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 */
208 struct rb_root_cached rbr;
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 */
228 #ifdef CONFIG_NET_RX_BUSY_POLL
229 /* used to track busy poll napi_id */
230 unsigned int napi_id;
234 /* Wait structure used by the poll hooks */
235 struct eppoll_entry {
236 /* List header used to link this structure to the "struct epitem" */
237 struct list_head llink;
239 /* The "base" pointer is set to the container "struct epitem" */
243 * Wait queue item that will be linked to the target file wait
246 wait_queue_entry_t wait;
248 /* The wait queue head that linked the "wait" wait queue item */
249 wait_queue_head_t *whead;
252 /* Wrapper struct used by poll queueing */
258 /* Used by the ep_send_events() function as callback private data */
259 struct ep_send_events_data {
261 struct epoll_event __user *events;
265 * Configuration options available inside /proc/sys/fs/epoll/
267 /* Maximum number of epoll watched descriptors, per user */
268 static long max_user_watches __read_mostly;
271 * This mutex is used to serialize ep_free() and eventpoll_release_file().
273 static DEFINE_MUTEX(epmutex);
275 static u64 loop_check_gen = 0;
277 /* Used to check for epoll file descriptor inclusion loops */
278 static struct nested_calls poll_loop_ncalls;
280 /* Used for safe wake up implementation */
281 static struct nested_calls poll_safewake_ncalls;
283 /* Used to call file's f_op->poll() under the nested calls boundaries */
284 static struct nested_calls poll_readywalk_ncalls;
286 /* Slab cache used to allocate "struct epitem" */
287 static struct kmem_cache *epi_cache __read_mostly;
289 /* Slab cache used to allocate "struct eppoll_entry" */
290 static struct kmem_cache *pwq_cache __read_mostly;
293 * List of files with newly added links, where we may need to limit the number
294 * of emanating paths. Protected by the epmutex.
296 static LIST_HEAD(tfile_check_list);
300 #include <linux/sysctl.h>
303 static long long_max = LONG_MAX;
305 struct ctl_table epoll_table[] = {
307 .procname = "max_user_watches",
308 .data = &max_user_watches,
309 .maxlen = sizeof(max_user_watches),
311 .proc_handler = proc_doulongvec_minmax,
317 #endif /* CONFIG_SYSCTL */
319 static const struct file_operations eventpoll_fops;
321 static inline int is_file_epoll(struct file *f)
323 return f->f_op == &eventpoll_fops;
326 /* Setup the structure that is used as key for the RB tree */
327 static inline void ep_set_ffd(struct epoll_filefd *ffd,
328 struct file *file, int fd)
334 /* Compare RB tree keys */
335 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
336 struct epoll_filefd *p2)
338 return (p1->file > p2->file ? +1:
339 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
342 /* Tells us if the item is currently linked */
343 static inline int ep_is_linked(struct list_head *p)
345 return !list_empty(p);
348 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
350 return container_of(p, struct eppoll_entry, wait);
353 /* Get the "struct epitem" from a wait queue pointer */
354 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
356 return container_of(p, struct eppoll_entry, wait)->base;
359 /* Get the "struct epitem" from an epoll queue wrapper */
360 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
362 return container_of(p, struct ep_pqueue, pt)->epi;
365 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
366 static inline int ep_op_has_event(int op)
368 return op != EPOLL_CTL_DEL;
371 /* Initialize the poll safe wake up structure */
372 static void ep_nested_calls_init(struct nested_calls *ncalls)
374 INIT_LIST_HEAD(&ncalls->tasks_call_list);
375 spin_lock_init(&ncalls->lock);
379 * ep_events_available - Checks if ready events might be available.
381 * @ep: Pointer to the eventpoll context.
383 * Returns: Returns a value different than zero if ready events are available,
386 static inline int ep_events_available(struct eventpoll *ep)
388 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
391 #ifdef CONFIG_NET_RX_BUSY_POLL
392 static bool ep_busy_loop_end(void *p, unsigned long start_time)
394 struct eventpoll *ep = p;
396 return ep_events_available(ep) || busy_loop_timeout(start_time);
398 #endif /* CONFIG_NET_RX_BUSY_POLL */
401 * Busy poll if globally on and supporting sockets found && no events,
402 * busy loop will return if need_resched or ep_events_available.
404 * we must do our busy polling with irqs enabled
406 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
408 #ifdef CONFIG_NET_RX_BUSY_POLL
409 unsigned int napi_id = READ_ONCE(ep->napi_id);
411 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
412 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
416 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
418 #ifdef CONFIG_NET_RX_BUSY_POLL
425 * Set epoll busy poll NAPI ID from sk.
427 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
429 #ifdef CONFIG_NET_RX_BUSY_POLL
430 struct eventpoll *ep;
431 unsigned int napi_id;
436 if (!net_busy_loop_on())
439 sock = sock_from_file(epi->ffd.file, &err);
447 napi_id = READ_ONCE(sk->sk_napi_id);
450 /* Non-NAPI IDs can be rejected
452 * Nothing to do if we already have this ID
454 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
457 /* record NAPI ID for use in next busy poll */
458 ep->napi_id = napi_id;
463 * ep_call_nested - Perform a bound (possibly) nested call, by checking
464 * that the recursion limit is not exceeded, and that
465 * the same nested call (by the meaning of same cookie) is
468 * @ncalls: Pointer to the nested_calls structure to be used for this call.
469 * @max_nests: Maximum number of allowed nesting calls.
470 * @nproc: Nested call core function pointer.
471 * @priv: Opaque data to be passed to the @nproc callback.
472 * @cookie: Cookie to be used to identify this nested call.
473 * @ctx: This instance context.
475 * Returns: Returns the code returned by the @nproc callback, or -1 if
476 * the maximum recursion limit has been exceeded.
478 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
479 int (*nproc)(void *, void *, int), void *priv,
480 void *cookie, void *ctx)
482 int error, call_nests = 0;
484 struct list_head *lsthead = &ncalls->tasks_call_list;
485 struct nested_call_node *tncur;
486 struct nested_call_node tnode;
488 spin_lock_irqsave(&ncalls->lock, flags);
491 * Try to see if the current task is already inside this wakeup call.
492 * We use a list here, since the population inside this set is always
495 list_for_each_entry(tncur, lsthead, llink) {
496 if (tncur->ctx == ctx &&
497 (tncur->cookie == cookie || ++call_nests > max_nests)) {
499 * Ops ... loop detected or maximum nest level reached.
500 * We abort this wake by breaking the cycle itself.
507 /* Add the current task and cookie to the list */
509 tnode.cookie = cookie;
510 list_add(&tnode.llink, lsthead);
512 spin_unlock_irqrestore(&ncalls->lock, flags);
514 /* Call the nested function */
515 error = (*nproc)(priv, cookie, call_nests);
517 /* Remove the current task from the list */
518 spin_lock_irqsave(&ncalls->lock, flags);
519 list_del(&tnode.llink);
521 spin_unlock_irqrestore(&ncalls->lock, flags);
527 * As described in commit 0ccf831cb lockdep: annotate epoll
528 * the use of wait queues used by epoll is done in a very controlled
529 * manner. Wake ups can nest inside each other, but are never done
530 * with the same locking. For example:
533 * efd1 = epoll_create();
534 * efd2 = epoll_create();
535 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
536 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
538 * When a packet arrives to the device underneath "dfd", the net code will
539 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
540 * callback wakeup entry on that queue, and the wake_up() performed by the
541 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
542 * (efd1) notices that it may have some event ready, so it needs to wake up
543 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
544 * that ends up in another wake_up(), after having checked about the
545 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
546 * avoid stack blasting.
548 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
549 * this special case of epoll.
551 #ifdef CONFIG_DEBUG_LOCK_ALLOC
552 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
553 unsigned long events, int subclass)
557 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
558 wake_up_locked_poll(wqueue, events);
559 spin_unlock_irqrestore(&wqueue->lock, flags);
562 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
563 unsigned long events, int subclass)
565 wake_up_poll(wqueue, events);
569 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
571 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
577 * Perform a safe wake up of the poll wait list. The problem is that
578 * with the new callback'd wake up system, it is possible that the
579 * poll callback is reentered from inside the call to wake_up() done
580 * on the poll wait queue head. The rule is that we cannot reenter the
581 * wake up code from the same task more than EP_MAX_NESTS times,
582 * and we cannot reenter the same wait queue head at all. This will
583 * enable to have a hierarchy of epoll file descriptor of no more than
586 static void ep_poll_safewake(wait_queue_head_t *wq)
588 int this_cpu = get_cpu();
590 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
591 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
596 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
598 wait_queue_head_t *whead;
602 * If it is cleared by POLLFREE, it should be rcu-safe.
603 * If we read NULL we need a barrier paired with
604 * smp_store_release() in ep_poll_callback(), otherwise
605 * we rely on whead->lock.
607 whead = smp_load_acquire(&pwq->whead);
609 remove_wait_queue(whead, &pwq->wait);
614 * This function unregisters poll callbacks from the associated file
615 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
618 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
620 struct list_head *lsthead = &epi->pwqlist;
621 struct eppoll_entry *pwq;
623 while (!list_empty(lsthead)) {
624 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
626 list_del(&pwq->llink);
627 ep_remove_wait_queue(pwq);
628 kmem_cache_free(pwq_cache, pwq);
632 /* call only when ep->mtx is held */
633 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
635 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
638 /* call only when ep->mtx is held */
639 static inline void ep_pm_stay_awake(struct epitem *epi)
641 struct wakeup_source *ws = ep_wakeup_source(epi);
647 static inline bool ep_has_wakeup_source(struct epitem *epi)
649 return rcu_access_pointer(epi->ws) ? true : false;
652 /* call when ep->mtx cannot be held (ep_poll_callback) */
653 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
655 struct wakeup_source *ws;
658 ws = rcu_dereference(epi->ws);
665 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
666 * the scan code, to call f_op->poll(). Also allows for
667 * O(NumReady) performance.
669 * @ep: Pointer to the epoll private data structure.
670 * @sproc: Pointer to the scan callback.
671 * @priv: Private opaque data passed to the @sproc callback.
672 * @depth: The current depth of recursive f_op->poll calls.
673 * @ep_locked: caller already holds ep->mtx
675 * Returns: The same integer error code returned by the @sproc callback.
677 static int ep_scan_ready_list(struct eventpoll *ep,
678 int (*sproc)(struct eventpoll *,
679 struct list_head *, void *),
680 void *priv, int depth, bool ep_locked)
682 int error, pwake = 0;
684 struct epitem *epi, *nepi;
688 * We need to lock this because we could be hit by
689 * eventpoll_release_file() and epoll_ctl().
693 mutex_lock_nested(&ep->mtx, depth);
696 * Steal the ready list, and re-init the original one to the
697 * empty list. Also, set ep->ovflist to NULL so that events
698 * happening while looping w/out locks, are not lost. We cannot
699 * have the poll callback to queue directly on ep->rdllist,
700 * because we want the "sproc" callback to be able to do it
703 spin_lock_irqsave(&ep->lock, flags);
704 list_splice_init(&ep->rdllist, &txlist);
706 spin_unlock_irqrestore(&ep->lock, flags);
709 * Now call the callback function.
711 error = (*sproc)(ep, &txlist, priv);
713 spin_lock_irqsave(&ep->lock, flags);
715 * During the time we spent inside the "sproc" callback, some
716 * other events might have been queued by the poll callback.
717 * We re-insert them inside the main ready-list here.
719 for (nepi = ep->ovflist; (epi = nepi) != NULL;
720 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
722 * We need to check if the item is already in the list.
723 * During the "sproc" callback execution time, items are
724 * queued into ->ovflist but the "txlist" might already
725 * contain them, and the list_splice() below takes care of them.
727 if (!ep_is_linked(&epi->rdllink)) {
728 list_add_tail(&epi->rdllink, &ep->rdllist);
729 ep_pm_stay_awake(epi);
733 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
734 * releasing the lock, events will be queued in the normal way inside
737 ep->ovflist = EP_UNACTIVE_PTR;
740 * Quickly re-inject items left on "txlist".
742 list_splice(&txlist, &ep->rdllist);
745 if (!list_empty(&ep->rdllist)) {
747 * Wake up (if active) both the eventpoll wait list and
748 * the ->poll() wait list (delayed after we release the lock).
750 if (waitqueue_active(&ep->wq))
751 wake_up_locked(&ep->wq);
752 if (waitqueue_active(&ep->poll_wait))
755 spin_unlock_irqrestore(&ep->lock, flags);
758 mutex_unlock(&ep->mtx);
760 /* We have to call this outside the lock */
762 ep_poll_safewake(&ep->poll_wait);
767 static void epi_rcu_free(struct rcu_head *head)
769 struct epitem *epi = container_of(head, struct epitem, rcu);
770 kmem_cache_free(epi_cache, epi);
774 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
775 * all the associated resources. Must be called with "mtx" held.
777 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
780 struct file *file = epi->ffd.file;
783 * Removes poll wait queue hooks. We _have_ to do this without holding
784 * the "ep->lock" otherwise a deadlock might occur. This because of the
785 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
786 * queue head lock when unregistering the wait queue. The wakeup callback
787 * will run by holding the wait queue head lock and will call our callback
788 * that will try to get "ep->lock".
790 ep_unregister_pollwait(ep, epi);
792 /* Remove the current item from the list of epoll hooks */
793 spin_lock(&file->f_lock);
794 list_del_rcu(&epi->fllink);
795 spin_unlock(&file->f_lock);
797 rb_erase_cached(&epi->rbn, &ep->rbr);
799 spin_lock_irqsave(&ep->lock, flags);
800 if (ep_is_linked(&epi->rdllink))
801 list_del_init(&epi->rdllink);
802 spin_unlock_irqrestore(&ep->lock, flags);
804 wakeup_source_unregister(ep_wakeup_source(epi));
806 * At this point it is safe to free the eventpoll item. Use the union
807 * field epi->rcu, since we are trying to minimize the size of
808 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
809 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
810 * use of the rbn field.
812 call_rcu(&epi->rcu, epi_rcu_free);
814 atomic_long_dec(&ep->user->epoll_watches);
819 static void ep_free(struct eventpoll *ep)
824 /* We need to release all tasks waiting for these file */
825 if (waitqueue_active(&ep->poll_wait))
826 ep_poll_safewake(&ep->poll_wait);
829 * We need to lock this because we could be hit by
830 * eventpoll_release_file() while we're freeing the "struct eventpoll".
831 * We do not need to hold "ep->mtx" here because the epoll file
832 * is on the way to be removed and no one has references to it
833 * anymore. The only hit might come from eventpoll_release_file() but
834 * holding "epmutex" is sufficient here.
836 mutex_lock(&epmutex);
839 * Walks through the whole tree by unregistering poll callbacks.
841 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
842 epi = rb_entry(rbp, struct epitem, rbn);
844 ep_unregister_pollwait(ep, epi);
849 * Walks through the whole tree by freeing each "struct epitem". At this
850 * point we are sure no poll callbacks will be lingering around, and also by
851 * holding "epmutex" we can be sure that no file cleanup code will hit
852 * us during this operation. So we can avoid the lock on "ep->lock".
853 * We do not need to lock ep->mtx, either, we only do it to prevent
856 mutex_lock(&ep->mtx);
857 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
858 epi = rb_entry(rbp, struct epitem, rbn);
862 mutex_unlock(&ep->mtx);
864 mutex_unlock(&epmutex);
865 mutex_destroy(&ep->mtx);
867 wakeup_source_unregister(ep->ws);
871 static int ep_eventpoll_release(struct inode *inode, struct file *file)
873 struct eventpoll *ep = file->private_data;
881 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
883 pt->_key = epi->event.events;
885 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
888 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
891 struct epitem *epi, *tmp;
894 init_poll_funcptr(&pt, NULL);
896 list_for_each_entry_safe(epi, tmp, head, rdllink) {
897 if (ep_item_poll(epi, &pt))
898 return POLLIN | POLLRDNORM;
901 * Item has been dropped into the ready list by the poll
902 * callback, but it's not actually ready, as far as
903 * caller requested events goes. We can remove it here.
905 __pm_relax(ep_wakeup_source(epi));
906 list_del_init(&epi->rdllink);
913 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
916 struct readyevents_arg {
917 struct eventpoll *ep;
921 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
923 struct readyevents_arg *arg = priv;
925 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
926 call_nests + 1, arg->locked);
929 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
932 struct eventpoll *ep = file->private_data;
933 struct readyevents_arg arg;
936 * During ep_insert() we already hold the ep->mtx for the tfile.
937 * Prevent re-aquisition.
939 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
942 /* Insert inside our poll wait queue */
943 poll_wait(file, &ep->poll_wait, wait);
946 * Proceed to find out if wanted events are really available inside
947 * the ready list. This need to be done under ep_call_nested()
948 * supervision, since the call to f_op->poll() done on listed files
949 * could re-enter here.
951 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
952 ep_poll_readyevents_proc, &arg, ep, current);
954 return pollflags != -1 ? pollflags : 0;
957 #ifdef CONFIG_PROC_FS
958 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
960 struct eventpoll *ep = f->private_data;
963 mutex_lock(&ep->mtx);
964 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
965 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
966 struct inode *inode = file_inode(epi->ffd.file);
968 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
969 " pos:%lli ino:%lx sdev:%x\n",
970 epi->ffd.fd, epi->event.events,
971 (long long)epi->event.data,
972 (long long)epi->ffd.file->f_pos,
973 inode->i_ino, inode->i_sb->s_dev);
974 if (seq_has_overflowed(m))
977 mutex_unlock(&ep->mtx);
981 /* File callbacks that implement the eventpoll file behaviour */
982 static const struct file_operations eventpoll_fops = {
983 #ifdef CONFIG_PROC_FS
984 .show_fdinfo = ep_show_fdinfo,
986 .release = ep_eventpoll_release,
987 .poll = ep_eventpoll_poll,
988 .llseek = noop_llseek,
992 * This is called from eventpoll_release() to unlink files from the eventpoll
993 * interface. We need to have this facility to cleanup correctly files that are
994 * closed without being removed from the eventpoll interface.
996 void eventpoll_release_file(struct file *file)
998 struct eventpoll *ep;
999 struct epitem *epi, *next;
1002 * We don't want to get "file->f_lock" because it is not
1003 * necessary. It is not necessary because we're in the "struct file"
1004 * cleanup path, and this means that no one is using this file anymore.
1005 * So, for example, epoll_ctl() cannot hit here since if we reach this
1006 * point, the file counter already went to zero and fget() would fail.
1007 * The only hit might come from ep_free() but by holding the mutex
1008 * will correctly serialize the operation. We do need to acquire
1009 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
1010 * from anywhere but ep_free().
1012 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1014 mutex_lock(&epmutex);
1015 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1017 mutex_lock_nested(&ep->mtx, 0);
1019 mutex_unlock(&ep->mtx);
1021 mutex_unlock(&epmutex);
1024 static int ep_alloc(struct eventpoll **pep)
1027 struct user_struct *user;
1028 struct eventpoll *ep;
1030 user = get_current_user();
1032 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1036 spin_lock_init(&ep->lock);
1037 mutex_init(&ep->mtx);
1038 init_waitqueue_head(&ep->wq);
1039 init_waitqueue_head(&ep->poll_wait);
1040 INIT_LIST_HEAD(&ep->rdllist);
1041 ep->rbr = RB_ROOT_CACHED;
1042 ep->ovflist = EP_UNACTIVE_PTR;
1055 * Search the file inside the eventpoll tree. The RB tree operations
1056 * are protected by the "mtx" mutex, and ep_find() must be called with
1059 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1062 struct rb_node *rbp;
1063 struct epitem *epi, *epir = NULL;
1064 struct epoll_filefd ffd;
1066 ep_set_ffd(&ffd, file, fd);
1067 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1068 epi = rb_entry(rbp, struct epitem, rbn);
1069 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1071 rbp = rbp->rb_right;
1083 #ifdef CONFIG_CHECKPOINT_RESTORE
1084 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1086 struct rb_node *rbp;
1089 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1090 epi = rb_entry(rbp, struct epitem, rbn);
1091 if (epi->ffd.fd == tfd) {
1103 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1106 struct file *file_raw;
1107 struct eventpoll *ep;
1110 if (!is_file_epoll(file))
1111 return ERR_PTR(-EINVAL);
1113 ep = file->private_data;
1115 mutex_lock(&ep->mtx);
1116 epi = ep_find_tfd(ep, tfd, toff);
1118 file_raw = epi->ffd.file;
1120 file_raw = ERR_PTR(-ENOENT);
1121 mutex_unlock(&ep->mtx);
1125 #endif /* CONFIG_CHECKPOINT_RESTORE */
1128 * This is the callback that is passed to the wait queue wakeup
1129 * mechanism. It is called by the stored file descriptors when they
1130 * have events to report.
1132 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1135 unsigned long flags;
1136 struct epitem *epi = ep_item_from_wait(wait);
1137 struct eventpoll *ep = epi->ep;
1140 spin_lock_irqsave(&ep->lock, flags);
1142 ep_set_busy_poll_napi_id(epi);
1145 * If the event mask does not contain any poll(2) event, we consider the
1146 * descriptor to be disabled. This condition is likely the effect of the
1147 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1148 * until the next EPOLL_CTL_MOD will be issued.
1150 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1154 * Check the events coming with the callback. At this stage, not
1155 * every device reports the events in the "key" parameter of the
1156 * callback. We need to be able to handle both cases here, hence the
1157 * test for "key" != NULL before the event match test.
1159 if (key && !((unsigned long) key & epi->event.events))
1163 * If we are transferring events to userspace, we can hold no locks
1164 * (because we're accessing user memory, and because of linux f_op->poll()
1165 * semantics). All the events that happen during that period of time are
1166 * chained in ep->ovflist and requeued later on.
1168 if (ep->ovflist != EP_UNACTIVE_PTR) {
1169 if (epi->next == EP_UNACTIVE_PTR) {
1170 epi->next = ep->ovflist;
1174 * Activate ep->ws since epi->ws may get
1175 * deactivated at any time.
1177 __pm_stay_awake(ep->ws);
1184 /* If this file is already in the ready list we exit soon */
1185 if (!ep_is_linked(&epi->rdllink)) {
1186 list_add_tail(&epi->rdllink, &ep->rdllist);
1187 ep_pm_stay_awake_rcu(epi);
1191 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1194 if (waitqueue_active(&ep->wq)) {
1195 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1196 !((unsigned long)key & POLLFREE)) {
1197 switch ((unsigned long)key & EPOLLINOUT_BITS) {
1199 if (epi->event.events & POLLIN)
1203 if (epi->event.events & POLLOUT)
1211 wake_up_locked(&ep->wq);
1213 if (waitqueue_active(&ep->poll_wait))
1217 spin_unlock_irqrestore(&ep->lock, flags);
1219 /* We have to call this outside the lock */
1221 ep_poll_safewake(&ep->poll_wait);
1223 if (!(epi->event.events & EPOLLEXCLUSIVE))
1226 if ((unsigned long)key & POLLFREE) {
1228 * If we race with ep_remove_wait_queue() it can miss
1229 * ->whead = NULL and do another remove_wait_queue() after
1230 * us, so we can't use __remove_wait_queue().
1232 list_del_init(&wait->entry);
1234 * ->whead != NULL protects us from the race with ep_free()
1235 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1236 * held by the caller. Once we nullify it, nothing protects
1237 * ep/epi or even wait.
1239 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1246 * This is the callback that is used to add our wait queue to the
1247 * target file wakeup lists.
1249 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1252 struct epitem *epi = ep_item_from_epqueue(pt);
1253 struct eppoll_entry *pwq;
1255 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1256 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1259 if (epi->event.events & EPOLLEXCLUSIVE)
1260 add_wait_queue_exclusive(whead, &pwq->wait);
1262 add_wait_queue(whead, &pwq->wait);
1263 list_add_tail(&pwq->llink, &epi->pwqlist);
1266 /* We have to signal that an error occurred */
1271 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1274 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1275 struct epitem *epic;
1276 bool leftmost = true;
1280 epic = rb_entry(parent, struct epitem, rbn);
1281 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1283 p = &parent->rb_right;
1286 p = &parent->rb_left;
1288 rb_link_node(&epi->rbn, parent, p);
1289 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1294 #define PATH_ARR_SIZE 5
1296 * These are the number paths of length 1 to 5, that we are allowing to emanate
1297 * from a single file of interest. For example, we allow 1000 paths of length
1298 * 1, to emanate from each file of interest. This essentially represents the
1299 * potential wakeup paths, which need to be limited in order to avoid massive
1300 * uncontrolled wakeup storms. The common use case should be a single ep which
1301 * is connected to n file sources. In this case each file source has 1 path
1302 * of length 1. Thus, the numbers below should be more than sufficient. These
1303 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1304 * and delete can't add additional paths. Protected by the epmutex.
1306 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1307 static int path_count[PATH_ARR_SIZE];
1309 static int path_count_inc(int nests)
1311 /* Allow an arbitrary number of depth 1 paths */
1315 if (++path_count[nests] > path_limits[nests])
1320 static void path_count_init(void)
1324 for (i = 0; i < PATH_ARR_SIZE; i++)
1328 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1331 struct file *file = priv;
1332 struct file *child_file;
1335 /* CTL_DEL can remove links here, but that can't increase our count */
1337 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1338 child_file = epi->ep->file;
1339 if (is_file_epoll(child_file)) {
1340 if (list_empty(&child_file->f_ep_links)) {
1341 if (path_count_inc(call_nests)) {
1346 error = ep_call_nested(&poll_loop_ncalls,
1348 reverse_path_check_proc,
1349 child_file, child_file,
1355 printk(KERN_ERR "reverse_path_check_proc: "
1356 "file is not an ep!\n");
1364 * reverse_path_check - The tfile_check_list is list of file *, which have
1365 * links that are proposed to be newly added. We need to
1366 * make sure that those added links don't add too many
1367 * paths such that we will spend all our time waking up
1368 * eventpoll objects.
1370 * Returns: Returns zero if the proposed links don't create too many paths,
1373 static int reverse_path_check(void)
1376 struct file *current_file;
1378 /* let's call this for all tfiles */
1379 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1381 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1382 reverse_path_check_proc, current_file,
1383 current_file, current);
1390 static int ep_create_wakeup_source(struct epitem *epi)
1392 struct name_snapshot n;
1393 struct wakeup_source *ws;
1396 epi->ep->ws = wakeup_source_register("eventpoll");
1401 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1402 ws = wakeup_source_register(n.name);
1403 release_dentry_name_snapshot(&n);
1407 rcu_assign_pointer(epi->ws, ws);
1412 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1413 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1415 struct wakeup_source *ws = ep_wakeup_source(epi);
1417 RCU_INIT_POINTER(epi->ws, NULL);
1420 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1421 * used internally by wakeup_source_remove, too (called by
1422 * wakeup_source_unregister), so we cannot use call_rcu
1425 wakeup_source_unregister(ws);
1429 * Must be called with "mtx" held.
1431 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1432 struct file *tfile, int fd, int full_check)
1434 int error, revents, pwake = 0;
1435 unsigned long flags;
1438 struct ep_pqueue epq;
1440 user_watches = atomic_long_read(&ep->user->epoll_watches);
1441 if (unlikely(user_watches >= max_user_watches))
1443 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1446 /* Item initialization follow here ... */
1447 INIT_LIST_HEAD(&epi->rdllink);
1448 INIT_LIST_HEAD(&epi->fllink);
1449 INIT_LIST_HEAD(&epi->pwqlist);
1451 ep_set_ffd(&epi->ffd, tfile, fd);
1452 epi->event = *event;
1454 epi->next = EP_UNACTIVE_PTR;
1455 if (epi->event.events & EPOLLWAKEUP) {
1456 error = ep_create_wakeup_source(epi);
1458 goto error_create_wakeup_source;
1460 RCU_INIT_POINTER(epi->ws, NULL);
1463 /* Add the current item to the list of active epoll hook for this file */
1464 spin_lock(&tfile->f_lock);
1465 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1466 spin_unlock(&tfile->f_lock);
1469 * Add the current item to the RB tree. All RB tree operations are
1470 * protected by "mtx", and ep_insert() is called with "mtx" held.
1472 ep_rbtree_insert(ep, epi);
1474 /* now check if we've created too many backpaths */
1476 if (full_check && reverse_path_check())
1477 goto error_remove_epi;
1479 /* Initialize the poll table using the queue callback */
1481 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1484 * Attach the item to the poll hooks and get current event bits.
1485 * We can safely use the file* here because its usage count has
1486 * been increased by the caller of this function. Note that after
1487 * this operation completes, the poll callback can start hitting
1490 revents = ep_item_poll(epi, &epq.pt);
1493 * We have to check if something went wrong during the poll wait queue
1494 * install process. Namely an allocation for a wait queue failed due
1495 * high memory pressure.
1499 goto error_unregister;
1501 /* We have to drop the new item inside our item list to keep track of it */
1502 spin_lock_irqsave(&ep->lock, flags);
1504 /* record NAPI ID of new item if present */
1505 ep_set_busy_poll_napi_id(epi);
1507 /* If the file is already "ready" we drop it inside the ready list */
1508 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1509 list_add_tail(&epi->rdllink, &ep->rdllist);
1510 ep_pm_stay_awake(epi);
1512 /* Notify waiting tasks that events are available */
1513 if (waitqueue_active(&ep->wq))
1514 wake_up_locked(&ep->wq);
1515 if (waitqueue_active(&ep->poll_wait))
1519 spin_unlock_irqrestore(&ep->lock, flags);
1521 atomic_long_inc(&ep->user->epoll_watches);
1523 /* We have to call this outside the lock */
1525 ep_poll_safewake(&ep->poll_wait);
1530 ep_unregister_pollwait(ep, epi);
1532 spin_lock(&tfile->f_lock);
1533 list_del_rcu(&epi->fllink);
1534 spin_unlock(&tfile->f_lock);
1536 rb_erase_cached(&epi->rbn, &ep->rbr);
1539 * We need to do this because an event could have been arrived on some
1540 * allocated wait queue. Note that we don't care about the ep->ovflist
1541 * list, since that is used/cleaned only inside a section bound by "mtx".
1542 * And ep_insert() is called with "mtx" held.
1544 spin_lock_irqsave(&ep->lock, flags);
1545 if (ep_is_linked(&epi->rdllink))
1546 list_del_init(&epi->rdllink);
1547 spin_unlock_irqrestore(&ep->lock, flags);
1549 wakeup_source_unregister(ep_wakeup_source(epi));
1551 error_create_wakeup_source:
1552 kmem_cache_free(epi_cache, epi);
1558 * Modify the interest event mask by dropping an event if the new mask
1559 * has a match in the current file status. Must be called with "mtx" held.
1561 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1564 unsigned int revents;
1567 init_poll_funcptr(&pt, NULL);
1570 * Set the new event interest mask before calling f_op->poll();
1571 * otherwise we might miss an event that happens between the
1572 * f_op->poll() call and the new event set registering.
1574 epi->event.events = event->events; /* need barrier below */
1575 epi->event.data = event->data; /* protected by mtx */
1576 if (epi->event.events & EPOLLWAKEUP) {
1577 if (!ep_has_wakeup_source(epi))
1578 ep_create_wakeup_source(epi);
1579 } else if (ep_has_wakeup_source(epi)) {
1580 ep_destroy_wakeup_source(epi);
1584 * The following barrier has two effects:
1586 * 1) Flush epi changes above to other CPUs. This ensures
1587 * we do not miss events from ep_poll_callback if an
1588 * event occurs immediately after we call f_op->poll().
1589 * We need this because we did not take ep->lock while
1590 * changing epi above (but ep_poll_callback does take
1593 * 2) We also need to ensure we do not miss _past_ events
1594 * when calling f_op->poll(). This barrier also
1595 * pairs with the barrier in wq_has_sleeper (see
1596 * comments for wq_has_sleeper).
1598 * This barrier will now guarantee ep_poll_callback or f_op->poll
1599 * (or both) will notice the readiness of an item.
1604 * Get current event bits. We can safely use the file* here because
1605 * its usage count has been increased by the caller of this function.
1607 revents = ep_item_poll(epi, &pt);
1610 * If the item is "hot" and it is not registered inside the ready
1611 * list, push it inside.
1613 if (revents & event->events) {
1614 spin_lock_irq(&ep->lock);
1615 if (!ep_is_linked(&epi->rdllink)) {
1616 list_add_tail(&epi->rdllink, &ep->rdllist);
1617 ep_pm_stay_awake(epi);
1619 /* Notify waiting tasks that events are available */
1620 if (waitqueue_active(&ep->wq))
1621 wake_up_locked(&ep->wq);
1622 if (waitqueue_active(&ep->poll_wait))
1625 spin_unlock_irq(&ep->lock);
1628 /* We have to call this outside the lock */
1630 ep_poll_safewake(&ep->poll_wait);
1635 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1638 struct ep_send_events_data *esed = priv;
1640 unsigned int revents;
1642 struct epoll_event __user *uevent;
1643 struct wakeup_source *ws;
1646 init_poll_funcptr(&pt, NULL);
1649 * We can loop without lock because we are passed a task private list.
1650 * Items cannot vanish during the loop because ep_scan_ready_list() is
1651 * holding "mtx" during this call.
1653 for (eventcnt = 0, uevent = esed->events;
1654 !list_empty(head) && eventcnt < esed->maxevents;) {
1655 epi = list_first_entry(head, struct epitem, rdllink);
1658 * Activate ep->ws before deactivating epi->ws to prevent
1659 * triggering auto-suspend here (in case we reactive epi->ws
1662 * This could be rearranged to delay the deactivation of epi->ws
1663 * instead, but then epi->ws would temporarily be out of sync
1664 * with ep_is_linked().
1666 ws = ep_wakeup_source(epi);
1669 __pm_stay_awake(ep->ws);
1673 list_del_init(&epi->rdllink);
1675 revents = ep_item_poll(epi, &pt);
1678 * If the event mask intersect the caller-requested one,
1679 * deliver the event to userspace. Again, ep_scan_ready_list()
1680 * is holding "mtx", so no operations coming from userspace
1681 * can change the item.
1684 if (__put_user(revents, &uevent->events) ||
1685 __put_user(epi->event.data, &uevent->data)) {
1686 list_add(&epi->rdllink, head);
1687 ep_pm_stay_awake(epi);
1688 return eventcnt ? eventcnt : -EFAULT;
1692 if (epi->event.events & EPOLLONESHOT)
1693 epi->event.events &= EP_PRIVATE_BITS;
1694 else if (!(epi->event.events & EPOLLET)) {
1696 * If this file has been added with Level
1697 * Trigger mode, we need to insert back inside
1698 * the ready list, so that the next call to
1699 * epoll_wait() will check again the events
1700 * availability. At this point, no one can insert
1701 * into ep->rdllist besides us. The epoll_ctl()
1702 * callers are locked out by
1703 * ep_scan_ready_list() holding "mtx" and the
1704 * poll callback will queue them in ep->ovflist.
1706 list_add_tail(&epi->rdllink, &ep->rdllist);
1707 ep_pm_stay_awake(epi);
1715 static int ep_send_events(struct eventpoll *ep,
1716 struct epoll_event __user *events, int maxevents)
1718 struct ep_send_events_data esed;
1720 esed.maxevents = maxevents;
1721 esed.events = events;
1723 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1726 static inline struct timespec64 ep_set_mstimeout(long ms)
1728 struct timespec64 now, ts = {
1729 .tv_sec = ms / MSEC_PER_SEC,
1730 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1733 ktime_get_ts64(&now);
1734 return timespec64_add_safe(now, ts);
1738 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1741 * @ep: Pointer to the eventpoll context.
1742 * @events: Pointer to the userspace buffer where the ready events should be
1744 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1745 * @timeout: Maximum timeout for the ready events fetch operation, in
1746 * milliseconds. If the @timeout is zero, the function will not block,
1747 * while if the @timeout is less than zero, the function will block
1748 * until at least one event has been retrieved (or an error
1751 * Returns: Returns the number of ready events which have been fetched, or an
1752 * error code, in case of error.
1754 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1755 int maxevents, long timeout)
1757 int res = 0, eavail, timed_out = 0;
1758 unsigned long flags;
1760 wait_queue_entry_t wait;
1761 ktime_t expires, *to = NULL;
1764 struct timespec64 end_time = ep_set_mstimeout(timeout);
1766 slack = select_estimate_accuracy(&end_time);
1768 *to = timespec64_to_ktime(end_time);
1769 } else if (timeout == 0) {
1771 * Avoid the unnecessary trip to the wait queue loop, if the
1772 * caller specified a non blocking operation.
1775 spin_lock_irqsave(&ep->lock, flags);
1781 if (!ep_events_available(ep))
1782 ep_busy_loop(ep, timed_out);
1784 spin_lock_irqsave(&ep->lock, flags);
1786 if (!ep_events_available(ep)) {
1788 * Busy poll timed out. Drop NAPI ID for now, we can add
1789 * it back in when we have moved a socket with a valid NAPI
1790 * ID onto the ready list.
1792 ep_reset_busy_poll_napi_id(ep);
1795 * We don't have any available event to return to the caller.
1796 * We need to sleep here, and we will be wake up by
1797 * ep_poll_callback() when events will become available.
1799 init_waitqueue_entry(&wait, current);
1800 __add_wait_queue_exclusive(&ep->wq, &wait);
1804 * We don't want to sleep if the ep_poll_callback() sends us
1805 * a wakeup in between. That's why we set the task state
1806 * to TASK_INTERRUPTIBLE before doing the checks.
1808 set_current_state(TASK_INTERRUPTIBLE);
1810 * Always short-circuit for fatal signals to allow
1811 * threads to make a timely exit without the chance of
1812 * finding more events available and fetching
1815 if (fatal_signal_pending(current)) {
1819 if (ep_events_available(ep) || timed_out)
1821 if (signal_pending(current)) {
1826 spin_unlock_irqrestore(&ep->lock, flags);
1827 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1830 spin_lock_irqsave(&ep->lock, flags);
1833 __remove_wait_queue(&ep->wq, &wait);
1834 __set_current_state(TASK_RUNNING);
1837 /* Is it worth to try to dig for events ? */
1838 eavail = ep_events_available(ep);
1840 spin_unlock_irqrestore(&ep->lock, flags);
1843 * Try to transfer events to user space. In case we get 0 events and
1844 * there's still timeout left over, we go trying again in search of
1847 if (!res && eavail &&
1848 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1855 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1856 * API, to verify that adding an epoll file inside another
1857 * epoll structure, does not violate the constraints, in
1858 * terms of closed loops, or too deep chains (which can
1859 * result in excessive stack usage).
1861 * @priv: Pointer to the epoll file to be currently checked.
1862 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1863 * data structure pointer.
1864 * @call_nests: Current dept of the @ep_call_nested() call stack.
1866 * Returns: Returns zero if adding the epoll @file inside current epoll
1867 * structure @ep does not violate the constraints, or -1 otherwise.
1869 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1872 struct file *file = priv;
1873 struct eventpoll *ep = file->private_data;
1874 struct eventpoll *ep_tovisit;
1875 struct rb_node *rbp;
1878 mutex_lock_nested(&ep->mtx, call_nests + 1);
1879 ep->gen = loop_check_gen;
1880 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1881 epi = rb_entry(rbp, struct epitem, rbn);
1882 if (unlikely(is_file_epoll(epi->ffd.file))) {
1883 ep_tovisit = epi->ffd.file->private_data;
1884 if (ep_tovisit->gen == loop_check_gen)
1886 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1887 ep_loop_check_proc, epi->ffd.file,
1888 ep_tovisit, current);
1893 * If we've reached a file that is not associated with
1894 * an ep, then we need to check if the newly added
1895 * links are going to add too many wakeup paths. We do
1896 * this by adding it to the tfile_check_list, if it's
1897 * not already there, and calling reverse_path_check()
1898 * during ep_insert().
1900 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
1901 if (get_file_rcu(epi->ffd.file))
1902 list_add(&epi->ffd.file->f_tfile_llink,
1907 mutex_unlock(&ep->mtx);
1913 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1914 * another epoll file (represented by @ep) does not create
1915 * closed loops or too deep chains.
1917 * @ep: Pointer to the epoll private data structure.
1918 * @file: Pointer to the epoll file to be checked.
1920 * Returns: Returns zero if adding the epoll @file inside current epoll
1921 * structure @ep does not violate the constraints, or -1 otherwise.
1923 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1925 return ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1926 ep_loop_check_proc, file, ep, current);
1929 static void clear_tfile_check_list(void)
1933 /* first clear the tfile_check_list */
1934 while (!list_empty(&tfile_check_list)) {
1935 file = list_first_entry(&tfile_check_list, struct file,
1937 list_del_init(&file->f_tfile_llink);
1940 INIT_LIST_HEAD(&tfile_check_list);
1944 * Open an eventpoll file descriptor.
1946 SYSCALL_DEFINE1(epoll_create1, int, flags)
1949 struct eventpoll *ep = NULL;
1952 /* Check the EPOLL_* constant for consistency. */
1953 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1955 if (flags & ~EPOLL_CLOEXEC)
1958 * Create the internal data structure ("struct eventpoll").
1960 error = ep_alloc(&ep);
1964 * Creates all the items needed to setup an eventpoll file. That is,
1965 * a file structure and a free file descriptor.
1967 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1972 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1973 O_RDWR | (flags & O_CLOEXEC));
1975 error = PTR_ERR(file);
1979 fd_install(fd, file);
1989 SYSCALL_DEFINE1(epoll_create, int, size)
1994 return sys_epoll_create1(0);
1998 * The following function implements the controller interface for
1999 * the eventpoll file that enables the insertion/removal/change of
2000 * file descriptors inside the interest set.
2002 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2003 struct epoll_event __user *, event)
2008 struct eventpoll *ep;
2010 struct epoll_event epds;
2011 struct eventpoll *tep = NULL;
2014 if (ep_op_has_event(op) &&
2015 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2023 /* Get the "struct file *" for the target file */
2028 /* The target file descriptor must support poll */
2030 if (!tf.file->f_op->poll)
2031 goto error_tgt_fput;
2033 /* Check if EPOLLWAKEUP is allowed */
2034 if (ep_op_has_event(op))
2035 ep_take_care_of_epollwakeup(&epds);
2038 * We have to check that the file structure underneath the file descriptor
2039 * the user passed to us _is_ an eventpoll file. And also we do not permit
2040 * adding an epoll file descriptor inside itself.
2043 if (f.file == tf.file || !is_file_epoll(f.file))
2044 goto error_tgt_fput;
2047 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2048 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2049 * Also, we do not currently supported nested exclusive wakeups.
2051 if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2052 if (op == EPOLL_CTL_MOD)
2053 goto error_tgt_fput;
2054 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2055 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2056 goto error_tgt_fput;
2060 * At this point it is safe to assume that the "private_data" contains
2061 * our own data structure.
2063 ep = f.file->private_data;
2066 * When we insert an epoll file descriptor, inside another epoll file
2067 * descriptor, there is the change of creating closed loops, which are
2068 * better be handled here, than in more critical paths. While we are
2069 * checking for loops we also determine the list of files reachable
2070 * and hang them on the tfile_check_list, so we can check that we
2071 * haven't created too many possible wakeup paths.
2073 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2074 * the epoll file descriptor is attaching directly to a wakeup source,
2075 * unless the epoll file descriptor is nested. The purpose of taking the
2076 * 'epmutex' on add is to prevent complex toplogies such as loops and
2077 * deep wakeup paths from forming in parallel through multiple
2078 * EPOLL_CTL_ADD operations.
2080 mutex_lock_nested(&ep->mtx, 0);
2081 if (op == EPOLL_CTL_ADD) {
2082 if (!list_empty(&f.file->f_ep_links) ||
2083 ep->gen == loop_check_gen ||
2084 is_file_epoll(tf.file)) {
2086 mutex_unlock(&ep->mtx);
2087 mutex_lock(&epmutex);
2088 if (is_file_epoll(tf.file)) {
2090 if (ep_loop_check(ep, tf.file) != 0)
2091 goto error_tgt_fput;
2094 list_add(&tf.file->f_tfile_llink,
2097 mutex_lock_nested(&ep->mtx, 0);
2098 if (is_file_epoll(tf.file)) {
2099 tep = tf.file->private_data;
2100 mutex_lock_nested(&tep->mtx, 1);
2106 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2107 * above, we can be sure to be able to use the item looked up by
2108 * ep_find() till we release the mutex.
2110 epi = ep_find(ep, tf.file, fd);
2116 epds.events |= POLLERR | POLLHUP;
2117 error = ep_insert(ep, &epds, tf.file, fd, full_check);
2123 error = ep_remove(ep, epi);
2129 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2130 epds.events |= POLLERR | POLLHUP;
2131 error = ep_modify(ep, epi, &epds);
2138 mutex_unlock(&tep->mtx);
2139 mutex_unlock(&ep->mtx);
2143 clear_tfile_check_list();
2145 mutex_unlock(&epmutex);
2157 * Implement the event wait interface for the eventpoll file. It is the kernel
2158 * part of the user space epoll_wait(2).
2160 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2161 int, maxevents, int, timeout)
2165 struct eventpoll *ep;
2167 /* The maximum number of event must be greater than zero */
2168 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2171 /* Verify that the area passed by the user is writeable */
2172 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2175 /* Get the "struct file *" for the eventpoll file */
2181 * We have to check that the file structure underneath the fd
2182 * the user passed to us _is_ an eventpoll file.
2185 if (!is_file_epoll(f.file))
2189 * At this point it is safe to assume that the "private_data" contains
2190 * our own data structure.
2192 ep = f.file->private_data;
2194 /* Time to fish for events ... */
2195 error = ep_poll(ep, events, maxevents, timeout);
2203 * Implement the event wait interface for the eventpoll file. It is the kernel
2204 * part of the user space epoll_pwait(2).
2206 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2207 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2211 sigset_t ksigmask, sigsaved;
2214 * If the caller wants a certain signal mask to be set during the wait,
2218 if (sigsetsize != sizeof(sigset_t))
2220 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2222 sigsaved = current->blocked;
2223 set_current_blocked(&ksigmask);
2226 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2229 * If we changed the signal mask, we need to restore the original one.
2230 * In case we've got a signal while waiting, we do not restore the
2231 * signal mask yet, and we allow do_signal() to deliver the signal on
2232 * the way back to userspace, before the signal mask is restored.
2235 if (error == -EINTR) {
2236 memcpy(¤t->saved_sigmask, &sigsaved,
2238 set_restore_sigmask();
2240 set_current_blocked(&sigsaved);
2246 #ifdef CONFIG_COMPAT
2247 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2248 struct epoll_event __user *, events,
2249 int, maxevents, int, timeout,
2250 const compat_sigset_t __user *, sigmask,
2251 compat_size_t, sigsetsize)
2254 compat_sigset_t csigmask;
2255 sigset_t ksigmask, sigsaved;
2258 * If the caller wants a certain signal mask to be set during the wait,
2262 if (sigsetsize != sizeof(compat_sigset_t))
2264 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2266 sigset_from_compat(&ksigmask, &csigmask);
2267 sigsaved = current->blocked;
2268 set_current_blocked(&ksigmask);
2271 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2274 * If we changed the signal mask, we need to restore the original one.
2275 * In case we've got a signal while waiting, we do not restore the
2276 * signal mask yet, and we allow do_signal() to deliver the signal on
2277 * the way back to userspace, before the signal mask is restored.
2280 if (err == -EINTR) {
2281 memcpy(¤t->saved_sigmask, &sigsaved,
2283 set_restore_sigmask();
2285 set_current_blocked(&sigsaved);
2292 static int __init eventpoll_init(void)
2298 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2300 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2302 BUG_ON(max_user_watches < 0);
2305 * Initialize the structure used to perform epoll file descriptor
2306 * inclusion loops checks.
2308 ep_nested_calls_init(&poll_loop_ncalls);
2310 /* Initialize the structure used to perform safe poll wait head wake ups */
2311 ep_nested_calls_init(&poll_safewake_ncalls);
2313 /* Initialize the structure used to perform file's f_op->poll() calls */
2314 ep_nested_calls_init(&poll_readywalk_ncalls);
2317 * We can have many thousands of epitems, so prevent this from
2318 * using an extra cache line on 64-bit (and smaller) CPUs
2320 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2322 /* Allocates slab cache used to allocate "struct epitem" items */
2323 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2324 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2326 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2327 pwq_cache = kmem_cache_create("eventpoll_pwq",
2328 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2332 fs_initcall(eventpoll_init);