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/signal.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/device.h>
37 #include <linux/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>
45 #include <net/busy_poll.h>
49 * There are three level of locking required by epoll :
53 * 3) ep->lock (spinlock)
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
63 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
67 * This mutex is acquired by ep_free() during the epoll file
68 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
78 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
87 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
90 * Events that require holding "epmutex" are very rare, while for
91 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
95 /* Epoll private bits inside the event mask */
96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
98 #define EPOLLINOUT_BITS (POLLIN | POLLOUT)
100 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | POLLERR | POLLHUP | \
101 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
103 /* Maximum number of nesting allowed inside epoll sets */
104 #define EP_MAX_NESTS 4
106 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
108 #define EP_UNACTIVE_PTR ((void *) -1L)
110 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
112 struct epoll_filefd {
118 * Structure used to track possible nested calls, for too deep recursions
121 struct nested_call_node {
122 struct list_head llink;
128 * This structure is used as collector for nested calls, to check for
129 * maximum recursion dept and loop cycles.
131 struct nested_calls {
132 struct list_head tasks_call_list;
137 * Each file descriptor added to the eventpoll interface will
138 * have an entry of this type linked to the "rbr" RB tree.
139 * Avoid increasing the size of this struct, there can be many thousands
140 * of these on a server and we do not want this to take another cache line.
144 /* RB tree node links this structure to the eventpoll RB tree */
146 /* Used to free the struct epitem */
150 /* List header used to link this structure to the eventpoll ready list */
151 struct list_head rdllink;
154 * Works together "struct eventpoll"->ovflist in keeping the
155 * single linked chain of items.
159 /* The file descriptor information this item refers to */
160 struct epoll_filefd ffd;
162 /* Number of active wait queue attached to poll operations */
165 /* List containing poll wait queues */
166 struct list_head pwqlist;
168 /* The "container" of this item */
169 struct eventpoll *ep;
171 /* List header used to link this item to the "struct file" items list */
172 struct list_head fllink;
174 /* wakeup_source used when EPOLLWAKEUP is set */
175 struct wakeup_source __rcu *ws;
177 /* The structure that describe the interested events and the source fd */
178 struct epoll_event event;
182 * This structure is stored inside the "private_data" member of the file
183 * structure and represents the main data structure for the eventpoll
187 /* Protect the access to this structure */
191 * This mutex is used to ensure that files are not removed
192 * while epoll is using them. This is held during the event
193 * collection loop, the file cleanup path, the epoll file exit
194 * code and the ctl operations.
198 /* Wait queue used by sys_epoll_wait() */
199 wait_queue_head_t wq;
201 /* Wait queue used by file->poll() */
202 wait_queue_head_t poll_wait;
204 /* List of ready file descriptors */
205 struct list_head rdllist;
207 /* RB tree root used to store monitored fd structs */
211 * This is a single linked list that chains all the "struct epitem" that
212 * happened while transferring ready events to userspace w/out
215 struct epitem *ovflist;
217 /* wakeup_source used when ep_scan_ready_list is running */
218 struct wakeup_source *ws;
220 /* The user that created the eventpoll descriptor */
221 struct user_struct *user;
225 /* used to optimize loop detection check */
227 struct list_head visited_list_link;
229 #ifdef CONFIG_NET_RX_BUSY_POLL
230 /* used to track busy poll napi_id */
231 unsigned int napi_id;
235 /* Wait structure used by the poll hooks */
236 struct eppoll_entry {
237 /* List header used to link this structure to the "struct epitem" */
238 struct list_head llink;
240 /* The "base" pointer is set to the container "struct epitem" */
244 * Wait queue item that will be linked to the target file wait
247 wait_queue_entry_t wait;
249 /* The wait queue head that linked the "wait" wait queue item */
250 wait_queue_head_t *whead;
253 /* Wrapper struct used by poll queueing */
259 /* Used by the ep_send_events() function as callback private data */
260 struct ep_send_events_data {
262 struct epoll_event __user *events;
266 * Configuration options available inside /proc/sys/fs/epoll/
268 /* Maximum number of epoll watched descriptors, per user */
269 static long max_user_watches __read_mostly;
272 * This mutex is used to serialize ep_free() and eventpoll_release_file().
274 static DEFINE_MUTEX(epmutex);
276 /* Used to check for epoll file descriptor inclusion loops */
277 static struct nested_calls poll_loop_ncalls;
279 /* Used for safe wake up implementation */
280 static struct nested_calls poll_safewake_ncalls;
282 /* Used to call file's f_op->poll() under the nested calls boundaries */
283 static struct nested_calls poll_readywalk_ncalls;
285 /* Slab cache used to allocate "struct epitem" */
286 static struct kmem_cache *epi_cache __read_mostly;
288 /* Slab cache used to allocate "struct eppoll_entry" */
289 static struct kmem_cache *pwq_cache __read_mostly;
291 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
292 static LIST_HEAD(visited_list);
295 * List of files with newly added links, where we may need to limit the number
296 * of emanating paths. Protected by the epmutex.
298 static LIST_HEAD(tfile_check_list);
302 #include <linux/sysctl.h>
305 static long long_max = LONG_MAX;
307 struct ctl_table epoll_table[] = {
309 .procname = "max_user_watches",
310 .data = &max_user_watches,
311 .maxlen = sizeof(max_user_watches),
313 .proc_handler = proc_doulongvec_minmax,
319 #endif /* CONFIG_SYSCTL */
321 static const struct file_operations eventpoll_fops;
323 static inline int is_file_epoll(struct file *f)
325 return f->f_op == &eventpoll_fops;
328 /* Setup the structure that is used as key for the RB tree */
329 static inline void ep_set_ffd(struct epoll_filefd *ffd,
330 struct file *file, int fd)
336 /* Compare RB tree keys */
337 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
338 struct epoll_filefd *p2)
340 return (p1->file > p2->file ? +1:
341 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
344 /* Tells us if the item is currently linked */
345 static inline int ep_is_linked(struct list_head *p)
347 return !list_empty(p);
350 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
352 return container_of(p, struct eppoll_entry, wait);
355 /* Get the "struct epitem" from a wait queue pointer */
356 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
358 return container_of(p, struct eppoll_entry, wait)->base;
361 /* Get the "struct epitem" from an epoll queue wrapper */
362 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
364 return container_of(p, struct ep_pqueue, pt)->epi;
367 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
368 static inline int ep_op_has_event(int op)
370 return op != EPOLL_CTL_DEL;
373 /* Initialize the poll safe wake up structure */
374 static void ep_nested_calls_init(struct nested_calls *ncalls)
376 INIT_LIST_HEAD(&ncalls->tasks_call_list);
377 spin_lock_init(&ncalls->lock);
381 * ep_events_available - Checks if ready events might be available.
383 * @ep: Pointer to the eventpoll context.
385 * Returns: Returns a value different than zero if ready events are available,
388 static inline int ep_events_available(struct eventpoll *ep)
390 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
393 #ifdef CONFIG_NET_RX_BUSY_POLL
394 static bool ep_busy_loop_end(void *p, unsigned long start_time)
396 struct eventpoll *ep = p;
398 return ep_events_available(ep) || busy_loop_timeout(start_time);
400 #endif /* CONFIG_NET_RX_BUSY_POLL */
403 * Busy poll if globally on and supporting sockets found && no events,
404 * busy loop will return if need_resched or ep_events_available.
406 * we must do our busy polling with irqs enabled
408 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
410 #ifdef CONFIG_NET_RX_BUSY_POLL
411 unsigned int napi_id = READ_ONCE(ep->napi_id);
413 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
414 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
418 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
420 #ifdef CONFIG_NET_RX_BUSY_POLL
427 * Set epoll busy poll NAPI ID from sk.
429 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
431 #ifdef CONFIG_NET_RX_BUSY_POLL
432 struct eventpoll *ep;
433 unsigned int napi_id;
438 if (!net_busy_loop_on())
441 sock = sock_from_file(epi->ffd.file, &err);
449 napi_id = READ_ONCE(sk->sk_napi_id);
452 /* Non-NAPI IDs can be rejected
454 * Nothing to do if we already have this ID
456 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
459 /* record NAPI ID for use in next busy poll */
460 ep->napi_id = napi_id;
465 * ep_call_nested - Perform a bound (possibly) nested call, by checking
466 * that the recursion limit is not exceeded, and that
467 * the same nested call (by the meaning of same cookie) is
470 * @ncalls: Pointer to the nested_calls structure to be used for this call.
471 * @max_nests: Maximum number of allowed nesting calls.
472 * @nproc: Nested call core function pointer.
473 * @priv: Opaque data to be passed to the @nproc callback.
474 * @cookie: Cookie to be used to identify this nested call.
475 * @ctx: This instance context.
477 * Returns: Returns the code returned by the @nproc callback, or -1 if
478 * the maximum recursion limit has been exceeded.
480 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
481 int (*nproc)(void *, void *, int), void *priv,
482 void *cookie, void *ctx)
484 int error, call_nests = 0;
486 struct list_head *lsthead = &ncalls->tasks_call_list;
487 struct nested_call_node *tncur;
488 struct nested_call_node tnode;
490 spin_lock_irqsave(&ncalls->lock, flags);
493 * Try to see if the current task is already inside this wakeup call.
494 * We use a list here, since the population inside this set is always
497 list_for_each_entry(tncur, lsthead, llink) {
498 if (tncur->ctx == ctx &&
499 (tncur->cookie == cookie || ++call_nests > max_nests)) {
501 * Ops ... loop detected or maximum nest level reached.
502 * We abort this wake by breaking the cycle itself.
509 /* Add the current task and cookie to the list */
511 tnode.cookie = cookie;
512 list_add(&tnode.llink, lsthead);
514 spin_unlock_irqrestore(&ncalls->lock, flags);
516 /* Call the nested function */
517 error = (*nproc)(priv, cookie, call_nests);
519 /* Remove the current task from the list */
520 spin_lock_irqsave(&ncalls->lock, flags);
521 list_del(&tnode.llink);
523 spin_unlock_irqrestore(&ncalls->lock, flags);
529 * As described in commit 0ccf831cb lockdep: annotate epoll
530 * the use of wait queues used by epoll is done in a very controlled
531 * manner. Wake ups can nest inside each other, but are never done
532 * with the same locking. For example:
535 * efd1 = epoll_create();
536 * efd2 = epoll_create();
537 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
538 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
540 * When a packet arrives to the device underneath "dfd", the net code will
541 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
542 * callback wakeup entry on that queue, and the wake_up() performed by the
543 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
544 * (efd1) notices that it may have some event ready, so it needs to wake up
545 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
546 * that ends up in another wake_up(), after having checked about the
547 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
548 * avoid stack blasting.
550 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
551 * this special case of epoll.
553 #ifdef CONFIG_DEBUG_LOCK_ALLOC
554 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
555 unsigned long events, int subclass)
559 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
560 wake_up_locked_poll(wqueue, events);
561 spin_unlock_irqrestore(&wqueue->lock, flags);
564 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
565 unsigned long events, int subclass)
567 wake_up_poll(wqueue, events);
571 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
573 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
579 * Perform a safe wake up of the poll wait list. The problem is that
580 * with the new callback'd wake up system, it is possible that the
581 * poll callback is reentered from inside the call to wake_up() done
582 * on the poll wait queue head. The rule is that we cannot reenter the
583 * wake up code from the same task more than EP_MAX_NESTS times,
584 * and we cannot reenter the same wait queue head at all. This will
585 * enable to have a hierarchy of epoll file descriptor of no more than
588 static void ep_poll_safewake(wait_queue_head_t *wq)
590 int this_cpu = get_cpu();
592 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
593 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
598 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
600 wait_queue_head_t *whead;
603 /* If it is cleared by POLLFREE, it should be rcu-safe */
604 whead = rcu_dereference(pwq->whead);
606 remove_wait_queue(whead, &pwq->wait);
611 * This function unregisters poll callbacks from the associated file
612 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
615 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
617 struct list_head *lsthead = &epi->pwqlist;
618 struct eppoll_entry *pwq;
620 while (!list_empty(lsthead)) {
621 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
623 list_del(&pwq->llink);
624 ep_remove_wait_queue(pwq);
625 kmem_cache_free(pwq_cache, pwq);
629 /* call only when ep->mtx is held */
630 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
632 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
635 /* call only when ep->mtx is held */
636 static inline void ep_pm_stay_awake(struct epitem *epi)
638 struct wakeup_source *ws = ep_wakeup_source(epi);
644 static inline bool ep_has_wakeup_source(struct epitem *epi)
646 return rcu_access_pointer(epi->ws) ? true : false;
649 /* call when ep->mtx cannot be held (ep_poll_callback) */
650 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
652 struct wakeup_source *ws;
655 ws = rcu_dereference(epi->ws);
662 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
663 * the scan code, to call f_op->poll(). Also allows for
664 * O(NumReady) performance.
666 * @ep: Pointer to the epoll private data structure.
667 * @sproc: Pointer to the scan callback.
668 * @priv: Private opaque data passed to the @sproc callback.
669 * @depth: The current depth of recursive f_op->poll calls.
670 * @ep_locked: caller already holds ep->mtx
672 * Returns: The same integer error code returned by the @sproc callback.
674 static int ep_scan_ready_list(struct eventpoll *ep,
675 int (*sproc)(struct eventpoll *,
676 struct list_head *, void *),
677 void *priv, int depth, bool ep_locked)
679 int error, pwake = 0;
681 struct epitem *epi, *nepi;
685 * We need to lock this because we could be hit by
686 * eventpoll_release_file() and epoll_ctl().
690 mutex_lock_nested(&ep->mtx, depth);
693 * Steal the ready list, and re-init the original one to the
694 * empty list. Also, set ep->ovflist to NULL so that events
695 * happening while looping w/out locks, are not lost. We cannot
696 * have the poll callback to queue directly on ep->rdllist,
697 * because we want the "sproc" callback to be able to do it
700 spin_lock_irqsave(&ep->lock, flags);
701 list_splice_init(&ep->rdllist, &txlist);
703 spin_unlock_irqrestore(&ep->lock, flags);
706 * Now call the callback function.
708 error = (*sproc)(ep, &txlist, priv);
710 spin_lock_irqsave(&ep->lock, flags);
712 * During the time we spent inside the "sproc" callback, some
713 * other events might have been queued by the poll callback.
714 * We re-insert them inside the main ready-list here.
716 for (nepi = ep->ovflist; (epi = nepi) != NULL;
717 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
719 * We need to check if the item is already in the list.
720 * During the "sproc" callback execution time, items are
721 * queued into ->ovflist but the "txlist" might already
722 * contain them, and the list_splice() below takes care of them.
724 if (!ep_is_linked(&epi->rdllink)) {
725 list_add_tail(&epi->rdllink, &ep->rdllist);
726 ep_pm_stay_awake(epi);
730 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
731 * releasing the lock, events will be queued in the normal way inside
734 ep->ovflist = EP_UNACTIVE_PTR;
737 * Quickly re-inject items left on "txlist".
739 list_splice(&txlist, &ep->rdllist);
742 if (!list_empty(&ep->rdllist)) {
744 * Wake up (if active) both the eventpoll wait list and
745 * the ->poll() wait list (delayed after we release the lock).
747 if (waitqueue_active(&ep->wq))
748 wake_up_locked(&ep->wq);
749 if (waitqueue_active(&ep->poll_wait))
752 spin_unlock_irqrestore(&ep->lock, flags);
755 mutex_unlock(&ep->mtx);
757 /* We have to call this outside the lock */
759 ep_poll_safewake(&ep->poll_wait);
764 static void epi_rcu_free(struct rcu_head *head)
766 struct epitem *epi = container_of(head, struct epitem, rcu);
767 kmem_cache_free(epi_cache, epi);
771 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
772 * all the associated resources. Must be called with "mtx" held.
774 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
777 struct file *file = epi->ffd.file;
780 * Removes poll wait queue hooks. We _have_ to do this without holding
781 * the "ep->lock" otherwise a deadlock might occur. This because of the
782 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
783 * queue head lock when unregistering the wait queue. The wakeup callback
784 * will run by holding the wait queue head lock and will call our callback
785 * that will try to get "ep->lock".
787 ep_unregister_pollwait(ep, epi);
789 /* Remove the current item from the list of epoll hooks */
790 spin_lock(&file->f_lock);
791 list_del_rcu(&epi->fllink);
792 spin_unlock(&file->f_lock);
794 rb_erase(&epi->rbn, &ep->rbr);
796 spin_lock_irqsave(&ep->lock, flags);
797 if (ep_is_linked(&epi->rdllink))
798 list_del_init(&epi->rdllink);
799 spin_unlock_irqrestore(&ep->lock, flags);
801 wakeup_source_unregister(ep_wakeup_source(epi));
803 * At this point it is safe to free the eventpoll item. Use the union
804 * field epi->rcu, since we are trying to minimize the size of
805 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
806 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
807 * use of the rbn field.
809 call_rcu(&epi->rcu, epi_rcu_free);
811 atomic_long_dec(&ep->user->epoll_watches);
816 static void ep_free(struct eventpoll *ep)
821 /* We need to release all tasks waiting for these file */
822 if (waitqueue_active(&ep->poll_wait))
823 ep_poll_safewake(&ep->poll_wait);
826 * We need to lock this because we could be hit by
827 * eventpoll_release_file() while we're freeing the "struct eventpoll".
828 * We do not need to hold "ep->mtx" here because the epoll file
829 * is on the way to be removed and no one has references to it
830 * anymore. The only hit might come from eventpoll_release_file() but
831 * holding "epmutex" is sufficient here.
833 mutex_lock(&epmutex);
836 * Walks through the whole tree by unregistering poll callbacks.
838 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
839 epi = rb_entry(rbp, struct epitem, rbn);
841 ep_unregister_pollwait(ep, epi);
846 * Walks through the whole tree by freeing each "struct epitem". At this
847 * point we are sure no poll callbacks will be lingering around, and also by
848 * holding "epmutex" we can be sure that no file cleanup code will hit
849 * us during this operation. So we can avoid the lock on "ep->lock".
850 * We do not need to lock ep->mtx, either, we only do it to prevent
853 mutex_lock(&ep->mtx);
854 while ((rbp = rb_first(&ep->rbr)) != NULL) {
855 epi = rb_entry(rbp, struct epitem, rbn);
859 mutex_unlock(&ep->mtx);
861 mutex_unlock(&epmutex);
862 mutex_destroy(&ep->mtx);
864 wakeup_source_unregister(ep->ws);
868 static int ep_eventpoll_release(struct inode *inode, struct file *file)
870 struct eventpoll *ep = file->private_data;
878 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
880 pt->_key = epi->event.events;
882 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
885 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
888 struct epitem *epi, *tmp;
891 init_poll_funcptr(&pt, NULL);
893 list_for_each_entry_safe(epi, tmp, head, rdllink) {
894 if (ep_item_poll(epi, &pt))
895 return POLLIN | POLLRDNORM;
898 * Item has been dropped into the ready list by the poll
899 * callback, but it's not actually ready, as far as
900 * caller requested events goes. We can remove it here.
902 __pm_relax(ep_wakeup_source(epi));
903 list_del_init(&epi->rdllink);
910 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
913 struct readyevents_arg {
914 struct eventpoll *ep;
918 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
920 struct readyevents_arg *arg = priv;
922 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
923 call_nests + 1, arg->locked);
926 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
929 struct eventpoll *ep = file->private_data;
930 struct readyevents_arg arg;
933 * During ep_insert() we already hold the ep->mtx for the tfile.
934 * Prevent re-aquisition.
936 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
939 /* Insert inside our poll wait queue */
940 poll_wait(file, &ep->poll_wait, wait);
943 * Proceed to find out if wanted events are really available inside
944 * the ready list. This need to be done under ep_call_nested()
945 * supervision, since the call to f_op->poll() done on listed files
946 * could re-enter here.
948 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
949 ep_poll_readyevents_proc, &arg, ep, current);
951 return pollflags != -1 ? pollflags : 0;
954 #ifdef CONFIG_PROC_FS
955 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
957 struct eventpoll *ep = f->private_data;
960 mutex_lock(&ep->mtx);
961 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
962 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
963 struct inode *inode = file_inode(epi->ffd.file);
965 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
966 " pos:%lli ino:%lx sdev:%x\n",
967 epi->ffd.fd, epi->event.events,
968 (long long)epi->event.data,
969 (long long)epi->ffd.file->f_pos,
970 inode->i_ino, inode->i_sb->s_dev);
971 if (seq_has_overflowed(m))
974 mutex_unlock(&ep->mtx);
978 /* File callbacks that implement the eventpoll file behaviour */
979 static const struct file_operations eventpoll_fops = {
980 #ifdef CONFIG_PROC_FS
981 .show_fdinfo = ep_show_fdinfo,
983 .release = ep_eventpoll_release,
984 .poll = ep_eventpoll_poll,
985 .llseek = noop_llseek,
989 * This is called from eventpoll_release() to unlink files from the eventpoll
990 * interface. We need to have this facility to cleanup correctly files that are
991 * closed without being removed from the eventpoll interface.
993 void eventpoll_release_file(struct file *file)
995 struct eventpoll *ep;
996 struct epitem *epi, *next;
999 * We don't want to get "file->f_lock" because it is not
1000 * necessary. It is not necessary because we're in the "struct file"
1001 * cleanup path, and this means that no one is using this file anymore.
1002 * So, for example, epoll_ctl() cannot hit here since if we reach this
1003 * point, the file counter already went to zero and fget() would fail.
1004 * The only hit might come from ep_free() but by holding the mutex
1005 * will correctly serialize the operation. We do need to acquire
1006 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
1007 * from anywhere but ep_free().
1009 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1011 mutex_lock(&epmutex);
1012 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1014 mutex_lock_nested(&ep->mtx, 0);
1016 mutex_unlock(&ep->mtx);
1018 mutex_unlock(&epmutex);
1021 static int ep_alloc(struct eventpoll **pep)
1024 struct user_struct *user;
1025 struct eventpoll *ep;
1027 user = get_current_user();
1029 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1033 spin_lock_init(&ep->lock);
1034 mutex_init(&ep->mtx);
1035 init_waitqueue_head(&ep->wq);
1036 init_waitqueue_head(&ep->poll_wait);
1037 INIT_LIST_HEAD(&ep->rdllist);
1039 ep->ovflist = EP_UNACTIVE_PTR;
1052 * Search the file inside the eventpoll tree. The RB tree operations
1053 * are protected by the "mtx" mutex, and ep_find() must be called with
1056 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1059 struct rb_node *rbp;
1060 struct epitem *epi, *epir = NULL;
1061 struct epoll_filefd ffd;
1063 ep_set_ffd(&ffd, file, fd);
1064 for (rbp = ep->rbr.rb_node; rbp; ) {
1065 epi = rb_entry(rbp, struct epitem, rbn);
1066 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1068 rbp = rbp->rb_right;
1081 * This is the callback that is passed to the wait queue wakeup
1082 * mechanism. It is called by the stored file descriptors when they
1083 * have events to report.
1085 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1088 unsigned long flags;
1089 struct epitem *epi = ep_item_from_wait(wait);
1090 struct eventpoll *ep = epi->ep;
1093 if ((unsigned long)key & POLLFREE) {
1094 ep_pwq_from_wait(wait)->whead = NULL;
1096 * whead = NULL above can race with ep_remove_wait_queue()
1097 * which can do another remove_wait_queue() after us, so we
1098 * can't use __remove_wait_queue(). whead->lock is held by
1101 list_del_init(&wait->entry);
1104 spin_lock_irqsave(&ep->lock, flags);
1106 ep_set_busy_poll_napi_id(epi);
1109 * If the event mask does not contain any poll(2) event, we consider the
1110 * descriptor to be disabled. This condition is likely the effect of the
1111 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1112 * until the next EPOLL_CTL_MOD will be issued.
1114 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1118 * Check the events coming with the callback. At this stage, not
1119 * every device reports the events in the "key" parameter of the
1120 * callback. We need to be able to handle both cases here, hence the
1121 * test for "key" != NULL before the event match test.
1123 if (key && !((unsigned long) key & epi->event.events))
1127 * If we are transferring events to userspace, we can hold no locks
1128 * (because we're accessing user memory, and because of linux f_op->poll()
1129 * semantics). All the events that happen during that period of time are
1130 * chained in ep->ovflist and requeued later on.
1132 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1133 if (epi->next == EP_UNACTIVE_PTR) {
1134 epi->next = ep->ovflist;
1138 * Activate ep->ws since epi->ws may get
1139 * deactivated at any time.
1141 __pm_stay_awake(ep->ws);
1148 /* If this file is already in the ready list we exit soon */
1149 if (!ep_is_linked(&epi->rdllink)) {
1150 list_add_tail(&epi->rdllink, &ep->rdllist);
1151 ep_pm_stay_awake_rcu(epi);
1155 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1158 if (waitqueue_active(&ep->wq)) {
1159 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1160 !((unsigned long)key & POLLFREE)) {
1161 switch ((unsigned long)key & EPOLLINOUT_BITS) {
1163 if (epi->event.events & POLLIN)
1167 if (epi->event.events & POLLOUT)
1175 wake_up_locked(&ep->wq);
1177 if (waitqueue_active(&ep->poll_wait))
1181 spin_unlock_irqrestore(&ep->lock, flags);
1183 /* We have to call this outside the lock */
1185 ep_poll_safewake(&ep->poll_wait);
1187 if (epi->event.events & EPOLLEXCLUSIVE)
1194 * This is the callback that is used to add our wait queue to the
1195 * target file wakeup lists.
1197 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1200 struct epitem *epi = ep_item_from_epqueue(pt);
1201 struct eppoll_entry *pwq;
1203 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1204 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1207 if (epi->event.events & EPOLLEXCLUSIVE)
1208 add_wait_queue_exclusive(whead, &pwq->wait);
1210 add_wait_queue(whead, &pwq->wait);
1211 list_add_tail(&pwq->llink, &epi->pwqlist);
1214 /* We have to signal that an error occurred */
1219 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1222 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1223 struct epitem *epic;
1227 epic = rb_entry(parent, struct epitem, rbn);
1228 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1230 p = &parent->rb_right;
1232 p = &parent->rb_left;
1234 rb_link_node(&epi->rbn, parent, p);
1235 rb_insert_color(&epi->rbn, &ep->rbr);
1240 #define PATH_ARR_SIZE 5
1242 * These are the number paths of length 1 to 5, that we are allowing to emanate
1243 * from a single file of interest. For example, we allow 1000 paths of length
1244 * 1, to emanate from each file of interest. This essentially represents the
1245 * potential wakeup paths, which need to be limited in order to avoid massive
1246 * uncontrolled wakeup storms. The common use case should be a single ep which
1247 * is connected to n file sources. In this case each file source has 1 path
1248 * of length 1. Thus, the numbers below should be more than sufficient. These
1249 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1250 * and delete can't add additional paths. Protected by the epmutex.
1252 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1253 static int path_count[PATH_ARR_SIZE];
1255 static int path_count_inc(int nests)
1257 /* Allow an arbitrary number of depth 1 paths */
1261 if (++path_count[nests] > path_limits[nests])
1266 static void path_count_init(void)
1270 for (i = 0; i < PATH_ARR_SIZE; i++)
1274 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1277 struct file *file = priv;
1278 struct file *child_file;
1281 /* CTL_DEL can remove links here, but that can't increase our count */
1283 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1284 child_file = epi->ep->file;
1285 if (is_file_epoll(child_file)) {
1286 if (list_empty(&child_file->f_ep_links)) {
1287 if (path_count_inc(call_nests)) {
1292 error = ep_call_nested(&poll_loop_ncalls,
1294 reverse_path_check_proc,
1295 child_file, child_file,
1301 printk(KERN_ERR "reverse_path_check_proc: "
1302 "file is not an ep!\n");
1310 * reverse_path_check - The tfile_check_list is list of file *, which have
1311 * links that are proposed to be newly added. We need to
1312 * make sure that those added links don't add too many
1313 * paths such that we will spend all our time waking up
1314 * eventpoll objects.
1316 * Returns: Returns zero if the proposed links don't create too many paths,
1319 static int reverse_path_check(void)
1322 struct file *current_file;
1324 /* let's call this for all tfiles */
1325 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1327 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1328 reverse_path_check_proc, current_file,
1329 current_file, current);
1336 static int ep_create_wakeup_source(struct epitem *epi)
1339 struct wakeup_source *ws;
1342 epi->ep->ws = wakeup_source_register("eventpoll");
1347 name = epi->ffd.file->f_path.dentry->d_name.name;
1348 ws = wakeup_source_register(name);
1352 rcu_assign_pointer(epi->ws, ws);
1357 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1358 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1360 struct wakeup_source *ws = ep_wakeup_source(epi);
1362 RCU_INIT_POINTER(epi->ws, NULL);
1365 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1366 * used internally by wakeup_source_remove, too (called by
1367 * wakeup_source_unregister), so we cannot use call_rcu
1370 wakeup_source_unregister(ws);
1374 * Must be called with "mtx" held.
1376 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1377 struct file *tfile, int fd, int full_check)
1379 int error, revents, pwake = 0;
1380 unsigned long flags;
1383 struct ep_pqueue epq;
1385 user_watches = atomic_long_read(&ep->user->epoll_watches);
1386 if (unlikely(user_watches >= max_user_watches))
1388 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1391 /* Item initialization follow here ... */
1392 INIT_LIST_HEAD(&epi->rdllink);
1393 INIT_LIST_HEAD(&epi->fllink);
1394 INIT_LIST_HEAD(&epi->pwqlist);
1396 ep_set_ffd(&epi->ffd, tfile, fd);
1397 epi->event = *event;
1399 epi->next = EP_UNACTIVE_PTR;
1400 if (epi->event.events & EPOLLWAKEUP) {
1401 error = ep_create_wakeup_source(epi);
1403 goto error_create_wakeup_source;
1405 RCU_INIT_POINTER(epi->ws, NULL);
1408 /* Initialize the poll table using the queue callback */
1410 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1413 * Attach the item to the poll hooks and get current event bits.
1414 * We can safely use the file* here because its usage count has
1415 * been increased by the caller of this function. Note that after
1416 * this operation completes, the poll callback can start hitting
1419 revents = ep_item_poll(epi, &epq.pt);
1422 * We have to check if something went wrong during the poll wait queue
1423 * install process. Namely an allocation for a wait queue failed due
1424 * high memory pressure.
1428 goto error_unregister;
1430 /* Add the current item to the list of active epoll hook for this file */
1431 spin_lock(&tfile->f_lock);
1432 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1433 spin_unlock(&tfile->f_lock);
1436 * Add the current item to the RB tree. All RB tree operations are
1437 * protected by "mtx", and ep_insert() is called with "mtx" held.
1439 ep_rbtree_insert(ep, epi);
1441 /* now check if we've created too many backpaths */
1443 if (full_check && reverse_path_check())
1444 goto error_remove_epi;
1446 /* We have to drop the new item inside our item list to keep track of it */
1447 spin_lock_irqsave(&ep->lock, flags);
1449 /* record NAPI ID of new item if present */
1450 ep_set_busy_poll_napi_id(epi);
1452 /* If the file is already "ready" we drop it inside the ready list */
1453 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1454 list_add_tail(&epi->rdllink, &ep->rdllist);
1455 ep_pm_stay_awake(epi);
1457 /* Notify waiting tasks that events are available */
1458 if (waitqueue_active(&ep->wq))
1459 wake_up_locked(&ep->wq);
1460 if (waitqueue_active(&ep->poll_wait))
1464 spin_unlock_irqrestore(&ep->lock, flags);
1466 atomic_long_inc(&ep->user->epoll_watches);
1468 /* We have to call this outside the lock */
1470 ep_poll_safewake(&ep->poll_wait);
1475 spin_lock(&tfile->f_lock);
1476 list_del_rcu(&epi->fllink);
1477 spin_unlock(&tfile->f_lock);
1479 rb_erase(&epi->rbn, &ep->rbr);
1482 ep_unregister_pollwait(ep, epi);
1485 * We need to do this because an event could have been arrived on some
1486 * allocated wait queue. Note that we don't care about the ep->ovflist
1487 * list, since that is used/cleaned only inside a section bound by "mtx".
1488 * And ep_insert() is called with "mtx" held.
1490 spin_lock_irqsave(&ep->lock, flags);
1491 if (ep_is_linked(&epi->rdllink))
1492 list_del_init(&epi->rdllink);
1493 spin_unlock_irqrestore(&ep->lock, flags);
1495 wakeup_source_unregister(ep_wakeup_source(epi));
1497 error_create_wakeup_source:
1498 kmem_cache_free(epi_cache, epi);
1504 * Modify the interest event mask by dropping an event if the new mask
1505 * has a match in the current file status. Must be called with "mtx" held.
1507 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1510 unsigned int revents;
1513 init_poll_funcptr(&pt, NULL);
1516 * Set the new event interest mask before calling f_op->poll();
1517 * otherwise we might miss an event that happens between the
1518 * f_op->poll() call and the new event set registering.
1520 epi->event.events = event->events; /* need barrier below */
1521 epi->event.data = event->data; /* protected by mtx */
1522 if (epi->event.events & EPOLLWAKEUP) {
1523 if (!ep_has_wakeup_source(epi))
1524 ep_create_wakeup_source(epi);
1525 } else if (ep_has_wakeup_source(epi)) {
1526 ep_destroy_wakeup_source(epi);
1530 * The following barrier has two effects:
1532 * 1) Flush epi changes above to other CPUs. This ensures
1533 * we do not miss events from ep_poll_callback if an
1534 * event occurs immediately after we call f_op->poll().
1535 * We need this because we did not take ep->lock while
1536 * changing epi above (but ep_poll_callback does take
1539 * 2) We also need to ensure we do not miss _past_ events
1540 * when calling f_op->poll(). This barrier also
1541 * pairs with the barrier in wq_has_sleeper (see
1542 * comments for wq_has_sleeper).
1544 * This barrier will now guarantee ep_poll_callback or f_op->poll
1545 * (or both) will notice the readiness of an item.
1550 * Get current event bits. We can safely use the file* here because
1551 * its usage count has been increased by the caller of this function.
1553 revents = ep_item_poll(epi, &pt);
1556 * If the item is "hot" and it is not registered inside the ready
1557 * list, push it inside.
1559 if (revents & event->events) {
1560 spin_lock_irq(&ep->lock);
1561 if (!ep_is_linked(&epi->rdllink)) {
1562 list_add_tail(&epi->rdllink, &ep->rdllist);
1563 ep_pm_stay_awake(epi);
1565 /* Notify waiting tasks that events are available */
1566 if (waitqueue_active(&ep->wq))
1567 wake_up_locked(&ep->wq);
1568 if (waitqueue_active(&ep->poll_wait))
1571 spin_unlock_irq(&ep->lock);
1574 /* We have to call this outside the lock */
1576 ep_poll_safewake(&ep->poll_wait);
1581 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1584 struct ep_send_events_data *esed = priv;
1586 unsigned int revents;
1588 struct epoll_event __user *uevent;
1589 struct wakeup_source *ws;
1592 init_poll_funcptr(&pt, NULL);
1595 * We can loop without lock because we are passed a task private list.
1596 * Items cannot vanish during the loop because ep_scan_ready_list() is
1597 * holding "mtx" during this call.
1599 for (eventcnt = 0, uevent = esed->events;
1600 !list_empty(head) && eventcnt < esed->maxevents;) {
1601 epi = list_first_entry(head, struct epitem, rdllink);
1604 * Activate ep->ws before deactivating epi->ws to prevent
1605 * triggering auto-suspend here (in case we reactive epi->ws
1608 * This could be rearranged to delay the deactivation of epi->ws
1609 * instead, but then epi->ws would temporarily be out of sync
1610 * with ep_is_linked().
1612 ws = ep_wakeup_source(epi);
1615 __pm_stay_awake(ep->ws);
1619 list_del_init(&epi->rdllink);
1621 revents = ep_item_poll(epi, &pt);
1624 * If the event mask intersect the caller-requested one,
1625 * deliver the event to userspace. Again, ep_scan_ready_list()
1626 * is holding "mtx", so no operations coming from userspace
1627 * can change the item.
1630 if (__put_user(revents, &uevent->events) ||
1631 __put_user(epi->event.data, &uevent->data)) {
1632 list_add(&epi->rdllink, head);
1633 ep_pm_stay_awake(epi);
1634 return eventcnt ? eventcnt : -EFAULT;
1638 if (epi->event.events & EPOLLONESHOT)
1639 epi->event.events &= EP_PRIVATE_BITS;
1640 else if (!(epi->event.events & EPOLLET)) {
1642 * If this file has been added with Level
1643 * Trigger mode, we need to insert back inside
1644 * the ready list, so that the next call to
1645 * epoll_wait() will check again the events
1646 * availability. At this point, no one can insert
1647 * into ep->rdllist besides us. The epoll_ctl()
1648 * callers are locked out by
1649 * ep_scan_ready_list() holding "mtx" and the
1650 * poll callback will queue them in ep->ovflist.
1652 list_add_tail(&epi->rdllink, &ep->rdllist);
1653 ep_pm_stay_awake(epi);
1661 static int ep_send_events(struct eventpoll *ep,
1662 struct epoll_event __user *events, int maxevents)
1664 struct ep_send_events_data esed;
1666 esed.maxevents = maxevents;
1667 esed.events = events;
1669 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1672 static inline struct timespec64 ep_set_mstimeout(long ms)
1674 struct timespec64 now, ts = {
1675 .tv_sec = ms / MSEC_PER_SEC,
1676 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1679 ktime_get_ts64(&now);
1680 return timespec64_add_safe(now, ts);
1684 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1687 * @ep: Pointer to the eventpoll context.
1688 * @events: Pointer to the userspace buffer where the ready events should be
1690 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1691 * @timeout: Maximum timeout for the ready events fetch operation, in
1692 * milliseconds. If the @timeout is zero, the function will not block,
1693 * while if the @timeout is less than zero, the function will block
1694 * until at least one event has been retrieved (or an error
1697 * Returns: Returns the number of ready events which have been fetched, or an
1698 * error code, in case of error.
1700 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1701 int maxevents, long timeout)
1703 int res = 0, eavail, timed_out = 0;
1704 unsigned long flags;
1706 wait_queue_entry_t wait;
1707 ktime_t expires, *to = NULL;
1710 struct timespec64 end_time = ep_set_mstimeout(timeout);
1712 slack = select_estimate_accuracy(&end_time);
1714 *to = timespec64_to_ktime(end_time);
1715 } else if (timeout == 0) {
1717 * Avoid the unnecessary trip to the wait queue loop, if the
1718 * caller specified a non blocking operation.
1721 spin_lock_irqsave(&ep->lock, flags);
1727 if (!ep_events_available(ep))
1728 ep_busy_loop(ep, timed_out);
1730 spin_lock_irqsave(&ep->lock, flags);
1732 if (!ep_events_available(ep)) {
1734 * Busy poll timed out. Drop NAPI ID for now, we can add
1735 * it back in when we have moved a socket with a valid NAPI
1736 * ID onto the ready list.
1738 ep_reset_busy_poll_napi_id(ep);
1741 * We don't have any available event to return to the caller.
1742 * We need to sleep here, and we will be wake up by
1743 * ep_poll_callback() when events will become available.
1745 init_waitqueue_entry(&wait, current);
1746 __add_wait_queue_exclusive(&ep->wq, &wait);
1750 * We don't want to sleep if the ep_poll_callback() sends us
1751 * a wakeup in between. That's why we set the task state
1752 * to TASK_INTERRUPTIBLE before doing the checks.
1754 set_current_state(TASK_INTERRUPTIBLE);
1756 * Always short-circuit for fatal signals to allow
1757 * threads to make a timely exit without the chance of
1758 * finding more events available and fetching
1761 if (fatal_signal_pending(current)) {
1765 if (ep_events_available(ep) || timed_out)
1767 if (signal_pending(current)) {
1772 spin_unlock_irqrestore(&ep->lock, flags);
1773 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1776 spin_lock_irqsave(&ep->lock, flags);
1779 __remove_wait_queue(&ep->wq, &wait);
1780 __set_current_state(TASK_RUNNING);
1783 /* Is it worth to try to dig for events ? */
1784 eavail = ep_events_available(ep);
1786 spin_unlock_irqrestore(&ep->lock, flags);
1789 * Try to transfer events to user space. In case we get 0 events and
1790 * there's still timeout left over, we go trying again in search of
1793 if (!res && eavail &&
1794 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1801 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1802 * API, to verify that adding an epoll file inside another
1803 * epoll structure, does not violate the constraints, in
1804 * terms of closed loops, or too deep chains (which can
1805 * result in excessive stack usage).
1807 * @priv: Pointer to the epoll file to be currently checked.
1808 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1809 * data structure pointer.
1810 * @call_nests: Current dept of the @ep_call_nested() call stack.
1812 * Returns: Returns zero if adding the epoll @file inside current epoll
1813 * structure @ep does not violate the constraints, or -1 otherwise.
1815 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1818 struct file *file = priv;
1819 struct eventpoll *ep = file->private_data;
1820 struct eventpoll *ep_tovisit;
1821 struct rb_node *rbp;
1824 mutex_lock_nested(&ep->mtx, call_nests + 1);
1826 list_add(&ep->visited_list_link, &visited_list);
1827 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1828 epi = rb_entry(rbp, struct epitem, rbn);
1829 if (unlikely(is_file_epoll(epi->ffd.file))) {
1830 ep_tovisit = epi->ffd.file->private_data;
1831 if (ep_tovisit->visited)
1833 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1834 ep_loop_check_proc, epi->ffd.file,
1835 ep_tovisit, current);
1840 * If we've reached a file that is not associated with
1841 * an ep, then we need to check if the newly added
1842 * links are going to add too many wakeup paths. We do
1843 * this by adding it to the tfile_check_list, if it's
1844 * not already there, and calling reverse_path_check()
1845 * during ep_insert().
1847 if (list_empty(&epi->ffd.file->f_tfile_llink))
1848 list_add(&epi->ffd.file->f_tfile_llink,
1852 mutex_unlock(&ep->mtx);
1858 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1859 * another epoll file (represented by @ep) does not create
1860 * closed loops or too deep chains.
1862 * @ep: Pointer to the epoll private data structure.
1863 * @file: Pointer to the epoll file to be checked.
1865 * Returns: Returns zero if adding the epoll @file inside current epoll
1866 * structure @ep does not violate the constraints, or -1 otherwise.
1868 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1871 struct eventpoll *ep_cur, *ep_next;
1873 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1874 ep_loop_check_proc, file, ep, current);
1875 /* clear visited list */
1876 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1877 visited_list_link) {
1878 ep_cur->visited = 0;
1879 list_del(&ep_cur->visited_list_link);
1884 static void clear_tfile_check_list(void)
1888 /* first clear the tfile_check_list */
1889 while (!list_empty(&tfile_check_list)) {
1890 file = list_first_entry(&tfile_check_list, struct file,
1892 list_del_init(&file->f_tfile_llink);
1894 INIT_LIST_HEAD(&tfile_check_list);
1898 * Open an eventpoll file descriptor.
1900 SYSCALL_DEFINE1(epoll_create1, int, flags)
1903 struct eventpoll *ep = NULL;
1906 /* Check the EPOLL_* constant for consistency. */
1907 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1909 if (flags & ~EPOLL_CLOEXEC)
1912 * Create the internal data structure ("struct eventpoll").
1914 error = ep_alloc(&ep);
1918 * Creates all the items needed to setup an eventpoll file. That is,
1919 * a file structure and a free file descriptor.
1921 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1926 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1927 O_RDWR | (flags & O_CLOEXEC));
1929 error = PTR_ERR(file);
1933 fd_install(fd, file);
1943 SYSCALL_DEFINE1(epoll_create, int, size)
1948 return sys_epoll_create1(0);
1952 * The following function implements the controller interface for
1953 * the eventpoll file that enables the insertion/removal/change of
1954 * file descriptors inside the interest set.
1956 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1957 struct epoll_event __user *, event)
1962 struct eventpoll *ep;
1964 struct epoll_event epds;
1965 struct eventpoll *tep = NULL;
1968 if (ep_op_has_event(op) &&
1969 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1977 /* Get the "struct file *" for the target file */
1982 /* The target file descriptor must support poll */
1984 if (!tf.file->f_op->poll)
1985 goto error_tgt_fput;
1987 /* Check if EPOLLWAKEUP is allowed */
1988 if (ep_op_has_event(op))
1989 ep_take_care_of_epollwakeup(&epds);
1992 * We have to check that the file structure underneath the file descriptor
1993 * the user passed to us _is_ an eventpoll file. And also we do not permit
1994 * adding an epoll file descriptor inside itself.
1997 if (f.file == tf.file || !is_file_epoll(f.file))
1998 goto error_tgt_fput;
2001 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2002 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2003 * Also, we do not currently supported nested exclusive wakeups.
2005 if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2006 if (op == EPOLL_CTL_MOD)
2007 goto error_tgt_fput;
2008 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2009 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2010 goto error_tgt_fput;
2014 * At this point it is safe to assume that the "private_data" contains
2015 * our own data structure.
2017 ep = f.file->private_data;
2020 * When we insert an epoll file descriptor, inside another epoll file
2021 * descriptor, there is the change of creating closed loops, which are
2022 * better be handled here, than in more critical paths. While we are
2023 * checking for loops we also determine the list of files reachable
2024 * and hang them on the tfile_check_list, so we can check that we
2025 * haven't created too many possible wakeup paths.
2027 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2028 * the epoll file descriptor is attaching directly to a wakeup source,
2029 * unless the epoll file descriptor is nested. The purpose of taking the
2030 * 'epmutex' on add is to prevent complex toplogies such as loops and
2031 * deep wakeup paths from forming in parallel through multiple
2032 * EPOLL_CTL_ADD operations.
2034 mutex_lock_nested(&ep->mtx, 0);
2035 if (op == EPOLL_CTL_ADD) {
2036 if (!list_empty(&f.file->f_ep_links) ||
2037 is_file_epoll(tf.file)) {
2039 mutex_unlock(&ep->mtx);
2040 mutex_lock(&epmutex);
2041 if (is_file_epoll(tf.file)) {
2043 if (ep_loop_check(ep, tf.file) != 0) {
2044 clear_tfile_check_list();
2045 goto error_tgt_fput;
2048 list_add(&tf.file->f_tfile_llink,
2050 mutex_lock_nested(&ep->mtx, 0);
2051 if (is_file_epoll(tf.file)) {
2052 tep = tf.file->private_data;
2053 mutex_lock_nested(&tep->mtx, 1);
2059 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2060 * above, we can be sure to be able to use the item looked up by
2061 * ep_find() till we release the mutex.
2063 epi = ep_find(ep, tf.file, fd);
2069 epds.events |= POLLERR | POLLHUP;
2070 error = ep_insert(ep, &epds, tf.file, fd, full_check);
2074 clear_tfile_check_list();
2078 error = ep_remove(ep, epi);
2084 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2085 epds.events |= POLLERR | POLLHUP;
2086 error = ep_modify(ep, epi, &epds);
2093 mutex_unlock(&tep->mtx);
2094 mutex_unlock(&ep->mtx);
2098 mutex_unlock(&epmutex);
2109 * Implement the event wait interface for the eventpoll file. It is the kernel
2110 * part of the user space epoll_wait(2).
2112 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2113 int, maxevents, int, timeout)
2117 struct eventpoll *ep;
2119 /* The maximum number of event must be greater than zero */
2120 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2123 /* Verify that the area passed by the user is writeable */
2124 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2127 /* Get the "struct file *" for the eventpoll file */
2133 * We have to check that the file structure underneath the fd
2134 * the user passed to us _is_ an eventpoll file.
2137 if (!is_file_epoll(f.file))
2141 * At this point it is safe to assume that the "private_data" contains
2142 * our own data structure.
2144 ep = f.file->private_data;
2146 /* Time to fish for events ... */
2147 error = ep_poll(ep, events, maxevents, timeout);
2155 * Implement the event wait interface for the eventpoll file. It is the kernel
2156 * part of the user space epoll_pwait(2).
2158 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2159 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2163 sigset_t ksigmask, sigsaved;
2166 * If the caller wants a certain signal mask to be set during the wait,
2170 if (sigsetsize != sizeof(sigset_t))
2172 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2174 sigsaved = current->blocked;
2175 set_current_blocked(&ksigmask);
2178 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2181 * If we changed the signal mask, we need to restore the original one.
2182 * In case we've got a signal while waiting, we do not restore the
2183 * signal mask yet, and we allow do_signal() to deliver the signal on
2184 * the way back to userspace, before the signal mask is restored.
2187 if (error == -EINTR) {
2188 memcpy(¤t->saved_sigmask, &sigsaved,
2190 set_restore_sigmask();
2192 set_current_blocked(&sigsaved);
2198 #ifdef CONFIG_COMPAT
2199 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2200 struct epoll_event __user *, events,
2201 int, maxevents, int, timeout,
2202 const compat_sigset_t __user *, sigmask,
2203 compat_size_t, sigsetsize)
2206 compat_sigset_t csigmask;
2207 sigset_t ksigmask, sigsaved;
2210 * If the caller wants a certain signal mask to be set during the wait,
2214 if (sigsetsize != sizeof(compat_sigset_t))
2216 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2218 sigset_from_compat(&ksigmask, &csigmask);
2219 sigsaved = current->blocked;
2220 set_current_blocked(&ksigmask);
2223 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2226 * If we changed the signal mask, we need to restore the original one.
2227 * In case we've got a signal while waiting, we do not restore the
2228 * signal mask yet, and we allow do_signal() to deliver the signal on
2229 * the way back to userspace, before the signal mask is restored.
2232 if (err == -EINTR) {
2233 memcpy(¤t->saved_sigmask, &sigsaved,
2235 set_restore_sigmask();
2237 set_current_blocked(&sigsaved);
2244 static int __init eventpoll_init(void)
2250 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2252 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2254 BUG_ON(max_user_watches < 0);
2257 * Initialize the structure used to perform epoll file descriptor
2258 * inclusion loops checks.
2260 ep_nested_calls_init(&poll_loop_ncalls);
2262 /* Initialize the structure used to perform safe poll wait head wake ups */
2263 ep_nested_calls_init(&poll_safewake_ncalls);
2265 /* Initialize the structure used to perform file's f_op->poll() calls */
2266 ep_nested_calls_init(&poll_readywalk_ncalls);
2269 * We can have many thousands of epitems, so prevent this from
2270 * using an extra cache line on 64-bit (and smaller) CPUs
2272 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2274 /* Allocates slab cache used to allocate "struct epitem" items */
2275 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2276 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2278 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2279 pwq_cache = kmem_cache_create("eventpoll_pwq",
2280 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2284 fs_initcall(eventpoll_init);