1:
324 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
327 /* Tells us if the item is currently linked */
328 static inline int ep_is_linked(struct list_head
*p
)
330 return !list_empty(p
);
333 static inline struct eppoll_entry
*ep_pwq_from_wait(wait_queue_t
*p
)
335 return container_of(p
, struct eppoll_entry
, wait
);
338 /* Get the "struct epitem" from a wait queue pointer */
339 static inline struct epitem
*ep_item_from_wait(wait_queue_t
*p
)
341 return container_of(p
, struct eppoll_entry
, wait
)->base
;
344 /* Get the "struct epitem" from an epoll queue wrapper */
345 static inline struct epitem
*ep_item_from_epqueue(poll_table
*p
)
347 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
350 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
351 static inline int ep_op_has_event(int op
)
353 return op
!= EPOLL_CTL_DEL
;
356 /* Initialize the poll safe wake up structure */
357 static void ep_nested_calls_init(struct nested_calls
*ncalls
)
359 INIT_LIST_HEAD(&ncalls
->tasks_call_list
);
360 spin_lock_init(&ncalls
->lock
);
364 * ep_events_available - Checks if ready events might be available.
366 * @ep: Pointer to the eventpoll context.
368 * Returns: Returns a value different than zero if ready events are available,
371 static inline int ep_events_available(struct eventpoll
*ep
)
373 return !list_empty(&ep
->rdllist
) || ep
->ovflist
!= EP_UNACTIVE_PTR
;
377 * ep_call_nested - Perform a bound (possibly) nested call, by checking
378 * that the recursion limit is not exceeded, and that
379 * the same nested call (by the meaning of same cookie) is
382 * @ncalls: Pointer to the nested_calls structure to be used for this call.
383 * @max_nests: Maximum number of allowed nesting calls.
384 * @nproc: Nested call core function pointer.
385 * @priv: Opaque data to be passed to the @nproc callback.
386 * @cookie: Cookie to be used to identify this nested call.
387 * @ctx: This instance context.
389 * Returns: Returns the code returned by the @nproc callback, or -1 if
390 * the maximum recursion limit has been exceeded.
392 static int ep_call_nested(struct nested_calls
*ncalls
, int max_nests
,
393 int (*nproc
)(void *, void *, int), void *priv
,
394 void *cookie
, void *ctx
)
396 int error
, call_nests
= 0;
398 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
399 struct nested_call_node
*tncur
;
400 struct nested_call_node tnode
;
402 spin_lock_irqsave(&ncalls
->lock
, flags
);
405 * Try to see if the current task is already inside this wakeup call.
406 * We use a list here, since the population inside this set is always
409 list_for_each_entry(tncur
, lsthead
, llink
) {
410 if (tncur
->ctx
== ctx
&&
411 (tncur
->cookie
== cookie
|| ++call_nests
> max_nests
)) {
413 * Ops ... loop detected or maximum nest level reached.
414 * We abort this wake by breaking the cycle itself.
421 /* Add the current task and cookie to the list */
423 tnode
.cookie
= cookie
;
424 list_add(&tnode
.llink
, lsthead
);
426 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
428 /* Call the nested function */
429 error
= (*nproc
)(priv
, cookie
, call_nests
);
431 /* Remove the current task from the list */
432 spin_lock_irqsave(&ncalls
->lock
, flags
);
433 list_del(&tnode
.llink
);
435 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
441 * As described in commit 0ccf831cb lockdep: annotate epoll
442 * the use of wait queues used by epoll is done in a very controlled
443 * manner. Wake ups can nest inside each other, but are never done
444 * with the same locking. For example:
447 * efd1 = epoll_create();
448 * efd2 = epoll_create();
449 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
450 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
452 * When a packet arrives to the device underneath "dfd", the net code will
453 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
454 * callback wakeup entry on that queue, and the wake_up() performed by the
455 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
456 * (efd1) notices that it may have some event ready, so it needs to wake up
457 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
458 * that ends up in another wake_up(), after having checked about the
459 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
460 * avoid stack blasting.
462 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
463 * this special case of epoll.
465 #ifdef CONFIG_DEBUG_LOCK_ALLOC
466 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
467 unsigned long events
, int subclass
)
471 spin_lock_irqsave_nested(&wqueue
->lock
, flags
, subclass
);
472 wake_up_locked_poll(wqueue
, events
);
473 spin_unlock_irqrestore(&wqueue
->lock
, flags
);
476 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
477 unsigned long events
, int subclass
)
479 wake_up_poll(wqueue
, events
);
483 static int ep_poll_wakeup_proc(void *priv
, void *cookie
, int call_nests
)
485 ep_wake_up_nested((wait_queue_head_t
*) cookie
, POLLIN
,
491 * Perform a safe wake up of the poll wait list. The problem is that
492 * with the new callback'd wake up system, it is possible that the
493 * poll callback is reentered from inside the call to wake_up() done
494 * on the poll wait queue head. The rule is that we cannot reenter the
495 * wake up code from the same task more than EP_MAX_NESTS times,
496 * and we cannot reenter the same wait queue head at all. This will
497 * enable to have a hierarchy of epoll file descriptor of no more than
500 static void ep_poll_safewake(wait_queue_head_t
*wq
)
502 int this_cpu
= get_cpu();
504 ep_call_nested(&poll_safewake_ncalls
, EP_MAX_NESTS
,
505 ep_poll_wakeup_proc
, NULL
, wq
, (void *) (long) this_cpu
);
510 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
512 wait_queue_head_t
*whead
;
515 /* If it is cleared by POLLFREE, it should be rcu-safe */
516 whead
= rcu_dereference(pwq
->whead
);
518 remove_wait_queue(whead
, &pwq
->wait
);
523 * This function unregisters poll callbacks from the associated file
524 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
527 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
529 struct list_head
*lsthead
= &epi
->pwqlist
;
530 struct eppoll_entry
*pwq
;
532 while (!list_empty(lsthead
)) {
533 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
535 list_del(&pwq
->llink
);
536 ep_remove_wait_queue(pwq
);
537 kmem_cache_free(pwq_cache
, pwq
);
541 /* call only when ep->mtx is held */
542 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
544 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
547 /* call only when ep->mtx is held */
548 static inline void ep_pm_stay_awake(struct epitem
*epi
)
550 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
556 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
558 return rcu_access_pointer(epi
->ws
) ? true : false;
561 /* call when ep->mtx cannot be held (ep_poll_callback) */
562 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
564 struct wakeup_source
*ws
;
567 ws
= rcu_dereference(epi
->ws
);
574 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
575 * the scan code, to call f_op->poll(). Also allows for
576 * O(NumReady) performance.
578 * @ep: Pointer to the epoll private data structure.
579 * @sproc: Pointer to the scan callback.
580 * @priv: Private opaque data passed to the @sproc callback.
581 * @depth: The current depth of recursive f_op->poll calls.
583 * Returns: The same integer error code returned by the @sproc callback.
585 static int ep_scan_ready_list(struct eventpoll
*ep
,
586 int (*sproc
)(struct eventpoll
*,
587 struct list_head
*, void *),
591 int error
, pwake
= 0;
593 struct epitem
*epi
, *nepi
;
597 * We need to lock this because we could be hit by
598 * eventpoll_release_file() and epoll_ctl().
600 mutex_lock_nested(&ep
->mtx
, depth
);
603 * Steal the ready list, and re-init the original one to the
604 * empty list. Also, set ep->ovflist to NULL so that events
605 * happening while looping w/out locks, are not lost. We cannot
606 * have the poll callback to queue directly on ep->rdllist,
607 * because we want the "sproc" callback to be able to do it
610 spin_lock_irqsave(&ep
->lock
, flags
);
611 list_splice_init(&ep
->rdllist
, &txlist
);
613 spin_unlock_irqrestore(&ep
->lock
, flags
);
616 * Now call the callback function.
618 error
= (*sproc
)(ep
, &txlist
, priv
);
620 spin_lock_irqsave(&ep
->lock
, flags
);
622 * During the time we spent inside the "sproc" callback, some
623 * other events might have been queued by the poll callback.
624 * We re-insert them inside the main ready-list here.
626 for (nepi
= ep
->ovflist
; (epi
= nepi
) != NULL
;
627 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
629 * We need to check if the item is already in the list.
630 * During the "sproc" callback execution time, items are
631 * queued into ->ovflist but the "txlist" might already
632 * contain them, and the list_splice() below takes care of them.
634 if (!ep_is_linked(&epi
->rdllink
)) {
635 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
636 ep_pm_stay_awake(epi
);
640 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
641 * releasing the lock, events will be queued in the normal way inside
644 ep
->ovflist
= EP_UNACTIVE_PTR
;
647 * Quickly re-inject items left on "txlist".
649 list_splice(&txlist
, &ep
->rdllist
);
652 if (!list_empty(&ep
->rdllist
)) {
654 * Wake up (if active) both the eventpoll wait list and
655 * the ->poll() wait list (delayed after we release the lock).
657 if (waitqueue_active(&ep
->wq
))
658 wake_up_locked(&ep
->wq
);
659 if (waitqueue_active(&ep
->poll_wait
))
662 spin_unlock_irqrestore(&ep
->lock
, flags
);
664 mutex_unlock(&ep
->mtx
);
666 /* We have to call this outside the lock */
668 ep_poll_safewake(&ep
->poll_wait
);
674 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
675 * all the associated resources. Must be called with "mtx" held.
677 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
680 struct file
*file
= epi
->ffd
.file
;
683 * Removes poll wait queue hooks. We _have_ to do this without holding
684 * the "ep->lock" otherwise a deadlock might occur. This because of the
685 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
686 * queue head lock when unregistering the wait queue. The wakeup callback
687 * will run by holding the wait queue head lock and will call our callback
688 * that will try to get "ep->lock".
690 ep_unregister_pollwait(ep
, epi
);
692 /* Remove the current item from the list of epoll hooks */
693 spin_lock(&file
->f_lock
);
694 if (ep_is_linked(&epi
->fllink
))
695 list_del_init(&epi
->fllink
);
696 spin_unlock(&file
->f_lock
);
698 rb_erase(&epi
->rbn
, &ep
->rbr
);
700 spin_lock_irqsave(&ep
->lock
, flags
);
701 if (ep_is_linked(&epi
->rdllink
))
702 list_del_init(&epi
->rdllink
);
703 spin_unlock_irqrestore(&ep
->lock
, flags
);
705 wakeup_source_unregister(ep_wakeup_source(epi
));
707 /* At this point it is safe to free the eventpoll item */
708 kmem_cache_free(epi_cache
, epi
);
710 atomic_long_dec(&ep
->user
->epoll_watches
);
715 static void ep_free(struct eventpoll
*ep
)
720 /* We need to release all tasks waiting for these file */
721 if (waitqueue_active(&ep
->poll_wait
))
722 ep_poll_safewake(&ep
->poll_wait
);
725 * We need to lock this because we could be hit by
726 * eventpoll_release_file() while we're freeing the "struct eventpoll".
727 * We do not need to hold "ep->mtx" here because the epoll file
728 * is on the way to be removed and no one has references to it
729 * anymore. The only hit might come from eventpoll_release_file() but
730 * holding "epmutex" is sufficient here.
732 mutex_lock(&epmutex
);
735 * Walks through the whole tree by unregistering poll callbacks.
737 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
738 epi
= rb_entry(rbp
, struct epitem
, rbn
);
740 ep_unregister_pollwait(ep
, epi
);
744 * Walks through the whole tree by freeing each "struct epitem". At this
745 * point we are sure no poll callbacks will be lingering around, and also by
746 * holding "epmutex" we can be sure that no file cleanup code will hit
747 * us during this operation. So we can avoid the lock on "ep->lock".
748 * We do not need to lock ep->mtx, either, we only do it to prevent
751 mutex_lock(&ep
->mtx
);
752 while ((rbp
= rb_first(&ep
->rbr
)) != NULL
) {
753 epi
= rb_entry(rbp
, struct epitem
, rbn
);
756 mutex_unlock(&ep
->mtx
);
758 mutex_unlock(&epmutex
);
759 mutex_destroy(&ep
->mtx
);
761 wakeup_source_unregister(ep
->ws
);
765 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
767 struct eventpoll
*ep
= file
->private_data
;
775 static inline unsigned int ep_item_poll(struct epitem
*epi
, poll_table
*pt
)
777 pt
->_key
= epi
->event
.events
;
779 return epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, pt
) & epi
->event
.events
;
782 static int ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
785 struct epitem
*epi
, *tmp
;
788 init_poll_funcptr(&pt
, NULL
);
790 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
791 if (ep_item_poll(epi
, &pt
))
792 return POLLIN
| POLLRDNORM
;
795 * Item has been dropped into the ready list by the poll
796 * callback, but it's not actually ready, as far as
797 * caller requested events goes. We can remove it here.
799 __pm_relax(ep_wakeup_source(epi
));
800 list_del_init(&epi
->rdllink
);
807 static int ep_poll_readyevents_proc(void *priv
, void *cookie
, int call_nests
)
809 return ep_scan_ready_list(priv
, ep_read_events_proc
, NULL
, call_nests
+ 1);
812 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
815 struct eventpoll
*ep
= file
->private_data
;
817 /* Insert inside our poll wait queue */
818 poll_wait(file
, &ep
->poll_wait
, wait
);
821 * Proceed to find out if wanted events are really available inside
822 * the ready list. This need to be done under ep_call_nested()
823 * supervision, since the call to f_op->poll() done on listed files
824 * could re-enter here.
826 pollflags
= ep_call_nested(&poll_readywalk_ncalls
, EP_MAX_NESTS
,
827 ep_poll_readyevents_proc
, ep
, ep
, current
);
829 return pollflags
!= -1 ? pollflags
: 0;
832 #ifdef CONFIG_PROC_FS
833 static int ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
835 struct eventpoll
*ep
= f
->private_data
;
839 mutex_lock(&ep
->mtx
);
840 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
841 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
843 ret
= seq_printf(m
, "tfd: %8d events: %8x data: %16llx\n",
844 epi
->ffd
.fd
, epi
->event
.events
,
845 (long long)epi
->event
.data
);
849 mutex_unlock(&ep
->mtx
);
855 /* File callbacks that implement the eventpoll file behaviour */
856 static const struct file_operations eventpoll_fops
= {
857 #ifdef CONFIG_PROC_FS
858 .show_fdinfo
= ep_show_fdinfo
,
860 .release
= ep_eventpoll_release
,
861 .poll
= ep_eventpoll_poll
,
862 .llseek
= noop_llseek
,
866 * This is called from eventpoll_release() to unlink files from the eventpoll
867 * interface. We need to have this facility to cleanup correctly files that are
868 * closed without being removed from the eventpoll interface.
870 void eventpoll_release_file(struct file
*file
)
872 struct list_head
*lsthead
= &file
->f_ep_links
;
873 struct eventpoll
*ep
;
877 * We don't want to get "file->f_lock" because it is not
878 * necessary. It is not necessary because we're in the "struct file"
879 * cleanup path, and this means that no one is using this file anymore.
880 * So, for example, epoll_ctl() cannot hit here since if we reach this
881 * point, the file counter already went to zero and fget() would fail.
882 * The only hit might come from ep_free() but by holding the mutex
883 * will correctly serialize the operation. We do need to acquire
884 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
885 * from anywhere but ep_free().
887 * Besides, ep_remove() acquires the lock, so we can't hold it here.
889 mutex_lock(&epmutex
);
891 while (!list_empty(lsthead
)) {
892 epi
= list_first_entry(lsthead
, struct epitem
, fllink
);
895 list_del_init(&epi
->fllink
);
896 mutex_lock_nested(&ep
->mtx
, 0);
898 mutex_unlock(&ep
->mtx
);
901 mutex_unlock(&epmutex
);
904 static int ep_alloc(struct eventpoll
**pep
)
907 struct user_struct
*user
;
908 struct eventpoll
*ep
;
910 user
= get_current_user();
912 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
916 spin_lock_init(&ep
->lock
);
917 mutex_init(&ep
->mtx
);
918 init_waitqueue_head(&ep
->wq
);
919 init_waitqueue_head(&ep
->poll_wait
);
920 INIT_LIST_HEAD(&ep
->rdllist
);
922 ep
->ovflist
= EP_UNACTIVE_PTR
;
935 * Search the file inside the eventpoll tree. The RB tree operations
936 * are protected by the "mtx" mutex, and ep_find() must be called with
939 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
943 struct epitem
*epi
, *epir
= NULL
;
944 struct epoll_filefd ffd
;
946 ep_set_ffd(&ffd
, file
, fd
);
947 for (rbp
= ep
->rbr
.rb_node
; rbp
; ) {
948 epi
= rb_entry(rbp
, struct epitem
, rbn
);
949 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
964 * This is the callback that is passed to the wait queue wakeup
965 * mechanism. It is called by the stored file descriptors when they
966 * have events to report.
968 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
972 struct epitem
*epi
= ep_item_from_wait(wait
);
973 struct eventpoll
*ep
= epi
->ep
;
975 if ((unsigned long)key
& POLLFREE
) {
976 ep_pwq_from_wait(wait
)->whead
= NULL
;
978 * whead = NULL above can race with ep_remove_wait_queue()
979 * which can do another remove_wait_queue() after us, so we
980 * can't use __remove_wait_queue(). whead->lock is held by
983 list_del_init(&wait
->task_list
);
986 spin_lock_irqsave(&ep
->lock
, flags
);
989 * If the event mask does not contain any poll(2) event, we consider the
990 * descriptor to be disabled. This condition is likely the effect of the
991 * EPOLLONESHOT bit that disables the descriptor when an event is received,
992 * until the next EPOLL_CTL_MOD will be issued.
994 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
998 * Check the events coming with the callback. At this stage, not
999 * every device reports the events in the "key" parameter of the
1000 * callback. We need to be able to handle both cases here, hence the
1001 * test for "key" != NULL before the event match test.
1003 if (key
&& !((unsigned long) key
& epi
->event
.events
))
1007 * If we are transferring events to userspace, we can hold no locks
1008 * (because we're accessing user memory, and because of linux f_op->poll()
1009 * semantics). All the events that happen during that period of time are
1010 * chained in ep->ovflist and requeued later on.
1012 if (unlikely(ep
->ovflist
!= EP_UNACTIVE_PTR
)) {
1013 if (epi
->next
== EP_UNACTIVE_PTR
) {
1014 epi
->next
= ep
->ovflist
;
1018 * Activate ep->ws since epi->ws may get
1019 * deactivated at any time.
1021 __pm_stay_awake(ep
->ws
);
1028 /* If this file is already in the ready list we exit soon */
1029 if (!ep_is_linked(&epi
->rdllink
)) {
1030 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1031 ep_pm_stay_awake_rcu(epi
);
1035 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1038 if (waitqueue_active(&ep
->wq
))
1039 wake_up_locked(&ep
->wq
);
1040 if (waitqueue_active(&ep
->poll_wait
))
1044 spin_unlock_irqrestore(&ep
->lock
, flags
);
1046 /* We have to call this outside the lock */
1048 ep_poll_safewake(&ep
->poll_wait
);
1054 * This is the callback that is used to add our wait queue to the
1055 * target file wakeup lists.
1057 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1060 struct epitem
*epi
= ep_item_from_epqueue(pt
);
1061 struct eppoll_entry
*pwq
;
1063 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
1064 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1067 add_wait_queue(whead
, &pwq
->wait
);
1068 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
1071 /* We have to signal that an error occurred */
1076 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1079 struct rb_node
**p
= &ep
->rbr
.rb_node
, *parent
= NULL
;
1080 struct epitem
*epic
;
1084 epic
= rb_entry(parent
, struct epitem
, rbn
);
1085 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1087 p
= &parent
->rb_right
;
1089 p
= &parent
->rb_left
;
1091 rb_link_node(&epi
->rbn
, parent
, p
);
1092 rb_insert_color(&epi
->rbn
, &ep
->rbr
);
1097 #define PATH_ARR_SIZE 5
1099 * These are the number paths of length 1 to 5, that we are allowing to emanate
1100 * from a single file of interest. For example, we allow 1000 paths of length
1101 * 1, to emanate from each file of interest. This essentially represents the
1102 * potential wakeup paths, which need to be limited in order to avoid massive
1103 * uncontrolled wakeup storms. The common use case should be a single ep which
1104 * is connected to n file sources. In this case each file source has 1 path
1105 * of length 1. Thus, the numbers below should be more than sufficient. These
1106 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1107 * and delete can't add additional paths. Protected by the epmutex.
1109 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1110 static int path_count
[PATH_ARR_SIZE
];
1112 static int path_count_inc(int nests
)
1114 /* Allow an arbitrary number of depth 1 paths */
1118 if (++path_count
[nests
] > path_limits
[nests
])
1123 static void path_count_init(void)
1127 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1131 static int reverse_path_check_proc(void *priv
, void *cookie
, int call_nests
)
1134 struct file
*file
= priv
;
1135 struct file
*child_file
;
1138 list_for_each_entry(epi
, &file
->f_ep_links
, fllink
) {
1139 child_file
= epi
->ep
->file
;
1140 if (is_file_epoll(child_file
)) {
1141 if (list_empty(&child_file
->f_ep_links
)) {
1142 if (path_count_inc(call_nests
)) {
1147 error
= ep_call_nested(&poll_loop_ncalls
,
1149 reverse_path_check_proc
,
1150 child_file
, child_file
,
1156 printk(KERN_ERR
"reverse_path_check_proc: "
1157 "file is not an ep!\n");
1164 * reverse_path_check - The tfile_check_list is list of file *, which have
1165 * links that are proposed to be newly added. We need to
1166 * make sure that those added links don't add too many
1167 * paths such that we will spend all our time waking up
1168 * eventpoll objects.
1170 * Returns: Returns zero if the proposed links don't create too many paths,
1173 static int reverse_path_check(void)
1176 struct file
*current_file
;
1178 /* let's call this for all tfiles */
1179 list_for_each_entry(current_file
, &tfile_check_list
, f_tfile_llink
) {
1181 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1182 reverse_path_check_proc
, current_file
,
1183 current_file
, current
);
1190 static int ep_create_wakeup_source(struct epitem
*epi
)
1193 struct wakeup_source
*ws
;
1196 epi
->ep
->ws
= wakeup_source_register("eventpoll");
1201 name
= epi
->ffd
.file
->f_path
.dentry
->d_name
.name
;
1202 ws
= wakeup_source_register(name
);
1206 rcu_assign_pointer(epi
->ws
, ws
);
1211 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1212 static noinline
void ep_destroy_wakeup_source(struct epitem
*epi
)
1214 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
1216 RCU_INIT_POINTER(epi
->ws
, NULL
);
1219 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1220 * used internally by wakeup_source_remove, too (called by
1221 * wakeup_source_unregister), so we cannot use call_rcu
1224 wakeup_source_unregister(ws
);
1228 * Must be called with "mtx" held.
1230 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
1231 struct file
*tfile
, int fd
)
1233 int error
, revents
, pwake
= 0;
1234 unsigned long flags
;
1237 struct ep_pqueue epq
;
1239 user_watches
= atomic_long_read(&ep
->user
->epoll_watches
);
1240 if (unlikely(user_watches
>= max_user_watches
))
1242 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
1245 /* Item initialization follow here ... */
1246 INIT_LIST_HEAD(&epi
->rdllink
);
1247 INIT_LIST_HEAD(&epi
->fllink
);
1248 INIT_LIST_HEAD(&epi
->pwqlist
);
1250 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
1251 epi
->event
= *event
;
1253 epi
->next
= EP_UNACTIVE_PTR
;
1254 if (epi
->event
.events
& EPOLLWAKEUP
) {
1255 error
= ep_create_wakeup_source(epi
);
1257 goto error_create_wakeup_source
;
1259 RCU_INIT_POINTER(epi
->ws
, NULL
);
1262 /* Initialize the poll table using the queue callback */
1264 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1267 * Attach the item to the poll hooks and get current event bits.
1268 * We can safely use the file* here because its usage count has
1269 * been increased by the caller of this function. Note that after
1270 * this operation completes, the poll callback can start hitting
1273 revents
= ep_item_poll(epi
, &epq
.pt
);
1276 * We have to check if something went wrong during the poll wait queue
1277 * install process. Namely an allocation for a wait queue failed due
1278 * high memory pressure.
1282 goto error_unregister
;
1284 /* Add the current item to the list of active epoll hook for this file */
1285 spin_lock(&tfile
->f_lock
);
1286 list_add_tail(&epi
->fllink
, &tfile
->f_ep_links
);
1287 spin_unlock(&tfile
->f_lock
);
1290 * Add the current item to the RB tree. All RB tree operations are
1291 * protected by "mtx", and ep_insert() is called with "mtx" held.
1293 ep_rbtree_insert(ep
, epi
);
1295 /* now check if we've created too many backpaths */
1297 if (reverse_path_check())
1298 goto error_remove_epi
;
1300 /* We have to drop the new item inside our item list to keep track of it */
1301 spin_lock_irqsave(&ep
->lock
, flags
);
1303 /* If the file is already "ready" we drop it inside the ready list */
1304 if ((revents
& event
->events
) && !ep_is_linked(&epi
->rdllink
)) {
1305 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1306 ep_pm_stay_awake(epi
);
1308 /* Notify waiting tasks that events are available */
1309 if (waitqueue_active(&ep
->wq
))
1310 wake_up_locked(&ep
->wq
);
1311 if (waitqueue_active(&ep
->poll_wait
))
1315 spin_unlock_irqrestore(&ep
->lock
, flags
);
1317 atomic_long_inc(&ep
->user
->epoll_watches
);
1319 /* We have to call this outside the lock */
1321 ep_poll_safewake(&ep
->poll_wait
);
1326 spin_lock(&tfile
->f_lock
);
1327 if (ep_is_linked(&epi
->fllink
))
1328 list_del_init(&epi
->fllink
);
1329 spin_unlock(&tfile
->f_lock
);
1331 rb_erase(&epi
->rbn
, &ep
->rbr
);
1334 ep_unregister_pollwait(ep
, epi
);
1337 * We need to do this because an event could have been arrived on some
1338 * allocated wait queue. Note that we don't care about the ep->ovflist
1339 * list, since that is used/cleaned only inside a section bound by "mtx".
1340 * And ep_insert() is called with "mtx" held.
1342 spin_lock_irqsave(&ep
->lock
, flags
);
1343 if (ep_is_linked(&epi
->rdllink
))
1344 list_del_init(&epi
->rdllink
);
1345 spin_unlock_irqrestore(&ep
->lock
, flags
);
1347 wakeup_source_unregister(ep_wakeup_source(epi
));
1349 error_create_wakeup_source
:
1350 kmem_cache_free(epi_cache
, epi
);
1356 * Modify the interest event mask by dropping an event if the new mask
1357 * has a match in the current file status. Must be called with "mtx" held.
1359 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
, struct epoll_event
*event
)
1362 unsigned int revents
;
1365 init_poll_funcptr(&pt
, NULL
);
1368 * Set the new event interest mask before calling f_op->poll();
1369 * otherwise we might miss an event that happens between the
1370 * f_op->poll() call and the new event set registering.
1372 epi
->event
.events
= event
->events
; /* need barrier below */
1373 epi
->event
.data
= event
->data
; /* protected by mtx */
1374 if (epi
->event
.events
& EPOLLWAKEUP
) {
1375 if (!ep_has_wakeup_source(epi
))
1376 ep_create_wakeup_source(epi
);
1377 } else if (ep_has_wakeup_source(epi
)) {
1378 ep_destroy_wakeup_source(epi
);
1382 * The following barrier has two effects:
1384 * 1) Flush epi changes above to other CPUs. This ensures
1385 * we do not miss events from ep_poll_callback if an
1386 * event occurs immediately after we call f_op->poll().
1387 * We need this because we did not take ep->lock while
1388 * changing epi above (but ep_poll_callback does take
1391 * 2) We also need to ensure we do not miss _past_ events
1392 * when calling f_op->poll(). This barrier also
1393 * pairs with the barrier in wq_has_sleeper (see
1394 * comments for wq_has_sleeper).
1396 * This barrier will now guarantee ep_poll_callback or f_op->poll
1397 * (or both) will notice the readiness of an item.
1402 * Get current event bits. We can safely use the file* here because
1403 * its usage count has been increased by the caller of this function.
1405 revents
= ep_item_poll(epi
, &pt
);
1408 * If the item is "hot" and it is not registered inside the ready
1409 * list, push it inside.
1411 if (revents
& event
->events
) {
1412 spin_lock_irq(&ep
->lock
);
1413 if (!ep_is_linked(&epi
->rdllink
)) {
1414 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1415 ep_pm_stay_awake(epi
);
1417 /* Notify waiting tasks that events are available */
1418 if (waitqueue_active(&ep
->wq
))
1419 wake_up_locked(&ep
->wq
);
1420 if (waitqueue_active(&ep
->poll_wait
))
1423 spin_unlock_irq(&ep
->lock
);
1426 /* We have to call this outside the lock */
1428 ep_poll_safewake(&ep
->poll_wait
);
1433 static int ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1436 struct ep_send_events_data
*esed
= priv
;
1438 unsigned int revents
;
1440 struct epoll_event __user
*uevent
;
1441 struct wakeup_source
*ws
;
1444 init_poll_funcptr(&pt
, NULL
);
1447 * We can loop without lock because we are passed a task private list.
1448 * Items cannot vanish during the loop because ep_scan_ready_list() is
1449 * holding "mtx" during this call.
1451 for (eventcnt
= 0, uevent
= esed
->events
;
1452 !list_empty(head
) && eventcnt
< esed
->maxevents
;) {
1453 epi
= list_first_entry(head
, struct epitem
, rdllink
);
1456 * Activate ep->ws before deactivating epi->ws to prevent
1457 * triggering auto-suspend here (in case we reactive epi->ws
1460 * This could be rearranged to delay the deactivation of epi->ws
1461 * instead, but then epi->ws would temporarily be out of sync
1462 * with ep_is_linked().
1464 ws
= ep_wakeup_source(epi
);
1467 __pm_stay_awake(ep
->ws
);
1471 list_del_init(&epi
->rdllink
);
1473 revents
= ep_item_poll(epi
, &pt
);
1476 * If the event mask intersect the caller-requested one,
1477 * deliver the event to userspace. Again, ep_scan_ready_list()
1478 * is holding "mtx", so no operations coming from userspace
1479 * can change the item.
1482 if (__put_user(revents
, &uevent
->events
) ||
1483 __put_user(epi
->event
.data
, &uevent
->data
)) {
1484 list_add(&epi
->rdllink
, head
);
1485 ep_pm_stay_awake(epi
);
1486 return eventcnt
? eventcnt
: -EFAULT
;
1490 if (epi
->event
.events
& EPOLLONESHOT
)
1491 epi
->event
.events
&= EP_PRIVATE_BITS
;
1492 else if (!(epi
->event
.events
& EPOLLET
)) {
1494 * If this file has been added with Level
1495 * Trigger mode, we need to insert back inside
1496 * the ready list, so that the next call to
1497 * epoll_wait() will check again the events
1498 * availability. At this point, no one can insert
1499 * into ep->rdllist besides us. The epoll_ctl()
1500 * callers are locked out by
1501 * ep_scan_ready_list() holding "mtx" and the
1502 * poll callback will queue them in ep->ovflist.
1504 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1505 ep_pm_stay_awake(epi
);
1513 static int ep_send_events(struct eventpoll
*ep
,
1514 struct epoll_event __user
*events
, int maxevents
)
1516 struct ep_send_events_data esed
;
1518 esed
.maxevents
= maxevents
;
1519 esed
.events
= events
;
1521 return ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
, 0);
1524 static inline struct timespec
ep_set_mstimeout(long ms
)
1526 struct timespec now
, ts
= {
1527 .tv_sec
= ms
/ MSEC_PER_SEC
,
1528 .tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
),
1532 return timespec_add_safe(now
, ts
);
1536 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1539 * @ep: Pointer to the eventpoll context.
1540 * @events: Pointer to the userspace buffer where the ready events should be
1542 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1543 * @timeout: Maximum timeout for the ready events fetch operation, in
1544 * milliseconds. If the @timeout is zero, the function will not block,
1545 * while if the @timeout is less than zero, the function will block
1546 * until at least one event has been retrieved (or an error
1549 * Returns: Returns the number of ready events which have been fetched, or an
1550 * error code, in case of error.
1552 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1553 int maxevents
, long timeout
)
1555 int res
= 0, eavail
, timed_out
= 0;
1556 unsigned long flags
;
1559 ktime_t expires
, *to
= NULL
;
1562 struct timespec end_time
= ep_set_mstimeout(timeout
);
1564 slack
= select_estimate_accuracy(&end_time
);
1566 *to
= timespec_to_ktime(end_time
);
1567 } else if (timeout
== 0) {
1569 * Avoid the unnecessary trip to the wait queue loop, if the
1570 * caller specified a non blocking operation.
1573 spin_lock_irqsave(&ep
->lock
, flags
);
1578 spin_lock_irqsave(&ep
->lock
, flags
);
1580 if (!ep_events_available(ep
)) {
1582 * We don't have any available event to return to the caller.
1583 * We need to sleep here, and we will be wake up by
1584 * ep_poll_callback() when events will become available.
1586 init_waitqueue_entry(&wait
, current
);
1587 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1591 * We don't want to sleep if the ep_poll_callback() sends us
1592 * a wakeup in between. That's why we set the task state
1593 * to TASK_INTERRUPTIBLE before doing the checks.
1595 set_current_state(TASK_INTERRUPTIBLE
);
1596 if (ep_events_available(ep
) || timed_out
)
1598 if (signal_pending(current
)) {
1603 spin_unlock_irqrestore(&ep
->lock
, flags
);
1604 if (!schedule_hrtimeout_range(to
, slack
, HRTIMER_MODE_ABS
))
1607 spin_lock_irqsave(&ep
->lock
, flags
);
1609 __remove_wait_queue(&ep
->wq
, &wait
);
1611 set_current_state(TASK_RUNNING
);
1614 /* Is it worth to try to dig for events ? */
1615 eavail
= ep_events_available(ep
);
1617 spin_unlock_irqrestore(&ep
->lock
, flags
);
1620 * Try to transfer events to user space. In case we get 0 events and
1621 * there's still timeout left over, we go trying again in search of
1624 if (!res
&& eavail
&&
1625 !(res
= ep_send_events(ep
, events
, maxevents
)) && !timed_out
)
1632 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1633 * API, to verify that adding an epoll file inside another
1634 * epoll structure, does not violate the constraints, in
1635 * terms of closed loops, or too deep chains (which can
1636 * result in excessive stack usage).
1638 * @priv: Pointer to the epoll file to be currently checked.
1639 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1640 * data structure pointer.
1641 * @call_nests: Current dept of the @ep_call_nested() call stack.
1643 * Returns: Returns zero if adding the epoll @file inside current epoll
1644 * structure @ep does not violate the constraints, or -1 otherwise.
1646 static int ep_loop_check_proc(void *priv
, void *cookie
, int call_nests
)
1649 struct file
*file
= priv
;
1650 struct eventpoll
*ep
= file
->private_data
;
1651 struct eventpoll
*ep_tovisit
;
1652 struct rb_node
*rbp
;
1655 mutex_lock_nested(&ep
->mtx
, call_nests
+ 1);
1657 list_add(&ep
->visited_list_link
, &visited_list
);
1658 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1659 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1660 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1661 ep_tovisit
= epi
->ffd
.file
->private_data
;
1662 if (ep_tovisit
->visited
)
1664 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1665 ep_loop_check_proc
, epi
->ffd
.file
,
1666 ep_tovisit
, current
);
1671 * If we've reached a file that is not associated with
1672 * an ep, then we need to check if the newly added
1673 * links are going to add too many wakeup paths. We do
1674 * this by adding it to the tfile_check_list, if it's
1675 * not already there, and calling reverse_path_check()
1676 * during ep_insert().
1678 if (list_empty(&epi
->ffd
.file
->f_tfile_llink
))
1679 list_add(&epi
->ffd
.file
->f_tfile_llink
,
1683 mutex_unlock(&ep
->mtx
);
1689 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1690 * another epoll file (represented by @ep) does not create
1691 * closed loops or too deep chains.
1693 * @ep: Pointer to the epoll private data structure.
1694 * @file: Pointer to the epoll file to be checked.
1696 * Returns: Returns zero if adding the epoll @file inside current epoll
1697 * structure @ep does not violate the constraints, or -1 otherwise.
1699 static int ep_loop_check(struct eventpoll
*ep
, struct file
*file
)
1702 struct eventpoll
*ep_cur
, *ep_next
;
1704 ret
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1705 ep_loop_check_proc
, file
, ep
, current
);
1706 /* clear visited list */
1707 list_for_each_entry_safe(ep_cur
, ep_next
, &visited_list
,
1708 visited_list_link
) {
1709 ep_cur
->visited
= 0;
1710 list_del(&ep_cur
->visited_list_link
);
1715 static void clear_tfile_check_list(void)
1719 /* first clear the tfile_check_list */
1720 while (!list_empty(&tfile_check_list
)) {
1721 file
= list_first_entry(&tfile_check_list
, struct file
,
1723 list_del_init(&file
->f_tfile_llink
);
1725 INIT_LIST_HEAD(&tfile_check_list
);
1729 * Open an eventpoll file descriptor.
1731 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
1734 struct eventpoll
*ep
= NULL
;
1737 /* Check the EPOLL_* constant for consistency. */
1738 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
1740 if (flags
& ~EPOLL_CLOEXEC
)
1743 * Create the internal data structure ("struct eventpoll").
1745 error
= ep_alloc(&ep
);
1749 * Creates all the items needed to setup an eventpoll file. That is,
1750 * a file structure and a free file descriptor.
1752 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
1757 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
1758 O_RDWR
| (flags
& O_CLOEXEC
));
1760 error
= PTR_ERR(file
);
1764 fd_install(fd
, file
);
1774 SYSCALL_DEFINE1(epoll_create
, int, size
)
1779 return sys_epoll_create1(0);
1783 * The following function implements the controller interface for
1784 * the eventpoll file that enables the insertion/removal/change of
1785 * file descriptors inside the interest set.
1787 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
1788 struct epoll_event __user
*, event
)
1791 int did_lock_epmutex
= 0;
1792 struct file
*file
, *tfile
;
1793 struct eventpoll
*ep
;
1795 struct epoll_event epds
;
1798 if (ep_op_has_event(op
) &&
1799 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
1802 /* Get the "struct file *" for the eventpoll file */
1808 /* Get the "struct file *" for the target file */
1813 /* The target file descriptor must support poll */
1815 if (!tfile
->f_op
|| !tfile
->f_op
->poll
)
1816 goto error_tgt_fput
;
1818 /* Check if EPOLLWAKEUP is allowed */
1819 if ((epds
.events
& EPOLLWAKEUP
) && !capable(CAP_BLOCK_SUSPEND
))
1820 epds
.events
&= ~EPOLLWAKEUP
;
1823 * We have to check that the file structure underneath the file descriptor
1824 * the user passed to us _is_ an eventpoll file. And also we do not permit
1825 * adding an epoll file descriptor inside itself.
1828 if (file
== tfile
|| !is_file_epoll(file
))
1829 goto error_tgt_fput
;
1832 * At this point it is safe to assume that the "private_data" contains
1833 * our own data structure.
1835 ep
= file
->private_data
;
1838 * When we insert an epoll file descriptor, inside another epoll file
1839 * descriptor, there is the change of creating closed loops, which are
1840 * better be handled here, than in more critical paths. While we are
1841 * checking for loops we also determine the list of files reachable
1842 * and hang them on the tfile_check_list, so we can check that we
1843 * haven't created too many possible wakeup paths.
1845 * We need to hold the epmutex across both ep_insert and ep_remove
1846 * b/c we want to make sure we are looking at a coherent view of
1849 if (op
== EPOLL_CTL_ADD
|| op
== EPOLL_CTL_DEL
) {
1850 mutex_lock(&epmutex
);
1851 did_lock_epmutex
= 1;
1853 if (op
== EPOLL_CTL_ADD
) {
1854 if (is_file_epoll(tfile
)) {
1856 if (ep_loop_check(ep
, tfile
) != 0) {
1857 clear_tfile_check_list();
1858 goto error_tgt_fput
;
1861 list_add(&tfile
->f_tfile_llink
, &tfile_check_list
);
1864 mutex_lock_nested(&ep
->mtx
, 0);
1867 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1868 * above, we can be sure to be able to use the item looked up by
1869 * ep_find() till we release the mutex.
1871 epi
= ep_find(ep
, tfile
, fd
);
1877 epds
.events
|= POLLERR
| POLLHUP
;
1878 error
= ep_insert(ep
, &epds
, tfile
, fd
);
1881 clear_tfile_check_list();
1885 error
= ep_remove(ep
, epi
);
1891 epds
.events
|= POLLERR
| POLLHUP
;
1892 error
= ep_modify(ep
, epi
, &epds
);
1897 mutex_unlock(&ep
->mtx
);
1900 if (did_lock_epmutex
)
1901 mutex_unlock(&epmutex
);
1912 * Implement the event wait interface for the eventpoll file. It is the kernel
1913 * part of the user space epoll_wait(2).
1915 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
1916 int, maxevents
, int, timeout
)
1920 struct eventpoll
*ep
;
1922 /* The maximum number of event must be greater than zero */
1923 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
1926 /* Verify that the area passed by the user is writeable */
1927 if (!access_ok(VERIFY_WRITE
, events
, maxevents
* sizeof(struct epoll_event
)))
1930 /* Get the "struct file *" for the eventpoll file */
1936 * We have to check that the file structure underneath the fd
1937 * the user passed to us _is_ an eventpoll file.
1940 if (!is_file_epoll(f
.file
))
1944 * At this point it is safe to assume that the "private_data" contains
1945 * our own data structure.
1947 ep
= f
.file
->private_data
;
1949 /* Time to fish for events ... */
1950 error
= ep_poll(ep
, events
, maxevents
, timeout
);
1958 * Implement the event wait interface for the eventpoll file. It is the kernel
1959 * part of the user space epoll_pwait(2).
1961 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
1962 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
1966 sigset_t ksigmask
, sigsaved
;
1969 * If the caller wants a certain signal mask to be set during the wait,
1973 if (sigsetsize
!= sizeof(sigset_t
))
1975 if (copy_from_user(&ksigmask
, sigmask
, sizeof(ksigmask
)))
1977 sigdelsetmask(&ksigmask
, sigmask(SIGKILL
) | sigmask(SIGSTOP
));
1978 sigprocmask(SIG_SETMASK
, &ksigmask
, &sigsaved
);
1981 error
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
1984 * If we changed the signal mask, we need to restore the original one.
1985 * In case we've got a signal while waiting, we do not restore the
1986 * signal mask yet, and we allow do_signal() to deliver the signal on
1987 * the way back to userspace, before the signal mask is restored.
1990 if (error
== -EINTR
) {
1991 memcpy(¤t
->saved_sigmask
, &sigsaved
,
1993 set_restore_sigmask();
1995 sigprocmask(SIG_SETMASK
, &sigsaved
, NULL
);
2001 static int __init
eventpoll_init(void)
2007 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2009 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2011 BUG_ON(max_user_watches
< 0);
2014 * Initialize the structure used to perform epoll file descriptor
2015 * inclusion loops checks.
2017 ep_nested_calls_init(&poll_loop_ncalls
);
2019 /* Initialize the structure used to perform safe poll wait head wake ups */
2020 ep_nested_calls_init(&poll_safewake_ncalls
);
2022 /* Initialize the structure used to perform file's f_op->poll() calls */
2023 ep_nested_calls_init(&poll_readywalk_ncalls
);
2026 * We can have many thousands of epitems, so prevent this from
2027 * using an extra cache line on 64-bit (and smaller) CPUs
2029 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2031 /* Allocates slab cache used to allocate "struct epitem" items */
2032 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2033 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
2035 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2036 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2037 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
, NULL
);
2041 fs_initcall(eventpoll_init
);