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 possible to drop the "ep->mtx" and to use the global
67 * mutex "epmutex" (together with "ep->lock") to have it working,
68 * but having "ep->mtx" will make the interface more scalable.
69 * Events that require holding "epmutex" are very rare, while for
70 * normal operations the epoll private "ep->mtx" will guarantee
71 * a better scalability.
74 /* Epoll private bits inside the event mask */
75 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
77 /* Maximum number of nesting allowed inside epoll sets */
78 #define EP_MAX_NESTS 4
80 /* Maximum msec timeout value storeable in a long int */
81 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
83 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
85 #define EP_UNACTIVE_PTR ((void *) -1L)
87 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
95 * Structure used to track possible nested calls, for too deep recursions
98 struct nested_call_node
{
99 struct list_head llink
;
105 * This structure is used as collector for nested calls, to check for
106 * maximum recursion dept and loop cycles.
108 struct nested_calls
{
109 struct list_head tasks_call_list
;
114 * Each file descriptor added to the eventpoll interface will
115 * have an entry of this type linked to the "rbr" RB tree.
118 /* RB tree node used to link this structure to the eventpoll RB tree */
121 /* List header used to link this structure to the eventpoll ready list */
122 struct list_head rdllink
;
125 * Works together "struct eventpoll"->ovflist in keeping the
126 * single linked chain of items.
130 /* The file descriptor information this item refers to */
131 struct epoll_filefd ffd
;
133 /* Number of active wait queue attached to poll operations */
136 /* List containing poll wait queues */
137 struct list_head pwqlist
;
139 /* The "container" of this item */
140 struct eventpoll
*ep
;
142 /* List header used to link this item to the "struct file" items list */
143 struct list_head fllink
;
145 /* The structure that describe the interested events and the source fd */
146 struct epoll_event event
;
150 * This structure is stored inside the "private_data" member of the file
151 * structure and rapresent the main data sructure for the eventpoll
155 /* Protect the this structure access */
159 * This mutex is used to ensure that files are not removed
160 * while epoll is using them. This is held during the event
161 * collection loop, the file cleanup path, the epoll file exit
162 * code and the ctl operations.
166 /* Wait queue used by sys_epoll_wait() */
167 wait_queue_head_t wq
;
169 /* Wait queue used by file->poll() */
170 wait_queue_head_t poll_wait
;
172 /* List of ready file descriptors */
173 struct list_head rdllist
;
175 /* RB tree root used to store monitored fd structs */
179 * This is a single linked list that chains all the "struct epitem" that
180 * happened while transfering ready events to userspace w/out
183 struct epitem
*ovflist
;
185 /* The user that created the eventpoll descriptor */
186 struct user_struct
*user
;
189 /* Wait structure used by the poll hooks */
190 struct eppoll_entry
{
191 /* List header used to link this structure to the "struct epitem" */
192 struct list_head llink
;
194 /* The "base" pointer is set to the container "struct epitem" */
198 * Wait queue item that will be linked to the target file wait
203 /* The wait queue head that linked the "wait" wait queue item */
204 wait_queue_head_t
*whead
;
207 /* Wrapper struct used by poll queueing */
213 /* Used by the ep_send_events() function as callback private data */
214 struct ep_send_events_data
{
216 struct epoll_event __user
*events
;
220 * Configuration options available inside /proc/sys/fs/epoll/
222 /* Maximum number of epoll watched descriptors, per user */
223 static int max_user_watches __read_mostly
;
226 * This mutex is used to serialize ep_free() and eventpoll_release_file().
228 static DEFINE_MUTEX(epmutex
);
230 /* Used for safe wake up implementation */
231 static struct nested_calls poll_safewake_ncalls
;
233 /* Used to call file's f_op->poll() under the nested calls boundaries */
234 static struct nested_calls poll_readywalk_ncalls
;
236 /* Slab cache used to allocate "struct epitem" */
237 static struct kmem_cache
*epi_cache __read_mostly
;
239 /* Slab cache used to allocate "struct eppoll_entry" */
240 static struct kmem_cache
*pwq_cache __read_mostly
;
244 #include <linux/sysctl.h>
248 ctl_table epoll_table
[] = {
250 .procname
= "max_user_watches",
251 .data
= &max_user_watches
,
252 .maxlen
= sizeof(int),
254 .proc_handler
= &proc_dointvec_minmax
,
259 #endif /* CONFIG_SYSCTL */
262 /* Setup the structure that is used as key for the RB tree */
263 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
264 struct file
*file
, int fd
)
270 /* Compare RB tree keys */
271 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
272 struct epoll_filefd
*p2
)
274 return (p1
->file
> p2
->file
? +1:
275 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
278 /* Tells us if the item is currently linked */
279 static inline int ep_is_linked(struct list_head
*p
)
281 return !list_empty(p
);
284 /* Get the "struct epitem" from a wait queue pointer */
285 static inline struct epitem
*ep_item_from_wait(wait_queue_t
*p
)
287 return container_of(p
, struct eppoll_entry
, wait
)->base
;
290 /* Get the "struct epitem" from an epoll queue wrapper */
291 static inline struct epitem
*ep_item_from_epqueue(poll_table
*p
)
293 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
296 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
297 static inline int ep_op_has_event(int op
)
299 return op
!= EPOLL_CTL_DEL
;
302 /* Initialize the poll safe wake up structure */
303 static void ep_nested_calls_init(struct nested_calls
*ncalls
)
305 INIT_LIST_HEAD(&ncalls
->tasks_call_list
);
306 spin_lock_init(&ncalls
->lock
);
310 * ep_call_nested - Perform a bound (possibly) nested call, by checking
311 * that the recursion limit is not exceeded, and that
312 * the same nested call (by the meaning of same cookie) is
315 * @ncalls: Pointer to the nested_calls structure to be used for this call.
316 * @max_nests: Maximum number of allowed nesting calls.
317 * @nproc: Nested call core function pointer.
318 * @priv: Opaque data to be passed to the @nproc callback.
319 * @cookie: Cookie to be used to identify this nested call.
321 * Returns: Returns the code returned by the @nproc callback, or -1 if
322 * the maximum recursion limit has been exceeded.
324 static int ep_call_nested(struct nested_calls
*ncalls
, int max_nests
,
325 int (*nproc
)(void *, void *, int), void *priv
,
328 int error
, call_nests
= 0;
330 int this_cpu
= get_cpu();
331 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
332 struct nested_call_node
*tncur
;
333 struct nested_call_node tnode
;
335 spin_lock_irqsave(&ncalls
->lock
, flags
);
338 * Try to see if the current task is already inside this wakeup call.
339 * We use a list here, since the population inside this set is always
342 list_for_each_entry(tncur
, lsthead
, llink
) {
343 if (tncur
->cpu
== this_cpu
&&
344 (tncur
->cookie
== cookie
|| ++call_nests
> max_nests
)) {
346 * Ops ... loop detected or maximum nest level reached.
347 * We abort this wake by breaking the cycle itself.
354 /* Add the current task and cookie to the list */
355 tnode
.cpu
= this_cpu
;
356 tnode
.cookie
= cookie
;
357 list_add(&tnode
.llink
, lsthead
);
359 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
361 /* Call the nested function */
362 error
= (*nproc
)(priv
, cookie
, call_nests
);
364 /* Remove the current task from the list */
365 spin_lock_irqsave(&ncalls
->lock
, flags
);
366 list_del(&tnode
.llink
);
368 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
374 static int ep_poll_wakeup_proc(void *priv
, void *cookie
, int call_nests
)
376 wake_up_nested((wait_queue_head_t
*) cookie
, 1 + call_nests
);
381 * Perform a safe wake up of the poll wait list. The problem is that
382 * with the new callback'd wake up system, it is possible that the
383 * poll callback is reentered from inside the call to wake_up() done
384 * on the poll wait queue head. The rule is that we cannot reenter the
385 * wake up code from the same task more than EP_MAX_NESTS times,
386 * and we cannot reenter the same wait queue head at all. This will
387 * enable to have a hierarchy of epoll file descriptor of no more than
390 static void ep_poll_safewake(wait_queue_head_t
*wq
)
392 ep_call_nested(&poll_safewake_ncalls
, EP_MAX_NESTS
,
393 ep_poll_wakeup_proc
, NULL
, wq
);
397 * This function unregisters poll callbacks from the associated file
398 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
401 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
403 struct list_head
*lsthead
= &epi
->pwqlist
;
404 struct eppoll_entry
*pwq
;
406 while (!list_empty(lsthead
)) {
407 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
409 list_del(&pwq
->llink
);
410 remove_wait_queue(pwq
->whead
, &pwq
->wait
);
411 kmem_cache_free(pwq_cache
, pwq
);
416 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
417 * the scan code, to call f_op->poll(). Also allows for
418 * O(NumReady) performance.
420 * @ep: Pointer to the epoll private data structure.
421 * @sproc: Pointer to the scan callback.
422 * @priv: Private opaque data passed to the @sproc callback.
424 * Returns: The same integer error code returned by the @sproc callback.
426 static int ep_scan_ready_list(struct eventpoll
*ep
,
427 int (*sproc
)(struct eventpoll
*,
428 struct list_head
*, void *),
431 int error
, pwake
= 0;
433 struct epitem
*epi
, *nepi
;
437 * We need to lock this because we could be hit by
438 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
440 mutex_lock(&ep
->mtx
);
443 * Steal the ready list, and re-init the original one to the
444 * empty list. Also, set ep->ovflist to NULL so that events
445 * happening while looping w/out locks, are not lost. We cannot
446 * have the poll callback to queue directly on ep->rdllist,
447 * because we want the "sproc" callback to be able to do it
450 spin_lock_irqsave(&ep
->lock
, flags
);
451 list_splice_init(&ep
->rdllist
, &txlist
);
453 spin_unlock_irqrestore(&ep
->lock
, flags
);
456 * Now call the callback function.
458 error
= (*sproc
)(ep
, &txlist
, priv
);
460 spin_lock_irqsave(&ep
->lock
, flags
);
462 * During the time we spent inside the "sproc" callback, some
463 * other events might have been queued by the poll callback.
464 * We re-insert them inside the main ready-list here.
466 for (nepi
= ep
->ovflist
; (epi
= nepi
) != NULL
;
467 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
469 * We need to check if the item is already in the list.
470 * During the "sproc" callback execution time, items are
471 * queued into ->ovflist but the "txlist" might already
472 * contain them, and the list_splice() below takes care of them.
474 if (!ep_is_linked(&epi
->rdllink
))
475 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
478 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
479 * releasing the lock, events will be queued in the normal way inside
482 ep
->ovflist
= EP_UNACTIVE_PTR
;
485 * Quickly re-inject items left on "txlist".
487 list_splice(&txlist
, &ep
->rdllist
);
489 if (!list_empty(&ep
->rdllist
)) {
491 * Wake up (if active) both the eventpoll wait list and
492 * the ->poll() wait list (delayed after we release the lock).
494 if (waitqueue_active(&ep
->wq
))
495 wake_up_locked(&ep
->wq
);
496 if (waitqueue_active(&ep
->poll_wait
))
499 spin_unlock_irqrestore(&ep
->lock
, flags
);
501 mutex_unlock(&ep
->mtx
);
503 /* We have to call this outside the lock */
505 ep_poll_safewake(&ep
->poll_wait
);
511 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
512 * all the associated resources. Must be called with "mtx" held.
514 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
517 struct file
*file
= epi
->ffd
.file
;
520 * Removes poll wait queue hooks. We _have_ to do this without holding
521 * the "ep->lock" otherwise a deadlock might occur. This because of the
522 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
523 * queue head lock when unregistering the wait queue. The wakeup callback
524 * will run by holding the wait queue head lock and will call our callback
525 * that will try to get "ep->lock".
527 ep_unregister_pollwait(ep
, epi
);
529 /* Remove the current item from the list of epoll hooks */
530 spin_lock(&file
->f_lock
);
531 if (ep_is_linked(&epi
->fllink
))
532 list_del_init(&epi
->fllink
);
533 spin_unlock(&file
->f_lock
);
535 rb_erase(&epi
->rbn
, &ep
->rbr
);
537 spin_lock_irqsave(&ep
->lock
, flags
);
538 if (ep_is_linked(&epi
->rdllink
))
539 list_del_init(&epi
->rdllink
);
540 spin_unlock_irqrestore(&ep
->lock
, flags
);
542 /* At this point it is safe to free the eventpoll item */
543 kmem_cache_free(epi_cache
, epi
);
545 atomic_dec(&ep
->user
->epoll_watches
);
550 static void ep_free(struct eventpoll
*ep
)
555 /* We need to release all tasks waiting for these file */
556 if (waitqueue_active(&ep
->poll_wait
))
557 ep_poll_safewake(&ep
->poll_wait
);
560 * We need to lock this because we could be hit by
561 * eventpoll_release_file() while we're freeing the "struct eventpoll".
562 * We do not need to hold "ep->mtx" here because the epoll file
563 * is on the way to be removed and no one has references to it
564 * anymore. The only hit might come from eventpoll_release_file() but
565 * holding "epmutex" is sufficent here.
567 mutex_lock(&epmutex
);
570 * Walks through the whole tree by unregistering poll callbacks.
572 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
573 epi
= rb_entry(rbp
, struct epitem
, rbn
);
575 ep_unregister_pollwait(ep
, epi
);
579 * Walks through the whole tree by freeing each "struct epitem". At this
580 * point we are sure no poll callbacks will be lingering around, and also by
581 * holding "epmutex" we can be sure that no file cleanup code will hit
582 * us during this operation. So we can avoid the lock on "ep->lock".
584 while ((rbp
= rb_first(&ep
->rbr
)) != NULL
) {
585 epi
= rb_entry(rbp
, struct epitem
, rbn
);
589 mutex_unlock(&epmutex
);
590 mutex_destroy(&ep
->mtx
);
595 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
597 struct eventpoll
*ep
= file
->private_data
;
605 static int ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
608 struct epitem
*epi
, *tmp
;
610 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
611 if (epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
) &
613 return POLLIN
| POLLRDNORM
;
616 * Item has been dropped into the ready list by the poll
617 * callback, but it's not actually ready, as far as
618 * caller requested events goes. We can remove it here.
620 list_del_init(&epi
->rdllink
);
627 static int ep_poll_readyevents_proc(void *priv
, void *cookie
, int call_nests
)
629 return ep_scan_ready_list(priv
, ep_read_events_proc
, NULL
);
632 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
635 struct eventpoll
*ep
= file
->private_data
;
637 /* Insert inside our poll wait queue */
638 poll_wait(file
, &ep
->poll_wait
, wait
);
641 * Proceed to find out if wanted events are really available inside
642 * the ready list. This need to be done under ep_call_nested()
643 * supervision, since the call to f_op->poll() done on listed files
644 * could re-enter here.
646 pollflags
= ep_call_nested(&poll_readywalk_ncalls
, EP_MAX_NESTS
,
647 ep_poll_readyevents_proc
, ep
, ep
);
649 return pollflags
!= -1 ? pollflags
: 0;
652 /* File callbacks that implement the eventpoll file behaviour */
653 static const struct file_operations eventpoll_fops
= {
654 .release
= ep_eventpoll_release
,
655 .poll
= ep_eventpoll_poll
658 /* Fast test to see if the file is an evenpoll file */
659 static inline int is_file_epoll(struct file
*f
)
661 return f
->f_op
== &eventpoll_fops
;
665 * This is called from eventpoll_release() to unlink files from the eventpoll
666 * interface. We need to have this facility to cleanup correctly files that are
667 * closed without being removed from the eventpoll interface.
669 void eventpoll_release_file(struct file
*file
)
671 struct list_head
*lsthead
= &file
->f_ep_links
;
672 struct eventpoll
*ep
;
676 * We don't want to get "file->f_lock" because it is not
677 * necessary. It is not necessary because we're in the "struct file"
678 * cleanup path, and this means that noone is using this file anymore.
679 * So, for example, epoll_ctl() cannot hit here since if we reach this
680 * point, the file counter already went to zero and fget() would fail.
681 * The only hit might come from ep_free() but by holding the mutex
682 * will correctly serialize the operation. We do need to acquire
683 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
684 * from anywhere but ep_free().
686 * Besides, ep_remove() acquires the lock, so we can't hold it here.
688 mutex_lock(&epmutex
);
690 while (!list_empty(lsthead
)) {
691 epi
= list_first_entry(lsthead
, struct epitem
, fllink
);
694 list_del_init(&epi
->fllink
);
695 mutex_lock(&ep
->mtx
);
697 mutex_unlock(&ep
->mtx
);
700 mutex_unlock(&epmutex
);
703 static int ep_alloc(struct eventpoll
**pep
)
706 struct user_struct
*user
;
707 struct eventpoll
*ep
;
709 user
= get_current_user();
711 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
715 spin_lock_init(&ep
->lock
);
716 mutex_init(&ep
->mtx
);
717 init_waitqueue_head(&ep
->wq
);
718 init_waitqueue_head(&ep
->poll_wait
);
719 INIT_LIST_HEAD(&ep
->rdllist
);
721 ep
->ovflist
= EP_UNACTIVE_PTR
;
734 * Search the file inside the eventpoll tree. The RB tree operations
735 * are protected by the "mtx" mutex, and ep_find() must be called with
738 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
742 struct epitem
*epi
, *epir
= NULL
;
743 struct epoll_filefd ffd
;
745 ep_set_ffd(&ffd
, file
, fd
);
746 for (rbp
= ep
->rbr
.rb_node
; rbp
; ) {
747 epi
= rb_entry(rbp
, struct epitem
, rbn
);
748 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
763 * This is the callback that is passed to the wait queue wakeup
764 * machanism. It is called by the stored file descriptors when they
765 * have events to report.
767 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
771 struct epitem
*epi
= ep_item_from_wait(wait
);
772 struct eventpoll
*ep
= epi
->ep
;
774 spin_lock_irqsave(&ep
->lock
, flags
);
777 * If the event mask does not contain any poll(2) event, we consider the
778 * descriptor to be disabled. This condition is likely the effect of the
779 * EPOLLONESHOT bit that disables the descriptor when an event is received,
780 * until the next EPOLL_CTL_MOD will be issued.
782 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
786 * If we are trasfering events to userspace, we can hold no locks
787 * (because we're accessing user memory, and because of linux f_op->poll()
788 * semantics). All the events that happens during that period of time are
789 * chained in ep->ovflist and requeued later on.
791 if (unlikely(ep
->ovflist
!= EP_UNACTIVE_PTR
)) {
792 if (epi
->next
== EP_UNACTIVE_PTR
) {
793 epi
->next
= ep
->ovflist
;
799 /* If this file is already in the ready list we exit soon */
800 if (!ep_is_linked(&epi
->rdllink
))
801 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
804 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
807 if (waitqueue_active(&ep
->wq
))
808 wake_up_locked(&ep
->wq
);
809 if (waitqueue_active(&ep
->poll_wait
))
813 spin_unlock_irqrestore(&ep
->lock
, flags
);
815 /* We have to call this outside the lock */
817 ep_poll_safewake(&ep
->poll_wait
);
823 * This is the callback that is used to add our wait queue to the
824 * target file wakeup lists.
826 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
829 struct epitem
*epi
= ep_item_from_epqueue(pt
);
830 struct eppoll_entry
*pwq
;
832 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
833 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
836 add_wait_queue(whead
, &pwq
->wait
);
837 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
840 /* We have to signal that an error occurred */
845 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
848 struct rb_node
**p
= &ep
->rbr
.rb_node
, *parent
= NULL
;
853 epic
= rb_entry(parent
, struct epitem
, rbn
);
854 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
856 p
= &parent
->rb_right
;
858 p
= &parent
->rb_left
;
860 rb_link_node(&epi
->rbn
, parent
, p
);
861 rb_insert_color(&epi
->rbn
, &ep
->rbr
);
865 * Must be called with "mtx" held.
867 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
868 struct file
*tfile
, int fd
)
870 int error
, revents
, pwake
= 0;
873 struct ep_pqueue epq
;
875 if (unlikely(atomic_read(&ep
->user
->epoll_watches
) >=
878 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
881 /* Item initialization follow here ... */
882 INIT_LIST_HEAD(&epi
->rdllink
);
883 INIT_LIST_HEAD(&epi
->fllink
);
884 INIT_LIST_HEAD(&epi
->pwqlist
);
886 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
889 epi
->next
= EP_UNACTIVE_PTR
;
891 /* Initialize the poll table using the queue callback */
893 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
896 * Attach the item to the poll hooks and get current event bits.
897 * We can safely use the file* here because its usage count has
898 * been increased by the caller of this function. Note that after
899 * this operation completes, the poll callback can start hitting
902 revents
= tfile
->f_op
->poll(tfile
, &epq
.pt
);
905 * We have to check if something went wrong during the poll wait queue
906 * install process. Namely an allocation for a wait queue failed due
907 * high memory pressure.
911 goto error_unregister
;
913 /* Add the current item to the list of active epoll hook for this file */
914 spin_lock(&tfile
->f_lock
);
915 list_add_tail(&epi
->fllink
, &tfile
->f_ep_links
);
916 spin_unlock(&tfile
->f_lock
);
919 * Add the current item to the RB tree. All RB tree operations are
920 * protected by "mtx", and ep_insert() is called with "mtx" held.
922 ep_rbtree_insert(ep
, epi
);
924 /* We have to drop the new item inside our item list to keep track of it */
925 spin_lock_irqsave(&ep
->lock
, flags
);
927 /* If the file is already "ready" we drop it inside the ready list */
928 if ((revents
& event
->events
) && !ep_is_linked(&epi
->rdllink
)) {
929 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
931 /* Notify waiting tasks that events are available */
932 if (waitqueue_active(&ep
->wq
))
933 wake_up_locked(&ep
->wq
);
934 if (waitqueue_active(&ep
->poll_wait
))
938 spin_unlock_irqrestore(&ep
->lock
, flags
);
940 atomic_inc(&ep
->user
->epoll_watches
);
942 /* We have to call this outside the lock */
944 ep_poll_safewake(&ep
->poll_wait
);
949 ep_unregister_pollwait(ep
, epi
);
952 * We need to do this because an event could have been arrived on some
953 * allocated wait queue. Note that we don't care about the ep->ovflist
954 * list, since that is used/cleaned only inside a section bound by "mtx".
955 * And ep_insert() is called with "mtx" held.
957 spin_lock_irqsave(&ep
->lock
, flags
);
958 if (ep_is_linked(&epi
->rdllink
))
959 list_del_init(&epi
->rdllink
);
960 spin_unlock_irqrestore(&ep
->lock
, flags
);
962 kmem_cache_free(epi_cache
, epi
);
968 * Modify the interest event mask by dropping an event if the new mask
969 * has a match in the current file status. Must be called with "mtx" held.
971 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
, struct epoll_event
*event
)
974 unsigned int revents
;
978 * Set the new event interest mask before calling f_op->poll(), otherwise
979 * a potential race might occur. In fact if we do this operation inside
980 * the lock, an event might happen between the f_op->poll() call and the
981 * new event set registering.
983 epi
->event
.events
= event
->events
;
986 * Get current event bits. We can safely use the file* here because
987 * its usage count has been increased by the caller of this function.
989 revents
= epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
);
991 spin_lock_irqsave(&ep
->lock
, flags
);
993 /* Copy the data member from inside the lock */
994 epi
->event
.data
= event
->data
;
997 * If the item is "hot" and it is not registered inside the ready
998 * list, push it inside.
1000 if (revents
& event
->events
) {
1001 if (!ep_is_linked(&epi
->rdllink
)) {
1002 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1004 /* Notify waiting tasks that events are available */
1005 if (waitqueue_active(&ep
->wq
))
1006 wake_up_locked(&ep
->wq
);
1007 if (waitqueue_active(&ep
->poll_wait
))
1011 spin_unlock_irqrestore(&ep
->lock
, flags
);
1013 /* We have to call this outside the lock */
1015 ep_poll_safewake(&ep
->poll_wait
);
1020 static int ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1023 struct ep_send_events_data
*esed
= priv
;
1025 unsigned int revents
;
1027 struct epoll_event __user
*uevent
;
1030 * We can loop without lock because we are passed a task private list.
1031 * Items cannot vanish during the loop because ep_scan_ready_list() is
1032 * holding "mtx" during this call.
1034 for (eventcnt
= 0, uevent
= esed
->events
;
1035 !list_empty(head
) && eventcnt
< esed
->maxevents
;) {
1036 epi
= list_first_entry(head
, struct epitem
, rdllink
);
1038 list_del_init(&epi
->rdllink
);
1040 revents
= epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
) &
1044 * If the event mask intersect the caller-requested one,
1045 * deliver the event to userspace. Again, ep_scan_ready_list()
1046 * is holding "mtx", so no operations coming from userspace
1047 * can change the item.
1050 if (__put_user(revents
, &uevent
->events
) ||
1051 __put_user(epi
->event
.data
, &uevent
->data
)) {
1052 list_add(&epi
->rdllink
, head
);
1053 return eventcnt
? eventcnt
: -EFAULT
;
1057 if (epi
->event
.events
& EPOLLONESHOT
)
1058 epi
->event
.events
&= EP_PRIVATE_BITS
;
1059 else if (!(epi
->event
.events
& EPOLLET
)) {
1061 * If this file has been added with Level
1062 * Trigger mode, we need to insert back inside
1063 * the ready list, so that the next call to
1064 * epoll_wait() will check again the events
1065 * availability. At this point, noone can insert
1066 * into ep->rdllist besides us. The epoll_ctl()
1067 * callers are locked out by
1068 * ep_scan_ready_list() holding "mtx" and the
1069 * poll callback will queue them in ep->ovflist.
1071 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1079 static int ep_send_events(struct eventpoll
*ep
,
1080 struct epoll_event __user
*events
, int maxevents
)
1082 struct ep_send_events_data esed
;
1084 esed
.maxevents
= maxevents
;
1085 esed
.events
= events
;
1087 return ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
);
1090 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1091 int maxevents
, long timeout
)
1094 unsigned long flags
;
1099 * Calculate the timeout by checking for the "infinite" value (-1)
1100 * and the overflow condition. The passed timeout is in milliseconds,
1101 * that why (t * HZ) / 1000.
1103 jtimeout
= (timeout
< 0 || timeout
>= EP_MAX_MSTIMEO
) ?
1104 MAX_SCHEDULE_TIMEOUT
: (timeout
* HZ
+ 999) / 1000;
1107 spin_lock_irqsave(&ep
->lock
, flags
);
1110 if (list_empty(&ep
->rdllist
)) {
1112 * We don't have any available event to return to the caller.
1113 * We need to sleep here, and we will be wake up by
1114 * ep_poll_callback() when events will become available.
1116 init_waitqueue_entry(&wait
, current
);
1117 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
1118 __add_wait_queue(&ep
->wq
, &wait
);
1122 * We don't want to sleep if the ep_poll_callback() sends us
1123 * a wakeup in between. That's why we set the task state
1124 * to TASK_INTERRUPTIBLE before doing the checks.
1126 set_current_state(TASK_INTERRUPTIBLE
);
1127 if (!list_empty(&ep
->rdllist
) || !jtimeout
)
1129 if (signal_pending(current
)) {
1134 spin_unlock_irqrestore(&ep
->lock
, flags
);
1135 jtimeout
= schedule_timeout(jtimeout
);
1136 spin_lock_irqsave(&ep
->lock
, flags
);
1138 __remove_wait_queue(&ep
->wq
, &wait
);
1140 set_current_state(TASK_RUNNING
);
1142 /* Is it worth to try to dig for events ? */
1143 eavail
= !list_empty(&ep
->rdllist
) || ep
->ovflist
!= EP_UNACTIVE_PTR
;
1145 spin_unlock_irqrestore(&ep
->lock
, flags
);
1148 * Try to transfer events to user space. In case we get 0 events and
1149 * there's still timeout left over, we go trying again in search of
1152 if (!res
&& eavail
&&
1153 !(res
= ep_send_events(ep
, events
, maxevents
)) && jtimeout
)
1160 * Open an eventpoll file descriptor.
1162 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
1165 struct eventpoll
*ep
= NULL
;
1167 /* Check the EPOLL_* constant for consistency. */
1168 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
1170 if (flags
& ~EPOLL_CLOEXEC
)
1173 * Create the internal data structure ("struct eventpoll").
1175 error
= ep_alloc(&ep
);
1179 * Creates all the items needed to setup an eventpoll file. That is,
1180 * a file structure and a free file descriptor.
1182 error
= anon_inode_getfd("[eventpoll]", &eventpoll_fops
, ep
,
1190 SYSCALL_DEFINE1(epoll_create
, int, size
)
1195 return sys_epoll_create1(0);
1199 * The following function implements the controller interface for
1200 * the eventpoll file that enables the insertion/removal/change of
1201 * file descriptors inside the interest set.
1203 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
1204 struct epoll_event __user
*, event
)
1207 struct file
*file
, *tfile
;
1208 struct eventpoll
*ep
;
1210 struct epoll_event epds
;
1213 if (ep_op_has_event(op
) &&
1214 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
1217 /* Get the "struct file *" for the eventpoll file */
1223 /* Get the "struct file *" for the target file */
1228 /* The target file descriptor must support poll */
1230 if (!tfile
->f_op
|| !tfile
->f_op
->poll
)
1231 goto error_tgt_fput
;
1234 * We have to check that the file structure underneath the file descriptor
1235 * the user passed to us _is_ an eventpoll file. And also we do not permit
1236 * adding an epoll file descriptor inside itself.
1239 if (file
== tfile
|| !is_file_epoll(file
))
1240 goto error_tgt_fput
;
1243 * At this point it is safe to assume that the "private_data" contains
1244 * our own data structure.
1246 ep
= file
->private_data
;
1248 mutex_lock(&ep
->mtx
);
1251 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1252 * above, we can be sure to be able to use the item looked up by
1253 * ep_find() till we release the mutex.
1255 epi
= ep_find(ep
, tfile
, fd
);
1261 epds
.events
|= POLLERR
| POLLHUP
;
1263 error
= ep_insert(ep
, &epds
, tfile
, fd
);
1269 error
= ep_remove(ep
, epi
);
1275 epds
.events
|= POLLERR
| POLLHUP
;
1276 error
= ep_modify(ep
, epi
, &epds
);
1281 mutex_unlock(&ep
->mtx
);
1293 * Implement the event wait interface for the eventpoll file. It is the kernel
1294 * part of the user space epoll_wait(2).
1296 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
1297 int, maxevents
, int, timeout
)
1301 struct eventpoll
*ep
;
1303 /* The maximum number of event must be greater than zero */
1304 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
1307 /* Verify that the area passed by the user is writeable */
1308 if (!access_ok(VERIFY_WRITE
, events
, maxevents
* sizeof(struct epoll_event
))) {
1313 /* Get the "struct file *" for the eventpoll file */
1320 * We have to check that the file structure underneath the fd
1321 * the user passed to us _is_ an eventpoll file.
1324 if (!is_file_epoll(file
))
1328 * At this point it is safe to assume that the "private_data" contains
1329 * our own data structure.
1331 ep
= file
->private_data
;
1333 /* Time to fish for events ... */
1334 error
= ep_poll(ep
, events
, maxevents
, timeout
);
1343 #ifdef HAVE_SET_RESTORE_SIGMASK
1346 * Implement the event wait interface for the eventpoll file. It is the kernel
1347 * part of the user space epoll_pwait(2).
1349 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
1350 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
1354 sigset_t ksigmask
, sigsaved
;
1357 * If the caller wants a certain signal mask to be set during the wait,
1361 if (sigsetsize
!= sizeof(sigset_t
))
1363 if (copy_from_user(&ksigmask
, sigmask
, sizeof(ksigmask
)))
1365 sigdelsetmask(&ksigmask
, sigmask(SIGKILL
) | sigmask(SIGSTOP
));
1366 sigprocmask(SIG_SETMASK
, &ksigmask
, &sigsaved
);
1369 error
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
1372 * If we changed the signal mask, we need to restore the original one.
1373 * In case we've got a signal while waiting, we do not restore the
1374 * signal mask yet, and we allow do_signal() to deliver the signal on
1375 * the way back to userspace, before the signal mask is restored.
1378 if (error
== -EINTR
) {
1379 memcpy(¤t
->saved_sigmask
, &sigsaved
,
1381 set_restore_sigmask();
1383 sigprocmask(SIG_SETMASK
, &sigsaved
, NULL
);
1389 #endif /* HAVE_SET_RESTORE_SIGMASK */
1391 static int __init
eventpoll_init(void)
1397 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1399 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
1402 /* Initialize the structure used to perform safe poll wait head wake ups */
1403 ep_nested_calls_init(&poll_safewake_ncalls
);
1405 /* Initialize the structure used to perform file's f_op->poll() calls */
1406 ep_nested_calls_init(&poll_readywalk_ncalls
);
1408 /* Allocates slab cache used to allocate "struct epitem" items */
1409 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
1410 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
1412 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1413 pwq_cache
= kmem_cache_create("eventpoll_pwq",
1414 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
, NULL
);
1418 fs_initcall(eventpoll_init
);