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>
12 * Copyright (c) 2013, NVIDIA CORPORATION. All rights reserved.
16 #include <linux/init.h>
17 #include <linux/kernel.h>
18 #include <linux/sched.h>
20 #include <linux/file.h>
21 #include <linux/signal.h>
22 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/poll.h>
26 #include <linux/string.h>
27 #include <linux/list.h>
28 #include <linux/hash.h>
29 #include <linux/spinlock.h>
30 #include <linux/syscalls.h>
31 #include <linux/rbtree.h>
32 #include <linux/wait.h>
33 #include <linux/eventpoll.h>
34 #include <linux/mount.h>
35 #include <linux/bitops.h>
36 #include <linux/mutex.h>
37 #include <linux/anon_inodes.h>
38 #include <linux/device.h>
39 #include <linux/freezer.h>
40 #include <asm/uaccess.h>
43 #include <linux/atomic.h>
44 #include <linux/proc_fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/compat.h>
50 * There are three level of locking required by epoll :
54 * 3) ep->lock (spinlock)
56 * The acquire order is the one listed above, from 1 to 3.
57 * We need a spinlock (ep->lock) because we manipulate objects
58 * from inside the poll callback, that might be triggered from
59 * a wake_up() that in turn might be called from IRQ context.
60 * So we can't sleep inside the poll callback and hence we need
61 * a spinlock. During the event transfer loop (from kernel to
62 * user space) we could end up sleeping due a copy_to_user(), so
63 * we need a lock that will allow us to sleep. This lock is a
64 * mutex (ep->mtx). It is acquired during the event transfer loop,
65 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
66 * Then we also need a global mutex to serialize eventpoll_release_file()
68 * This mutex is acquired by ep_free() during the epoll file
69 * cleanup path and it is also acquired by eventpoll_release_file()
70 * if a file has been pushed inside an epoll set and it is then
71 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
72 * It is also acquired when inserting an epoll fd onto another epoll
73 * fd. We do this so that we walk the epoll tree and ensure that this
74 * insertion does not create a cycle of epoll file descriptors, which
75 * could lead to deadlock. We need a global mutex to prevent two
76 * simultaneous inserts (A into B and B into A) from racing and
77 * constructing a cycle without either insert observing that it is
79 * It is necessary to acquire multiple "ep->mtx"es at once in the
80 * case when one epoll fd is added to another. In this case, we
81 * always acquire the locks in the order of nesting (i.e. after
82 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
83 * before e2->mtx). Since we disallow cycles of epoll file
84 * descriptors, this ensures that the mutexes are well-ordered. In
85 * order to communicate this nesting to lockdep, when walking a tree
86 * of epoll file descriptors, we use the current recursion depth as
88 * It is possible to drop the "ep->mtx" and to use the global
89 * mutex "epmutex" (together with "ep->lock") to have it working,
90 * but having "ep->mtx" will make the interface more scalable.
91 * Events that require holding "epmutex" are very rare, while for
92 * normal operations the epoll private "ep->mtx" will guarantee
93 * a better scalability.
96 /* Epoll private bits inside the event mask */
97 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
99 /* Maximum number of nesting allowed inside epoll sets */
100 #define EP_MAX_NESTS 4
102 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
104 #define EP_UNACTIVE_PTR ((void *) -1L)
106 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
108 struct epoll_filefd {
114 * Structure used to track possible nested calls, for too deep recursions
117 struct nested_call_node {
118 struct list_head llink;
124 * This structure is used as collector for nested calls, to check for
125 * maximum recursion dept and loop cycles.
127 struct nested_calls {
128 struct list_head tasks_call_list;
133 * Each file descriptor added to the eventpoll interface will
134 * have an entry of this type linked to the "rbr" RB tree.
135 * Avoid increasing the size of this struct, there can be many thousands
136 * of these on a server and we do not want this to take another cache line.
139 /* RB tree node used to link this structure to the eventpoll RB tree */
142 /* List header used to link this structure to the eventpoll ready list */
143 struct list_head rdllink;
146 * Works together "struct eventpoll"->ovflist in keeping the
147 * single linked chain of items.
151 /* The file descriptor information this item refers to */
152 struct epoll_filefd ffd;
154 /* Number of active wait queue attached to poll operations */
157 /* List containing poll wait queues */
158 struct list_head pwqlist;
160 /* The "container" of this item */
161 struct eventpoll *ep;
163 /* List header used to link this item to the "struct file" items list */
164 struct list_head fllink;
166 /* wakeup_source used when EPOLLWAKEUP is set */
167 struct wakeup_source __rcu *ws;
169 /* The structure that describe the interested events and the source fd */
170 struct epoll_event event;
174 * This structure is stored inside the "private_data" member of the file
175 * structure and represents the main data structure for the eventpoll
179 /* Protect the access to this structure */
183 * This mutex is used to ensure that files are not removed
184 * while epoll is using them. This is held during the event
185 * collection loop, the file cleanup path, the epoll file exit
186 * code and the ctl operations.
190 /* Wait queue used by sys_epoll_wait() */
191 wait_queue_head_t wq;
193 /* Wait queue used by file->poll() */
194 wait_queue_head_t poll_wait;
196 /* List of ready file descriptors */
197 struct list_head rdllist;
199 /* RB tree root used to store monitored fd structs */
203 * This is a single linked list that chains all the "struct epitem" that
204 * happened while transferring ready events to userspace w/out
207 struct epitem *ovflist;
209 /* wakeup_source used when ep_scan_ready_list is running */
210 struct wakeup_source *ws;
212 /* The user that created the eventpoll descriptor */
213 struct user_struct *user;
217 /* used to optimize loop detection check */
219 struct list_head visited_list_link;
222 /* Wait structure used by the poll hooks */
223 struct eppoll_entry {
224 /* List header used to link this structure to the "struct epitem" */
225 struct list_head llink;
227 /* The "base" pointer is set to the container "struct epitem" */
231 * Wait queue item that will be linked to the target file wait
236 /* The wait queue head that linked the "wait" wait queue item */
237 wait_queue_head_t *whead;
240 /* Wrapper struct used by poll queueing */
246 /* Used by the ep_send_events() function as callback private data */
247 struct ep_send_events_data {
249 struct epoll_event __user *events;
253 * Configuration options available inside /proc/sys/fs/epoll/
255 /* Maximum number of epoll watched descriptors, per user */
256 static long max_user_watches __read_mostly;
259 * This mutex is used to serialize ep_free() and eventpoll_release_file().
261 static DEFINE_MUTEX(epmutex);
263 /* Used to check for epoll file descriptor inclusion loops */
264 static struct nested_calls poll_loop_ncalls;
266 /* Used for safe wake up implementation */
267 static struct nested_calls poll_safewake_ncalls;
269 /* Used to call file's f_op->poll() under the nested calls boundaries */
270 static struct nested_calls poll_readywalk_ncalls;
272 /* Slab cache used to allocate "struct epitem" */
273 static struct kmem_cache *epi_cache __read_mostly;
275 /* Slab cache used to allocate "struct eppoll_entry" */
276 static struct kmem_cache *pwq_cache __read_mostly;
278 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
279 static LIST_HEAD(visited_list);
282 * List of files with newly added links, where we may need to limit the number
283 * of emanating paths. Protected by the epmutex.
285 static LIST_HEAD(tfile_check_list);
289 #include <linux/sysctl.h>
292 static long long_max = LONG_MAX;
294 ctl_table epoll_table[] = {
296 .procname = "max_user_watches",
297 .data = &max_user_watches,
298 .maxlen = sizeof(max_user_watches),
300 .proc_handler = proc_doulongvec_minmax,
306 #endif /* CONFIG_SYSCTL */
308 static const struct file_operations eventpoll_fops;
310 static inline int is_file_epoll(struct file *f)
312 return f->f_op == &eventpoll_fops;
315 /* Setup the structure that is used as key for the RB tree */
316 static inline void ep_set_ffd(struct epoll_filefd *ffd,
317 struct file *file, int fd)
323 /* Compare RB tree keys */
324 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
325 struct epoll_filefd *p2)
327 return (p1->file > p2->file ? +1:
328 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
331 /* Tells us if the item is currently linked */
332 static inline int ep_is_linked(struct list_head *p)
334 return !list_empty(p);
337 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
339 return container_of(p, struct eppoll_entry, wait);
342 /* Get the "struct epitem" from a wait queue pointer */
343 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
345 return container_of(p, struct eppoll_entry, wait)->base;
348 /* Get the "struct epitem" from an epoll queue wrapper */
349 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
351 return container_of(p, struct ep_pqueue, pt)->epi;
354 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
355 static inline int ep_op_has_event(int op)
357 return op != EPOLL_CTL_DEL;
360 /* Initialize the poll safe wake up structure */
361 static void ep_nested_calls_init(struct nested_calls *ncalls)
363 INIT_LIST_HEAD(&ncalls->tasks_call_list);
364 spin_lock_init(&ncalls->lock);
368 * ep_events_available - Checks if ready events might be available.
370 * @ep: Pointer to the eventpoll context.
372 * Returns: Returns a value different than zero if ready events are available,
375 static inline int ep_events_available(struct eventpoll *ep)
377 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
381 * ep_call_nested - Perform a bound (possibly) nested call, by checking
382 * that the recursion limit is not exceeded, and that
383 * the same nested call (by the meaning of same cookie) is
386 * @ncalls: Pointer to the nested_calls structure to be used for this call.
387 * @max_nests: Maximum number of allowed nesting calls.
388 * @nproc: Nested call core function pointer.
389 * @priv: Opaque data to be passed to the @nproc callback.
390 * @cookie: Cookie to be used to identify this nested call.
391 * @ctx: This instance context.
393 * Returns: Returns the code returned by the @nproc callback, or -1 if
394 * the maximum recursion limit has been exceeded.
396 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
397 int (*nproc)(void *, void *, int), void *priv,
398 void *cookie, void *ctx)
400 int error, call_nests = 0;
402 struct list_head *lsthead = &ncalls->tasks_call_list;
403 struct nested_call_node *tncur;
404 struct nested_call_node tnode;
406 spin_lock_irqsave(&ncalls->lock, flags);
409 * Try to see if the current task is already inside this wakeup call.
410 * We use a list here, since the population inside this set is always
413 list_for_each_entry(tncur, lsthead, llink) {
414 if (tncur->ctx == ctx &&
415 (tncur->cookie == cookie || ++call_nests > max_nests)) {
417 * Ops ... loop detected or maximum nest level reached.
418 * We abort this wake by breaking the cycle itself.
425 /* Add the current task and cookie to the list */
427 tnode.cookie = cookie;
428 list_add(&tnode.llink, lsthead);
430 spin_unlock_irqrestore(&ncalls->lock, flags);
432 /* Call the nested function */
433 error = (*nproc)(priv, cookie, call_nests);
435 /* Remove the current task from the list */
436 spin_lock_irqsave(&ncalls->lock, flags);
437 list_del(&tnode.llink);
439 spin_unlock_irqrestore(&ncalls->lock, flags);
445 * As described in commit 0ccf831cb lockdep: annotate epoll
446 * the use of wait queues used by epoll is done in a very controlled
447 * manner. Wake ups can nest inside each other, but are never done
448 * with the same locking. For example:
451 * efd1 = epoll_create();
452 * efd2 = epoll_create();
453 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
454 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
456 * When a packet arrives to the device underneath "dfd", the net code will
457 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
458 * callback wakeup entry on that queue, and the wake_up() performed by the
459 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
460 * (efd1) notices that it may have some event ready, so it needs to wake up
461 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
462 * that ends up in another wake_up(), after having checked about the
463 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
464 * avoid stack blasting.
466 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
467 * this special case of epoll.
469 #ifdef CONFIG_DEBUG_LOCK_ALLOC
470 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
471 unsigned long events, int subclass)
475 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
476 wake_up_locked_poll(wqueue, events);
477 spin_unlock_irqrestore(&wqueue->lock, flags);
480 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
481 unsigned long events, int subclass)
483 wake_up_poll(wqueue, events);
487 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
489 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
495 * Perform a safe wake up of the poll wait list. The problem is that
496 * with the new callback'd wake up system, it is possible that the
497 * poll callback is reentered from inside the call to wake_up() done
498 * on the poll wait queue head. The rule is that we cannot reenter the
499 * wake up code from the same task more than EP_MAX_NESTS times,
500 * and we cannot reenter the same wait queue head at all. This will
501 * enable to have a hierarchy of epoll file descriptor of no more than
504 static void ep_poll_safewake(wait_queue_head_t *wq)
506 int this_cpu = get_cpu();
508 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
509 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
514 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
516 wait_queue_head_t *whead;
519 /* If it is cleared by POLLFREE, it should be rcu-safe */
520 whead = rcu_dereference(pwq->whead);
522 remove_wait_queue(whead, &pwq->wait);
527 * This function unregisters poll callbacks from the associated file
528 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
531 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
533 struct list_head *lsthead = &epi->pwqlist;
534 struct eppoll_entry *pwq;
536 while (!list_empty(lsthead)) {
537 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
539 list_del(&pwq->llink);
540 ep_remove_wait_queue(pwq);
541 kmem_cache_free(pwq_cache, pwq);
545 /* call only when ep->mtx is held */
546 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
548 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
551 /* call only when ep->mtx is held */
552 static inline void ep_pm_stay_awake(struct epitem *epi)
554 struct wakeup_source *ws = ep_wakeup_source(epi);
560 static inline bool ep_has_wakeup_source(struct epitem *epi)
562 return rcu_access_pointer(epi->ws) ? true : false;
565 /* call when ep->mtx cannot be held (ep_poll_callback) */
566 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
568 struct wakeup_source *ws;
571 ws = rcu_dereference(epi->ws);
578 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
579 * the scan code, to call f_op->poll(). Also allows for
580 * O(NumReady) performance.
582 * @ep: Pointer to the epoll private data structure.
583 * @sproc: Pointer to the scan callback.
584 * @priv: Private opaque data passed to the @sproc callback.
585 * @depth: The current depth of recursive f_op->poll calls.
587 * Returns: The same integer error code returned by the @sproc callback.
589 static int ep_scan_ready_list(struct eventpoll *ep,
590 int (*sproc)(struct eventpoll *,
591 struct list_head *, void *),
595 int error, pwake = 0;
597 struct epitem *epi, *nepi;
601 * We need to lock this because we could be hit by
602 * eventpoll_release_file() and epoll_ctl().
604 mutex_lock_nested(&ep->mtx, depth);
607 * Steal the ready list, and re-init the original one to the
608 * empty list. Also, set ep->ovflist to NULL so that events
609 * happening while looping w/out locks, are not lost. We cannot
610 * have the poll callback to queue directly on ep->rdllist,
611 * because we want the "sproc" callback to be able to do it
614 spin_lock_irqsave(&ep->lock, flags);
615 list_splice_init(&ep->rdllist, &txlist);
617 spin_unlock_irqrestore(&ep->lock, flags);
620 * Now call the callback function.
622 error = (*sproc)(ep, &txlist, priv);
624 spin_lock_irqsave(&ep->lock, flags);
626 * During the time we spent inside the "sproc" callback, some
627 * other events might have been queued by the poll callback.
628 * We re-insert them inside the main ready-list here.
630 for (nepi = ep->ovflist; (epi = nepi) != NULL;
631 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
633 * We need to check if the item is already in the list.
634 * During the "sproc" callback execution time, items are
635 * queued into ->ovflist but the "txlist" might already
636 * contain them, and the list_splice() below takes care of them.
638 if (!ep_is_linked(&epi->rdllink)) {
639 list_add_tail(&epi->rdllink, &ep->rdllist);
640 ep_pm_stay_awake(epi);
644 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
645 * releasing the lock, events will be queued in the normal way inside
648 ep->ovflist = EP_UNACTIVE_PTR;
651 * Quickly re-inject items left on "txlist".
653 list_splice(&txlist, &ep->rdllist);
656 if (!list_empty(&ep->rdllist)) {
658 * Wake up (if active) both the eventpoll wait list and
659 * the ->poll() wait list (delayed after we release the lock).
661 if (waitqueue_active(&ep->wq))
662 wake_up_locked(&ep->wq);
663 if (waitqueue_active(&ep->poll_wait))
666 spin_unlock_irqrestore(&ep->lock, flags);
668 mutex_unlock(&ep->mtx);
670 /* We have to call this outside the lock */
672 ep_poll_safewake(&ep->poll_wait);
678 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
679 * all the associated resources. Must be called with "mtx" held.
681 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
684 struct file *file = epi->ffd.file;
687 * Removes poll wait queue hooks. We _have_ to do this without holding
688 * the "ep->lock" otherwise a deadlock might occur. This because of the
689 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
690 * queue head lock when unregistering the wait queue. The wakeup callback
691 * will run by holding the wait queue head lock and will call our callback
692 * that will try to get "ep->lock".
694 ep_unregister_pollwait(ep, epi);
696 /* Remove the current item from the list of epoll hooks */
697 spin_lock(&file->f_lock);
698 if (ep_is_linked(&epi->fllink))
699 list_del_init(&epi->fllink);
700 spin_unlock(&file->f_lock);
702 rb_erase(&epi->rbn, &ep->rbr);
704 spin_lock_irqsave(&ep->lock, flags);
705 if (ep_is_linked(&epi->rdllink))
706 list_del_init(&epi->rdllink);
707 spin_unlock_irqrestore(&ep->lock, flags);
709 wakeup_source_unregister(ep_wakeup_source(epi));
711 /* At this point it is safe to free the eventpoll item */
712 kmem_cache_free(epi_cache, epi);
714 atomic_long_dec(&ep->user->epoll_watches);
719 static void ep_free(struct eventpoll *ep)
724 /* We need to release all tasks waiting for these file */
725 if (waitqueue_active(&ep->poll_wait))
726 ep_poll_safewake(&ep->poll_wait);
729 * We need to lock this because we could be hit by
730 * eventpoll_release_file() while we're freeing the "struct eventpoll".
731 * We do not need to hold "ep->mtx" here because the epoll file
732 * is on the way to be removed and no one has references to it
733 * anymore. The only hit might come from eventpoll_release_file() but
734 * holding "epmutex" is sufficient here.
736 mutex_lock(&epmutex);
739 * Walks through the whole tree by unregistering poll callbacks.
741 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
742 epi = rb_entry(rbp, struct epitem, rbn);
744 ep_unregister_pollwait(ep, epi);
748 * Walks through the whole tree by freeing each "struct epitem". At this
749 * point we are sure no poll callbacks will be lingering around, and also by
750 * holding "epmutex" we can be sure that no file cleanup code will hit
751 * us during this operation. So we can avoid the lock on "ep->lock".
752 * We do not need to lock ep->mtx, either, we only do it to prevent
755 mutex_lock(&ep->mtx);
756 while ((rbp = rb_first(&ep->rbr)) != NULL) {
757 epi = rb_entry(rbp, struct epitem, rbn);
760 mutex_unlock(&ep->mtx);
762 mutex_unlock(&epmutex);
763 mutex_destroy(&ep->mtx);
765 wakeup_source_unregister(ep->ws);
769 static int ep_eventpoll_release(struct inode *inode, struct file *file)
771 struct eventpoll *ep = file->private_data;
779 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
781 pt->_key = epi->event.events;
783 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
786 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
789 struct epitem *epi, *tmp;
792 init_poll_funcptr(&pt, NULL);
794 list_for_each_entry_safe(epi, tmp, head, rdllink) {
795 if (ep_item_poll(epi, &pt))
796 return POLLIN | POLLRDNORM;
799 * Item has been dropped into the ready list by the poll
800 * callback, but it's not actually ready, as far as
801 * caller requested events goes. We can remove it here.
803 __pm_relax(ep_wakeup_source(epi));
804 list_del_init(&epi->rdllink);
811 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
813 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
816 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
819 struct eventpoll *ep = file->private_data;
821 /* Insert inside our poll wait queue */
822 poll_wait(file, &ep->poll_wait, wait);
825 * Proceed to find out if wanted events are really available inside
826 * the ready list. This need to be done under ep_call_nested()
827 * supervision, since the call to f_op->poll() done on listed files
828 * could re-enter here.
830 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
831 ep_poll_readyevents_proc, ep, ep, current);
833 return pollflags != -1 ? pollflags : 0;
836 #ifdef CONFIG_PROC_FS
837 static int ep_show_fdinfo(struct seq_file *m, struct file *f)
839 struct eventpoll *ep = f->private_data;
843 mutex_lock(&ep->mtx);
844 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
845 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
847 ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
848 epi->ffd.fd, epi->event.events,
849 (long long)epi->event.data);
853 mutex_unlock(&ep->mtx);
859 /* File callbacks that implement the eventpoll file behaviour */
860 static const struct file_operations eventpoll_fops = {
861 #ifdef CONFIG_PROC_FS
862 .show_fdinfo = ep_show_fdinfo,
864 .release = ep_eventpoll_release,
865 .poll = ep_eventpoll_poll,
866 .llseek = noop_llseek,
870 * This is called from eventpoll_release() to unlink files from the eventpoll
871 * interface. We need to have this facility to cleanup correctly files that are
872 * closed without being removed from the eventpoll interface.
874 void eventpoll_release_file(struct file *file)
876 struct list_head *lsthead = &file->f_ep_links;
877 struct eventpoll *ep;
881 * We don't want to get "file->f_lock" because it is not
882 * necessary. It is not necessary because we're in the "struct file"
883 * cleanup path, and this means that no one is using this file anymore.
884 * So, for example, epoll_ctl() cannot hit here since if we reach this
885 * point, the file counter already went to zero and fget() would fail.
886 * The only hit might come from ep_free() but by holding the mutex
887 * will correctly serialize the operation. We do need to acquire
888 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
889 * from anywhere but ep_free().
891 * Besides, ep_remove() acquires the lock, so we can't hold it here.
893 mutex_lock(&epmutex);
895 while (!list_empty(lsthead)) {
896 epi = list_first_entry(lsthead, struct epitem, fllink);
899 list_del_init(&epi->fllink);
900 mutex_lock_nested(&ep->mtx, 0);
902 mutex_unlock(&ep->mtx);
905 mutex_unlock(&epmutex);
908 static int ep_alloc(struct eventpoll **pep)
911 struct user_struct *user;
912 struct eventpoll *ep;
914 user = get_current_user();
916 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
920 spin_lock_init(&ep->lock);
921 mutex_init(&ep->mtx);
922 init_waitqueue_head(&ep->wq);
923 init_waitqueue_head(&ep->poll_wait);
924 INIT_LIST_HEAD(&ep->rdllist);
926 ep->ovflist = EP_UNACTIVE_PTR;
939 * Search the file inside the eventpoll tree. The RB tree operations
940 * are protected by the "mtx" mutex, and ep_find() must be called with
943 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
947 struct epitem *epi, *epir = NULL;
948 struct epoll_filefd ffd;
950 ep_set_ffd(&ffd, file, fd);
951 for (rbp = ep->rbr.rb_node; rbp; ) {
952 epi = rb_entry(rbp, struct epitem, rbn);
953 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
968 * This is the callback that is passed to the wait queue wakeup
969 * mechanism. It is called by the stored file descriptors when they
970 * have events to report.
972 int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
976 struct epitem *epi = ep_item_from_wait(wait);
977 struct eventpoll *ep = epi->ep;
979 if ((unsigned long)key & POLLFREE) {
980 ep_pwq_from_wait(wait)->whead = NULL;
982 * whead = NULL above can race with ep_remove_wait_queue()
983 * which can do another remove_wait_queue() after us, so we
984 * can't use __remove_wait_queue(). whead->lock is held by
987 list_del_init(&wait->task_list);
990 spin_lock_irqsave(&ep->lock, flags);
993 * If the event mask does not contain any poll(2) event, we consider the
994 * descriptor to be disabled. This condition is likely the effect of the
995 * EPOLLONESHOT bit that disables the descriptor when an event is received,
996 * until the next EPOLL_CTL_MOD will be issued.
998 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1002 * Check the events coming with the callback. At this stage, not
1003 * every device reports the events in the "key" parameter of the
1004 * callback. We need to be able to handle both cases here, hence the
1005 * test for "key" != NULL before the event match test.
1007 if (key && !((unsigned long) key & epi->event.events))
1011 * If we are transferring events to userspace, we can hold no locks
1012 * (because we're accessing user memory, and because of linux f_op->poll()
1013 * semantics). All the events that happen during that period of time are
1014 * chained in ep->ovflist and requeued later on.
1016 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1017 if (epi->next == EP_UNACTIVE_PTR) {
1018 epi->next = ep->ovflist;
1022 * Activate ep->ws since epi->ws may get
1023 * deactivated at any time.
1025 __pm_stay_awake(ep->ws);
1032 /* If this file is already in the ready list we exit soon */
1033 if (!ep_is_linked(&epi->rdllink)) {
1034 list_add_tail(&epi->rdllink, &ep->rdllist);
1035 ep_pm_stay_awake_rcu(epi);
1039 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1042 if (waitqueue_active(&ep->wq))
1043 wake_up_locked(&ep->wq);
1044 if (waitqueue_active(&ep->poll_wait))
1048 spin_unlock_irqrestore(&ep->lock, flags);
1050 /* We have to call this outside the lock */
1052 ep_poll_safewake(&ep->poll_wait);
1058 * This is the callback that is used to add our wait queue to the
1059 * target file wakeup lists.
1061 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1064 struct epitem *epi = ep_item_from_epqueue(pt);
1065 struct eppoll_entry *pwq;
1067 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1068 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1069 pwq->wait.private = get_thread_process(current);
1072 add_wait_queue(whead, &pwq->wait);
1073 list_add_tail(&pwq->llink, &epi->pwqlist);
1076 /* We have to signal that an error occurred */
1081 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1084 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1085 struct epitem *epic;
1089 epic = rb_entry(parent, struct epitem, rbn);
1090 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1092 p = &parent->rb_right;
1094 p = &parent->rb_left;
1096 rb_link_node(&epi->rbn, parent, p);
1097 rb_insert_color(&epi->rbn, &ep->rbr);
1102 #define PATH_ARR_SIZE 5
1104 * These are the number paths of length 1 to 5, that we are allowing to emanate
1105 * from a single file of interest. For example, we allow 1000 paths of length
1106 * 1, to emanate from each file of interest. This essentially represents the
1107 * potential wakeup paths, which need to be limited in order to avoid massive
1108 * uncontrolled wakeup storms. The common use case should be a single ep which
1109 * is connected to n file sources. In this case each file source has 1 path
1110 * of length 1. Thus, the numbers below should be more than sufficient. These
1111 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1112 * and delete can't add additional paths. Protected by the epmutex.
1114 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1115 static int path_count[PATH_ARR_SIZE];
1117 static int path_count_inc(int nests)
1119 /* Allow an arbitrary number of depth 1 paths */
1123 if (++path_count[nests] > path_limits[nests])
1128 static void path_count_init(void)
1132 for (i = 0; i < PATH_ARR_SIZE; i++)
1136 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1139 struct file *file = priv;
1140 struct file *child_file;
1143 list_for_each_entry(epi, &file->f_ep_links, fllink) {
1144 child_file = epi->ep->file;
1145 if (is_file_epoll(child_file)) {
1146 if (list_empty(&child_file->f_ep_links)) {
1147 if (path_count_inc(call_nests)) {
1152 error = ep_call_nested(&poll_loop_ncalls,
1154 reverse_path_check_proc,
1155 child_file, child_file,
1161 printk(KERN_ERR "reverse_path_check_proc: "
1162 "file is not an ep!\n");
1169 * reverse_path_check - The tfile_check_list is list of file *, which have
1170 * links that are proposed to be newly added. We need to
1171 * make sure that those added links don't add too many
1172 * paths such that we will spend all our time waking up
1173 * eventpoll objects.
1175 * Returns: Returns zero if the proposed links don't create too many paths,
1178 static int reverse_path_check(void)
1181 struct file *current_file;
1183 /* let's call this for all tfiles */
1184 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1186 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1187 reverse_path_check_proc, current_file,
1188 current_file, current);
1195 static int ep_create_wakeup_source(struct epitem *epi)
1198 struct wakeup_source *ws;
1201 epi->ep->ws = wakeup_source_register("eventpoll");
1206 name = epi->ffd.file->f_path.dentry->d_name.name;
1207 ws = wakeup_source_register(name);
1211 rcu_assign_pointer(epi->ws, ws);
1216 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1217 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1219 struct wakeup_source *ws = ep_wakeup_source(epi);
1221 RCU_INIT_POINTER(epi->ws, NULL);
1224 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1225 * used internally by wakeup_source_remove, too (called by
1226 * wakeup_source_unregister), so we cannot use call_rcu
1229 wakeup_source_unregister(ws);
1233 * Must be called with "mtx" held.
1235 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1236 struct file *tfile, int fd)
1238 int error, revents, pwake = 0;
1239 unsigned long flags;
1242 struct ep_pqueue epq;
1244 user_watches = atomic_long_read(&ep->user->epoll_watches);
1245 if (unlikely(user_watches >= max_user_watches))
1247 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1250 /* Item initialization follow here ... */
1251 INIT_LIST_HEAD(&epi->rdllink);
1252 INIT_LIST_HEAD(&epi->fllink);
1253 INIT_LIST_HEAD(&epi->pwqlist);
1255 ep_set_ffd(&epi->ffd, tfile, fd);
1256 epi->event = *event;
1258 epi->next = EP_UNACTIVE_PTR;
1259 if (epi->event.events & EPOLLWAKEUP) {
1260 error = ep_create_wakeup_source(epi);
1262 goto error_create_wakeup_source;
1264 RCU_INIT_POINTER(epi->ws, NULL);
1267 /* Initialize the poll table using the queue callback */
1269 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1272 * Attach the item to the poll hooks and get current event bits.
1273 * We can safely use the file* here because its usage count has
1274 * been increased by the caller of this function. Note that after
1275 * this operation completes, the poll callback can start hitting
1278 revents = ep_item_poll(epi, &epq.pt);
1281 * We have to check if something went wrong during the poll wait queue
1282 * install process. Namely an allocation for a wait queue failed due
1283 * high memory pressure.
1287 goto error_unregister;
1289 /* Add the current item to the list of active epoll hook for this file */
1290 spin_lock(&tfile->f_lock);
1291 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1292 spin_unlock(&tfile->f_lock);
1293 tfile->f_path.dentry->d_inode->i_private = get_thread_process(current);
1296 * Add the current item to the RB tree. All RB tree operations are
1297 * protected by "mtx", and ep_insert() is called with "mtx" held.
1299 ep_rbtree_insert(ep, epi);
1301 /* now check if we've created too many backpaths */
1303 if (reverse_path_check())
1304 goto error_remove_epi;
1306 /* We have to drop the new item inside our item list to keep track of it */
1307 spin_lock_irqsave(&ep->lock, flags);
1309 /* If the file is already "ready" we drop it inside the ready list */
1310 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1311 list_add_tail(&epi->rdllink, &ep->rdllist);
1312 ep_pm_stay_awake(epi);
1314 /* Notify waiting tasks that events are available */
1315 if (waitqueue_active(&ep->wq))
1316 wake_up_locked(&ep->wq);
1317 if (waitqueue_active(&ep->poll_wait))
1321 spin_unlock_irqrestore(&ep->lock, flags);
1323 atomic_long_inc(&ep->user->epoll_watches);
1325 /* We have to call this outside the lock */
1327 ep_poll_safewake(&ep->poll_wait);
1332 spin_lock(&tfile->f_lock);
1333 if (ep_is_linked(&epi->fllink))
1334 list_del_init(&epi->fllink);
1335 spin_unlock(&tfile->f_lock);
1337 rb_erase(&epi->rbn, &ep->rbr);
1340 ep_unregister_pollwait(ep, epi);
1343 * We need to do this because an event could have been arrived on some
1344 * allocated wait queue. Note that we don't care about the ep->ovflist
1345 * list, since that is used/cleaned only inside a section bound by "mtx".
1346 * And ep_insert() is called with "mtx" held.
1348 spin_lock_irqsave(&ep->lock, flags);
1349 if (ep_is_linked(&epi->rdllink))
1350 list_del_init(&epi->rdllink);
1351 spin_unlock_irqrestore(&ep->lock, flags);
1353 wakeup_source_unregister(ep_wakeup_source(epi));
1355 error_create_wakeup_source:
1356 kmem_cache_free(epi_cache, epi);
1362 * Modify the interest event mask by dropping an event if the new mask
1363 * has a match in the current file status. Must be called with "mtx" held.
1365 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1368 unsigned int revents;
1371 init_poll_funcptr(&pt, NULL);
1374 * Set the new event interest mask before calling f_op->poll();
1375 * otherwise we might miss an event that happens between the
1376 * f_op->poll() call and the new event set registering.
1378 epi->event.events = event->events; /* need barrier below */
1379 epi->event.data = event->data; /* protected by mtx */
1380 if (epi->event.events & EPOLLWAKEUP) {
1381 if (!ep_has_wakeup_source(epi))
1382 ep_create_wakeup_source(epi);
1383 } else if (ep_has_wakeup_source(epi)) {
1384 ep_destroy_wakeup_source(epi);
1388 * The following barrier has two effects:
1390 * 1) Flush epi changes above to other CPUs. This ensures
1391 * we do not miss events from ep_poll_callback if an
1392 * event occurs immediately after we call f_op->poll().
1393 * We need this because we did not take ep->lock while
1394 * changing epi above (but ep_poll_callback does take
1397 * 2) We also need to ensure we do not miss _past_ events
1398 * when calling f_op->poll(). This barrier also
1399 * pairs with the barrier in wq_has_sleeper (see
1400 * comments for wq_has_sleeper).
1402 * This barrier will now guarantee ep_poll_callback or f_op->poll
1403 * (or both) will notice the readiness of an item.
1408 * Get current event bits. We can safely use the file* here because
1409 * its usage count has been increased by the caller of this function.
1411 revents = ep_item_poll(epi, &pt);
1414 * If the item is "hot" and it is not registered inside the ready
1415 * list, push it inside.
1417 if (revents & event->events) {
1418 spin_lock_irq(&ep->lock);
1419 if (!ep_is_linked(&epi->rdllink)) {
1420 list_add_tail(&epi->rdllink, &ep->rdllist);
1421 ep_pm_stay_awake(epi);
1423 /* Notify waiting tasks that events are available */
1424 if (waitqueue_active(&ep->wq))
1425 wake_up_locked(&ep->wq);
1426 if (waitqueue_active(&ep->poll_wait))
1429 spin_unlock_irq(&ep->lock);
1432 /* We have to call this outside the lock */
1434 ep_poll_safewake(&ep->poll_wait);
1439 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1442 struct ep_send_events_data *esed = priv;
1444 unsigned int revents;
1446 struct epoll_event __user *uevent;
1447 struct wakeup_source *ws;
1450 init_poll_funcptr(&pt, NULL);
1453 * We can loop without lock because we are passed a task private list.
1454 * Items cannot vanish during the loop because ep_scan_ready_list() is
1455 * holding "mtx" during this call.
1457 for (eventcnt = 0, uevent = esed->events;
1458 !list_empty(head) && eventcnt < esed->maxevents;) {
1459 epi = list_first_entry(head, struct epitem, rdllink);
1462 * Activate ep->ws before deactivating epi->ws to prevent
1463 * triggering auto-suspend here (in case we reactive epi->ws
1466 * This could be rearranged to delay the deactivation of epi->ws
1467 * instead, but then epi->ws would temporarily be out of sync
1468 * with ep_is_linked().
1470 ws = ep_wakeup_source(epi);
1473 __pm_stay_awake(ep->ws);
1477 list_del_init(&epi->rdllink);
1479 revents = ep_item_poll(epi, &pt);
1482 * If the event mask intersect the caller-requested one,
1483 * deliver the event to userspace. Again, ep_scan_ready_list()
1484 * is holding "mtx", so no operations coming from userspace
1485 * can change the item.
1488 if (__put_user(revents, &uevent->events) ||
1489 __put_user(epi->event.data, &uevent->data)) {
1490 list_add(&epi->rdllink, head);
1491 ep_pm_stay_awake(epi);
1492 return eventcnt ? eventcnt : -EFAULT;
1496 if (epi->event.events & EPOLLONESHOT)
1497 epi->event.events &= EP_PRIVATE_BITS;
1498 else if (!(epi->event.events & EPOLLET)) {
1500 * If this file has been added with Level
1501 * Trigger mode, we need to insert back inside
1502 * the ready list, so that the next call to
1503 * epoll_wait() will check again the events
1504 * availability. At this point, no one can insert
1505 * into ep->rdllist besides us. The epoll_ctl()
1506 * callers are locked out by
1507 * ep_scan_ready_list() holding "mtx" and the
1508 * poll callback will queue them in ep->ovflist.
1510 list_add_tail(&epi->rdllink, &ep->rdllist);
1511 ep_pm_stay_awake(epi);
1519 static int ep_send_events(struct eventpoll *ep,
1520 struct epoll_event __user *events, int maxevents)
1522 struct ep_send_events_data esed;
1524 esed.maxevents = maxevents;
1525 esed.events = events;
1527 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
1530 static inline struct timespec ep_set_mstimeout(long ms)
1532 struct timespec now, ts = {
1533 .tv_sec = ms / MSEC_PER_SEC,
1534 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1538 return timespec_add_safe(now, ts);
1542 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1545 * @ep: Pointer to the eventpoll context.
1546 * @events: Pointer to the userspace buffer where the ready events should be
1548 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1549 * @timeout: Maximum timeout for the ready events fetch operation, in
1550 * milliseconds. If the @timeout is zero, the function will not block,
1551 * while if the @timeout is less than zero, the function will block
1552 * until at least one event has been retrieved (or an error
1555 * Returns: Returns the number of ready events which have been fetched, or an
1556 * error code, in case of error.
1558 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1559 int maxevents, long timeout)
1561 int res = 0, eavail, timed_out = 0;
1562 unsigned long flags;
1565 ktime_t expires, *to = NULL;
1568 struct timespec end_time = ep_set_mstimeout(timeout);
1570 slack = select_estimate_accuracy(&end_time);
1572 *to = timespec_to_ktime(end_time);
1573 } else if (timeout == 0) {
1575 * Avoid the unnecessary trip to the wait queue loop, if the
1576 * caller specified a non blocking operation.
1579 spin_lock_irqsave(&ep->lock, flags);
1584 spin_lock_irqsave(&ep->lock, flags);
1586 if (!ep_events_available(ep)) {
1588 * We don't have any available event to return to the caller.
1589 * We need to sleep here, and we will be wake up by
1590 * ep_poll_callback() when events will become available.
1592 init_waitqueue_entry(&wait, current);
1593 __add_wait_queue_exclusive(&ep->wq, &wait);
1597 * We don't want to sleep if the ep_poll_callback() sends us
1598 * a wakeup in between. That's why we set the task state
1599 * to TASK_INTERRUPTIBLE before doing the checks.
1601 set_current_state(TASK_INTERRUPTIBLE);
1602 if (ep_events_available(ep) || timed_out)
1604 if (signal_pending(current)) {
1609 spin_unlock_irqrestore(&ep->lock, flags);
1610 if (!freezable_schedule_hrtimeout_range(to, slack,
1614 spin_lock_irqsave(&ep->lock, flags);
1616 __remove_wait_queue(&ep->wq, &wait);
1618 set_current_state(TASK_RUNNING);
1621 /* Is it worth to try to dig for events ? */
1622 eavail = ep_events_available(ep);
1624 spin_unlock_irqrestore(&ep->lock, flags);
1627 * Try to transfer events to user space. In case we get 0 events and
1628 * there's still timeout left over, we go trying again in search of
1631 if (!res && eavail &&
1632 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1639 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1640 * API, to verify that adding an epoll file inside another
1641 * epoll structure, does not violate the constraints, in
1642 * terms of closed loops, or too deep chains (which can
1643 * result in excessive stack usage).
1645 * @priv: Pointer to the epoll file to be currently checked.
1646 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1647 * data structure pointer.
1648 * @call_nests: Current dept of the @ep_call_nested() call stack.
1650 * Returns: Returns zero if adding the epoll @file inside current epoll
1651 * structure @ep does not violate the constraints, or -1 otherwise.
1653 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1656 struct file *file = priv;
1657 struct eventpoll *ep = file->private_data;
1658 struct eventpoll *ep_tovisit;
1659 struct rb_node *rbp;
1662 mutex_lock_nested(&ep->mtx, call_nests + 1);
1664 list_add(&ep->visited_list_link, &visited_list);
1665 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1666 epi = rb_entry(rbp, struct epitem, rbn);
1667 if (unlikely(is_file_epoll(epi->ffd.file))) {
1668 ep_tovisit = epi->ffd.file->private_data;
1669 if (ep_tovisit->visited)
1671 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1672 ep_loop_check_proc, epi->ffd.file,
1673 ep_tovisit, current);
1678 * If we've reached a file that is not associated with
1679 * an ep, then we need to check if the newly added
1680 * links are going to add too many wakeup paths. We do
1681 * this by adding it to the tfile_check_list, if it's
1682 * not already there, and calling reverse_path_check()
1683 * during ep_insert().
1685 if (list_empty(&epi->ffd.file->f_tfile_llink))
1686 list_add(&epi->ffd.file->f_tfile_llink,
1690 mutex_unlock(&ep->mtx);
1696 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1697 * another epoll file (represented by @ep) does not create
1698 * closed loops or too deep chains.
1700 * @ep: Pointer to the epoll private data structure.
1701 * @file: Pointer to the epoll file to be checked.
1703 * Returns: Returns zero if adding the epoll @file inside current epoll
1704 * structure @ep does not violate the constraints, or -1 otherwise.
1706 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1709 struct eventpoll *ep_cur, *ep_next;
1711 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1712 ep_loop_check_proc, file, ep, current);
1713 /* clear visited list */
1714 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1715 visited_list_link) {
1716 ep_cur->visited = 0;
1717 list_del(&ep_cur->visited_list_link);
1722 static void clear_tfile_check_list(void)
1726 /* first clear the tfile_check_list */
1727 while (!list_empty(&tfile_check_list)) {
1728 file = list_first_entry(&tfile_check_list, struct file,
1730 list_del_init(&file->f_tfile_llink);
1732 INIT_LIST_HEAD(&tfile_check_list);
1735 struct task_struct *get_epoll_file_task(struct file *file)
1737 struct list_head *lh;
1738 struct epitem *epi = NULL;
1739 struct eppoll_entry *pwq = NULL;
1740 struct task_struct *task = NULL;
1741 wait_queue_t *wq = NULL;
1743 lh = &file->f_ep_links;
1744 if (!list_empty(lh)) {
1746 epi = list_entry(lh, struct epitem, fllink);
1748 if (!list_empty(lh)) {
1750 pwq = list_entry(lh, struct eppoll_entry, llink);
1751 lh = &pwq->whead->task_list;
1752 if (!list_empty(lh)) {
1754 wq = list_entry(lh, wait_queue_t, task_list);
1764 * Open an eventpoll file descriptor.
1766 SYSCALL_DEFINE1(epoll_create1, int, flags)
1769 struct eventpoll *ep = NULL;
1772 /* Check the EPOLL_* constant for consistency. */
1773 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1775 if (flags & ~EPOLL_CLOEXEC)
1778 * Create the internal data structure ("struct eventpoll").
1780 error = ep_alloc(&ep);
1784 * Creates all the items needed to setup an eventpoll file. That is,
1785 * a file structure and a free file descriptor.
1787 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1792 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1793 O_RDWR | (flags & O_CLOEXEC));
1795 error = PTR_ERR(file);
1799 fd_install(fd, file);
1809 SYSCALL_DEFINE1(epoll_create, int, size)
1814 return sys_epoll_create1(0);
1818 * The following function implements the controller interface for
1819 * the eventpoll file that enables the insertion/removal/change of
1820 * file descriptors inside the interest set.
1822 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1823 struct epoll_event __user *, event)
1826 int did_lock_epmutex = 0;
1827 struct file *file, *tfile;
1828 struct eventpoll *ep;
1830 struct epoll_event epds;
1833 if (ep_op_has_event(op) &&
1834 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1837 /* Get the "struct file *" for the eventpoll file */
1843 /* Get the "struct file *" for the target file */
1848 /* The target file descriptor must support poll */
1850 if (!tfile->f_op || !tfile->f_op->poll)
1851 goto error_tgt_fput;
1853 /* Check if EPOLLWAKEUP is allowed */
1854 if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
1855 epds.events &= ~EPOLLWAKEUP;
1858 * We have to check that the file structure underneath the file descriptor
1859 * the user passed to us _is_ an eventpoll file. And also we do not permit
1860 * adding an epoll file descriptor inside itself.
1863 if (file == tfile || !is_file_epoll(file))
1864 goto error_tgt_fput;
1867 * At this point it is safe to assume that the "private_data" contains
1868 * our own data structure.
1870 ep = file->private_data;
1873 * When we insert an epoll file descriptor, inside another epoll file
1874 * descriptor, there is the change of creating closed loops, which are
1875 * better be handled here, than in more critical paths. While we are
1876 * checking for loops we also determine the list of files reachable
1877 * and hang them on the tfile_check_list, so we can check that we
1878 * haven't created too many possible wakeup paths.
1880 * We need to hold the epmutex across both ep_insert and ep_remove
1881 * b/c we want to make sure we are looking at a coherent view of
1884 if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
1885 mutex_lock(&epmutex);
1886 did_lock_epmutex = 1;
1888 if (op == EPOLL_CTL_ADD) {
1889 if (is_file_epoll(tfile)) {
1891 if (ep_loop_check(ep, tfile) != 0) {
1892 clear_tfile_check_list();
1893 goto error_tgt_fput;
1896 list_add(&tfile->f_tfile_llink, &tfile_check_list);
1899 mutex_lock_nested(&ep->mtx, 0);
1902 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1903 * above, we can be sure to be able to use the item looked up by
1904 * ep_find() till we release the mutex.
1906 epi = ep_find(ep, tfile, fd);
1912 epds.events |= POLLERR | POLLHUP;
1913 error = ep_insert(ep, &epds, tfile, fd);
1916 clear_tfile_check_list();
1920 error = ep_remove(ep, epi);
1926 epds.events |= POLLERR | POLLHUP;
1927 error = ep_modify(ep, epi, &epds);
1932 mutex_unlock(&ep->mtx);
1935 if (did_lock_epmutex)
1936 mutex_unlock(&epmutex);
1947 * Implement the event wait interface for the eventpoll file. It is the kernel
1948 * part of the user space epoll_wait(2).
1950 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1951 int, maxevents, int, timeout)
1955 struct eventpoll *ep;
1957 /* The maximum number of event must be greater than zero */
1958 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1961 /* Verify that the area passed by the user is writeable */
1962 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1965 /* Get the "struct file *" for the eventpoll file */
1971 * We have to check that the file structure underneath the fd
1972 * the user passed to us _is_ an eventpoll file.
1975 if (!is_file_epoll(f.file))
1979 * At this point it is safe to assume that the "private_data" contains
1980 * our own data structure.
1982 ep = f.file->private_data;
1984 /* Time to fish for events ... */
1985 error = ep_poll(ep, events, maxevents, timeout);
1993 * Implement the event wait interface for the eventpoll file. It is the kernel
1994 * part of the user space epoll_pwait(2).
1996 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1997 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2001 sigset_t ksigmask, sigsaved;
2004 * If the caller wants a certain signal mask to be set during the wait,
2008 if (sigsetsize != sizeof(sigset_t))
2010 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2012 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2013 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2016 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2019 * If we changed the signal mask, we need to restore the original one.
2020 * In case we've got a signal while waiting, we do not restore the
2021 * signal mask yet, and we allow do_signal() to deliver the signal on
2022 * the way back to userspace, before the signal mask is restored.
2025 if (error == -EINTR) {
2026 memcpy(¤t->saved_sigmask, &sigsaved,
2028 set_restore_sigmask();
2030 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
2036 #ifdef CONFIG_COMPAT
2037 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2038 struct epoll_event __user *, events,
2039 int, maxevents, int, timeout,
2040 const compat_sigset_t __user *, sigmask,
2041 compat_size_t, sigsetsize)
2044 compat_sigset_t csigmask;
2045 sigset_t ksigmask, sigsaved;
2048 * If the caller wants a certain signal mask to be set during the wait,
2052 if (sigsetsize != sizeof(compat_sigset_t))
2054 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2056 sigset_from_compat(&ksigmask, &csigmask);
2057 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2058 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2061 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2064 * If we changed the signal mask, we need to restore the original one.
2065 * In case we've got a signal while waiting, we do not restore the
2066 * signal mask yet, and we allow do_signal() to deliver the signal on
2067 * the way back to userspace, before the signal mask is restored.
2070 if (err == -EINTR) {
2071 memcpy(¤t->saved_sigmask, &sigsaved,
2073 set_restore_sigmask();
2075 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
2082 static int __init eventpoll_init(void)
2088 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2090 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2092 BUG_ON(max_user_watches < 0);
2095 * Initialize the structure used to perform epoll file descriptor
2096 * inclusion loops checks.
2098 ep_nested_calls_init(&poll_loop_ncalls);
2100 /* Initialize the structure used to perform safe poll wait head wake ups */
2101 ep_nested_calls_init(&poll_safewake_ncalls);
2103 /* Initialize the structure used to perform file's f_op->poll() calls */
2104 ep_nested_calls_init(&poll_readywalk_ncalls);
2107 * We can have many thousands of epitems, so prevent this from
2108 * using an extra cache line on 64-bit (and smaller) CPUs
2110 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2112 /* Allocates slab cache used to allocate "struct epitem" items */
2113 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2114 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2116 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2117 pwq_cache = kmem_cache_create("eventpoll_pwq",
2118 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2122 fs_initcall(eventpoll_init);