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 <asm/uaccess.h>
37 #include <asm/system.h>
40 #include <asm/atomic.h>
44 * There are three level of locking required by epoll :
48 * 3) ep->lock (spinlock)
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a spinlock (ep->lock) because we manipulate objects
52 * from inside the poll callback, that might be triggered from
53 * a wake_up() that in turn might be called from IRQ context.
54 * So we can't sleep inside the poll callback and hence we need
55 * a spinlock. During the event transfer loop (from kernel to
56 * user space) we could end up sleeping due a copy_to_user(), so
57 * we need a lock that will allow us to sleep. This lock is a
58 * mutex (ep->mtx). It is acquired during the event transfer loop,
59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * This mutex is acquired by ep_free() during the epoll file
63 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
66 * It is also acquired when inserting an epoll fd onto another epoll
67 * fd. We do this so that we walk the epoll tree and ensure that this
68 * insertion does not create a cycle of epoll file descriptors, which
69 * could lead to deadlock. We need a global mutex to prevent two
70 * simultaneous inserts (A into B and B into A) from racing and
71 * constructing a cycle without either insert observing that it is
73 * It is necessary to acquire multiple "ep->mtx"es at once in the
74 * case when one epoll fd is added to another. In this case, we
75 * always acquire the locks in the order of nesting (i.e. after
76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77 * before e2->mtx). Since we disallow cycles of epoll file
78 * descriptors, this ensures that the mutexes are well-ordered. In
79 * order to communicate this nesting to lockdep, when walking a tree
80 * of epoll file descriptors, we use the current recursion depth as
82 * It is possible to drop the "ep->mtx" and to use the global
83 * mutex "epmutex" (together with "ep->lock") to have it working,
84 * but having "ep->mtx" will make the interface more scalable.
85 * Events that require holding "epmutex" are very rare, while for
86 * normal operations the epoll private "ep->mtx" will guarantee
87 * a better scalability.
90 /* Epoll private bits inside the event mask */
91 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
93 /* Maximum number of nesting allowed inside epoll sets */
94 #define EP_MAX_NESTS 4
96 /* Maximum msec timeout value storeable in a long int */
97 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
99 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
101 #define EP_UNACTIVE_PTR ((void *) -1L)
103 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
105 struct epoll_filefd
{
111 * Structure used to track possible nested calls, for too deep recursions
114 struct nested_call_node
{
115 struct list_head llink
;
121 * This structure is used as collector for nested calls, to check for
122 * maximum recursion dept and loop cycles.
124 struct nested_calls
{
125 struct list_head tasks_call_list
;
130 * Each file descriptor added to the eventpoll interface will
131 * have an entry of this type linked to the "rbr" RB tree.
134 /* RB tree node used to link this structure to the eventpoll RB tree */
137 /* List header used to link this structure to the eventpoll ready list */
138 struct list_head rdllink
;
141 * Works together "struct eventpoll"->ovflist in keeping the
142 * single linked chain of items.
146 /* The file descriptor information this item refers to */
147 struct epoll_filefd ffd
;
149 /* Number of active wait queue attached to poll operations */
152 /* List containing poll wait queues */
153 struct list_head pwqlist
;
155 /* The "container" of this item */
156 struct eventpoll
*ep
;
158 /* List header used to link this item to the "struct file" items list */
159 struct list_head fllink
;
161 /* The structure that describe the interested events and the source fd */
162 struct epoll_event event
;
166 * This structure is stored inside the "private_data" member of the file
167 * structure and rapresent the main data sructure for the eventpoll
171 /* Protect the this structure access */
175 * This mutex is used to ensure that files are not removed
176 * while epoll is using them. This is held during the event
177 * collection loop, the file cleanup path, the epoll file exit
178 * code and the ctl operations.
182 /* Wait queue used by sys_epoll_wait() */
183 wait_queue_head_t wq
;
185 /* Wait queue used by file->poll() */
186 wait_queue_head_t poll_wait
;
188 /* List of ready file descriptors */
189 struct list_head rdllist
;
191 /* RB tree root used to store monitored fd structs */
195 * This is a single linked list that chains all the "struct epitem" that
196 * happened while transfering ready events to userspace w/out
199 struct epitem
*ovflist
;
201 /* The user that created the eventpoll descriptor */
202 struct user_struct
*user
;
205 /* Wait structure used by the poll hooks */
206 struct eppoll_entry
{
207 /* List header used to link this structure to the "struct epitem" */
208 struct list_head llink
;
210 /* The "base" pointer is set to the container "struct epitem" */
214 * Wait queue item that will be linked to the target file wait
219 /* The wait queue head that linked the "wait" wait queue item */
220 wait_queue_head_t
*whead
;
223 /* Wrapper struct used by poll queueing */
229 /* Used by the ep_send_events() function as callback private data */
230 struct ep_send_events_data
{
232 struct epoll_event __user
*events
;
236 * Configuration options available inside /proc/sys/fs/epoll/
238 /* Maximum number of epoll watched descriptors, per user */
239 static long max_user_watches __read_mostly
;
242 * This mutex is used to serialize ep_free() and eventpoll_release_file().
244 static DEFINE_MUTEX(epmutex
);
246 /* Used to check for epoll file descriptor inclusion loops */
247 static struct nested_calls poll_loop_ncalls
;
249 /* Used for safe wake up implementation */
250 static struct nested_calls poll_safewake_ncalls
;
252 /* Used to call file's f_op->poll() under the nested calls boundaries */
253 static struct nested_calls poll_readywalk_ncalls
;
255 /* Slab cache used to allocate "struct epitem" */
256 static struct kmem_cache
*epi_cache __read_mostly
;
258 /* Slab cache used to allocate "struct eppoll_entry" */
259 static struct kmem_cache
*pwq_cache __read_mostly
;
263 #include <linux/sysctl.h>
266 static long long_max
= LONG_MAX
;
268 ctl_table epoll_table
[] = {
270 .procname
= "max_user_watches",
271 .data
= &max_user_watches
,
272 .maxlen
= sizeof(max_user_watches
),
274 .proc_handler
= proc_doulongvec_minmax
,
280 #endif /* CONFIG_SYSCTL */
283 /* Setup the structure that is used as key for the RB tree */
284 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
285 struct file
*file
, int fd
)
291 /* Compare RB tree keys */
292 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
293 struct epoll_filefd
*p2
)
295 return (p1
->file
> p2
->file
? +1:
296 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
299 /* Tells us if the item is currently linked */
300 static inline int ep_is_linked(struct list_head
*p
)
302 return !list_empty(p
);
305 /* Get the "struct epitem" from a wait queue pointer */
306 static inline struct epitem
*ep_item_from_wait(wait_queue_t
*p
)
308 return container_of(p
, struct eppoll_entry
, wait
)->base
;
311 /* Get the "struct epitem" from an epoll queue wrapper */
312 static inline struct epitem
*ep_item_from_epqueue(poll_table
*p
)
314 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
317 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
318 static inline int ep_op_has_event(int op
)
320 return op
!= EPOLL_CTL_DEL
;
323 /* Initialize the poll safe wake up structure */
324 static void ep_nested_calls_init(struct nested_calls
*ncalls
)
326 INIT_LIST_HEAD(&ncalls
->tasks_call_list
);
327 spin_lock_init(&ncalls
->lock
);
331 * ep_call_nested - Perform a bound (possibly) nested call, by checking
332 * that the recursion limit is not exceeded, and that
333 * the same nested call (by the meaning of same cookie) is
336 * @ncalls: Pointer to the nested_calls structure to be used for this call.
337 * @max_nests: Maximum number of allowed nesting calls.
338 * @nproc: Nested call core function pointer.
339 * @priv: Opaque data to be passed to the @nproc callback.
340 * @cookie: Cookie to be used to identify this nested call.
341 * @ctx: This instance context.
343 * Returns: Returns the code returned by the @nproc callback, or -1 if
344 * the maximum recursion limit has been exceeded.
346 static int ep_call_nested(struct nested_calls
*ncalls
, int max_nests
,
347 int (*nproc
)(void *, void *, int), void *priv
,
348 void *cookie
, void *ctx
)
350 int error
, call_nests
= 0;
352 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
353 struct nested_call_node
*tncur
;
354 struct nested_call_node tnode
;
356 spin_lock_irqsave(&ncalls
->lock
, flags
);
359 * Try to see if the current task is already inside this wakeup call.
360 * We use a list here, since the population inside this set is always
363 list_for_each_entry(tncur
, lsthead
, llink
) {
364 if (tncur
->ctx
== ctx
&&
365 (tncur
->cookie
== cookie
|| ++call_nests
> max_nests
)) {
367 * Ops ... loop detected or maximum nest level reached.
368 * We abort this wake by breaking the cycle itself.
375 /* Add the current task and cookie to the list */
377 tnode
.cookie
= cookie
;
378 list_add(&tnode
.llink
, lsthead
);
380 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
382 /* Call the nested function */
383 error
= (*nproc
)(priv
, cookie
, call_nests
);
385 /* Remove the current task from the list */
386 spin_lock_irqsave(&ncalls
->lock
, flags
);
387 list_del(&tnode
.llink
);
389 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
394 #ifdef CONFIG_DEBUG_LOCK_ALLOC
395 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
396 unsigned long events
, int subclass
)
400 spin_lock_irqsave_nested(&wqueue
->lock
, flags
, subclass
);
401 wake_up_locked_poll(wqueue
, events
);
402 spin_unlock_irqrestore(&wqueue
->lock
, flags
);
405 static inline void ep_wake_up_nested(wait_queue_head_t
*wqueue
,
406 unsigned long events
, int subclass
)
408 wake_up_poll(wqueue
, events
);
412 static int ep_poll_wakeup_proc(void *priv
, void *cookie
, int call_nests
)
414 ep_wake_up_nested((wait_queue_head_t
*) cookie
, POLLIN
,
420 * Perform a safe wake up of the poll wait list. The problem is that
421 * with the new callback'd wake up system, it is possible that the
422 * poll callback is reentered from inside the call to wake_up() done
423 * on the poll wait queue head. The rule is that we cannot reenter the
424 * wake up code from the same task more than EP_MAX_NESTS times,
425 * and we cannot reenter the same wait queue head at all. This will
426 * enable to have a hierarchy of epoll file descriptor of no more than
429 static void ep_poll_safewake(wait_queue_head_t
*wq
)
431 int this_cpu
= get_cpu();
433 ep_call_nested(&poll_safewake_ncalls
, EP_MAX_NESTS
,
434 ep_poll_wakeup_proc
, NULL
, wq
, (void *) (long) this_cpu
);
440 * This function unregisters poll callbacks from the associated file
441 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
444 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
446 struct list_head
*lsthead
= &epi
->pwqlist
;
447 struct eppoll_entry
*pwq
;
449 while (!list_empty(lsthead
)) {
450 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
452 list_del(&pwq
->llink
);
453 remove_wait_queue(pwq
->whead
, &pwq
->wait
);
454 kmem_cache_free(pwq_cache
, pwq
);
459 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
460 * the scan code, to call f_op->poll(). Also allows for
461 * O(NumReady) performance.
463 * @ep: Pointer to the epoll private data structure.
464 * @sproc: Pointer to the scan callback.
465 * @priv: Private opaque data passed to the @sproc callback.
466 * @depth: The current depth of recursive f_op->poll calls.
468 * Returns: The same integer error code returned by the @sproc callback.
470 static int ep_scan_ready_list(struct eventpoll
*ep
,
471 int (*sproc
)(struct eventpoll
*,
472 struct list_head
*, void *),
476 int error
, pwake
= 0;
478 struct epitem
*epi
, *nepi
;
482 * We need to lock this because we could be hit by
483 * eventpoll_release_file() and epoll_ctl().
485 mutex_lock_nested(&ep
->mtx
, depth
);
488 * Steal the ready list, and re-init the original one to the
489 * empty list. Also, set ep->ovflist to NULL so that events
490 * happening while looping w/out locks, are not lost. We cannot
491 * have the poll callback to queue directly on ep->rdllist,
492 * because we want the "sproc" callback to be able to do it
495 spin_lock_irqsave(&ep
->lock
, flags
);
496 list_splice_init(&ep
->rdllist
, &txlist
);
498 spin_unlock_irqrestore(&ep
->lock
, flags
);
501 * Now call the callback function.
503 error
= (*sproc
)(ep
, &txlist
, priv
);
505 spin_lock_irqsave(&ep
->lock
, flags
);
507 * During the time we spent inside the "sproc" callback, some
508 * other events might have been queued by the poll callback.
509 * We re-insert them inside the main ready-list here.
511 for (nepi
= ep
->ovflist
; (epi
= nepi
) != NULL
;
512 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
514 * We need to check if the item is already in the list.
515 * During the "sproc" callback execution time, items are
516 * queued into ->ovflist but the "txlist" might already
517 * contain them, and the list_splice() below takes care of them.
519 if (!ep_is_linked(&epi
->rdllink
))
520 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
523 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
524 * releasing the lock, events will be queued in the normal way inside
527 ep
->ovflist
= EP_UNACTIVE_PTR
;
530 * Quickly re-inject items left on "txlist".
532 list_splice(&txlist
, &ep
->rdllist
);
534 if (!list_empty(&ep
->rdllist
)) {
536 * Wake up (if active) both the eventpoll wait list and
537 * the ->poll() wait list (delayed after we release the lock).
539 if (waitqueue_active(&ep
->wq
))
540 wake_up_locked(&ep
->wq
);
541 if (waitqueue_active(&ep
->poll_wait
))
544 spin_unlock_irqrestore(&ep
->lock
, flags
);
546 mutex_unlock(&ep
->mtx
);
548 /* We have to call this outside the lock */
550 ep_poll_safewake(&ep
->poll_wait
);
556 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
557 * all the associated resources. Must be called with "mtx" held.
559 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
562 struct file
*file
= epi
->ffd
.file
;
565 * Removes poll wait queue hooks. We _have_ to do this without holding
566 * the "ep->lock" otherwise a deadlock might occur. This because of the
567 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
568 * queue head lock when unregistering the wait queue. The wakeup callback
569 * will run by holding the wait queue head lock and will call our callback
570 * that will try to get "ep->lock".
572 ep_unregister_pollwait(ep
, epi
);
574 /* Remove the current item from the list of epoll hooks */
575 spin_lock(&file
->f_lock
);
576 if (ep_is_linked(&epi
->fllink
))
577 list_del_init(&epi
->fllink
);
578 spin_unlock(&file
->f_lock
);
580 rb_erase(&epi
->rbn
, &ep
->rbr
);
582 spin_lock_irqsave(&ep
->lock
, flags
);
583 if (ep_is_linked(&epi
->rdllink
))
584 list_del_init(&epi
->rdllink
);
585 spin_unlock_irqrestore(&ep
->lock
, flags
);
587 /* At this point it is safe to free the eventpoll item */
588 kmem_cache_free(epi_cache
, epi
);
590 atomic_long_dec(&ep
->user
->epoll_watches
);
595 static void ep_free(struct eventpoll
*ep
)
600 /* We need to release all tasks waiting for these file */
601 if (waitqueue_active(&ep
->poll_wait
))
602 ep_poll_safewake(&ep
->poll_wait
);
605 * We need to lock this because we could be hit by
606 * eventpoll_release_file() while we're freeing the "struct eventpoll".
607 * We do not need to hold "ep->mtx" here because the epoll file
608 * is on the way to be removed and no one has references to it
609 * anymore. The only hit might come from eventpoll_release_file() but
610 * holding "epmutex" is sufficent here.
612 mutex_lock(&epmutex
);
615 * Walks through the whole tree by unregistering poll callbacks.
617 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
618 epi
= rb_entry(rbp
, struct epitem
, rbn
);
620 ep_unregister_pollwait(ep
, epi
);
624 * Walks through the whole tree by freeing each "struct epitem". At this
625 * point we are sure no poll callbacks will be lingering around, and also by
626 * holding "epmutex" we can be sure that no file cleanup code will hit
627 * us during this operation. So we can avoid the lock on "ep->lock".
629 while ((rbp
= rb_first(&ep
->rbr
)) != NULL
) {
630 epi
= rb_entry(rbp
, struct epitem
, rbn
);
634 mutex_unlock(&epmutex
);
635 mutex_destroy(&ep
->mtx
);
640 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
642 struct eventpoll
*ep
= file
->private_data
;
650 static int ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
653 struct epitem
*epi
, *tmp
;
655 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
656 if (epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
) &
658 return POLLIN
| POLLRDNORM
;
661 * Item has been dropped into the ready list by the poll
662 * callback, but it's not actually ready, as far as
663 * caller requested events goes. We can remove it here.
665 list_del_init(&epi
->rdllink
);
672 static int ep_poll_readyevents_proc(void *priv
, void *cookie
, int call_nests
)
674 return ep_scan_ready_list(priv
, ep_read_events_proc
, NULL
, call_nests
+ 1);
677 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
680 struct eventpoll
*ep
= file
->private_data
;
682 /* Insert inside our poll wait queue */
683 poll_wait(file
, &ep
->poll_wait
, wait
);
686 * Proceed to find out if wanted events are really available inside
687 * the ready list. This need to be done under ep_call_nested()
688 * supervision, since the call to f_op->poll() done on listed files
689 * could re-enter here.
691 pollflags
= ep_call_nested(&poll_readywalk_ncalls
, EP_MAX_NESTS
,
692 ep_poll_readyevents_proc
, ep
, ep
, current
);
694 return pollflags
!= -1 ? pollflags
: 0;
697 /* File callbacks that implement the eventpoll file behaviour */
698 static const struct file_operations eventpoll_fops
= {
699 .release
= ep_eventpoll_release
,
700 .poll
= ep_eventpoll_poll
703 /* Fast test to see if the file is an evenpoll file */
704 static inline int is_file_epoll(struct file
*f
)
706 return f
->f_op
== &eventpoll_fops
;
710 * This is called from eventpoll_release() to unlink files from the eventpoll
711 * interface. We need to have this facility to cleanup correctly files that are
712 * closed without being removed from the eventpoll interface.
714 void eventpoll_release_file(struct file
*file
)
716 struct list_head
*lsthead
= &file
->f_ep_links
;
717 struct eventpoll
*ep
;
721 * We don't want to get "file->f_lock" because it is not
722 * necessary. It is not necessary because we're in the "struct file"
723 * cleanup path, and this means that noone is using this file anymore.
724 * So, for example, epoll_ctl() cannot hit here since if we reach this
725 * point, the file counter already went to zero and fget() would fail.
726 * The only hit might come from ep_free() but by holding the mutex
727 * will correctly serialize the operation. We do need to acquire
728 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
729 * from anywhere but ep_free().
731 * Besides, ep_remove() acquires the lock, so we can't hold it here.
733 mutex_lock(&epmutex
);
735 while (!list_empty(lsthead
)) {
736 epi
= list_first_entry(lsthead
, struct epitem
, fllink
);
739 list_del_init(&epi
->fllink
);
740 mutex_lock_nested(&ep
->mtx
, 0);
742 mutex_unlock(&ep
->mtx
);
745 mutex_unlock(&epmutex
);
748 static int ep_alloc(struct eventpoll
**pep
)
751 struct user_struct
*user
;
752 struct eventpoll
*ep
;
754 user
= get_current_user();
756 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
760 spin_lock_init(&ep
->lock
);
761 mutex_init(&ep
->mtx
);
762 init_waitqueue_head(&ep
->wq
);
763 init_waitqueue_head(&ep
->poll_wait
);
764 INIT_LIST_HEAD(&ep
->rdllist
);
766 ep
->ovflist
= EP_UNACTIVE_PTR
;
779 * Search the file inside the eventpoll tree. The RB tree operations
780 * are protected by the "mtx" mutex, and ep_find() must be called with
783 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
787 struct epitem
*epi
, *epir
= NULL
;
788 struct epoll_filefd ffd
;
790 ep_set_ffd(&ffd
, file
, fd
);
791 for (rbp
= ep
->rbr
.rb_node
; rbp
; ) {
792 epi
= rb_entry(rbp
, struct epitem
, rbn
);
793 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
808 * This is the callback that is passed to the wait queue wakeup
809 * machanism. It is called by the stored file descriptors when they
810 * have events to report.
812 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
816 struct epitem
*epi
= ep_item_from_wait(wait
);
817 struct eventpoll
*ep
= epi
->ep
;
819 spin_lock_irqsave(&ep
->lock
, flags
);
822 * If the event mask does not contain any poll(2) event, we consider the
823 * descriptor to be disabled. This condition is likely the effect of the
824 * EPOLLONESHOT bit that disables the descriptor when an event is received,
825 * until the next EPOLL_CTL_MOD will be issued.
827 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
831 * Check the events coming with the callback. At this stage, not
832 * every device reports the events in the "key" parameter of the
833 * callback. We need to be able to handle both cases here, hence the
834 * test for "key" != NULL before the event match test.
836 if (key
&& !((unsigned long) key
& epi
->event
.events
))
840 * If we are trasfering events to userspace, we can hold no locks
841 * (because we're accessing user memory, and because of linux f_op->poll()
842 * semantics). All the events that happens during that period of time are
843 * chained in ep->ovflist and requeued later on.
845 if (unlikely(ep
->ovflist
!= EP_UNACTIVE_PTR
)) {
846 if (epi
->next
== EP_UNACTIVE_PTR
) {
847 epi
->next
= ep
->ovflist
;
853 /* If this file is already in the ready list we exit soon */
854 if (!ep_is_linked(&epi
->rdllink
))
855 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
858 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
861 if (waitqueue_active(&ep
->wq
))
862 wake_up_locked(&ep
->wq
);
863 if (waitqueue_active(&ep
->poll_wait
))
867 spin_unlock_irqrestore(&ep
->lock
, flags
);
869 /* We have to call this outside the lock */
871 ep_poll_safewake(&ep
->poll_wait
);
877 * This is the callback that is used to add our wait queue to the
878 * target file wakeup lists.
880 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
883 struct epitem
*epi
= ep_item_from_epqueue(pt
);
884 struct eppoll_entry
*pwq
;
886 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
887 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
890 add_wait_queue(whead
, &pwq
->wait
);
891 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
894 /* We have to signal that an error occurred */
899 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
902 struct rb_node
**p
= &ep
->rbr
.rb_node
, *parent
= NULL
;
907 epic
= rb_entry(parent
, struct epitem
, rbn
);
908 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
910 p
= &parent
->rb_right
;
912 p
= &parent
->rb_left
;
914 rb_link_node(&epi
->rbn
, parent
, p
);
915 rb_insert_color(&epi
->rbn
, &ep
->rbr
);
919 * Must be called with "mtx" held.
921 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
922 struct file
*tfile
, int fd
)
924 int error
, revents
, pwake
= 0;
928 struct ep_pqueue epq
;
930 user_watches
= atomic_long_read(&ep
->user
->epoll_watches
);
931 if (unlikely(user_watches
>= max_user_watches
))
933 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
936 /* Item initialization follow here ... */
937 INIT_LIST_HEAD(&epi
->rdllink
);
938 INIT_LIST_HEAD(&epi
->fllink
);
939 INIT_LIST_HEAD(&epi
->pwqlist
);
941 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
944 epi
->next
= EP_UNACTIVE_PTR
;
946 /* Initialize the poll table using the queue callback */
948 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
951 * Attach the item to the poll hooks and get current event bits.
952 * We can safely use the file* here because its usage count has
953 * been increased by the caller of this function. Note that after
954 * this operation completes, the poll callback can start hitting
957 revents
= tfile
->f_op
->poll(tfile
, &epq
.pt
);
960 * We have to check if something went wrong during the poll wait queue
961 * install process. Namely an allocation for a wait queue failed due
962 * high memory pressure.
966 goto error_unregister
;
968 /* Add the current item to the list of active epoll hook for this file */
969 spin_lock(&tfile
->f_lock
);
970 list_add_tail(&epi
->fllink
, &tfile
->f_ep_links
);
971 spin_unlock(&tfile
->f_lock
);
974 * Add the current item to the RB tree. All RB tree operations are
975 * protected by "mtx", and ep_insert() is called with "mtx" held.
977 ep_rbtree_insert(ep
, epi
);
979 /* We have to drop the new item inside our item list to keep track of it */
980 spin_lock_irqsave(&ep
->lock
, flags
);
982 /* If the file is already "ready" we drop it inside the ready list */
983 if ((revents
& event
->events
) && !ep_is_linked(&epi
->rdllink
)) {
984 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
986 /* Notify waiting tasks that events are available */
987 if (waitqueue_active(&ep
->wq
))
988 wake_up_locked(&ep
->wq
);
989 if (waitqueue_active(&ep
->poll_wait
))
993 spin_unlock_irqrestore(&ep
->lock
, flags
);
995 atomic_long_inc(&ep
->user
->epoll_watches
);
997 /* We have to call this outside the lock */
999 ep_poll_safewake(&ep
->poll_wait
);
1004 ep_unregister_pollwait(ep
, epi
);
1007 * We need to do this because an event could have been arrived on some
1008 * allocated wait queue. Note that we don't care about the ep->ovflist
1009 * list, since that is used/cleaned only inside a section bound by "mtx".
1010 * And ep_insert() is called with "mtx" held.
1012 spin_lock_irqsave(&ep
->lock
, flags
);
1013 if (ep_is_linked(&epi
->rdllink
))
1014 list_del_init(&epi
->rdllink
);
1015 spin_unlock_irqrestore(&ep
->lock
, flags
);
1017 kmem_cache_free(epi_cache
, epi
);
1023 * Modify the interest event mask by dropping an event if the new mask
1024 * has a match in the current file status. Must be called with "mtx" held.
1026 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
, struct epoll_event
*event
)
1029 unsigned int revents
;
1032 * Set the new event interest mask before calling f_op->poll();
1033 * otherwise we might miss an event that happens between the
1034 * f_op->poll() call and the new event set registering.
1036 epi
->event
.events
= event
->events
;
1037 epi
->event
.data
= event
->data
; /* protected by mtx */
1040 * Get current event bits. We can safely use the file* here because
1041 * its usage count has been increased by the caller of this function.
1043 revents
= epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
);
1046 * If the item is "hot" and it is not registered inside the ready
1047 * list, push it inside.
1049 if (revents
& event
->events
) {
1050 spin_lock_irq(&ep
->lock
);
1051 if (!ep_is_linked(&epi
->rdllink
)) {
1052 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1054 /* Notify waiting tasks that events are available */
1055 if (waitqueue_active(&ep
->wq
))
1056 wake_up_locked(&ep
->wq
);
1057 if (waitqueue_active(&ep
->poll_wait
))
1060 spin_unlock_irq(&ep
->lock
);
1063 /* We have to call this outside the lock */
1065 ep_poll_safewake(&ep
->poll_wait
);
1070 static int ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1073 struct ep_send_events_data
*esed
= priv
;
1075 unsigned int revents
;
1077 struct epoll_event __user
*uevent
;
1080 * We can loop without lock because we are passed a task private list.
1081 * Items cannot vanish during the loop because ep_scan_ready_list() is
1082 * holding "mtx" during this call.
1084 for (eventcnt
= 0, uevent
= esed
->events
;
1085 !list_empty(head
) && eventcnt
< esed
->maxevents
;) {
1086 epi
= list_first_entry(head
, struct epitem
, rdllink
);
1088 list_del_init(&epi
->rdllink
);
1090 revents
= epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
) &
1094 * If the event mask intersect the caller-requested one,
1095 * deliver the event to userspace. Again, ep_scan_ready_list()
1096 * is holding "mtx", so no operations coming from userspace
1097 * can change the item.
1100 if (__put_user(revents
, &uevent
->events
) ||
1101 __put_user(epi
->event
.data
, &uevent
->data
)) {
1102 list_add(&epi
->rdllink
, head
);
1103 return eventcnt
? eventcnt
: -EFAULT
;
1107 if (epi
->event
.events
& EPOLLONESHOT
)
1108 epi
->event
.events
&= EP_PRIVATE_BITS
;
1109 else if (!(epi
->event
.events
& EPOLLET
)) {
1111 * If this file has been added with Level
1112 * Trigger mode, we need to insert back inside
1113 * the ready list, so that the next call to
1114 * epoll_wait() will check again the events
1115 * availability. At this point, noone can insert
1116 * into ep->rdllist besides us. The epoll_ctl()
1117 * callers are locked out by
1118 * ep_scan_ready_list() holding "mtx" and the
1119 * poll callback will queue them in ep->ovflist.
1121 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1129 static int ep_send_events(struct eventpoll
*ep
,
1130 struct epoll_event __user
*events
, int maxevents
)
1132 struct ep_send_events_data esed
;
1134 esed
.maxevents
= maxevents
;
1135 esed
.events
= events
;
1137 return ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
, 0);
1140 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1141 int maxevents
, long timeout
)
1144 unsigned long flags
;
1149 * Calculate the timeout by checking for the "infinite" value (-1)
1150 * and the overflow condition. The passed timeout is in milliseconds,
1151 * that why (t * HZ) / 1000.
1153 jtimeout
= (timeout
< 0 || timeout
>= EP_MAX_MSTIMEO
) ?
1154 MAX_SCHEDULE_TIMEOUT
: (timeout
* HZ
+ 999) / 1000;
1157 spin_lock_irqsave(&ep
->lock
, flags
);
1160 if (list_empty(&ep
->rdllist
)) {
1162 * We don't have any available event to return to the caller.
1163 * We need to sleep here, and we will be wake up by
1164 * ep_poll_callback() when events will become available.
1166 init_waitqueue_entry(&wait
, current
);
1167 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
1168 __add_wait_queue(&ep
->wq
, &wait
);
1172 * We don't want to sleep if the ep_poll_callback() sends us
1173 * a wakeup in between. That's why we set the task state
1174 * to TASK_INTERRUPTIBLE before doing the checks.
1176 set_current_state(TASK_INTERRUPTIBLE
);
1177 if (!list_empty(&ep
->rdllist
) || !jtimeout
)
1179 if (signal_pending(current
)) {
1184 spin_unlock_irqrestore(&ep
->lock
, flags
);
1185 jtimeout
= schedule_timeout(jtimeout
);
1186 spin_lock_irqsave(&ep
->lock
, flags
);
1188 __remove_wait_queue(&ep
->wq
, &wait
);
1190 set_current_state(TASK_RUNNING
);
1192 /* Is it worth to try to dig for events ? */
1193 eavail
= !list_empty(&ep
->rdllist
) || ep
->ovflist
!= EP_UNACTIVE_PTR
;
1195 spin_unlock_irqrestore(&ep
->lock
, flags
);
1198 * Try to transfer events to user space. In case we get 0 events and
1199 * there's still timeout left over, we go trying again in search of
1202 if (!res
&& eavail
&&
1203 !(res
= ep_send_events(ep
, events
, maxevents
)) && jtimeout
)
1210 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1211 * API, to verify that adding an epoll file inside another
1212 * epoll structure, does not violate the constraints, in
1213 * terms of closed loops, or too deep chains (which can
1214 * result in excessive stack usage).
1216 * @priv: Pointer to the epoll file to be currently checked.
1217 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1218 * data structure pointer.
1219 * @call_nests: Current dept of the @ep_call_nested() call stack.
1221 * Returns: Returns zero if adding the epoll @file inside current epoll
1222 * structure @ep does not violate the constraints, or -1 otherwise.
1224 static int ep_loop_check_proc(void *priv
, void *cookie
, int call_nests
)
1227 struct file
*file
= priv
;
1228 struct eventpoll
*ep
= file
->private_data
;
1229 struct rb_node
*rbp
;
1232 mutex_lock_nested(&ep
->mtx
, call_nests
+ 1);
1233 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1234 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1235 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1236 error
= ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1237 ep_loop_check_proc
, epi
->ffd
.file
,
1238 epi
->ffd
.file
->private_data
, current
);
1243 mutex_unlock(&ep
->mtx
);
1249 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1250 * another epoll file (represented by @ep) does not create
1251 * closed loops or too deep chains.
1253 * @ep: Pointer to the epoll private data structure.
1254 * @file: Pointer to the epoll file to be checked.
1256 * Returns: Returns zero if adding the epoll @file inside current epoll
1257 * structure @ep does not violate the constraints, or -1 otherwise.
1259 static int ep_loop_check(struct eventpoll
*ep
, struct file
*file
)
1261 return ep_call_nested(&poll_loop_ncalls
, EP_MAX_NESTS
,
1262 ep_loop_check_proc
, file
, ep
, current
);
1266 * Open an eventpoll file descriptor.
1268 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
1271 struct eventpoll
*ep
= NULL
;
1273 /* Check the EPOLL_* constant for consistency. */
1274 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
1276 if (flags
& ~EPOLL_CLOEXEC
)
1279 * Create the internal data structure ("struct eventpoll").
1281 error
= ep_alloc(&ep
);
1285 * Creates all the items needed to setup an eventpoll file. That is,
1286 * a file structure and a free file descriptor.
1288 error
= anon_inode_getfd("[eventpoll]", &eventpoll_fops
, ep
,
1289 O_RDWR
| (flags
& O_CLOEXEC
));
1296 SYSCALL_DEFINE1(epoll_create
, int, size
)
1301 return sys_epoll_create1(0);
1305 * The following function implements the controller interface for
1306 * the eventpoll file that enables the insertion/removal/change of
1307 * file descriptors inside the interest set.
1309 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
1310 struct epoll_event __user
*, event
)
1313 int did_lock_epmutex
= 0;
1314 struct file
*file
, *tfile
;
1315 struct eventpoll
*ep
;
1317 struct epoll_event epds
;
1320 if (ep_op_has_event(op
) &&
1321 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
1324 /* Get the "struct file *" for the eventpoll file */
1330 /* Get the "struct file *" for the target file */
1335 /* The target file descriptor must support poll */
1337 if (!tfile
->f_op
|| !tfile
->f_op
->poll
)
1338 goto error_tgt_fput
;
1341 * We have to check that the file structure underneath the file descriptor
1342 * the user passed to us _is_ an eventpoll file. And also we do not permit
1343 * adding an epoll file descriptor inside itself.
1346 if (file
== tfile
|| !is_file_epoll(file
))
1347 goto error_tgt_fput
;
1350 * At this point it is safe to assume that the "private_data" contains
1351 * our own data structure.
1353 ep
= file
->private_data
;
1356 * When we insert an epoll file descriptor, inside another epoll file
1357 * descriptor, there is the change of creating closed loops, which are
1358 * better be handled here, than in more critical paths.
1360 * We hold epmutex across the loop check and the insert in this case, in
1361 * order to prevent two separate inserts from racing and each doing the
1362 * insert "at the same time" such that ep_loop_check passes on both
1363 * before either one does the insert, thereby creating a cycle.
1365 if (unlikely(is_file_epoll(tfile
) && op
== EPOLL_CTL_ADD
)) {
1366 mutex_lock(&epmutex
);
1367 did_lock_epmutex
= 1;
1369 if (ep_loop_check(ep
, tfile
) != 0)
1370 goto error_tgt_fput
;
1374 mutex_lock_nested(&ep
->mtx
, 0);
1377 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1378 * above, we can be sure to be able to use the item looked up by
1379 * ep_find() till we release the mutex.
1381 epi
= ep_find(ep
, tfile
, fd
);
1387 epds
.events
|= POLLERR
| POLLHUP
;
1388 error
= ep_insert(ep
, &epds
, tfile
, fd
);
1394 error
= ep_remove(ep
, epi
);
1400 epds
.events
|= POLLERR
| POLLHUP
;
1401 error
= ep_modify(ep
, epi
, &epds
);
1406 mutex_unlock(&ep
->mtx
);
1409 if (unlikely(did_lock_epmutex
))
1410 mutex_unlock(&epmutex
);
1421 * Implement the event wait interface for the eventpoll file. It is the kernel
1422 * part of the user space epoll_wait(2).
1424 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
1425 int, maxevents
, int, timeout
)
1429 struct eventpoll
*ep
;
1431 /* The maximum number of event must be greater than zero */
1432 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
1435 /* Verify that the area passed by the user is writeable */
1436 if (!access_ok(VERIFY_WRITE
, events
, maxevents
* sizeof(struct epoll_event
))) {
1441 /* Get the "struct file *" for the eventpoll file */
1448 * We have to check that the file structure underneath the fd
1449 * the user passed to us _is_ an eventpoll file.
1452 if (!is_file_epoll(file
))
1456 * At this point it is safe to assume that the "private_data" contains
1457 * our own data structure.
1459 ep
= file
->private_data
;
1461 /* Time to fish for events ... */
1462 error
= ep_poll(ep
, events
, maxevents
, timeout
);
1471 #ifdef HAVE_SET_RESTORE_SIGMASK
1474 * Implement the event wait interface for the eventpoll file. It is the kernel
1475 * part of the user space epoll_pwait(2).
1477 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
1478 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
1482 sigset_t ksigmask
, sigsaved
;
1485 * If the caller wants a certain signal mask to be set during the wait,
1489 if (sigsetsize
!= sizeof(sigset_t
))
1491 if (copy_from_user(&ksigmask
, sigmask
, sizeof(ksigmask
)))
1493 sigdelsetmask(&ksigmask
, sigmask(SIGKILL
) | sigmask(SIGSTOP
));
1494 sigprocmask(SIG_SETMASK
, &ksigmask
, &sigsaved
);
1497 error
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
1500 * If we changed the signal mask, we need to restore the original one.
1501 * In case we've got a signal while waiting, we do not restore the
1502 * signal mask yet, and we allow do_signal() to deliver the signal on
1503 * the way back to userspace, before the signal mask is restored.
1506 if (error
== -EINTR
) {
1507 memcpy(¤t
->saved_sigmask
, &sigsaved
,
1509 set_restore_sigmask();
1511 sigprocmask(SIG_SETMASK
, &sigsaved
, NULL
);
1517 #endif /* HAVE_SET_RESTORE_SIGMASK */
1519 static int __init
eventpoll_init(void)
1525 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1527 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
1529 BUG_ON(max_user_watches
< 0);
1532 * Initialize the structure used to perform epoll file descriptor
1533 * inclusion loops checks.
1535 ep_nested_calls_init(&poll_loop_ncalls
);
1537 /* Initialize the structure used to perform safe poll wait head wake ups */
1538 ep_nested_calls_init(&poll_safewake_ncalls
);
1540 /* Initialize the structure used to perform file's f_op->poll() calls */
1541 ep_nested_calls_init(&poll_readywalk_ncalls
);
1543 /* Allocates slab cache used to allocate "struct epitem" items */
1544 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
1545 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
1547 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1548 pwq_cache
= kmem_cache_create("eventpoll_pwq",
1549 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
, NULL
);
1553 fs_initcall(eventpoll_init
);