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 <linux/freezer.h>
38 #include <asm/uaccess.h>
41 #include <linux/atomic.h>
42 #include <linux/proc_fs.h>
43 #include <linux/seq_file.h>
44 #include <linux/compat.h>
45 #include <linux/rculist.h>
49 * There are three level of locking required by epoll :
53 * 3) ep->lock (spinlock)
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
63 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
67 * This mutex is acquired by ep_free() during the epoll file
68 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
78 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
87 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
90 * Events that require holding "epmutex" are very rare, while for
91 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
95 /* Epoll private bits inside the event mask */
96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EP_MAX_NESTS 4
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
103 #define EP_UNACTIVE_PTR ((void *) -1L)
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
107 struct epoll_filefd
{
113 * Structure used to track possible nested calls, for too deep recursions
116 struct nested_call_node
{
117 struct list_head llink
;
123 * This structure is used as collector for nested calls, to check for
124 * maximum recursion dept and loop cycles.
126 struct nested_calls
{
127 struct list_head tasks_call_list
;
132 * Each file descriptor added to the eventpoll interface will
133 * have an entry of this type linked to the "rbr" RB tree.
134 * Avoid increasing the size of this struct, there can be many thousands
135 * of these on a server and we do not want this to take another cache line.
139 /* RB tree node links this structure to the eventpoll RB tree */
141 /* Used to free the struct epitem */
145 /* List header used to link this structure to the eventpoll ready list */
146 struct list_head rdllink
;
149 * Works together "struct eventpoll"->ovflist in keeping the
150 * single linked chain of items.
154 /* The file descriptor information this item refers to */
155 struct epoll_filefd ffd
;
157 /* Number of active wait queue attached to poll operations */
160 /* List containing poll wait queues */
161 struct list_head pwqlist
;
163 /* The "container" of this item */
164 struct eventpoll
*ep
;
166 /* List header used to link this item to the "struct file" items list */
167 struct list_head fllink
;
169 /* wakeup_source used when EPOLLWAKEUP is set */
170 struct wakeup_source __rcu
*ws
;
172 /* The structure that describe the interested events and the source fd */
173 struct epoll_event event
;
177 * This structure is stored inside the "private_data" member of the file
178 * structure and represents the main data structure for the eventpoll
182 /* Protect the access to this structure */
186 * This mutex is used to ensure that files are not removed
187 * while epoll is using them. This is held during the event
188 * collection loop, the file cleanup path, the epoll file exit
189 * code and the ctl operations.
193 /* Wait queue used by sys_epoll_wait() */
194 wait_queue_head_t wq
;
196 /* Wait queue used by file->poll() */
197 wait_queue_head_t poll_wait
;
199 /* List of ready file descriptors */
200 struct list_head rdllist
;
202 /* RB tree root used to store monitored fd structs */
206 * This is a single linked list that chains all the "struct epitem" that
207 * happened while transferring ready events to userspace w/out
210 struct epitem
*ovflist
;
212 /* wakeup_source used when ep_scan_ready_list is running */
213 struct wakeup_source
*ws
;
215 /* The user that created the eventpoll descriptor */
216 struct user_struct
*user
;
220 /* used to optimize loop detection check */
222 struct list_head visited_list_link
;
225 /* Wait structure used by the poll hooks */
226 struct eppoll_entry
{
227 /* List header used to link this structure to the "struct epitem" */
228 struct list_head llink
;
230 /* The "base" pointer is set to the container "struct epitem" */
234 * Wait queue item that will be linked to the target file wait
239 /* The wait queue head that linked the "wait" wait queue item */
240 wait_queue_head_t
*whead
;
243 /* Wrapper struct used by poll queueing */
249 /* Used by the ep_send_events() function as callback private data */
250 struct ep_send_events_data
{
252 struct epoll_event __user
*events
;
256 * Configuration options available inside /proc/sys/fs/epoll/
258 /* Maximum number of epoll watched descriptors, per user */
259 static long max_user_watches __read_mostly
;
262 * This mutex is used to serialize ep_free() and eventpoll_release_file().
264 static DEFINE_MUTEX(epmutex
);
266 /* Used to check for epoll file descriptor inclusion loops */
267 static struct nested_calls poll_loop_ncalls
;
269 /* Used for safe wake up implementation */
270 static struct nested_calls poll_safewake_ncalls
;
272 /* Used to call file's f_op->poll() under the nested calls boundaries */
273 static struct nested_calls poll_readywalk_ncalls
;
275 /* Slab cache used to allocate "struct epitem" */
276 static struct kmem_cache
*epi_cache __read_mostly
;
278 /* Slab cache used to allocate "struct eppoll_entry" */
279 static struct kmem_cache
*pwq_cache __read_mostly
;
281 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
282 static LIST_HEAD(visited_list
);
285 * List of files with newly added links, where we may need to limit the number
286 * of emanating paths. Protected by the epmutex.
288 static LIST_HEAD(tfile_check_list
);
292 #include <linux/sysctl.h>
295 static long long_max
= LONG_MAX
;
297 struct ctl_table epoll_table
[] = {
299 .procname
= "max_user_watches",
300 .data
= &max_user_watches
,
301 .maxlen
= sizeof(max_user_watches
),
303 .proc_handler
= proc_doulongvec_minmax
,
309 #endif /* CONFIG_SYSCTL */
311 static const struct file_operations eventpoll_fops
;
313 #if defined(CONFIG_SEC_FD_DETECT)
314 int is_file_epoll(struct file
*f
)
316 static inline int is_file_epoll(struct file
*f
)
319 return f
->f_op
== &eventpoll_fops
;
322 /* Setup the structure that is used as key for the RB tree */
323 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
324 struct file
*file
, int fd
)
330 /* Compare RB tree keys */
331 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
332 struct epoll_filefd
*p2
)
334 return (p1
->file
> p2
->file
? +1:
335 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
338 /* Tells us if the item is currently linked */
339 static inline int ep_is_linked(struct list_head
*p
)
341 return !list_empty(p
);
344 static inline struct eppoll_entry
*ep_pwq_from_wait(wait_queue_t
*p
)
346 return container_of(p
, struct eppoll_entry
, wait
);
349 /* Get the "struct epitem" from a wait queue pointer */
350 static inline struct epitem
*ep_item_from_wait(wait_queue_t
*p
)
352 return container_of(p
, struct eppoll_entry
, wait
)->base
;
355 /* Get the "struct epitem" from an epoll queue wrapper */
356 static inline struct epitem
*ep_item_from_epqueue(poll_table
*p
)
358 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
361 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
362 static inline int ep_op_has_event(int op
)
364 return op
!= EPOLL_CTL_DEL
;
367 /* Initialize the poll safe wake up structure */
368 static void ep_nested_calls_init(struct nested_calls
*ncalls
)
370 INIT_LIST_HEAD(&ncalls
->tasks_call_list
);
371 spin_lock_init(&ncalls
->lock
);
375 * ep_events_available - Checks if ready events might be available.
377 * @ep: Pointer to the eventpoll context.
379 * Returns: Returns a value different than zero if ready events are available,
382 static inline int ep_events_available(struct eventpoll
*ep
)
384 return !list_empty(&ep
->rdllist
) || ep
->ovflist
!= EP_UNACTIVE_PTR
;
388 * ep_call_nested - Perform a bound (possibly) nested call, by checking
389 * that the recursion limit is not exceeded, and that
390 * the same nested call (by the meaning of same cookie) is
393 * @ncalls: Pointer to the nested_calls structure to be used for this call.
394 * @max_nests: Maximum number of allowed nesting calls.
395 * @nproc: Nested call core function pointer.
396 * @priv: Opaque data to be passed to the @nproc callback.
397 * @cookie: Cookie to be used to identify this nested call.
398 * @ctx: This instance context.
400 * Returns: Returns the code returned by the @nproc callback, or -1 if
401 * the maximum recursion limit has been exceeded.
403 static int ep_call_nested(struct nested_calls
*ncalls
, int max_nests
,
404 int (*nproc
)(void *, void *, int), void *priv
,
405 void *cookie
, void *ctx
)
407 int error
, call_nests
= 0;
409 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
410 struct nested_call_node
*tncur
;
411 struct nested_call_node tnode
;
413 spin_lock_irqsave(&ncalls
->lock
, flags
);
416 * Try to see if the current task is already inside this wakeup call.
417 * We use a list here, since the population inside this set is always
420 list_for_each_entry(tncur
, lsthead
, llink
) {
421 if (tncur
->ctx
== ctx
&&
422 (tncur
->cookie
== cookie
|| ++call_nests
> max_nests
)) {
424 * Ops ... loop detected or maximum nest level reached.
425 * We abort this wake by breaking the cycle itself.
432 /* Add the current task and cookie to the list */
434 tnode
.cookie
= cookie
;
435 list_add(&tnode
.llink
, lsthead
);
437 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
439 /* Call the nested function */
440 error
= (*nproc
)(priv
, cookie
, call_nests
);
442 /* Remove the current task from the list */
443 spin_lock_irqsave(&ncalls
->lock
, flags
);
444 list_del(&tnode
.llink
);
446 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
452 * As described in commit 0ccf831cb lockdep: annotate epoll
453 * the use of wait queues used by epoll is done in a very controlled
454 * manner. Wake ups can nest inside each other, but are never done
455 * with the same locking. For example:
458 * efd1 = epoll_create();
459 * efd2 = epoll_create();
460 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
461 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
463 * When a packet arrives to the device underneath "dfd", the net code will
464 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
465 * callback wakeup entry on that queue, and the wake_up() performed by the
466 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
467 * (efd1) notices that it may have some event ready, so it needs to wake up
468 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
469 * that ends up in another wake_up(), after having checked about the
470 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
471 * avoid stack blasting.
473 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
474 * this special case of epoll.
476 #ifdef CONFIG_DEBUG_LOCK_ALLOC
477 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
478 unsigned long events
, int subclass
)
482 spin_lock_irqsave_nested(&wqueue
->lock
, flags
, subclass
);
483 wake_up_locked_poll(wqueue
, events
);
484 spin_unlock_irqrestore(&wqueue
->lock
, flags
);
487 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
488 unsigned long events
, int subclass
)
490 wake_up_poll(wqueue
, events
);
494 static int ep_poll_wakeup_proc(void *priv
, void *cookie
, int call_nests
)
496 ep_wake_up_nested((wait_queue_head_t
*) cookie
, POLLIN
,
502 * Perform a safe wake up of the poll wait list. The problem is that
503 * with the new callback'd wake up system, it is possible that the
504 * poll callback is reentered from inside the call to wake_up() done
505 * on the poll wait queue head. The rule is that we cannot reenter the
506 * wake up code from the same task more than EP_MAX_NESTS times,
507 * and we cannot reenter the same wait queue head at all. This will
508 * enable to have a hierarchy of epoll file descriptor of no more than
511 static void ep_poll_safewake(wait_queue_head_t
*wq
)
513 int this_cpu
= get_cpu();
515 ep_call_nested(&poll_safewake_ncalls
, EP_MAX_NESTS
,
516 ep_poll_wakeup_proc
, NULL
, wq
, (void *) (long) this_cpu
);
521 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
523 wait_queue_head_t
*whead
;
527 * If it is cleared by POLLFREE, it should be rcu-safe.
528 * If we read NULL we need a barrier paired with
529 * smp_store_release() in ep_poll_callback(), otherwise
530 * we rely on whead->lock.
532 whead
= smp_load_acquire(&pwq
->whead
);
534 remove_wait_queue(whead
, &pwq
->wait
);
539 * This function unregisters poll callbacks from the associated file
540 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
543 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
545 struct list_head
*lsthead
= &epi
->pwqlist
;
546 struct eppoll_entry
*pwq
;
548 while (!list_empty(lsthead
)) {
549 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
551 list_del(&pwq
->llink
);
552 ep_remove_wait_queue(pwq
);
553 kmem_cache_free(pwq_cache
, pwq
);
557 /* call only when ep->mtx is held */
558 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
560 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
563 /* call only when ep->mtx is held */
564 static inline void ep_pm_stay_awake(struct epitem
*epi
)
566 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
572 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
574 return rcu_access_pointer(epi
->ws
) ? true : false;
577 /* call when ep->mtx cannot be held (ep_poll_callback) */
578 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
580 struct wakeup_source
*ws
;
583 ws
= rcu_dereference(epi
->ws
);
590 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
591 * the scan code, to call f_op->poll(). Also allows for
592 * O(NumReady) performance.
594 * @ep: Pointer to the epoll private data structure.
595 * @sproc: Pointer to the scan callback.
596 * @priv: Private opaque data passed to the @sproc callback.
597 * @depth: The current depth of recursive f_op->poll calls.
598 * @ep_locked: caller already holds ep->mtx
600 * Returns: The same integer error code returned by the @sproc callback.
602 static int ep_scan_ready_list(struct eventpoll
*ep
,
603 int (*sproc
)(struct eventpoll
*,
604 struct list_head
*, void *),
605 void *priv
, int depth
, bool ep_locked
)
607 int error
, pwake
= 0;
609 struct epitem
*epi
, *nepi
;
613 * We need to lock this because we could be hit by
614 * eventpoll_release_file() and epoll_ctl().
618 mutex_lock_nested(&ep
->mtx
, depth
);
621 * Steal the ready list, and re-init the original one to the
622 * empty list. Also, set ep->ovflist to NULL so that events
623 * happening while looping w/out locks, are not lost. We cannot
624 * have the poll callback to queue directly on ep->rdllist,
625 * because we want the "sproc" callback to be able to do it
628 spin_lock_irqsave(&ep
->lock
, flags
);
629 list_splice_init(&ep
->rdllist
, &txlist
);
631 spin_unlock_irqrestore(&ep
->lock
, flags
);
634 * Now call the callback function.
636 error
= (*sproc
)(ep
, &txlist
, priv
);
638 spin_lock_irqsave(&ep
->lock
, flags
);
640 * During the time we spent inside the "sproc" callback, some
641 * other events might have been queued by the poll callback.
642 * We re-insert them inside the main ready-list here.
644 for (nepi
= ep
->ovflist
; (epi
= nepi
) != NULL
;
645 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
647 * We need to check if the item is already in the list.
648 * During the "sproc" callback execution time, items are
649 * queued into ->ovflist but the "txlist" might already
650 * contain them, and the list_splice() below takes care of them.
652 if (!ep_is_linked(&epi
->rdllink
)) {
653 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
654 ep_pm_stay_awake(epi
);
658 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
659 * releasing the lock, events will be queued in the normal way inside
662 ep
->ovflist
= EP_UNACTIVE_PTR
;
665 * Quickly re-inject items left on "txlist".
667 list_splice(&txlist
, &ep
->rdllist
);
670 if (!list_empty(&ep
->rdllist
)) {
672 * Wake up (if active) both the eventpoll wait list and
673 * the ->poll() wait list (delayed after we release the lock).
675 if (waitqueue_active(&ep
->wq
))
676 wake_up_locked(&ep
->wq
);
677 if (waitqueue_active(&ep
->poll_wait
))
680 spin_unlock_irqrestore(&ep
->lock
, flags
);
683 mutex_unlock(&ep
->mtx
);
685 /* We have to call this outside the lock */
687 ep_poll_safewake(&ep
->poll_wait
);
692 static void epi_rcu_free(struct rcu_head
*head
)
694 struct epitem
*epi
= container_of(head
, struct epitem
, rcu
);
695 kmem_cache_free(epi_cache
, epi
);
699 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
700 * all the associated resources. Must be called with "mtx" held.
702 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
705 struct file
*file
= epi
->ffd
.file
;
708 * Removes poll wait queue hooks. We _have_ to do this without holding
709 * the "ep->lock" otherwise a deadlock might occur. This because of the
710 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
711 * queue head lock when unregistering the wait queue. The wakeup callback
712 * will run by holding the wait queue head lock and will call our callback
713 * that will try to get "ep->lock".
715 ep_unregister_pollwait(ep
, epi
);
717 /* Remove the current item from the list of epoll hooks */
718 spin_lock(&file
->f_lock
);
719 list_del_rcu(&epi
->fllink
);
720 spin_unlock(&file
->f_lock
);
722 rb_erase(&epi
->rbn
, &ep
->rbr
);
724 spin_lock_irqsave(&ep
->lock
, flags
);
725 if (ep_is_linked(&epi
->rdllink
))
726 list_del_init(&epi
->rdllink
);
727 spin_unlock_irqrestore(&ep
->lock
, flags
);
729 wakeup_source_unregister(ep_wakeup_source(epi
));
731 * At this point it is safe to free the eventpoll item. Use the union
732 * field epi->rcu, since we are trying to minimize the size of
733 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
734 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
735 * use of the rbn field.
737 call_rcu(&epi
->rcu
, epi_rcu_free
);
739 atomic_long_dec(&ep
->user
->epoll_watches
);
744 static void ep_free(struct eventpoll
*ep
)
749 /* We need to release all tasks waiting for these file */
750 if (waitqueue_active(&ep
->poll_wait
))
751 ep_poll_safewake(&ep
->poll_wait
);
754 * We need to lock this because we could be hit by
755 * eventpoll_release_file() while we're freeing the "struct eventpoll".
756 * We do not need to hold "ep->mtx" here because the epoll file
757 * is on the way to be removed and no one has references to it
758 * anymore. The only hit might come from eventpoll_release_file() but
759 * holding "epmutex" is sufficient here.
761 mutex_lock(&epmutex
);
764 * Walks through the whole tree by unregistering poll callbacks.
766 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
767 epi
= rb_entry(rbp
, struct epitem
, rbn
);
769 ep_unregister_pollwait(ep
, epi
);
774 * Walks through the whole tree by freeing each "struct epitem". At this
775 * point we are sure no poll callbacks will be lingering around, and also by
776 * holding "epmutex" we can be sure that no file cleanup code will hit
777 * us during this operation. So we can avoid the lock on "ep->lock".
778 * We do not need to lock ep->mtx, either, we only do it to prevent
781 mutex_lock(&ep
->mtx
);
782 while ((rbp
= rb_first(&ep
->rbr
)) != NULL
) {
783 epi
= rb_entry(rbp
, struct epitem
, rbn
);
787 mutex_unlock(&ep
->mtx
);
789 mutex_unlock(&epmutex
);
790 mutex_destroy(&ep
->mtx
);
792 wakeup_source_unregister(ep
->ws
);
796 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
798 struct eventpoll
*ep
= file
->private_data
;
806 static inline unsigned int ep_item_poll(struct epitem
*epi
, poll_table
*pt
)
808 pt
->_key
= epi
->event
.events
;
810 return epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, pt
) & epi
->event
.events
;
813 static int ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
816 struct epitem
*epi
, *tmp
;
819 init_poll_funcptr(&pt
, NULL
);
821 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
822 if (ep_item_poll(epi
, &pt
))
823 return POLLIN
| POLLRDNORM
;
826 * Item has been dropped into the ready list by the poll
827 * callback, but it's not actually ready, as far as
828 * caller requested events goes. We can remove it here.
830 __pm_relax(ep_wakeup_source(epi
));
831 list_del_init(&epi
->rdllink
);
838 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
841 struct readyevents_arg
{
842 struct eventpoll
*ep
;
846 static int ep_poll_readyevents_proc(void *priv
, void *cookie
, int call_nests
)
848 struct readyevents_arg
*arg
= priv
;
850 return ep_scan_ready_list(arg
->ep
, ep_read_events_proc
, NULL
,
851 call_nests
+ 1, arg
->locked
);
854 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
857 struct eventpoll
*ep
= file
->private_data
;
858 struct readyevents_arg arg
;
861 * During ep_insert() we already hold the ep->mtx for the tfile.
862 * Prevent re-aquisition.
864 arg
.locked
= wait
&& (wait
->_qproc
== ep_ptable_queue_proc
);
867 /* Insert inside our poll wait queue */
868 poll_wait(file
, &ep
->poll_wait
, wait
);
871 * Proceed to find out if wanted events are really available inside
872 * the ready list. This need to be done under ep_call_nested()
873 * supervision, since the call to f_op->poll() done on listed files
874 * could re-enter here.
876 pollflags
= ep_call_nested(&poll_readywalk_ncalls
, EP_MAX_NESTS
,
877 ep_poll_readyevents_proc
, &arg
, ep
, current
);
879 return pollflags
!= -1 ? pollflags
: 0;
882 #ifdef CONFIG_PROC_FS
883 static void ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
885 struct eventpoll
*ep
= f
->private_data
;
888 mutex_lock(&ep
->mtx
);
889 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
890 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
892 seq_printf(m
, "tfd: %8d events: %8x data: %16llx\n",
893 epi
->ffd
.fd
, epi
->event
.events
,
894 (long long)epi
->event
.data
);
895 if (seq_has_overflowed(m
))
898 mutex_unlock(&ep
->mtx
);
902 /* File callbacks that implement the eventpoll file behaviour */
903 static const struct file_operations eventpoll_fops
= {
904 #ifdef CONFIG_PROC_FS
905 .show_fdinfo
= ep_show_fdinfo
,
907 .release
= ep_eventpoll_release
,
908 .poll
= ep_eventpoll_poll
,
909 .llseek
= noop_llseek
,
913 * This is called from eventpoll_release() to unlink files from the eventpoll
914 * interface. We need to have this facility to cleanup correctly files that are
915 * closed without being removed from the eventpoll interface.
917 void eventpoll_release_file(struct file
*file
)
919 struct eventpoll
*ep
;
920 struct epitem
*epi
, *next
;
923 * We don't want to get "file->f_lock" because it is not
924 * necessary. It is not necessary because we're in the "struct file"
925 * cleanup path, and this means that no one is using this file anymore.
926 * So, for example, epoll_ctl() cannot hit here since if we reach this
927 * point, the file counter already went to zero and fget() would fail.
928 * The only hit might come from ep_free() but by holding the mutex
929 * will correctly serialize the operation. We do need to acquire
930 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
931 * from anywhere but ep_free().
933 * Besides, ep_remove() acquires the lock, so we can't hold it here.
935 mutex_lock(&epmutex
);
936 list_for_each_entry_safe(epi
, next
, &file
->f_ep_links
, fllink
) {
938 mutex_lock_nested(&ep
->mtx
, 0);
940 mutex_unlock(&ep
->mtx
);
942 mutex_unlock(&epmutex
);
945 static int ep_alloc(struct eventpoll
**pep
)
948 struct user_struct
*user
;
949 struct eventpoll
*ep
;
951 user
= get_current_user();
953 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
957 spin_lock_init(&ep
->lock
);
958 mutex_init(&ep
->mtx
);
959 init_waitqueue_head(&ep
->wq
);
960 init_waitqueue_head(&ep
->poll_wait
);
961 INIT_LIST_HEAD(&ep
->rdllist
);
963 ep
->ovflist
= EP_UNACTIVE_PTR
;
976 * Search the file inside the eventpoll tree. The RB tree operations
977 * are protected by the "mtx" mutex, and ep_find() must be called with
980 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
984 struct epitem
*epi
, *epir
= NULL
;
985 struct epoll_filefd ffd
;
987 ep_set_ffd(&ffd
, file
, fd
);
988 for (rbp
= ep
->rbr
.rb_node
; rbp
; ) {
989 epi
= rb_entry(rbp
, struct epitem
, rbn
);
990 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
1005 * This is the callback that is passed to the wait queue wakeup
1006 * mechanism. It is called by the stored file descriptors when they
1007 * have events to report.
1009 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
1012 unsigned long flags
;
1013 struct epitem
*epi
= ep_item_from_wait(wait
);
1014 struct eventpoll
*ep
= epi
->ep
;
1016 spin_lock_irqsave(&ep
->lock
, flags
);
1019 * If the event mask does not contain any poll(2) event, we consider the
1020 * descriptor to be disabled. This condition is likely the effect of the
1021 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1022 * until the next EPOLL_CTL_MOD will be issued.
1024 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1028 * Check the events coming with the callback. At this stage, not
1029 * every device reports the events in the "key" parameter of the
1030 * callback. We need to be able to handle both cases here, hence the
1031 * test for "key" != NULL before the event match test.
1033 if (key
&& !((unsigned long) key
& epi
->event
.events
))
1037 * If we are transferring events to userspace, we can hold no locks
1038 * (because we're accessing user memory, and because of linux f_op->poll()
1039 * semantics). All the events that happen during that period of time are
1040 * chained in ep->ovflist and requeued later on.
1042 if (unlikely(ep
->ovflist
!= EP_UNACTIVE_PTR
)) {
1043 if (epi
->next
== EP_UNACTIVE_PTR
) {
1044 epi
->next
= ep
->ovflist
;
1048 * Activate ep->ws since epi->ws may get
1049 * deactivated at any time.
1051 __pm_stay_awake(ep
->ws
);
1058 /* If this file is already in the ready list we exit soon */
1059 if (!ep_is_linked(&epi
->rdllink
)) {
1060 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1061 ep_pm_stay_awake_rcu(epi
);
1065 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1068 if (waitqueue_active(&ep
->wq
))
1069 wake_up_locked(&ep
->wq
);
1070 if (waitqueue_active(&ep
->poll_wait
))
1074 spin_unlock_irqrestore(&ep
->lock
, flags
);
1076 /* We have to call this outside the lock */
1078 ep_poll_safewake(&ep
->poll_wait
);
1081 if ((unsigned long)key
& POLLFREE
) {
1083 * If we race with ep_remove_wait_queue() it can miss
1084 * ->whead = NULL and do another remove_wait_queue() after
1085 * us, so we can't use __remove_wait_queue().
1087 list_del_init(&wait
->task_list
);
1089 * ->whead != NULL protects us from the race with ep_free()
1090 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1091 * held by the caller. Once we nullify it, nothing protects
1092 * ep/epi or even wait.
1094 smp_store_release(&ep_pwq_from_wait(wait
)->whead
, NULL
);
1101 * This is the callback that is used to add our wait queue to the
1102 * target file wakeup lists.
1104 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1107 struct epitem
*epi
= ep_item_from_epqueue(pt
);
1108 struct eppoll_entry
*pwq
;
1110 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
1111 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1114 add_wait_queue(whead
, &pwq
->wait
);
1115 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
1118 /* We have to signal that an error occurred */
1123 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1126 struct rb_node
**p
= &ep
->rbr
.rb_node
, *parent
= NULL
;
1127 struct epitem
*epic
;
1131 epic
= rb_entry(parent
, struct epitem
, rbn
);
1132 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1134 p
= &parent
->rb_right
;
1136 p
= &parent
->rb_left
;
1138 rb_link_node(&epi
->rbn
, parent
, p
);
1139 rb_insert_color(&epi
->rbn
, &ep
->rbr
);
1144 #define PATH_ARR_SIZE 5
1146 * These are the number paths of length 1 to 5, that we are allowing to emanate
1147 * from a single file of interest. For example, we allow 1000 paths of length
1148 * 1, to emanate from each file of interest. This essentially represents the
1149 * potential wakeup paths, which need to be limited in order to avoid massive
1150 * uncontrolled wakeup storms. The common use case should be a single ep which
1151 * is connected to n file sources. In this case each file source has 1 path
1152 * of length 1. Thus, the numbers below should be more than sufficient. These
1153 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1154 * and delete can't add additional paths. Protected by the epmutex.
1156 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1157 static int path_count
[PATH_ARR_SIZE
];
1159 static int path_count_inc(int nests
)
1161 /* Allow an arbitrary number of depth 1 paths */
1165 if (++path_count
[nests
] > path_limits
[nests
])
1170 static void path_count_init(void)
1174 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1178 static int reverse_path_check_proc(void *priv
, void *cookie
, int call_nests
)
1181 struct file
*file
= priv
;
1182 struct file
*child_file
;
1185 /* CTL_DEL can remove links here, but that can't increase our count */
1187 list_for_each_entry_rcu(epi
, &file
->f_ep_links
, fllink
) {
1188 child_file
= epi
->ep
->file
;
1189 if (is_file_epoll(child_file
)) {
1190 if (list_empty(&child_file
->f_ep_links
)) {
1191 if (path_count_inc(call_nests
)) {
1196 error
= ep_call_nested(&poll_loop_ncalls
,
1198 reverse_path_check_proc
,
1199 child_file
, child_file
,
1205 printk(KERN_ERR
"reverse_path_check_proc: "
1206 "file is not an ep!\n");
1214 * reverse_path_check - The tfile_check_list is list of file *, which have
1215 * links that are proposed to be newly added. We need to
1216 * make sure that those added links don't add too many
1217 * paths such that we will spend all our time waking up
1218 * eventpoll objects.
1220 * Returns: Returns zero if the proposed links don't create too many paths,
1223 static int reverse_path_check(void)
1226 struct file
*current_file
;
1228 /* let's call this for all tfiles */
1229 list_for_each_entry(current_file
, &tfile_check_list
, f_tfile_llink
) {
1231 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1232 reverse_path_check_proc
, current_file
,
1233 current_file
, current
);
1240 static int ep_create_wakeup_source(struct epitem
*epi
)
1243 struct wakeup_source
*ws
;
1246 epi
->ep
->ws
= wakeup_source_register("eventpoll");
1251 name
= epi
->ffd
.file
->f_path
.dentry
->d_name
.name
;
1252 ws
= wakeup_source_register(name
);
1256 rcu_assign_pointer(epi
->ws
, ws
);
1261 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1262 static noinline
void ep_destroy_wakeup_source(struct epitem
*epi
)
1264 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
1266 RCU_INIT_POINTER(epi
->ws
, NULL
);
1269 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1270 * used internally by wakeup_source_remove, too (called by
1271 * wakeup_source_unregister), so we cannot use call_rcu
1274 wakeup_source_unregister(ws
);
1278 * Must be called with "mtx" held.
1280 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
1281 struct file
*tfile
, int fd
, int full_check
)
1283 int error
, revents
, pwake
= 0;
1284 unsigned long flags
;
1287 struct ep_pqueue epq
;
1289 user_watches
= atomic_long_read(&ep
->user
->epoll_watches
);
1290 if (unlikely(user_watches
>= max_user_watches
))
1292 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
1295 /* Item initialization follow here ... */
1296 INIT_LIST_HEAD(&epi
->rdllink
);
1297 INIT_LIST_HEAD(&epi
->fllink
);
1298 INIT_LIST_HEAD(&epi
->pwqlist
);
1300 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
1301 epi
->event
= *event
;
1303 epi
->next
= EP_UNACTIVE_PTR
;
1304 if (epi
->event
.events
& EPOLLWAKEUP
) {
1305 error
= ep_create_wakeup_source(epi
);
1307 goto error_create_wakeup_source
;
1309 RCU_INIT_POINTER(epi
->ws
, NULL
);
1312 /* Initialize the poll table using the queue callback */
1314 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1317 * Attach the item to the poll hooks and get current event bits.
1318 * We can safely use the file* here because its usage count has
1319 * been increased by the caller of this function. Note that after
1320 * this operation completes, the poll callback can start hitting
1323 revents
= ep_item_poll(epi
, &epq
.pt
);
1326 * We have to check if something went wrong during the poll wait queue
1327 * install process. Namely an allocation for a wait queue failed due
1328 * high memory pressure.
1332 goto error_unregister
;
1334 /* Add the current item to the list of active epoll hook for this file */
1335 spin_lock(&tfile
->f_lock
);
1336 list_add_tail_rcu(&epi
->fllink
, &tfile
->f_ep_links
);
1337 spin_unlock(&tfile
->f_lock
);
1340 * Add the current item to the RB tree. All RB tree operations are
1341 * protected by "mtx", and ep_insert() is called with "mtx" held.
1343 ep_rbtree_insert(ep
, epi
);
1345 /* now check if we've created too many backpaths */
1347 if (full_check
&& reverse_path_check())
1348 goto error_remove_epi
;
1350 /* We have to drop the new item inside our item list to keep track of it */
1351 spin_lock_irqsave(&ep
->lock
, flags
);
1353 /* If the file is already "ready" we drop it inside the ready list */
1354 if ((revents
& event
->events
) && !ep_is_linked(&epi
->rdllink
)) {
1355 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1356 ep_pm_stay_awake(epi
);
1358 /* Notify waiting tasks that events are available */
1359 if (waitqueue_active(&ep
->wq
))
1360 wake_up_locked(&ep
->wq
);
1361 if (waitqueue_active(&ep
->poll_wait
))
1365 spin_unlock_irqrestore(&ep
->lock
, flags
);
1367 atomic_long_inc(&ep
->user
->epoll_watches
);
1369 /* We have to call this outside the lock */
1371 ep_poll_safewake(&ep
->poll_wait
);
1376 spin_lock(&tfile
->f_lock
);
1377 list_del_rcu(&epi
->fllink
);
1378 spin_unlock(&tfile
->f_lock
);
1380 rb_erase(&epi
->rbn
, &ep
->rbr
);
1383 ep_unregister_pollwait(ep
, epi
);
1386 * We need to do this because an event could have been arrived on some
1387 * allocated wait queue. Note that we don't care about the ep->ovflist
1388 * list, since that is used/cleaned only inside a section bound by "mtx".
1389 * And ep_insert() is called with "mtx" held.
1391 spin_lock_irqsave(&ep
->lock
, flags
);
1392 if (ep_is_linked(&epi
->rdllink
))
1393 list_del_init(&epi
->rdllink
);
1394 spin_unlock_irqrestore(&ep
->lock
, flags
);
1396 wakeup_source_unregister(ep_wakeup_source(epi
));
1398 error_create_wakeup_source
:
1399 kmem_cache_free(epi_cache
, epi
);
1405 * Modify the interest event mask by dropping an event if the new mask
1406 * has a match in the current file status. Must be called with "mtx" held.
1408 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
, struct epoll_event
*event
)
1411 unsigned int revents
;
1414 init_poll_funcptr(&pt
, NULL
);
1417 * Set the new event interest mask before calling f_op->poll();
1418 * otherwise we might miss an event that happens between the
1419 * f_op->poll() call and the new event set registering.
1421 epi
->event
.events
= event
->events
; /* need barrier below */
1422 epi
->event
.data
= event
->data
; /* protected by mtx */
1423 if (epi
->event
.events
& EPOLLWAKEUP
) {
1424 if (!ep_has_wakeup_source(epi
))
1425 ep_create_wakeup_source(epi
);
1426 } else if (ep_has_wakeup_source(epi
)) {
1427 ep_destroy_wakeup_source(epi
);
1431 * The following barrier has two effects:
1433 * 1) Flush epi changes above to other CPUs. This ensures
1434 * we do not miss events from ep_poll_callback if an
1435 * event occurs immediately after we call f_op->poll().
1436 * We need this because we did not take ep->lock while
1437 * changing epi above (but ep_poll_callback does take
1440 * 2) We also need to ensure we do not miss _past_ events
1441 * when calling f_op->poll(). This barrier also
1442 * pairs with the barrier in wq_has_sleeper (see
1443 * comments for wq_has_sleeper).
1445 * This barrier will now guarantee ep_poll_callback or f_op->poll
1446 * (or both) will notice the readiness of an item.
1451 * Get current event bits. We can safely use the file* here because
1452 * its usage count has been increased by the caller of this function.
1454 revents
= ep_item_poll(epi
, &pt
);
1457 * If the item is "hot" and it is not registered inside the ready
1458 * list, push it inside.
1460 if (revents
& event
->events
) {
1461 spin_lock_irq(&ep
->lock
);
1462 if (!ep_is_linked(&epi
->rdllink
)) {
1463 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1464 ep_pm_stay_awake(epi
);
1466 /* Notify waiting tasks that events are available */
1467 if (waitqueue_active(&ep
->wq
))
1468 wake_up_locked(&ep
->wq
);
1469 if (waitqueue_active(&ep
->poll_wait
))
1472 spin_unlock_irq(&ep
->lock
);
1475 /* We have to call this outside the lock */
1477 ep_poll_safewake(&ep
->poll_wait
);
1482 static int ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1485 struct ep_send_events_data
*esed
= priv
;
1487 unsigned int revents
;
1489 struct epoll_event __user
*uevent
;
1490 struct wakeup_source
*ws
;
1493 init_poll_funcptr(&pt
, NULL
);
1496 * We can loop without lock because we are passed a task private list.
1497 * Items cannot vanish during the loop because ep_scan_ready_list() is
1498 * holding "mtx" during this call.
1500 for (eventcnt
= 0, uevent
= esed
->events
;
1501 !list_empty(head
) && eventcnt
< esed
->maxevents
;) {
1502 epi
= list_first_entry(head
, struct epitem
, rdllink
);
1505 * Activate ep->ws before deactivating epi->ws to prevent
1506 * triggering auto-suspend here (in case we reactive epi->ws
1509 * This could be rearranged to delay the deactivation of epi->ws
1510 * instead, but then epi->ws would temporarily be out of sync
1511 * with ep_is_linked().
1513 ws
= ep_wakeup_source(epi
);
1516 __pm_stay_awake(ep
->ws
);
1520 list_del_init(&epi
->rdllink
);
1522 revents
= ep_item_poll(epi
, &pt
);
1525 * If the event mask intersect the caller-requested one,
1526 * deliver the event to userspace. Again, ep_scan_ready_list()
1527 * is holding "mtx", so no operations coming from userspace
1528 * can change the item.
1531 if (__put_user(revents
, &uevent
->events
) ||
1532 __put_user(epi
->event
.data
, &uevent
->data
)) {
1533 list_add(&epi
->rdllink
, head
);
1534 ep_pm_stay_awake(epi
);
1535 return eventcnt
? eventcnt
: -EFAULT
;
1539 if (epi
->event
.events
& EPOLLONESHOT
)
1540 epi
->event
.events
&= EP_PRIVATE_BITS
;
1541 else if (!(epi
->event
.events
& EPOLLET
)) {
1543 * If this file has been added with Level
1544 * Trigger mode, we need to insert back inside
1545 * the ready list, so that the next call to
1546 * epoll_wait() will check again the events
1547 * availability. At this point, no one can insert
1548 * into ep->rdllist besides us. The epoll_ctl()
1549 * callers are locked out by
1550 * ep_scan_ready_list() holding "mtx" and the
1551 * poll callback will queue them in ep->ovflist.
1553 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1554 ep_pm_stay_awake(epi
);
1562 static int ep_send_events(struct eventpoll
*ep
,
1563 struct epoll_event __user
*events
, int maxevents
)
1565 struct ep_send_events_data esed
;
1567 esed
.maxevents
= maxevents
;
1568 esed
.events
= events
;
1570 return ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
, 0, false);
1573 static inline struct timespec
ep_set_mstimeout(long ms
)
1575 struct timespec now
, ts
= {
1576 .tv_sec
= ms
/ MSEC_PER_SEC
,
1577 .tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
),
1581 return timespec_add_safe(now
, ts
);
1585 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1588 * @ep: Pointer to the eventpoll context.
1589 * @events: Pointer to the userspace buffer where the ready events should be
1591 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1592 * @timeout: Maximum timeout for the ready events fetch operation, in
1593 * milliseconds. If the @timeout is zero, the function will not block,
1594 * while if the @timeout is less than zero, the function will block
1595 * until at least one event has been retrieved (or an error
1598 * Returns: Returns the number of ready events which have been fetched, or an
1599 * error code, in case of error.
1601 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1602 int maxevents
, long timeout
)
1604 int res
= 0, eavail
, timed_out
= 0;
1605 unsigned long flags
;
1608 ktime_t expires
, *to
= NULL
;
1611 struct timespec end_time
= ep_set_mstimeout(timeout
);
1613 slack
= select_estimate_accuracy(&end_time
);
1615 *to
= timespec_to_ktime(end_time
);
1616 } else if (timeout
== 0) {
1618 * Avoid the unnecessary trip to the wait queue loop, if the
1619 * caller specified a non blocking operation.
1622 spin_lock_irqsave(&ep
->lock
, flags
);
1627 spin_lock_irqsave(&ep
->lock
, flags
);
1629 if (!ep_events_available(ep
)) {
1631 * We don't have any available event to return to the caller.
1632 * We need to sleep here, and we will be wake up by
1633 * ep_poll_callback() when events will become available.
1635 init_waitqueue_entry(&wait
, current
);
1636 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1640 * We don't want to sleep if the ep_poll_callback() sends us
1641 * a wakeup in between. That's why we set the task state
1642 * to TASK_INTERRUPTIBLE before doing the checks.
1644 set_current_state(TASK_INTERRUPTIBLE
);
1645 if (ep_events_available(ep
) || timed_out
)
1647 if (signal_pending(current
)) {
1652 spin_unlock_irqrestore(&ep
->lock
, flags
);
1653 if (!freezable_schedule_hrtimeout_range(to
, slack
,
1657 spin_lock_irqsave(&ep
->lock
, flags
);
1660 __remove_wait_queue(&ep
->wq
, &wait
);
1661 __set_current_state(TASK_RUNNING
);
1664 /* Is it worth to try to dig for events ? */
1665 eavail
= ep_events_available(ep
);
1667 spin_unlock_irqrestore(&ep
->lock
, flags
);
1670 * Try to transfer events to user space. In case we get 0 events and
1671 * there's still timeout left over, we go trying again in search of
1674 if (!res
&& eavail
&&
1675 !(res
= ep_send_events(ep
, events
, maxevents
)) && !timed_out
)
1682 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1683 * API, to verify that adding an epoll file inside another
1684 * epoll structure, does not violate the constraints, in
1685 * terms of closed loops, or too deep chains (which can
1686 * result in excessive stack usage).
1688 * @priv: Pointer to the epoll file to be currently checked.
1689 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1690 * data structure pointer.
1691 * @call_nests: Current dept of the @ep_call_nested() call stack.
1693 * Returns: Returns zero if adding the epoll @file inside current epoll
1694 * structure @ep does not violate the constraints, or -1 otherwise.
1696 static int ep_loop_check_proc(void *priv
, void *cookie
, int call_nests
)
1699 struct file
*file
= priv
;
1700 struct eventpoll
*ep
= file
->private_data
;
1701 struct eventpoll
*ep_tovisit
;
1702 struct rb_node
*rbp
;
1705 mutex_lock_nested(&ep
->mtx
, call_nests
+ 1);
1707 list_add(&ep
->visited_list_link
, &visited_list
);
1708 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1709 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1710 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1711 ep_tovisit
= epi
->ffd
.file
->private_data
;
1712 if (ep_tovisit
->visited
)
1714 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1715 ep_loop_check_proc
, epi
->ffd
.file
,
1716 ep_tovisit
, current
);
1721 * If we've reached a file that is not associated with
1722 * an ep, then we need to check if the newly added
1723 * links are going to add too many wakeup paths. We do
1724 * this by adding it to the tfile_check_list, if it's
1725 * not already there, and calling reverse_path_check()
1726 * during ep_insert().
1728 if (list_empty(&epi
->ffd
.file
->f_tfile_llink
))
1729 list_add(&epi
->ffd
.file
->f_tfile_llink
,
1733 mutex_unlock(&ep
->mtx
);
1739 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1740 * another epoll file (represented by @ep) does not create
1741 * closed loops or too deep chains.
1743 * @ep: Pointer to the epoll private data structure.
1744 * @file: Pointer to the epoll file to be checked.
1746 * Returns: Returns zero if adding the epoll @file inside current epoll
1747 * structure @ep does not violate the constraints, or -1 otherwise.
1749 static int ep_loop_check(struct eventpoll
*ep
, struct file
*file
)
1752 struct eventpoll
*ep_cur
, *ep_next
;
1754 ret
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1755 ep_loop_check_proc
, file
, ep
, current
);
1756 /* clear visited list */
1757 list_for_each_entry_safe(ep_cur
, ep_next
, &visited_list
,
1758 visited_list_link
) {
1759 ep_cur
->visited
= 0;
1760 list_del(&ep_cur
->visited_list_link
);
1765 static void clear_tfile_check_list(void)
1769 /* first clear the tfile_check_list */
1770 while (!list_empty(&tfile_check_list
)) {
1771 file
= list_first_entry(&tfile_check_list
, struct file
,
1773 list_del_init(&file
->f_tfile_llink
);
1775 INIT_LIST_HEAD(&tfile_check_list
);
1779 * Open an eventpoll file descriptor.
1781 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
1784 struct eventpoll
*ep
= NULL
;
1787 /* Check the EPOLL_* constant for consistency. */
1788 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
1790 if (flags
& ~EPOLL_CLOEXEC
)
1793 * Create the internal data structure ("struct eventpoll").
1795 error
= ep_alloc(&ep
);
1799 * Creates all the items needed to setup an eventpoll file. That is,
1800 * a file structure and a free file descriptor.
1802 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
1807 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
1808 O_RDWR
| (flags
& O_CLOEXEC
));
1810 error
= PTR_ERR(file
);
1814 fd_install(fd
, file
);
1824 SYSCALL_DEFINE1(epoll_create
, int, size
)
1829 return sys_epoll_create1(0);
1833 * The following function implements the controller interface for
1834 * the eventpoll file that enables the insertion/removal/change of
1835 * file descriptors inside the interest set.
1837 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
1838 struct epoll_event __user
*, event
)
1843 struct eventpoll
*ep
;
1845 struct epoll_event epds
;
1846 struct eventpoll
*tep
= NULL
;
1849 if (ep_op_has_event(op
) &&
1850 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
1858 /* Get the "struct file *" for the target file */
1863 /* The target file descriptor must support poll */
1865 if (!tf
.file
->f_op
->poll
)
1866 goto error_tgt_fput
;
1868 /* Check if EPOLLWAKEUP is allowed */
1869 if (ep_op_has_event(op
))
1870 ep_take_care_of_epollwakeup(&epds
);
1873 * We have to check that the file structure underneath the file descriptor
1874 * the user passed to us _is_ an eventpoll file. And also we do not permit
1875 * adding an epoll file descriptor inside itself.
1878 if (f
.file
== tf
.file
|| !is_file_epoll(f
.file
))
1879 goto error_tgt_fput
;
1882 * At this point it is safe to assume that the "private_data" contains
1883 * our own data structure.
1885 ep
= f
.file
->private_data
;
1888 * When we insert an epoll file descriptor, inside another epoll file
1889 * descriptor, there is the change of creating closed loops, which are
1890 * better be handled here, than in more critical paths. While we are
1891 * checking for loops we also determine the list of files reachable
1892 * and hang them on the tfile_check_list, so we can check that we
1893 * haven't created too many possible wakeup paths.
1895 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1896 * the epoll file descriptor is attaching directly to a wakeup source,
1897 * unless the epoll file descriptor is nested. The purpose of taking the
1898 * 'epmutex' on add is to prevent complex toplogies such as loops and
1899 * deep wakeup paths from forming in parallel through multiple
1900 * EPOLL_CTL_ADD operations.
1902 mutex_lock_nested(&ep
->mtx
, 0);
1903 if (op
== EPOLL_CTL_ADD
) {
1904 if (!list_empty(&f
.file
->f_ep_links
) ||
1905 is_file_epoll(tf
.file
)) {
1907 mutex_unlock(&ep
->mtx
);
1908 mutex_lock(&epmutex
);
1909 if (is_file_epoll(tf
.file
)) {
1911 if (ep_loop_check(ep
, tf
.file
) != 0) {
1912 clear_tfile_check_list();
1913 goto error_tgt_fput
;
1916 list_add(&tf
.file
->f_tfile_llink
,
1918 mutex_lock_nested(&ep
->mtx
, 0);
1919 if (is_file_epoll(tf
.file
)) {
1920 tep
= tf
.file
->private_data
;
1921 mutex_lock_nested(&tep
->mtx
, 1);
1927 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1928 * above, we can be sure to be able to use the item looked up by
1929 * ep_find() till we release the mutex.
1931 epi
= ep_find(ep
, tf
.file
, fd
);
1937 epds
.events
|= POLLERR
| POLLHUP
;
1938 error
= ep_insert(ep
, &epds
, tf
.file
, fd
, full_check
);
1942 clear_tfile_check_list();
1946 error
= ep_remove(ep
, epi
);
1952 epds
.events
|= POLLERR
| POLLHUP
;
1953 error
= ep_modify(ep
, epi
, &epds
);
1959 mutex_unlock(&tep
->mtx
);
1960 mutex_unlock(&ep
->mtx
);
1964 mutex_unlock(&epmutex
);
1975 * Implement the event wait interface for the eventpoll file. It is the kernel
1976 * part of the user space epoll_wait(2).
1978 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
1979 int, maxevents
, int, timeout
)
1983 struct eventpoll
*ep
;
1985 /* The maximum number of event must be greater than zero */
1986 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
1989 /* Verify that the area passed by the user is writeable */
1990 if (!access_ok(VERIFY_WRITE
, events
, maxevents
* sizeof(struct epoll_event
)))
1993 /* Get the "struct file *" for the eventpoll file */
1999 * We have to check that the file structure underneath the fd
2000 * the user passed to us _is_ an eventpoll file.
2003 if (!is_file_epoll(f
.file
))
2007 * At this point it is safe to assume that the "private_data" contains
2008 * our own data structure.
2010 ep
= f
.file
->private_data
;
2012 /* Time to fish for events ... */
2013 error
= ep_poll(ep
, events
, maxevents
, timeout
);
2021 * Implement the event wait interface for the eventpoll file. It is the kernel
2022 * part of the user space epoll_pwait(2).
2024 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
2025 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
2029 sigset_t ksigmask
, sigsaved
;
2032 * If the caller wants a certain signal mask to be set during the wait,
2036 if (sigsetsize
!= sizeof(sigset_t
))
2038 if (copy_from_user(&ksigmask
, sigmask
, sizeof(ksigmask
)))
2040 sigsaved
= current
->blocked
;
2041 set_current_blocked(&ksigmask
);
2044 error
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
2047 * If we changed the signal mask, we need to restore the original one.
2048 * In case we've got a signal while waiting, we do not restore the
2049 * signal mask yet, and we allow do_signal() to deliver the signal on
2050 * the way back to userspace, before the signal mask is restored.
2053 if (error
== -EINTR
) {
2054 memcpy(¤t
->saved_sigmask
, &sigsaved
,
2056 set_restore_sigmask();
2058 set_current_blocked(&sigsaved
);
2064 #ifdef CONFIG_COMPAT
2065 COMPAT_SYSCALL_DEFINE6(epoll_pwait
, int, epfd
,
2066 struct epoll_event __user
*, events
,
2067 int, maxevents
, int, timeout
,
2068 const compat_sigset_t __user
*, sigmask
,
2069 compat_size_t
, sigsetsize
)
2072 compat_sigset_t csigmask
;
2073 sigset_t ksigmask
, sigsaved
;
2076 * If the caller wants a certain signal mask to be set during the wait,
2080 if (sigsetsize
!= sizeof(compat_sigset_t
))
2082 if (copy_from_user(&csigmask
, sigmask
, sizeof(csigmask
)))
2084 sigset_from_compat(&ksigmask
, &csigmask
);
2085 sigsaved
= current
->blocked
;
2086 set_current_blocked(&ksigmask
);
2089 err
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
2092 * If we changed the signal mask, we need to restore the original one.
2093 * In case we've got a signal while waiting, we do not restore the
2094 * signal mask yet, and we allow do_signal() to deliver the signal on
2095 * the way back to userspace, before the signal mask is restored.
2098 if (err
== -EINTR
) {
2099 memcpy(¤t
->saved_sigmask
, &sigsaved
,
2101 set_restore_sigmask();
2103 set_current_blocked(&sigsaved
);
2110 static int __init
eventpoll_init(void)
2116 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2118 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2120 BUG_ON(max_user_watches
< 0);
2123 * Initialize the structure used to perform epoll file descriptor
2124 * inclusion loops checks.
2126 ep_nested_calls_init(&poll_loop_ncalls
);
2128 /* Initialize the structure used to perform safe poll wait head wake ups */
2129 ep_nested_calls_init(&poll_safewake_ncalls
);
2131 /* Initialize the structure used to perform file's f_op->poll() calls */
2132 ep_nested_calls_init(&poll_readywalk_ncalls
);
2135 * We can have many thousands of epitems, so prevent this from
2136 * using an extra cache line on 64-bit (and smaller) CPUs
2138 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2140 /* Allocates slab cache used to allocate "struct epitem" items */
2141 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2142 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
2144 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2145 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2146 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
, NULL
);
2150 fs_initcall(eventpoll_init
);