2 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * Davide Libenzi <davidel@xmailserver.org>
14 #include <linux/init.h>
15 #include <linux/kernel.h>
16 #include <linux/sched.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/device.h>
37 #include <asm/uaccess.h>
40 #include <linux/atomic.h>
41 #include <linux/proc_fs.h>
42 #include <linux/seq_file.h>
43 #include <linux/compat.h>
44 #include <linux/rculist.h>
48 * There are three level of locking required by epoll :
52 * 3) ep->lock (spinlock)
54 * The acquire order is the one listed above, from 1 to 3.
55 * We need a spinlock (ep->lock) because we manipulate objects
56 * from inside the poll callback, that might be triggered from
57 * a wake_up() that in turn might be called from IRQ context.
58 * So we can't sleep inside the poll callback and hence we need
59 * a spinlock. During the event transfer loop (from kernel to
60 * user space) we could end up sleeping due a copy_to_user(), so
61 * we need a lock that will allow us to sleep. This lock is a
62 * mutex (ep->mtx). It is acquired during the event transfer loop,
63 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
64 * Then we also need a global mutex to serialize eventpoll_release_file()
66 * This mutex is acquired by ep_free() during the epoll file
67 * cleanup path and it is also acquired by eventpoll_release_file()
68 * if a file has been pushed inside an epoll set and it is then
69 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
70 * It is also acquired when inserting an epoll fd onto another epoll
71 * fd. We do this so that we walk the epoll tree and ensure that this
72 * insertion does not create a cycle of epoll file descriptors, which
73 * could lead to deadlock. We need a global mutex to prevent two
74 * simultaneous inserts (A into B and B into A) from racing and
75 * constructing a cycle without either insert observing that it is
77 * It is necessary to acquire multiple "ep->mtx"es at once in the
78 * case when one epoll fd is added to another. In this case, we
79 * always acquire the locks in the order of nesting (i.e. after
80 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
81 * before e2->mtx). Since we disallow cycles of epoll file
82 * descriptors, this ensures that the mutexes are well-ordered. In
83 * order to communicate this nesting to lockdep, when walking a tree
84 * of epoll file descriptors, we use the current recursion depth as
86 * It is possible to drop the "ep->mtx" and to use the global
87 * mutex "epmutex" (together with "ep->lock") to have it working,
88 * but having "ep->mtx" will make the interface more scalable.
89 * Events that require holding "epmutex" are very rare, while for
90 * normal operations the epoll private "ep->mtx" will guarantee
91 * a better scalability.
94 /* Epoll private bits inside the event mask */
95 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
97 /* Maximum number of nesting allowed inside epoll sets */
98 #define EP_MAX_NESTS 4
100 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
102 #define EP_UNACTIVE_PTR ((void *) -1L)
104 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
106 struct epoll_filefd {
112 * Structure used to track possible nested calls, for too deep recursions
115 struct nested_call_node {
116 struct list_head llink;
122 * This structure is used as collector for nested calls, to check for
123 * maximum recursion dept and loop cycles.
125 struct nested_calls {
126 struct list_head tasks_call_list;
131 * Each file descriptor added to the eventpoll interface will
132 * have an entry of this type linked to the "rbr" RB tree.
133 * Avoid increasing the size of this struct, there can be many thousands
134 * of these on a server and we do not want this to take another cache line.
138 /* RB tree node links this structure to the eventpoll RB tree */
140 /* Used to free the struct epitem */
144 /* List header used to link this structure to the eventpoll ready list */
145 struct list_head rdllink;
148 * Works together "struct eventpoll"->ovflist in keeping the
149 * single linked chain of items.
153 /* The file descriptor information this item refers to */
154 struct epoll_filefd ffd;
156 /* Number of active wait queue attached to poll operations */
159 /* List containing poll wait queues */
160 struct list_head pwqlist;
162 /* The "container" of this item */
163 struct eventpoll *ep;
165 /* List header used to link this item to the "struct file" items list */
166 struct list_head fllink;
168 /* wakeup_source used when EPOLLWAKEUP is set */
169 struct wakeup_source __rcu *ws;
171 /* The structure that describe the interested events and the source fd */
172 struct epoll_event event;
176 * This structure is stored inside the "private_data" member of the file
177 * structure and represents the main data structure for the eventpoll
181 /* Protect the access to this structure */
185 * This mutex is used to ensure that files are not removed
186 * while epoll is using them. This is held during the event
187 * collection loop, the file cleanup path, the epoll file exit
188 * code and the ctl operations.
192 /* Wait queue used by sys_epoll_wait() */
193 wait_queue_head_t wq;
195 /* Wait queue used by file->poll() */
196 wait_queue_head_t poll_wait;
198 /* List of ready file descriptors */
199 struct list_head rdllist;
201 /* RB tree root used to store monitored fd structs */
205 * This is a single linked list that chains all the "struct epitem" that
206 * happened while transferring ready events to userspace w/out
209 struct epitem *ovflist;
211 /* wakeup_source used when ep_scan_ready_list is running */
212 struct wakeup_source *ws;
214 /* The user that created the eventpoll descriptor */
215 struct user_struct *user;
219 /* used to optimize loop detection check */
221 struct list_head visited_list_link;
224 /* Wait structure used by the poll hooks */
225 struct eppoll_entry {
226 /* List header used to link this structure to the "struct epitem" */
227 struct list_head llink;
229 /* The "base" pointer is set to the container "struct epitem" */
233 * Wait queue item that will be linked to the target file wait
238 /* The wait queue head that linked the "wait" wait queue item */
239 wait_queue_head_t *whead;
242 /* Wrapper struct used by poll queueing */
248 /* Used by the ep_send_events() function as callback private data */
249 struct ep_send_events_data {
251 struct epoll_event __user *events;
255 * Configuration options available inside /proc/sys/fs/epoll/
257 /* Maximum number of epoll watched descriptors, per user */
258 static long max_user_watches __read_mostly;
261 * This mutex is used to serialize ep_free() and eventpoll_release_file().
263 static DEFINE_MUTEX(epmutex);
265 /* Used to check for epoll file descriptor inclusion loops */
266 static struct nested_calls poll_loop_ncalls;
268 /* Used for safe wake up implementation */
269 static struct nested_calls poll_safewake_ncalls;
271 /* Used to call file's f_op->poll() under the nested calls boundaries */
272 static struct nested_calls poll_readywalk_ncalls;
274 /* Slab cache used to allocate "struct epitem" */
275 static struct kmem_cache *epi_cache __read_mostly;
277 /* Slab cache used to allocate "struct eppoll_entry" */
278 static struct kmem_cache *pwq_cache __read_mostly;
280 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
281 static LIST_HEAD(visited_list);
284 * List of files with newly added links, where we may need to limit the number
285 * of emanating paths. Protected by the epmutex.
287 static LIST_HEAD(tfile_check_list);
291 #include <linux/sysctl.h>
294 static long long_max = LONG_MAX;
296 struct ctl_table epoll_table[] = {
298 .procname = "max_user_watches",
299 .data = &max_user_watches,
300 .maxlen = sizeof(max_user_watches),
302 .proc_handler = proc_doulongvec_minmax,
308 #endif /* CONFIG_SYSCTL */
310 static const struct file_operations eventpoll_fops;
312 static inline int is_file_epoll(struct file *f)
314 return f->f_op == &eventpoll_fops;
317 /* Setup the structure that is used as key for the RB tree */
318 static inline void ep_set_ffd(struct epoll_filefd *ffd,
319 struct file *file, int fd)
325 /* Compare RB tree keys */
326 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
327 struct epoll_filefd *p2)
329 return (p1->file > p2->file ? +1:
330 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
333 /* Tells us if the item is currently linked */
334 static inline int ep_is_linked(struct list_head *p)
336 return !list_empty(p);
339 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
341 return container_of(p, struct eppoll_entry, wait);
344 /* Get the "struct epitem" from a wait queue pointer */
345 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
347 return container_of(p, struct eppoll_entry, wait)->base;
350 /* Get the "struct epitem" from an epoll queue wrapper */
351 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
353 return container_of(p, struct ep_pqueue, pt)->epi;
356 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
357 static inline int ep_op_has_event(int op)
359 return op != EPOLL_CTL_DEL;
362 /* Initialize the poll safe wake up structure */
363 static void ep_nested_calls_init(struct nested_calls *ncalls)
365 INIT_LIST_HEAD(&ncalls->tasks_call_list);
366 spin_lock_init(&ncalls->lock);
370 * ep_events_available - Checks if ready events might be available.
372 * @ep: Pointer to the eventpoll context.
374 * Returns: Returns a value different than zero if ready events are available,
377 static inline int ep_events_available(struct eventpoll *ep)
379 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
383 * ep_call_nested - Perform a bound (possibly) nested call, by checking
384 * that the recursion limit is not exceeded, and that
385 * the same nested call (by the meaning of same cookie) is
388 * @ncalls: Pointer to the nested_calls structure to be used for this call.
389 * @max_nests: Maximum number of allowed nesting calls.
390 * @nproc: Nested call core function pointer.
391 * @priv: Opaque data to be passed to the @nproc callback.
392 * @cookie: Cookie to be used to identify this nested call.
393 * @ctx: This instance context.
395 * Returns: Returns the code returned by the @nproc callback, or -1 if
396 * the maximum recursion limit has been exceeded.
398 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
399 int (*nproc)(void *, void *, int), void *priv,
400 void *cookie, void *ctx)
402 int error, call_nests = 0;
404 struct list_head *lsthead = &ncalls->tasks_call_list;
405 struct nested_call_node *tncur;
406 struct nested_call_node tnode;
408 spin_lock_irqsave(&ncalls->lock, flags);
411 * Try to see if the current task is already inside this wakeup call.
412 * We use a list here, since the population inside this set is always
415 list_for_each_entry(tncur, lsthead, llink) {
416 if (tncur->ctx == ctx &&
417 (tncur->cookie == cookie || ++call_nests > max_nests)) {
419 * Ops ... loop detected or maximum nest level reached.
420 * We abort this wake by breaking the cycle itself.
427 /* Add the current task and cookie to the list */
429 tnode.cookie = cookie;
430 list_add(&tnode.llink, lsthead);
432 spin_unlock_irqrestore(&ncalls->lock, flags);
434 /* Call the nested function */
435 error = (*nproc)(priv, cookie, call_nests);
437 /* Remove the current task from the list */
438 spin_lock_irqsave(&ncalls->lock, flags);
439 list_del(&tnode.llink);
441 spin_unlock_irqrestore(&ncalls->lock, flags);
447 * As described in commit 0ccf831cb lockdep: annotate epoll
448 * the use of wait queues used by epoll is done in a very controlled
449 * manner. Wake ups can nest inside each other, but are never done
450 * with the same locking. For example:
453 * efd1 = epoll_create();
454 * efd2 = epoll_create();
455 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
456 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
458 * When a packet arrives to the device underneath "dfd", the net code will
459 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
460 * callback wakeup entry on that queue, and the wake_up() performed by the
461 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
462 * (efd1) notices that it may have some event ready, so it needs to wake up
463 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
464 * that ends up in another wake_up(), after having checked about the
465 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
466 * avoid stack blasting.
468 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
469 * this special case of epoll.
471 #ifdef CONFIG_DEBUG_LOCK_ALLOC
472 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
473 unsigned long events, int subclass)
477 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
478 wake_up_locked_poll(wqueue, events);
479 spin_unlock_irqrestore(&wqueue->lock, flags);
482 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
483 unsigned long events, int subclass)
485 wake_up_poll(wqueue, events);
489 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
491 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
497 * Perform a safe wake up of the poll wait list. The problem is that
498 * with the new callback'd wake up system, it is possible that the
499 * poll callback is reentered from inside the call to wake_up() done
500 * on the poll wait queue head. The rule is that we cannot reenter the
501 * wake up code from the same task more than EP_MAX_NESTS times,
502 * and we cannot reenter the same wait queue head at all. This will
503 * enable to have a hierarchy of epoll file descriptor of no more than
506 static void ep_poll_safewake(wait_queue_head_t *wq)
508 int this_cpu = get_cpu_light();
510 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
511 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
516 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
518 wait_queue_head_t *whead;
521 /* If it is cleared by POLLFREE, it should be rcu-safe */
522 whead = rcu_dereference(pwq->whead);
524 remove_wait_queue(whead, &pwq->wait);
529 * This function unregisters poll callbacks from the associated file
530 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
533 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
535 struct list_head *lsthead = &epi->pwqlist;
536 struct eppoll_entry *pwq;
538 while (!list_empty(lsthead)) {
539 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
541 list_del(&pwq->llink);
542 ep_remove_wait_queue(pwq);
543 kmem_cache_free(pwq_cache, pwq);
547 /* call only when ep->mtx is held */
548 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
550 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
553 /* call only when ep->mtx is held */
554 static inline void ep_pm_stay_awake(struct epitem *epi)
556 struct wakeup_source *ws = ep_wakeup_source(epi);
562 static inline bool ep_has_wakeup_source(struct epitem *epi)
564 return rcu_access_pointer(epi->ws) ? true : false;
567 /* call when ep->mtx cannot be held (ep_poll_callback) */
568 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
570 struct wakeup_source *ws;
573 ws = rcu_dereference(epi->ws);
580 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
581 * the scan code, to call f_op->poll(). Also allows for
582 * O(NumReady) performance.
584 * @ep: Pointer to the epoll private data structure.
585 * @sproc: Pointer to the scan callback.
586 * @priv: Private opaque data passed to the @sproc callback.
587 * @depth: The current depth of recursive f_op->poll calls.
588 * @ep_locked: caller already holds ep->mtx
590 * Returns: The same integer error code returned by the @sproc callback.
592 static int ep_scan_ready_list(struct eventpoll *ep,
593 int (*sproc)(struct eventpoll *,
594 struct list_head *, void *),
595 void *priv, int depth, bool ep_locked)
597 int error, pwake = 0;
599 struct epitem *epi, *nepi;
603 * We need to lock this because we could be hit by
604 * eventpoll_release_file() and epoll_ctl().
608 mutex_lock_nested(&ep->mtx, depth);
611 * Steal the ready list, and re-init the original one to the
612 * empty list. Also, set ep->ovflist to NULL so that events
613 * happening while looping w/out locks, are not lost. We cannot
614 * have the poll callback to queue directly on ep->rdllist,
615 * because we want the "sproc" callback to be able to do it
618 spin_lock_irqsave(&ep->lock, flags);
619 list_splice_init(&ep->rdllist, &txlist);
621 spin_unlock_irqrestore(&ep->lock, flags);
624 * Now call the callback function.
626 error = (*sproc)(ep, &txlist, priv);
628 spin_lock_irqsave(&ep->lock, flags);
630 * During the time we spent inside the "sproc" callback, some
631 * other events might have been queued by the poll callback.
632 * We re-insert them inside the main ready-list here.
634 for (nepi = ep->ovflist; (epi = nepi) != NULL;
635 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
637 * We need to check if the item is already in the list.
638 * During the "sproc" callback execution time, items are
639 * queued into ->ovflist but the "txlist" might already
640 * contain them, and the list_splice() below takes care of them.
642 if (!ep_is_linked(&epi->rdllink)) {
643 list_add_tail(&epi->rdllink, &ep->rdllist);
644 ep_pm_stay_awake(epi);
648 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
649 * releasing the lock, events will be queued in the normal way inside
652 ep->ovflist = EP_UNACTIVE_PTR;
655 * Quickly re-inject items left on "txlist".
657 list_splice(&txlist, &ep->rdllist);
660 if (!list_empty(&ep->rdllist)) {
662 * Wake up (if active) both the eventpoll wait list and
663 * the ->poll() wait list (delayed after we release the lock).
665 if (waitqueue_active(&ep->wq))
666 wake_up_locked(&ep->wq);
667 if (waitqueue_active(&ep->poll_wait))
670 spin_unlock_irqrestore(&ep->lock, flags);
673 mutex_unlock(&ep->mtx);
675 /* We have to call this outside the lock */
677 ep_poll_safewake(&ep->poll_wait);
682 static void epi_rcu_free(struct rcu_head *head)
684 struct epitem *epi = container_of(head, struct epitem, rcu);
685 kmem_cache_free(epi_cache, epi);
689 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
690 * all the associated resources. Must be called with "mtx" held.
692 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
695 struct file *file = epi->ffd.file;
698 * Removes poll wait queue hooks. We _have_ to do this without holding
699 * the "ep->lock" otherwise a deadlock might occur. This because of the
700 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
701 * queue head lock when unregistering the wait queue. The wakeup callback
702 * will run by holding the wait queue head lock and will call our callback
703 * that will try to get "ep->lock".
705 ep_unregister_pollwait(ep, epi);
707 /* Remove the current item from the list of epoll hooks */
708 spin_lock(&file->f_lock);
709 list_del_rcu(&epi->fllink);
710 spin_unlock(&file->f_lock);
712 rb_erase(&epi->rbn, &ep->rbr);
714 spin_lock_irqsave(&ep->lock, flags);
715 if (ep_is_linked(&epi->rdllink))
716 list_del_init(&epi->rdllink);
717 spin_unlock_irqrestore(&ep->lock, flags);
719 wakeup_source_unregister(ep_wakeup_source(epi));
721 * At this point it is safe to free the eventpoll item. Use the union
722 * field epi->rcu, since we are trying to minimize the size of
723 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
724 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
725 * use of the rbn field.
727 call_rcu(&epi->rcu, epi_rcu_free);
729 atomic_long_dec(&ep->user->epoll_watches);
734 static void ep_free(struct eventpoll *ep)
739 /* We need to release all tasks waiting for these file */
740 if (waitqueue_active(&ep->poll_wait))
741 ep_poll_safewake(&ep->poll_wait);
744 * We need to lock this because we could be hit by
745 * eventpoll_release_file() while we're freeing the "struct eventpoll".
746 * We do not need to hold "ep->mtx" here because the epoll file
747 * is on the way to be removed and no one has references to it
748 * anymore. The only hit might come from eventpoll_release_file() but
749 * holding "epmutex" is sufficient here.
751 mutex_lock(&epmutex);
754 * Walks through the whole tree by unregistering poll callbacks.
756 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
757 epi = rb_entry(rbp, struct epitem, rbn);
759 ep_unregister_pollwait(ep, epi);
764 * Walks through the whole tree by freeing each "struct epitem". At this
765 * point we are sure no poll callbacks will be lingering around, and also by
766 * holding "epmutex" we can be sure that no file cleanup code will hit
767 * us during this operation. So we can avoid the lock on "ep->lock".
768 * We do not need to lock ep->mtx, either, we only do it to prevent
771 mutex_lock(&ep->mtx);
772 while ((rbp = rb_first(&ep->rbr)) != NULL) {
773 epi = rb_entry(rbp, struct epitem, rbn);
777 mutex_unlock(&ep->mtx);
779 mutex_unlock(&epmutex);
780 mutex_destroy(&ep->mtx);
782 wakeup_source_unregister(ep->ws);
786 static int ep_eventpoll_release(struct inode *inode, struct file *file)
788 struct eventpoll *ep = file->private_data;
796 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
798 pt->_key = epi->event.events;
800 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
803 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
806 struct epitem *epi, *tmp;
809 init_poll_funcptr(&pt, NULL);
811 list_for_each_entry_safe(epi, tmp, head, rdllink) {
812 if (ep_item_poll(epi, &pt))
813 return POLLIN | POLLRDNORM;
816 * Item has been dropped into the ready list by the poll
817 * callback, but it's not actually ready, as far as
818 * caller requested events goes. We can remove it here.
820 __pm_relax(ep_wakeup_source(epi));
821 list_del_init(&epi->rdllink);
828 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
831 struct readyevents_arg {
832 struct eventpoll *ep;
836 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
838 struct readyevents_arg *arg = priv;
840 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
841 call_nests + 1, arg->locked);
844 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
847 struct eventpoll *ep = file->private_data;
848 struct readyevents_arg arg;
851 * During ep_insert() we already hold the ep->mtx for the tfile.
852 * Prevent re-aquisition.
854 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
857 /* Insert inside our poll wait queue */
858 poll_wait(file, &ep->poll_wait, wait);
861 * Proceed to find out if wanted events are really available inside
862 * the ready list. This need to be done under ep_call_nested()
863 * supervision, since the call to f_op->poll() done on listed files
864 * could re-enter here.
866 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
867 ep_poll_readyevents_proc, &arg, ep, current);
869 return pollflags != -1 ? pollflags : 0;
872 #ifdef CONFIG_PROC_FS
873 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
875 struct eventpoll *ep = f->private_data;
878 mutex_lock(&ep->mtx);
879 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
880 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
882 seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
883 epi->ffd.fd, epi->event.events,
884 (long long)epi->event.data);
885 if (seq_has_overflowed(m))
888 mutex_unlock(&ep->mtx);
892 /* File callbacks that implement the eventpoll file behaviour */
893 static const struct file_operations eventpoll_fops = {
894 #ifdef CONFIG_PROC_FS
895 .show_fdinfo = ep_show_fdinfo,
897 .release = ep_eventpoll_release,
898 .poll = ep_eventpoll_poll,
899 .llseek = noop_llseek,
903 * This is called from eventpoll_release() to unlink files from the eventpoll
904 * interface. We need to have this facility to cleanup correctly files that are
905 * closed without being removed from the eventpoll interface.
907 void eventpoll_release_file(struct file *file)
909 struct eventpoll *ep;
910 struct epitem *epi, *next;
913 * We don't want to get "file->f_lock" because it is not
914 * necessary. It is not necessary because we're in the "struct file"
915 * cleanup path, and this means that no one is using this file anymore.
916 * So, for example, epoll_ctl() cannot hit here since if we reach this
917 * point, the file counter already went to zero and fget() would fail.
918 * The only hit might come from ep_free() but by holding the mutex
919 * will correctly serialize the operation. We do need to acquire
920 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
921 * from anywhere but ep_free().
923 * Besides, ep_remove() acquires the lock, so we can't hold it here.
925 mutex_lock(&epmutex);
926 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
928 mutex_lock_nested(&ep->mtx, 0);
930 mutex_unlock(&ep->mtx);
932 mutex_unlock(&epmutex);
935 static int ep_alloc(struct eventpoll **pep)
938 struct user_struct *user;
939 struct eventpoll *ep;
941 user = get_current_user();
943 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
947 spin_lock_init(&ep->lock);
948 mutex_init(&ep->mtx);
949 init_waitqueue_head(&ep->wq);
950 init_waitqueue_head(&ep->poll_wait);
951 INIT_LIST_HEAD(&ep->rdllist);
953 ep->ovflist = EP_UNACTIVE_PTR;
966 * Search the file inside the eventpoll tree. The RB tree operations
967 * are protected by the "mtx" mutex, and ep_find() must be called with
970 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
974 struct epitem *epi, *epir = NULL;
975 struct epoll_filefd ffd;
977 ep_set_ffd(&ffd, file, fd);
978 for (rbp = ep->rbr.rb_node; rbp; ) {
979 epi = rb_entry(rbp, struct epitem, rbn);
980 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
995 * This is the callback that is passed to the wait queue wakeup
996 * mechanism. It is called by the stored file descriptors when they
997 * have events to report.
999 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1002 unsigned long flags;
1003 struct epitem *epi = ep_item_from_wait(wait);
1004 struct eventpoll *ep = epi->ep;
1006 if ((unsigned long)key & POLLFREE) {
1007 ep_pwq_from_wait(wait)->whead = NULL;
1009 * whead = NULL above can race with ep_remove_wait_queue()
1010 * which can do another remove_wait_queue() after us, so we
1011 * can't use __remove_wait_queue(). whead->lock is held by
1014 list_del_init(&wait->task_list);
1017 spin_lock_irqsave(&ep->lock, flags);
1020 * If the event mask does not contain any poll(2) event, we consider the
1021 * descriptor to be disabled. This condition is likely the effect of the
1022 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1023 * until the next EPOLL_CTL_MOD will be issued.
1025 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1029 * Check the events coming with the callback. At this stage, not
1030 * every device reports the events in the "key" parameter of the
1031 * callback. We need to be able to handle both cases here, hence the
1032 * test for "key" != NULL before the event match test.
1034 if (key && !((unsigned long) key & epi->event.events))
1038 * If we are transferring events to userspace, we can hold no locks
1039 * (because we're accessing user memory, and because of linux f_op->poll()
1040 * semantics). All the events that happen during that period of time are
1041 * chained in ep->ovflist and requeued later on.
1043 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1044 if (epi->next == EP_UNACTIVE_PTR) {
1045 epi->next = ep->ovflist;
1049 * Activate ep->ws since epi->ws may get
1050 * deactivated at any time.
1052 __pm_stay_awake(ep->ws);
1059 /* If this file is already in the ready list we exit soon */
1060 if (!ep_is_linked(&epi->rdllink)) {
1061 list_add_tail(&epi->rdllink, &ep->rdllist);
1062 ep_pm_stay_awake_rcu(epi);
1066 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1069 if (waitqueue_active(&ep->wq))
1070 wake_up_locked(&ep->wq);
1071 if (waitqueue_active(&ep->poll_wait))
1075 spin_unlock_irqrestore(&ep->lock, flags);
1077 /* We have to call this outside the lock */
1079 ep_poll_safewake(&ep->poll_wait);
1085 * This is the callback that is used to add our wait queue to the
1086 * target file wakeup lists.
1088 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1091 struct epitem *epi = ep_item_from_epqueue(pt);
1092 struct eppoll_entry *pwq;
1094 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1095 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1098 add_wait_queue(whead, &pwq->wait);
1099 list_add_tail(&pwq->llink, &epi->pwqlist);
1102 /* We have to signal that an error occurred */
1107 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1110 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1111 struct epitem *epic;
1115 epic = rb_entry(parent, struct epitem, rbn);
1116 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1118 p = &parent->rb_right;
1120 p = &parent->rb_left;
1122 rb_link_node(&epi->rbn, parent, p);
1123 rb_insert_color(&epi->rbn, &ep->rbr);
1128 #define PATH_ARR_SIZE 5
1130 * These are the number paths of length 1 to 5, that we are allowing to emanate
1131 * from a single file of interest. For example, we allow 1000 paths of length
1132 * 1, to emanate from each file of interest. This essentially represents the
1133 * potential wakeup paths, which need to be limited in order to avoid massive
1134 * uncontrolled wakeup storms. The common use case should be a single ep which
1135 * is connected to n file sources. In this case each file source has 1 path
1136 * of length 1. Thus, the numbers below should be more than sufficient. These
1137 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1138 * and delete can't add additional paths. Protected by the epmutex.
1140 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1141 static int path_count[PATH_ARR_SIZE];
1143 static int path_count_inc(int nests)
1145 /* Allow an arbitrary number of depth 1 paths */
1149 if (++path_count[nests] > path_limits[nests])
1154 static void path_count_init(void)
1158 for (i = 0; i < PATH_ARR_SIZE; i++)
1162 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1165 struct file *file = priv;
1166 struct file *child_file;
1169 /* CTL_DEL can remove links here, but that can't increase our count */
1171 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1172 child_file = epi->ep->file;
1173 if (is_file_epoll(child_file)) {
1174 if (list_empty(&child_file->f_ep_links)) {
1175 if (path_count_inc(call_nests)) {
1180 error = ep_call_nested(&poll_loop_ncalls,
1182 reverse_path_check_proc,
1183 child_file, child_file,
1189 printk(KERN_ERR "reverse_path_check_proc: "
1190 "file is not an ep!\n");
1198 * reverse_path_check - The tfile_check_list is list of file *, which have
1199 * links that are proposed to be newly added. We need to
1200 * make sure that those added links don't add too many
1201 * paths such that we will spend all our time waking up
1202 * eventpoll objects.
1204 * Returns: Returns zero if the proposed links don't create too many paths,
1207 static int reverse_path_check(void)
1210 struct file *current_file;
1212 /* let's call this for all tfiles */
1213 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1215 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1216 reverse_path_check_proc, current_file,
1217 current_file, current);
1224 static int ep_create_wakeup_source(struct epitem *epi)
1227 struct wakeup_source *ws;
1230 epi->ep->ws = wakeup_source_register("eventpoll");
1235 name = epi->ffd.file->f_path.dentry->d_name.name;
1236 ws = wakeup_source_register(name);
1240 rcu_assign_pointer(epi->ws, ws);
1245 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1246 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1248 struct wakeup_source *ws = ep_wakeup_source(epi);
1250 RCU_INIT_POINTER(epi->ws, NULL);
1253 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1254 * used internally by wakeup_source_remove, too (called by
1255 * wakeup_source_unregister), so we cannot use call_rcu
1258 wakeup_source_unregister(ws);
1262 * Must be called with "mtx" held.
1264 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1265 struct file *tfile, int fd, int full_check)
1267 int error, revents, pwake = 0;
1268 unsigned long flags;
1271 struct ep_pqueue epq;
1273 user_watches = atomic_long_read(&ep->user->epoll_watches);
1274 if (unlikely(user_watches >= max_user_watches))
1276 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1279 /* Item initialization follow here ... */
1280 INIT_LIST_HEAD(&epi->rdllink);
1281 INIT_LIST_HEAD(&epi->fllink);
1282 INIT_LIST_HEAD(&epi->pwqlist);
1284 ep_set_ffd(&epi->ffd, tfile, fd);
1285 epi->event = *event;
1287 epi->next = EP_UNACTIVE_PTR;
1288 if (epi->event.events & EPOLLWAKEUP) {
1289 error = ep_create_wakeup_source(epi);
1291 goto error_create_wakeup_source;
1293 RCU_INIT_POINTER(epi->ws, NULL);
1296 /* Initialize the poll table using the queue callback */
1298 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1301 * Attach the item to the poll hooks and get current event bits.
1302 * We can safely use the file* here because its usage count has
1303 * been increased by the caller of this function. Note that after
1304 * this operation completes, the poll callback can start hitting
1307 revents = ep_item_poll(epi, &epq.pt);
1310 * We have to check if something went wrong during the poll wait queue
1311 * install process. Namely an allocation for a wait queue failed due
1312 * high memory pressure.
1316 goto error_unregister;
1318 /* Add the current item to the list of active epoll hook for this file */
1319 spin_lock(&tfile->f_lock);
1320 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1321 spin_unlock(&tfile->f_lock);
1324 * Add the current item to the RB tree. All RB tree operations are
1325 * protected by "mtx", and ep_insert() is called with "mtx" held.
1327 ep_rbtree_insert(ep, epi);
1329 /* now check if we've created too many backpaths */
1331 if (full_check && reverse_path_check())
1332 goto error_remove_epi;
1334 /* We have to drop the new item inside our item list to keep track of it */
1335 spin_lock_irqsave(&ep->lock, flags);
1337 /* If the file is already "ready" we drop it inside the ready list */
1338 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1339 list_add_tail(&epi->rdllink, &ep->rdllist);
1340 ep_pm_stay_awake(epi);
1342 /* Notify waiting tasks that events are available */
1343 if (waitqueue_active(&ep->wq))
1344 wake_up_locked(&ep->wq);
1345 if (waitqueue_active(&ep->poll_wait))
1349 spin_unlock_irqrestore(&ep->lock, flags);
1351 atomic_long_inc(&ep->user->epoll_watches);
1353 /* We have to call this outside the lock */
1355 ep_poll_safewake(&ep->poll_wait);
1360 spin_lock(&tfile->f_lock);
1361 list_del_rcu(&epi->fllink);
1362 spin_unlock(&tfile->f_lock);
1364 rb_erase(&epi->rbn, &ep->rbr);
1367 ep_unregister_pollwait(ep, epi);
1370 * We need to do this because an event could have been arrived on some
1371 * allocated wait queue. Note that we don't care about the ep->ovflist
1372 * list, since that is used/cleaned only inside a section bound by "mtx".
1373 * And ep_insert() is called with "mtx" held.
1375 spin_lock_irqsave(&ep->lock, flags);
1376 if (ep_is_linked(&epi->rdllink))
1377 list_del_init(&epi->rdllink);
1378 spin_unlock_irqrestore(&ep->lock, flags);
1380 wakeup_source_unregister(ep_wakeup_source(epi));
1382 error_create_wakeup_source:
1383 kmem_cache_free(epi_cache, epi);
1389 * Modify the interest event mask by dropping an event if the new mask
1390 * has a match in the current file status. Must be called with "mtx" held.
1392 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1395 unsigned int revents;
1398 init_poll_funcptr(&pt, NULL);
1401 * Set the new event interest mask before calling f_op->poll();
1402 * otherwise we might miss an event that happens between the
1403 * f_op->poll() call and the new event set registering.
1405 epi->event.events = event->events; /* need barrier below */
1406 epi->event.data = event->data; /* protected by mtx */
1407 if (epi->event.events & EPOLLWAKEUP) {
1408 if (!ep_has_wakeup_source(epi))
1409 ep_create_wakeup_source(epi);
1410 } else if (ep_has_wakeup_source(epi)) {
1411 ep_destroy_wakeup_source(epi);
1415 * The following barrier has two effects:
1417 * 1) Flush epi changes above to other CPUs. This ensures
1418 * we do not miss events from ep_poll_callback if an
1419 * event occurs immediately after we call f_op->poll().
1420 * We need this because we did not take ep->lock while
1421 * changing epi above (but ep_poll_callback does take
1424 * 2) We also need to ensure we do not miss _past_ events
1425 * when calling f_op->poll(). This barrier also
1426 * pairs with the barrier in wq_has_sleeper (see
1427 * comments for wq_has_sleeper).
1429 * This barrier will now guarantee ep_poll_callback or f_op->poll
1430 * (or both) will notice the readiness of an item.
1435 * Get current event bits. We can safely use the file* here because
1436 * its usage count has been increased by the caller of this function.
1438 revents = ep_item_poll(epi, &pt);
1441 * If the item is "hot" and it is not registered inside the ready
1442 * list, push it inside.
1444 if (revents & event->events) {
1445 spin_lock_irq(&ep->lock);
1446 if (!ep_is_linked(&epi->rdllink)) {
1447 list_add_tail(&epi->rdllink, &ep->rdllist);
1448 ep_pm_stay_awake(epi);
1450 /* Notify waiting tasks that events are available */
1451 if (waitqueue_active(&ep->wq))
1452 wake_up_locked(&ep->wq);
1453 if (waitqueue_active(&ep->poll_wait))
1456 spin_unlock_irq(&ep->lock);
1459 /* We have to call this outside the lock */
1461 ep_poll_safewake(&ep->poll_wait);
1466 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1469 struct ep_send_events_data *esed = priv;
1471 unsigned int revents;
1473 struct epoll_event __user *uevent;
1474 struct wakeup_source *ws;
1477 init_poll_funcptr(&pt, NULL);
1480 * We can loop without lock because we are passed a task private list.
1481 * Items cannot vanish during the loop because ep_scan_ready_list() is
1482 * holding "mtx" during this call.
1484 for (eventcnt = 0, uevent = esed->events;
1485 !list_empty(head) && eventcnt < esed->maxevents;) {
1486 epi = list_first_entry(head, struct epitem, rdllink);
1489 * Activate ep->ws before deactivating epi->ws to prevent
1490 * triggering auto-suspend here (in case we reactive epi->ws
1493 * This could be rearranged to delay the deactivation of epi->ws
1494 * instead, but then epi->ws would temporarily be out of sync
1495 * with ep_is_linked().
1497 ws = ep_wakeup_source(epi);
1500 __pm_stay_awake(ep->ws);
1504 list_del_init(&epi->rdllink);
1506 revents = ep_item_poll(epi, &pt);
1509 * If the event mask intersect the caller-requested one,
1510 * deliver the event to userspace. Again, ep_scan_ready_list()
1511 * is holding "mtx", so no operations coming from userspace
1512 * can change the item.
1515 if (__put_user(revents, &uevent->events) ||
1516 __put_user(epi->event.data, &uevent->data)) {
1517 list_add(&epi->rdllink, head);
1518 ep_pm_stay_awake(epi);
1519 return eventcnt ? eventcnt : -EFAULT;
1523 if (epi->event.events & EPOLLONESHOT)
1524 epi->event.events &= EP_PRIVATE_BITS;
1525 else if (!(epi->event.events & EPOLLET)) {
1527 * If this file has been added with Level
1528 * Trigger mode, we need to insert back inside
1529 * the ready list, so that the next call to
1530 * epoll_wait() will check again the events
1531 * availability. At this point, no one can insert
1532 * into ep->rdllist besides us. The epoll_ctl()
1533 * callers are locked out by
1534 * ep_scan_ready_list() holding "mtx" and the
1535 * poll callback will queue them in ep->ovflist.
1537 list_add_tail(&epi->rdllink, &ep->rdllist);
1538 ep_pm_stay_awake(epi);
1546 static int ep_send_events(struct eventpoll *ep,
1547 struct epoll_event __user *events, int maxevents)
1549 struct ep_send_events_data esed;
1551 esed.maxevents = maxevents;
1552 esed.events = events;
1554 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1557 static inline struct timespec ep_set_mstimeout(long ms)
1559 struct timespec now, ts = {
1560 .tv_sec = ms / MSEC_PER_SEC,
1561 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1565 return timespec_add_safe(now, ts);
1569 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1572 * @ep: Pointer to the eventpoll context.
1573 * @events: Pointer to the userspace buffer where the ready events should be
1575 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1576 * @timeout: Maximum timeout for the ready events fetch operation, in
1577 * milliseconds. If the @timeout is zero, the function will not block,
1578 * while if the @timeout is less than zero, the function will block
1579 * until at least one event has been retrieved (or an error
1582 * Returns: Returns the number of ready events which have been fetched, or an
1583 * error code, in case of error.
1585 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1586 int maxevents, long timeout)
1588 int res = 0, eavail, timed_out = 0;
1589 unsigned long flags;
1592 ktime_t expires, *to = NULL;
1595 struct timespec end_time = ep_set_mstimeout(timeout);
1597 slack = select_estimate_accuracy(&end_time);
1599 *to = timespec_to_ktime(end_time);
1600 } else if (timeout == 0) {
1602 * Avoid the unnecessary trip to the wait queue loop, if the
1603 * caller specified a non blocking operation.
1606 spin_lock_irqsave(&ep->lock, flags);
1611 spin_lock_irqsave(&ep->lock, flags);
1613 if (!ep_events_available(ep)) {
1615 * We don't have any available event to return to the caller.
1616 * We need to sleep here, and we will be wake up by
1617 * ep_poll_callback() when events will become available.
1619 init_waitqueue_entry(&wait, current);
1620 __add_wait_queue_exclusive(&ep->wq, &wait);
1624 * We don't want to sleep if the ep_poll_callback() sends us
1625 * a wakeup in between. That's why we set the task state
1626 * to TASK_INTERRUPTIBLE before doing the checks.
1628 set_current_state(TASK_INTERRUPTIBLE);
1629 if (ep_events_available(ep) || timed_out)
1631 if (signal_pending(current)) {
1636 spin_unlock_irqrestore(&ep->lock, flags);
1637 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1640 spin_lock_irqsave(&ep->lock, flags);
1643 __remove_wait_queue(&ep->wq, &wait);
1644 __set_current_state(TASK_RUNNING);
1647 /* Is it worth to try to dig for events ? */
1648 eavail = ep_events_available(ep);
1650 spin_unlock_irqrestore(&ep->lock, flags);
1653 * Try to transfer events to user space. In case we get 0 events and
1654 * there's still timeout left over, we go trying again in search of
1657 if (!res && eavail &&
1658 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1665 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1666 * API, to verify that adding an epoll file inside another
1667 * epoll structure, does not violate the constraints, in
1668 * terms of closed loops, or too deep chains (which can
1669 * result in excessive stack usage).
1671 * @priv: Pointer to the epoll file to be currently checked.
1672 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1673 * data structure pointer.
1674 * @call_nests: Current dept of the @ep_call_nested() call stack.
1676 * Returns: Returns zero if adding the epoll @file inside current epoll
1677 * structure @ep does not violate the constraints, or -1 otherwise.
1679 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1682 struct file *file = priv;
1683 struct eventpoll *ep = file->private_data;
1684 struct eventpoll *ep_tovisit;
1685 struct rb_node *rbp;
1688 mutex_lock_nested(&ep->mtx, call_nests + 1);
1690 list_add(&ep->visited_list_link, &visited_list);
1691 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1692 epi = rb_entry(rbp, struct epitem, rbn);
1693 if (unlikely(is_file_epoll(epi->ffd.file))) {
1694 ep_tovisit = epi->ffd.file->private_data;
1695 if (ep_tovisit->visited)
1697 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1698 ep_loop_check_proc, epi->ffd.file,
1699 ep_tovisit, current);
1704 * If we've reached a file that is not associated with
1705 * an ep, then we need to check if the newly added
1706 * links are going to add too many wakeup paths. We do
1707 * this by adding it to the tfile_check_list, if it's
1708 * not already there, and calling reverse_path_check()
1709 * during ep_insert().
1711 if (list_empty(&epi->ffd.file->f_tfile_llink))
1712 list_add(&epi->ffd.file->f_tfile_llink,
1716 mutex_unlock(&ep->mtx);
1722 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1723 * another epoll file (represented by @ep) does not create
1724 * closed loops or too deep chains.
1726 * @ep: Pointer to the epoll private data structure.
1727 * @file: Pointer to the epoll file to be checked.
1729 * Returns: Returns zero if adding the epoll @file inside current epoll
1730 * structure @ep does not violate the constraints, or -1 otherwise.
1732 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1735 struct eventpoll *ep_cur, *ep_next;
1737 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1738 ep_loop_check_proc, file, ep, current);
1739 /* clear visited list */
1740 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1741 visited_list_link) {
1742 ep_cur->visited = 0;
1743 list_del(&ep_cur->visited_list_link);
1748 static void clear_tfile_check_list(void)
1752 /* first clear the tfile_check_list */
1753 while (!list_empty(&tfile_check_list)) {
1754 file = list_first_entry(&tfile_check_list, struct file,
1756 list_del_init(&file->f_tfile_llink);
1758 INIT_LIST_HEAD(&tfile_check_list);
1762 * Open an eventpoll file descriptor.
1764 SYSCALL_DEFINE1(epoll_create1, int, flags)
1767 struct eventpoll *ep = NULL;
1770 /* Check the EPOLL_* constant for consistency. */
1771 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1773 if (flags & ~EPOLL_CLOEXEC)
1776 * Create the internal data structure ("struct eventpoll").
1778 error = ep_alloc(&ep);
1782 * Creates all the items needed to setup an eventpoll file. That is,
1783 * a file structure and a free file descriptor.
1785 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1790 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1791 O_RDWR | (flags & O_CLOEXEC));
1793 error = PTR_ERR(file);
1797 fd_install(fd, file);
1807 SYSCALL_DEFINE1(epoll_create, int, size)
1812 return sys_epoll_create1(0);
1816 * The following function implements the controller interface for
1817 * the eventpoll file that enables the insertion/removal/change of
1818 * file descriptors inside the interest set.
1820 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1821 struct epoll_event __user *, event)
1826 struct eventpoll *ep;
1828 struct epoll_event epds;
1829 struct eventpoll *tep = NULL;
1832 if (ep_op_has_event(op) &&
1833 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1841 /* Get the "struct file *" for the target file */
1846 /* The target file descriptor must support poll */
1848 if (!tf.file->f_op->poll)
1849 goto error_tgt_fput;
1851 /* Check if EPOLLWAKEUP is allowed */
1852 if (ep_op_has_event(op))
1853 ep_take_care_of_epollwakeup(&epds);
1856 * We have to check that the file structure underneath the file descriptor
1857 * the user passed to us _is_ an eventpoll file. And also we do not permit
1858 * adding an epoll file descriptor inside itself.
1861 if (f.file == tf.file || !is_file_epoll(f.file))
1862 goto error_tgt_fput;
1865 * At this point it is safe to assume that the "private_data" contains
1866 * our own data structure.
1868 ep = f.file->private_data;
1871 * When we insert an epoll file descriptor, inside another epoll file
1872 * descriptor, there is the change of creating closed loops, which are
1873 * better be handled here, than in more critical paths. While we are
1874 * checking for loops we also determine the list of files reachable
1875 * and hang them on the tfile_check_list, so we can check that we
1876 * haven't created too many possible wakeup paths.
1878 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1879 * the epoll file descriptor is attaching directly to a wakeup source,
1880 * unless the epoll file descriptor is nested. The purpose of taking the
1881 * 'epmutex' on add is to prevent complex toplogies such as loops and
1882 * deep wakeup paths from forming in parallel through multiple
1883 * EPOLL_CTL_ADD operations.
1885 mutex_lock_nested(&ep->mtx, 0);
1886 if (op == EPOLL_CTL_ADD) {
1887 if (!list_empty(&f.file->f_ep_links) ||
1888 is_file_epoll(tf.file)) {
1890 mutex_unlock(&ep->mtx);
1891 mutex_lock(&epmutex);
1892 if (is_file_epoll(tf.file)) {
1894 if (ep_loop_check(ep, tf.file) != 0) {
1895 clear_tfile_check_list();
1896 goto error_tgt_fput;
1899 list_add(&tf.file->f_tfile_llink,
1901 mutex_lock_nested(&ep->mtx, 0);
1902 if (is_file_epoll(tf.file)) {
1903 tep = tf.file->private_data;
1904 mutex_lock_nested(&tep->mtx, 1);
1910 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1911 * above, we can be sure to be able to use the item looked up by
1912 * ep_find() till we release the mutex.
1914 epi = ep_find(ep, tf.file, fd);
1920 epds.events |= POLLERR | POLLHUP;
1921 error = ep_insert(ep, &epds, tf.file, fd, full_check);
1925 clear_tfile_check_list();
1929 error = ep_remove(ep, epi);
1935 epds.events |= POLLERR | POLLHUP;
1936 error = ep_modify(ep, epi, &epds);
1942 mutex_unlock(&tep->mtx);
1943 mutex_unlock(&ep->mtx);
1947 mutex_unlock(&epmutex);
1958 * Implement the event wait interface for the eventpoll file. It is the kernel
1959 * part of the user space epoll_wait(2).
1961 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1962 int, maxevents, int, timeout)
1966 struct eventpoll *ep;
1968 /* The maximum number of event must be greater than zero */
1969 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1972 /* Verify that the area passed by the user is writeable */
1973 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1976 /* Get the "struct file *" for the eventpoll file */
1982 * We have to check that the file structure underneath the fd
1983 * the user passed to us _is_ an eventpoll file.
1986 if (!is_file_epoll(f.file))
1990 * At this point it is safe to assume that the "private_data" contains
1991 * our own data structure.
1993 ep = f.file->private_data;
1995 /* Time to fish for events ... */
1996 error = ep_poll(ep, events, maxevents, timeout);
2004 * Implement the event wait interface for the eventpoll file. It is the kernel
2005 * part of the user space epoll_pwait(2).
2007 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2008 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2012 sigset_t ksigmask, sigsaved;
2015 * If the caller wants a certain signal mask to be set during the wait,
2019 if (sigsetsize != sizeof(sigset_t))
2021 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2023 sigsaved = current->blocked;
2024 set_current_blocked(&ksigmask);
2027 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2030 * If we changed the signal mask, we need to restore the original one.
2031 * In case we've got a signal while waiting, we do not restore the
2032 * signal mask yet, and we allow do_signal() to deliver the signal on
2033 * the way back to userspace, before the signal mask is restored.
2036 if (error == -EINTR) {
2037 memcpy(¤t->saved_sigmask, &sigsaved,
2039 set_restore_sigmask();
2041 set_current_blocked(&sigsaved);
2047 #ifdef CONFIG_COMPAT
2048 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2049 struct epoll_event __user *, events,
2050 int, maxevents, int, timeout,
2051 const compat_sigset_t __user *, sigmask,
2052 compat_size_t, sigsetsize)
2055 compat_sigset_t csigmask;
2056 sigset_t ksigmask, sigsaved;
2059 * If the caller wants a certain signal mask to be set during the wait,
2063 if (sigsetsize != sizeof(compat_sigset_t))
2065 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2067 sigset_from_compat(&ksigmask, &csigmask);
2068 sigsaved = current->blocked;
2069 set_current_blocked(&ksigmask);
2072 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2075 * If we changed the signal mask, we need to restore the original one.
2076 * In case we've got a signal while waiting, we do not restore the
2077 * signal mask yet, and we allow do_signal() to deliver the signal on
2078 * the way back to userspace, before the signal mask is restored.
2081 if (err == -EINTR) {
2082 memcpy(¤t->saved_sigmask, &sigsaved,
2084 set_restore_sigmask();
2086 set_current_blocked(&sigsaved);
2093 static int __init eventpoll_init(void)
2099 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2101 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2103 BUG_ON(max_user_watches < 0);
2106 * Initialize the structure used to perform epoll file descriptor
2107 * inclusion loops checks.
2109 ep_nested_calls_init(&poll_loop_ncalls);
2111 /* Initialize the structure used to perform safe poll wait head wake ups */
2112 ep_nested_calls_init(&poll_safewake_ncalls);
2114 /* Initialize the structure used to perform file's f_op->poll() calls */
2115 ep_nested_calls_init(&poll_readywalk_ncalls);
2118 * We can have many thousands of epitems, so prevent this from
2119 * using an extra cache line on 64-bit (and smaller) CPUs
2121 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2123 /* Allocates slab cache used to allocate "struct epitem" items */
2124 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2125 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2127 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2128 pwq_cache = kmem_cache_create("eventpoll_pwq",
2129 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2133 fs_initcall(eventpoll_init);