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->wq.lock (spinlock)
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->wq.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->wq.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 (EPOLLIN | EPOLLOUT)
100 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
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
186 * Access to it is protected by the lock inside wq.
190 * This mutex is used to ensure that files are not removed
191 * while epoll is using them. This is held during the event
192 * collection loop, the file cleanup path, the epoll file exit
193 * code and the ctl operations.
197 /* Wait queue used by sys_epoll_wait() */
198 wait_queue_head_t wq;
200 /* Wait queue used by file->poll() */
201 wait_queue_head_t poll_wait;
203 /* List of ready file descriptors */
204 struct list_head rdllist;
206 /* RB tree root used to store monitored fd structs */
207 struct rb_root_cached rbr;
210 * This is a single linked list that chains all the "struct epitem" that
211 * happened while transferring ready events to userspace w/out
214 struct epitem *ovflist;
216 /* wakeup_source used when ep_scan_ready_list is running */
217 struct wakeup_source *ws;
219 /* The user that created the eventpoll descriptor */
220 struct user_struct *user;
224 /* used to optimize loop detection check */
227 #ifdef CONFIG_NET_RX_BUSY_POLL
228 /* used to track busy poll napi_id */
229 unsigned int napi_id;
233 /* Wait structure used by the poll hooks */
234 struct eppoll_entry {
235 /* List header used to link this structure to the "struct epitem" */
236 struct list_head llink;
238 /* The "base" pointer is set to the container "struct epitem" */
242 * Wait queue item that will be linked to the target file wait
245 wait_queue_entry_t wait;
247 /* The wait queue head that linked the "wait" wait queue item */
248 wait_queue_head_t *whead;
251 /* Wrapper struct used by poll queueing */
257 /* Used by the ep_send_events() function as callback private data */
258 struct ep_send_events_data {
260 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 /* Slab cache used to allocate "struct epitem" */
281 static struct kmem_cache *epi_cache __read_mostly;
283 /* Slab cache used to allocate "struct eppoll_entry" */
284 static struct kmem_cache *pwq_cache __read_mostly;
287 * List of files with newly added links, where we may need to limit the number
288 * of emanating paths. Protected by the epmutex.
290 static LIST_HEAD(tfile_check_list);
294 #include <linux/sysctl.h>
297 static long long_max = LONG_MAX;
299 struct ctl_table epoll_table[] = {
301 .procname = "max_user_watches",
302 .data = &max_user_watches,
303 .maxlen = sizeof(max_user_watches),
305 .proc_handler = proc_doulongvec_minmax,
311 #endif /* CONFIG_SYSCTL */
313 static const struct file_operations eventpoll_fops;
315 static inline int is_file_epoll(struct file *f)
317 return f->f_op == &eventpoll_fops;
320 /* Setup the structure that is used as key for the RB tree */
321 static inline void ep_set_ffd(struct epoll_filefd *ffd,
322 struct file *file, int fd)
328 /* Compare RB tree keys */
329 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
330 struct epoll_filefd *p2)
332 return (p1->file > p2->file ? +1:
333 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
336 /* Tells us if the item is currently linked */
337 static inline int ep_is_linked(struct epitem *epi)
339 return !list_empty(&epi->rdllink);
342 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
344 return container_of(p, struct eppoll_entry, wait);
347 /* Get the "struct epitem" from a wait queue pointer */
348 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
350 return container_of(p, struct eppoll_entry, wait)->base;
353 /* Get the "struct epitem" from an epoll queue wrapper */
354 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
356 return container_of(p, struct ep_pqueue, pt)->epi;
359 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
360 static inline int ep_op_has_event(int op)
362 return op != EPOLL_CTL_DEL;
365 /* Initialize the poll safe wake up structure */
366 static void ep_nested_calls_init(struct nested_calls *ncalls)
368 INIT_LIST_HEAD(&ncalls->tasks_call_list);
369 spin_lock_init(&ncalls->lock);
373 * ep_events_available - Checks if ready events might be available.
375 * @ep: Pointer to the eventpoll context.
377 * Returns: Returns a value different than zero if ready events are available,
380 static inline int ep_events_available(struct eventpoll *ep)
382 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
385 #ifdef CONFIG_NET_RX_BUSY_POLL
386 static bool ep_busy_loop_end(void *p, unsigned long start_time)
388 struct eventpoll *ep = p;
390 return ep_events_available(ep) || busy_loop_timeout(start_time);
394 * Busy poll if globally on and supporting sockets found && no events,
395 * busy loop will return if need_resched or ep_events_available.
397 * we must do our busy polling with irqs enabled
399 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
401 unsigned int napi_id = READ_ONCE(ep->napi_id);
403 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
404 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
407 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
414 * Set epoll busy poll NAPI ID from sk.
416 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
418 struct eventpoll *ep;
419 unsigned int napi_id;
424 if (!net_busy_loop_on())
427 sock = sock_from_file(epi->ffd.file, &err);
435 napi_id = READ_ONCE(sk->sk_napi_id);
438 /* Non-NAPI IDs can be rejected
440 * Nothing to do if we already have this ID
442 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
445 /* record NAPI ID for use in next busy poll */
446 ep->napi_id = napi_id;
451 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
455 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
459 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
463 #endif /* CONFIG_NET_RX_BUSY_POLL */
466 * ep_call_nested - Perform a bound (possibly) nested call, by checking
467 * that the recursion limit is not exceeded, and that
468 * the same nested call (by the meaning of same cookie) is
471 * @ncalls: Pointer to the nested_calls structure to be used for this call.
472 * @max_nests: Maximum number of allowed nesting calls.
473 * @nproc: Nested call core function pointer.
474 * @priv: Opaque data to be passed to the @nproc callback.
475 * @cookie: Cookie to be used to identify this nested call.
476 * @ctx: This instance context.
478 * Returns: Returns the code returned by the @nproc callback, or -1 if
479 * the maximum recursion limit has been exceeded.
481 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
482 int (*nproc)(void *, void *, int), void *priv,
483 void *cookie, void *ctx)
485 int error, call_nests = 0;
487 struct list_head *lsthead = &ncalls->tasks_call_list;
488 struct nested_call_node *tncur;
489 struct nested_call_node tnode;
491 spin_lock_irqsave(&ncalls->lock, flags);
494 * Try to see if the current task is already inside this wakeup call.
495 * We use a list here, since the population inside this set is always
498 list_for_each_entry(tncur, lsthead, llink) {
499 if (tncur->ctx == ctx &&
500 (tncur->cookie == cookie || ++call_nests > max_nests)) {
502 * Ops ... loop detected or maximum nest level reached.
503 * We abort this wake by breaking the cycle itself.
510 /* Add the current task and cookie to the list */
512 tnode.cookie = cookie;
513 list_add(&tnode.llink, lsthead);
515 spin_unlock_irqrestore(&ncalls->lock, flags);
517 /* Call the nested function */
518 error = (*nproc)(priv, cookie, call_nests);
520 /* Remove the current task from the list */
521 spin_lock_irqsave(&ncalls->lock, flags);
522 list_del(&tnode.llink);
524 spin_unlock_irqrestore(&ncalls->lock, flags);
530 * As described in commit 0ccf831cb lockdep: annotate epoll
531 * the use of wait queues used by epoll is done in a very controlled
532 * manner. Wake ups can nest inside each other, but are never done
533 * with the same locking. For example:
536 * efd1 = epoll_create();
537 * efd2 = epoll_create();
538 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
539 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
541 * When a packet arrives to the device underneath "dfd", the net code will
542 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
543 * callback wakeup entry on that queue, and the wake_up() performed by the
544 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
545 * (efd1) notices that it may have some event ready, so it needs to wake up
546 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
547 * that ends up in another wake_up(), after having checked about the
548 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
549 * avoid stack blasting.
551 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
552 * this special case of epoll.
554 #ifdef CONFIG_DEBUG_LOCK_ALLOC
556 static struct nested_calls poll_safewake_ncalls;
558 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
561 wait_queue_head_t *wqueue = (wait_queue_head_t *)cookie;
563 spin_lock_irqsave_nested(&wqueue->lock, flags, call_nests + 1);
564 wake_up_locked_poll(wqueue, EPOLLIN);
565 spin_unlock_irqrestore(&wqueue->lock, flags);
570 static void ep_poll_safewake(wait_queue_head_t *wq)
572 int this_cpu = get_cpu();
574 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
575 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
582 static void ep_poll_safewake(wait_queue_head_t *wq)
584 wake_up_poll(wq, EPOLLIN);
589 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
591 wait_queue_head_t *whead;
595 * If it is cleared by POLLFREE, it should be rcu-safe.
596 * If we read NULL we need a barrier paired with
597 * smp_store_release() in ep_poll_callback(), otherwise
598 * we rely on whead->lock.
600 whead = smp_load_acquire(&pwq->whead);
602 remove_wait_queue(whead, &pwq->wait);
607 * This function unregisters poll callbacks from the associated file
608 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
611 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
613 struct list_head *lsthead = &epi->pwqlist;
614 struct eppoll_entry *pwq;
616 while (!list_empty(lsthead)) {
617 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
619 list_del(&pwq->llink);
620 ep_remove_wait_queue(pwq);
621 kmem_cache_free(pwq_cache, pwq);
625 /* call only when ep->mtx is held */
626 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
628 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
631 /* call only when ep->mtx is held */
632 static inline void ep_pm_stay_awake(struct epitem *epi)
634 struct wakeup_source *ws = ep_wakeup_source(epi);
640 static inline bool ep_has_wakeup_source(struct epitem *epi)
642 return rcu_access_pointer(epi->ws) ? true : false;
645 /* call when ep->mtx cannot be held (ep_poll_callback) */
646 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
648 struct wakeup_source *ws;
651 ws = rcu_dereference(epi->ws);
658 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
659 * the scan code, to call f_op->poll(). Also allows for
660 * O(NumReady) performance.
662 * @ep: Pointer to the epoll private data structure.
663 * @sproc: Pointer to the scan callback.
664 * @priv: Private opaque data passed to the @sproc callback.
665 * @depth: The current depth of recursive f_op->poll calls.
666 * @ep_locked: caller already holds ep->mtx
668 * Returns: The same integer error code returned by the @sproc callback.
670 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
671 __poll_t (*sproc)(struct eventpoll *,
672 struct list_head *, void *),
673 void *priv, int depth, bool ep_locked)
677 struct epitem *epi, *nepi;
680 lockdep_assert_irqs_enabled();
683 * We need to lock this because we could be hit by
684 * eventpoll_release_file() and epoll_ctl().
688 mutex_lock_nested(&ep->mtx, depth);
691 * Steal the ready list, and re-init the original one to the
692 * empty list. Also, set ep->ovflist to NULL so that events
693 * happening while looping w/out locks, are not lost. We cannot
694 * have the poll callback to queue directly on ep->rdllist,
695 * because we want the "sproc" callback to be able to do it
698 spin_lock_irq(&ep->wq.lock);
699 list_splice_init(&ep->rdllist, &txlist);
701 spin_unlock_irq(&ep->wq.lock);
704 * Now call the callback function.
706 res = (*sproc)(ep, &txlist, priv);
708 spin_lock_irq(&ep->wq.lock);
710 * During the time we spent inside the "sproc" callback, some
711 * other events might have been queued by the poll callback.
712 * We re-insert them inside the main ready-list here.
714 for (nepi = ep->ovflist; (epi = nepi) != NULL;
715 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
717 * We need to check if the item is already in the list.
718 * During the "sproc" callback execution time, items are
719 * queued into ->ovflist but the "txlist" might already
720 * contain them, and the list_splice() below takes care of them.
722 if (!ep_is_linked(epi)) {
723 list_add_tail(&epi->rdllink, &ep->rdllist);
724 ep_pm_stay_awake(epi);
728 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
729 * releasing the lock, events will be queued in the normal way inside
732 ep->ovflist = EP_UNACTIVE_PTR;
735 * Quickly re-inject items left on "txlist".
737 list_splice(&txlist, &ep->rdllist);
740 if (!list_empty(&ep->rdllist)) {
742 * Wake up (if active) both the eventpoll wait list and
743 * the ->poll() wait list (delayed after we release the lock).
745 if (waitqueue_active(&ep->wq))
746 wake_up_locked(&ep->wq);
747 if (waitqueue_active(&ep->poll_wait))
750 spin_unlock_irq(&ep->wq.lock);
753 mutex_unlock(&ep->mtx);
755 /* We have to call this outside the lock */
757 ep_poll_safewake(&ep->poll_wait);
762 static void epi_rcu_free(struct rcu_head *head)
764 struct epitem *epi = container_of(head, struct epitem, rcu);
765 kmem_cache_free(epi_cache, epi);
769 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
770 * all the associated resources. Must be called with "mtx" held.
772 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
774 struct file *file = epi->ffd.file;
776 lockdep_assert_irqs_enabled();
779 * Removes poll wait queue hooks.
781 ep_unregister_pollwait(ep, epi);
783 /* Remove the current item from the list of epoll hooks */
784 spin_lock(&file->f_lock);
785 list_del_rcu(&epi->fllink);
786 spin_unlock(&file->f_lock);
788 rb_erase_cached(&epi->rbn, &ep->rbr);
790 spin_lock_irq(&ep->wq.lock);
791 if (ep_is_linked(epi))
792 list_del_init(&epi->rdllink);
793 spin_unlock_irq(&ep->wq.lock);
795 wakeup_source_unregister(ep_wakeup_source(epi));
797 * At this point it is safe to free the eventpoll item. Use the union
798 * field epi->rcu, since we are trying to minimize the size of
799 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
800 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
801 * use of the rbn field.
803 call_rcu(&epi->rcu, epi_rcu_free);
805 atomic_long_dec(&ep->user->epoll_watches);
810 static void ep_free(struct eventpoll *ep)
815 /* We need to release all tasks waiting for these file */
816 if (waitqueue_active(&ep->poll_wait))
817 ep_poll_safewake(&ep->poll_wait);
820 * We need to lock this because we could be hit by
821 * eventpoll_release_file() while we're freeing the "struct eventpoll".
822 * We do not need to hold "ep->mtx" here because the epoll file
823 * is on the way to be removed and no one has references to it
824 * anymore. The only hit might come from eventpoll_release_file() but
825 * holding "epmutex" is sufficient here.
827 mutex_lock(&epmutex);
830 * Walks through the whole tree by unregistering poll callbacks.
832 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
833 epi = rb_entry(rbp, struct epitem, rbn);
835 ep_unregister_pollwait(ep, epi);
840 * Walks through the whole tree by freeing each "struct epitem". At this
841 * point we are sure no poll callbacks will be lingering around, and also by
842 * holding "epmutex" we can be sure that no file cleanup code will hit
843 * us during this operation. So we can avoid the lock on "ep->wq.lock".
844 * We do not need to lock ep->mtx, either, we only do it to prevent
847 mutex_lock(&ep->mtx);
848 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
849 epi = rb_entry(rbp, struct epitem, rbn);
853 mutex_unlock(&ep->mtx);
855 mutex_unlock(&epmutex);
856 mutex_destroy(&ep->mtx);
858 wakeup_source_unregister(ep->ws);
862 static int ep_eventpoll_release(struct inode *inode, struct file *file)
864 struct eventpoll *ep = file->private_data;
872 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
874 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
878 * Differs from ep_eventpoll_poll() in that internal callers already have
879 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
880 * is correctly annotated.
882 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
885 struct eventpoll *ep;
888 pt->_key = epi->event.events;
889 if (!is_file_epoll(epi->ffd.file))
890 return vfs_poll(epi->ffd.file, pt) & epi->event.events;
892 ep = epi->ffd.file->private_data;
893 poll_wait(epi->ffd.file, &ep->poll_wait, pt);
894 locked = pt && (pt->_qproc == ep_ptable_queue_proc);
896 return ep_scan_ready_list(epi->ffd.file->private_data,
897 ep_read_events_proc, &depth, depth,
898 locked) & epi->event.events;
901 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
904 struct epitem *epi, *tmp;
906 int depth = *(int *)priv;
908 init_poll_funcptr(&pt, NULL);
911 list_for_each_entry_safe(epi, tmp, head, rdllink) {
912 if (ep_item_poll(epi, &pt, depth)) {
913 return EPOLLIN | EPOLLRDNORM;
916 * Item has been dropped into the ready list by the poll
917 * callback, but it's not actually ready, as far as
918 * caller requested events goes. We can remove it here.
920 __pm_relax(ep_wakeup_source(epi));
921 list_del_init(&epi->rdllink);
928 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
930 struct eventpoll *ep = file->private_data;
933 /* Insert inside our poll wait queue */
934 poll_wait(file, &ep->poll_wait, wait);
937 * Proceed to find out if wanted events are really available inside
940 return ep_scan_ready_list(ep, ep_read_events_proc,
941 &depth, depth, false);
944 #ifdef CONFIG_PROC_FS
945 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
947 struct eventpoll *ep = f->private_data;
950 mutex_lock(&ep->mtx);
951 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
952 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
953 struct inode *inode = file_inode(epi->ffd.file);
955 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
956 " pos:%lli ino:%lx sdev:%x\n",
957 epi->ffd.fd, epi->event.events,
958 (long long)epi->event.data,
959 (long long)epi->ffd.file->f_pos,
960 inode->i_ino, inode->i_sb->s_dev);
961 if (seq_has_overflowed(m))
964 mutex_unlock(&ep->mtx);
968 /* File callbacks that implement the eventpoll file behaviour */
969 static const struct file_operations eventpoll_fops = {
970 #ifdef CONFIG_PROC_FS
971 .show_fdinfo = ep_show_fdinfo,
973 .release = ep_eventpoll_release,
974 .poll = ep_eventpoll_poll,
975 .llseek = noop_llseek,
979 * This is called from eventpoll_release() to unlink files from the eventpoll
980 * interface. We need to have this facility to cleanup correctly files that are
981 * closed without being removed from the eventpoll interface.
983 void eventpoll_release_file(struct file *file)
985 struct eventpoll *ep;
986 struct epitem *epi, *next;
989 * We don't want to get "file->f_lock" because it is not
990 * necessary. It is not necessary because we're in the "struct file"
991 * cleanup path, and this means that no one is using this file anymore.
992 * So, for example, epoll_ctl() cannot hit here since if we reach this
993 * point, the file counter already went to zero and fget() would fail.
994 * The only hit might come from ep_free() but by holding the mutex
995 * will correctly serialize the operation. We do need to acquire
996 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
997 * from anywhere but ep_free().
999 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1001 mutex_lock(&epmutex);
1002 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1004 mutex_lock_nested(&ep->mtx, 0);
1006 mutex_unlock(&ep->mtx);
1008 mutex_unlock(&epmutex);
1011 static int ep_alloc(struct eventpoll **pep)
1014 struct user_struct *user;
1015 struct eventpoll *ep;
1017 user = get_current_user();
1019 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1023 mutex_init(&ep->mtx);
1024 init_waitqueue_head(&ep->wq);
1025 init_waitqueue_head(&ep->poll_wait);
1026 INIT_LIST_HEAD(&ep->rdllist);
1027 ep->rbr = RB_ROOT_CACHED;
1028 ep->ovflist = EP_UNACTIVE_PTR;
1041 * Search the file inside the eventpoll tree. The RB tree operations
1042 * are protected by the "mtx" mutex, and ep_find() must be called with
1045 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1048 struct rb_node *rbp;
1049 struct epitem *epi, *epir = NULL;
1050 struct epoll_filefd ffd;
1052 ep_set_ffd(&ffd, file, fd);
1053 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1054 epi = rb_entry(rbp, struct epitem, rbn);
1055 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1057 rbp = rbp->rb_right;
1069 #ifdef CONFIG_CHECKPOINT_RESTORE
1070 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1072 struct rb_node *rbp;
1075 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1076 epi = rb_entry(rbp, struct epitem, rbn);
1077 if (epi->ffd.fd == tfd) {
1089 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1092 struct file *file_raw;
1093 struct eventpoll *ep;
1096 if (!is_file_epoll(file))
1097 return ERR_PTR(-EINVAL);
1099 ep = file->private_data;
1101 mutex_lock(&ep->mtx);
1102 epi = ep_find_tfd(ep, tfd, toff);
1104 file_raw = epi->ffd.file;
1106 file_raw = ERR_PTR(-ENOENT);
1107 mutex_unlock(&ep->mtx);
1111 #endif /* CONFIG_CHECKPOINT_RESTORE */
1114 * This is the callback that is passed to the wait queue wakeup
1115 * mechanism. It is called by the stored file descriptors when they
1116 * have events to report.
1118 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1121 unsigned long flags;
1122 struct epitem *epi = ep_item_from_wait(wait);
1123 struct eventpoll *ep = epi->ep;
1124 __poll_t pollflags = key_to_poll(key);
1127 spin_lock_irqsave(&ep->wq.lock, flags);
1129 ep_set_busy_poll_napi_id(epi);
1132 * If the event mask does not contain any poll(2) event, we consider the
1133 * descriptor to be disabled. This condition is likely the effect of the
1134 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1135 * until the next EPOLL_CTL_MOD will be issued.
1137 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1141 * Check the events coming with the callback. At this stage, not
1142 * every device reports the events in the "key" parameter of the
1143 * callback. We need to be able to handle both cases here, hence the
1144 * test for "key" != NULL before the event match test.
1146 if (pollflags && !(pollflags & epi->event.events))
1150 * If we are transferring events to userspace, we can hold no locks
1151 * (because we're accessing user memory, and because of linux f_op->poll()
1152 * semantics). All the events that happen during that period of time are
1153 * chained in ep->ovflist and requeued later on.
1155 if (ep->ovflist != EP_UNACTIVE_PTR) {
1156 if (epi->next == EP_UNACTIVE_PTR) {
1157 epi->next = ep->ovflist;
1161 * Activate ep->ws since epi->ws may get
1162 * deactivated at any time.
1164 __pm_stay_awake(ep->ws);
1171 /* If this file is already in the ready list we exit soon */
1172 if (!ep_is_linked(epi)) {
1173 list_add_tail(&epi->rdllink, &ep->rdllist);
1174 ep_pm_stay_awake_rcu(epi);
1178 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1181 if (waitqueue_active(&ep->wq)) {
1182 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1183 !(pollflags & POLLFREE)) {
1184 switch (pollflags & EPOLLINOUT_BITS) {
1186 if (epi->event.events & EPOLLIN)
1190 if (epi->event.events & EPOLLOUT)
1198 wake_up_locked(&ep->wq);
1200 if (waitqueue_active(&ep->poll_wait))
1204 spin_unlock_irqrestore(&ep->wq.lock, flags);
1206 /* We have to call this outside the lock */
1208 ep_poll_safewake(&ep->poll_wait);
1210 if (!(epi->event.events & EPOLLEXCLUSIVE))
1213 if (pollflags & POLLFREE) {
1215 * If we race with ep_remove_wait_queue() it can miss
1216 * ->whead = NULL and do another remove_wait_queue() after
1217 * us, so we can't use __remove_wait_queue().
1219 list_del_init(&wait->entry);
1221 * ->whead != NULL protects us from the race with ep_free()
1222 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1223 * held by the caller. Once we nullify it, nothing protects
1224 * ep/epi or even wait.
1226 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1233 * This is the callback that is used to add our wait queue to the
1234 * target file wakeup lists.
1236 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1239 struct epitem *epi = ep_item_from_epqueue(pt);
1240 struct eppoll_entry *pwq;
1242 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1243 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1246 if (epi->event.events & EPOLLEXCLUSIVE)
1247 add_wait_queue_exclusive(whead, &pwq->wait);
1249 add_wait_queue(whead, &pwq->wait);
1250 list_add_tail(&pwq->llink, &epi->pwqlist);
1253 /* We have to signal that an error occurred */
1258 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1261 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1262 struct epitem *epic;
1263 bool leftmost = true;
1267 epic = rb_entry(parent, struct epitem, rbn);
1268 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1270 p = &parent->rb_right;
1273 p = &parent->rb_left;
1275 rb_link_node(&epi->rbn, parent, p);
1276 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1281 #define PATH_ARR_SIZE 5
1283 * These are the number paths of length 1 to 5, that we are allowing to emanate
1284 * from a single file of interest. For example, we allow 1000 paths of length
1285 * 1, to emanate from each file of interest. This essentially represents the
1286 * potential wakeup paths, which need to be limited in order to avoid massive
1287 * uncontrolled wakeup storms. The common use case should be a single ep which
1288 * is connected to n file sources. In this case each file source has 1 path
1289 * of length 1. Thus, the numbers below should be more than sufficient. These
1290 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1291 * and delete can't add additional paths. Protected by the epmutex.
1293 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1294 static int path_count[PATH_ARR_SIZE];
1296 static int path_count_inc(int nests)
1298 /* Allow an arbitrary number of depth 1 paths */
1302 if (++path_count[nests] > path_limits[nests])
1307 static void path_count_init(void)
1311 for (i = 0; i < PATH_ARR_SIZE; i++)
1315 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1318 struct file *file = priv;
1319 struct file *child_file;
1322 /* CTL_DEL can remove links here, but that can't increase our count */
1324 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1325 child_file = epi->ep->file;
1326 if (is_file_epoll(child_file)) {
1327 if (list_empty(&child_file->f_ep_links)) {
1328 if (path_count_inc(call_nests)) {
1333 error = ep_call_nested(&poll_loop_ncalls,
1335 reverse_path_check_proc,
1336 child_file, child_file,
1342 printk(KERN_ERR "reverse_path_check_proc: "
1343 "file is not an ep!\n");
1351 * reverse_path_check - The tfile_check_list is list of file *, which have
1352 * links that are proposed to be newly added. We need to
1353 * make sure that those added links don't add too many
1354 * paths such that we will spend all our time waking up
1355 * eventpoll objects.
1357 * Returns: Returns zero if the proposed links don't create too many paths,
1360 static int reverse_path_check(void)
1363 struct file *current_file;
1365 /* let's call this for all tfiles */
1366 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1368 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1369 reverse_path_check_proc, current_file,
1370 current_file, current);
1377 static int ep_create_wakeup_source(struct epitem *epi)
1379 struct name_snapshot n;
1380 struct wakeup_source *ws;
1383 epi->ep->ws = wakeup_source_register("eventpoll");
1388 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1389 ws = wakeup_source_register(n.name);
1390 release_dentry_name_snapshot(&n);
1394 rcu_assign_pointer(epi->ws, ws);
1399 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1400 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1402 struct wakeup_source *ws = ep_wakeup_source(epi);
1404 RCU_INIT_POINTER(epi->ws, NULL);
1407 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1408 * used internally by wakeup_source_remove, too (called by
1409 * wakeup_source_unregister), so we cannot use call_rcu
1412 wakeup_source_unregister(ws);
1416 * Must be called with "mtx" held.
1418 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1419 struct file *tfile, int fd, int full_check)
1421 int error, pwake = 0;
1425 struct ep_pqueue epq;
1427 lockdep_assert_irqs_enabled();
1429 user_watches = atomic_long_read(&ep->user->epoll_watches);
1430 if (unlikely(user_watches >= max_user_watches))
1432 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1435 /* Item initialization follow here ... */
1436 INIT_LIST_HEAD(&epi->rdllink);
1437 INIT_LIST_HEAD(&epi->fllink);
1438 INIT_LIST_HEAD(&epi->pwqlist);
1440 ep_set_ffd(&epi->ffd, tfile, fd);
1441 epi->event = *event;
1443 epi->next = EP_UNACTIVE_PTR;
1444 if (epi->event.events & EPOLLWAKEUP) {
1445 error = ep_create_wakeup_source(epi);
1447 goto error_create_wakeup_source;
1449 RCU_INIT_POINTER(epi->ws, NULL);
1452 /* Add the current item to the list of active epoll hook for this file */
1453 spin_lock(&tfile->f_lock);
1454 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1455 spin_unlock(&tfile->f_lock);
1458 * Add the current item to the RB tree. All RB tree operations are
1459 * protected by "mtx", and ep_insert() is called with "mtx" held.
1461 ep_rbtree_insert(ep, epi);
1463 /* now check if we've created too many backpaths */
1465 if (full_check && reverse_path_check())
1466 goto error_remove_epi;
1468 /* Initialize the poll table using the queue callback */
1470 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1473 * Attach the item to the poll hooks and get current event bits.
1474 * We can safely use the file* here because its usage count has
1475 * been increased by the caller of this function. Note that after
1476 * this operation completes, the poll callback can start hitting
1479 revents = ep_item_poll(epi, &epq.pt, 1);
1482 * We have to check if something went wrong during the poll wait queue
1483 * install process. Namely an allocation for a wait queue failed due
1484 * high memory pressure.
1488 goto error_unregister;
1490 /* We have to drop the new item inside our item list to keep track of it */
1491 spin_lock_irq(&ep->wq.lock);
1493 /* record NAPI ID of new item if present */
1494 ep_set_busy_poll_napi_id(epi);
1496 /* If the file is already "ready" we drop it inside the ready list */
1497 if (revents && !ep_is_linked(epi)) {
1498 list_add_tail(&epi->rdllink, &ep->rdllist);
1499 ep_pm_stay_awake(epi);
1501 /* Notify waiting tasks that events are available */
1502 if (waitqueue_active(&ep->wq))
1503 wake_up_locked(&ep->wq);
1504 if (waitqueue_active(&ep->poll_wait))
1508 spin_unlock_irq(&ep->wq.lock);
1510 atomic_long_inc(&ep->user->epoll_watches);
1512 /* We have to call this outside the lock */
1514 ep_poll_safewake(&ep->poll_wait);
1519 ep_unregister_pollwait(ep, epi);
1521 spin_lock(&tfile->f_lock);
1522 list_del_rcu(&epi->fllink);
1523 spin_unlock(&tfile->f_lock);
1525 rb_erase_cached(&epi->rbn, &ep->rbr);
1528 * We need to do this because an event could have been arrived on some
1529 * allocated wait queue. Note that we don't care about the ep->ovflist
1530 * list, since that is used/cleaned only inside a section bound by "mtx".
1531 * And ep_insert() is called with "mtx" held.
1533 spin_lock_irq(&ep->wq.lock);
1534 if (ep_is_linked(epi))
1535 list_del_init(&epi->rdllink);
1536 spin_unlock_irq(&ep->wq.lock);
1538 wakeup_source_unregister(ep_wakeup_source(epi));
1540 error_create_wakeup_source:
1541 kmem_cache_free(epi_cache, epi);
1547 * Modify the interest event mask by dropping an event if the new mask
1548 * has a match in the current file status. Must be called with "mtx" held.
1550 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1551 const struct epoll_event *event)
1556 lockdep_assert_irqs_enabled();
1558 init_poll_funcptr(&pt, NULL);
1561 * Set the new event interest mask before calling f_op->poll();
1562 * otherwise we might miss an event that happens between the
1563 * f_op->poll() call and the new event set registering.
1565 epi->event.events = event->events; /* need barrier below */
1566 epi->event.data = event->data; /* protected by mtx */
1567 if (epi->event.events & EPOLLWAKEUP) {
1568 if (!ep_has_wakeup_source(epi))
1569 ep_create_wakeup_source(epi);
1570 } else if (ep_has_wakeup_source(epi)) {
1571 ep_destroy_wakeup_source(epi);
1575 * The following barrier has two effects:
1577 * 1) Flush epi changes above to other CPUs. This ensures
1578 * we do not miss events from ep_poll_callback if an
1579 * event occurs immediately after we call f_op->poll().
1580 * We need this because we did not take ep->wq.lock while
1581 * changing epi above (but ep_poll_callback does take
1584 * 2) We also need to ensure we do not miss _past_ events
1585 * when calling f_op->poll(). This barrier also
1586 * pairs with the barrier in wq_has_sleeper (see
1587 * comments for wq_has_sleeper).
1589 * This barrier will now guarantee ep_poll_callback or f_op->poll
1590 * (or both) will notice the readiness of an item.
1595 * Get current event bits. We can safely use the file* here because
1596 * its usage count has been increased by the caller of this function.
1597 * If the item is "hot" and it is not registered inside the ready
1598 * list, push it inside.
1600 if (ep_item_poll(epi, &pt, 1)) {
1601 spin_lock_irq(&ep->wq.lock);
1602 if (!ep_is_linked(epi)) {
1603 list_add_tail(&epi->rdllink, &ep->rdllist);
1604 ep_pm_stay_awake(epi);
1606 /* Notify waiting tasks that events are available */
1607 if (waitqueue_active(&ep->wq))
1608 wake_up_locked(&ep->wq);
1609 if (waitqueue_active(&ep->poll_wait))
1612 spin_unlock_irq(&ep->wq.lock);
1615 /* We have to call this outside the lock */
1617 ep_poll_safewake(&ep->poll_wait);
1622 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1625 struct ep_send_events_data *esed = priv;
1628 struct epoll_event __user *uevent;
1629 struct wakeup_source *ws;
1632 init_poll_funcptr(&pt, NULL);
1635 * We can loop without lock because we are passed a task private list.
1636 * Items cannot vanish during the loop because ep_scan_ready_list() is
1637 * holding "mtx" during this call.
1639 for (esed->res = 0, uevent = esed->events;
1640 !list_empty(head) && esed->res < esed->maxevents;) {
1641 epi = list_first_entry(head, struct epitem, rdllink);
1644 * Activate ep->ws before deactivating epi->ws to prevent
1645 * triggering auto-suspend here (in case we reactive epi->ws
1648 * This could be rearranged to delay the deactivation of epi->ws
1649 * instead, but then epi->ws would temporarily be out of sync
1650 * with ep_is_linked().
1652 ws = ep_wakeup_source(epi);
1655 __pm_stay_awake(ep->ws);
1659 list_del_init(&epi->rdllink);
1661 revents = ep_item_poll(epi, &pt, 1);
1664 * If the event mask intersect the caller-requested one,
1665 * deliver the event to userspace. Again, ep_scan_ready_list()
1666 * is holding "mtx", so no operations coming from userspace
1667 * can change the item.
1670 if (__put_user(revents, &uevent->events) ||
1671 __put_user(epi->event.data, &uevent->data)) {
1672 list_add(&epi->rdllink, head);
1673 ep_pm_stay_awake(epi);
1675 esed->res = -EFAULT;
1680 if (epi->event.events & EPOLLONESHOT)
1681 epi->event.events &= EP_PRIVATE_BITS;
1682 else if (!(epi->event.events & EPOLLET)) {
1684 * If this file has been added with Level
1685 * Trigger mode, we need to insert back inside
1686 * the ready list, so that the next call to
1687 * epoll_wait() will check again the events
1688 * availability. At this point, no one can insert
1689 * into ep->rdllist besides us. The epoll_ctl()
1690 * callers are locked out by
1691 * ep_scan_ready_list() holding "mtx" and the
1692 * poll callback will queue them in ep->ovflist.
1694 list_add_tail(&epi->rdllink, &ep->rdllist);
1695 ep_pm_stay_awake(epi);
1703 static int ep_send_events(struct eventpoll *ep,
1704 struct epoll_event __user *events, int maxevents)
1706 struct ep_send_events_data esed;
1708 esed.maxevents = maxevents;
1709 esed.events = events;
1711 ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1715 static inline struct timespec64 ep_set_mstimeout(long ms)
1717 struct timespec64 now, ts = {
1718 .tv_sec = ms / MSEC_PER_SEC,
1719 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1722 ktime_get_ts64(&now);
1723 return timespec64_add_safe(now, ts);
1727 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1730 * @ep: Pointer to the eventpoll context.
1731 * @events: Pointer to the userspace buffer where the ready events should be
1733 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1734 * @timeout: Maximum timeout for the ready events fetch operation, in
1735 * milliseconds. If the @timeout is zero, the function will not block,
1736 * while if the @timeout is less than zero, the function will block
1737 * until at least one event has been retrieved (or an error
1740 * Returns: Returns the number of ready events which have been fetched, or an
1741 * error code, in case of error.
1743 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1744 int maxevents, long timeout)
1746 int res = 0, eavail, timed_out = 0;
1748 wait_queue_entry_t wait;
1749 ktime_t expires, *to = NULL;
1751 lockdep_assert_irqs_enabled();
1754 struct timespec64 end_time = ep_set_mstimeout(timeout);
1756 slack = select_estimate_accuracy(&end_time);
1758 *to = timespec64_to_ktime(end_time);
1759 } else if (timeout == 0) {
1761 * Avoid the unnecessary trip to the wait queue loop, if the
1762 * caller specified a non blocking operation.
1765 spin_lock_irq(&ep->wq.lock);
1771 if (!ep_events_available(ep))
1772 ep_busy_loop(ep, timed_out);
1774 spin_lock_irq(&ep->wq.lock);
1776 if (!ep_events_available(ep)) {
1778 * Busy poll timed out. Drop NAPI ID for now, we can add
1779 * it back in when we have moved a socket with a valid NAPI
1780 * ID onto the ready list.
1782 ep_reset_busy_poll_napi_id(ep);
1785 * We don't have any available event to return to the caller.
1786 * We need to sleep here, and we will be wake up by
1787 * ep_poll_callback() when events will become available.
1789 init_waitqueue_entry(&wait, current);
1790 __add_wait_queue_exclusive(&ep->wq, &wait);
1794 * We don't want to sleep if the ep_poll_callback() sends us
1795 * a wakeup in between. That's why we set the task state
1796 * to TASK_INTERRUPTIBLE before doing the checks.
1798 set_current_state(TASK_INTERRUPTIBLE);
1800 * Always short-circuit for fatal signals to allow
1801 * threads to make a timely exit without the chance of
1802 * finding more events available and fetching
1805 if (fatal_signal_pending(current)) {
1809 if (ep_events_available(ep) || timed_out)
1811 if (signal_pending(current)) {
1816 spin_unlock_irq(&ep->wq.lock);
1817 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1820 spin_lock_irq(&ep->wq.lock);
1823 __remove_wait_queue(&ep->wq, &wait);
1824 __set_current_state(TASK_RUNNING);
1827 /* Is it worth to try to dig for events ? */
1828 eavail = ep_events_available(ep);
1830 spin_unlock_irq(&ep->wq.lock);
1833 * Try to transfer events to user space. In case we get 0 events and
1834 * there's still timeout left over, we go trying again in search of
1837 if (!res && eavail &&
1838 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1845 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1846 * API, to verify that adding an epoll file inside another
1847 * epoll structure, does not violate the constraints, in
1848 * terms of closed loops, or too deep chains (which can
1849 * result in excessive stack usage).
1851 * @priv: Pointer to the epoll file to be currently checked.
1852 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1853 * data structure pointer.
1854 * @call_nests: Current dept of the @ep_call_nested() call stack.
1856 * Returns: Returns zero if adding the epoll @file inside current epoll
1857 * structure @ep does not violate the constraints, or -1 otherwise.
1859 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1862 struct file *file = priv;
1863 struct eventpoll *ep = file->private_data;
1864 struct eventpoll *ep_tovisit;
1865 struct rb_node *rbp;
1868 mutex_lock_nested(&ep->mtx, call_nests + 1);
1869 ep->gen = loop_check_gen;
1870 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1871 epi = rb_entry(rbp, struct epitem, rbn);
1872 if (unlikely(is_file_epoll(epi->ffd.file))) {
1873 ep_tovisit = epi->ffd.file->private_data;
1874 if (ep_tovisit->gen == loop_check_gen)
1876 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1877 ep_loop_check_proc, epi->ffd.file,
1878 ep_tovisit, current);
1883 * If we've reached a file that is not associated with
1884 * an ep, then we need to check if the newly added
1885 * links are going to add too many wakeup paths. We do
1886 * this by adding it to the tfile_check_list, if it's
1887 * not already there, and calling reverse_path_check()
1888 * during ep_insert().
1890 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
1891 if (get_file_rcu(epi->ffd.file))
1892 list_add(&epi->ffd.file->f_tfile_llink,
1897 mutex_unlock(&ep->mtx);
1903 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1904 * another epoll file (represented by @ep) does not create
1905 * closed loops or too deep chains.
1907 * @ep: Pointer to the epoll private data structure.
1908 * @file: Pointer to the epoll file to be checked.
1910 * Returns: Returns zero if adding the epoll @file inside current epoll
1911 * structure @ep does not violate the constraints, or -1 otherwise.
1913 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1915 return ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1916 ep_loop_check_proc, file, ep, current);
1919 static void clear_tfile_check_list(void)
1923 /* first clear the tfile_check_list */
1924 while (!list_empty(&tfile_check_list)) {
1925 file = list_first_entry(&tfile_check_list, struct file,
1927 list_del_init(&file->f_tfile_llink);
1930 INIT_LIST_HEAD(&tfile_check_list);
1934 * Open an eventpoll file descriptor.
1936 static int do_epoll_create(int flags)
1939 struct eventpoll *ep = NULL;
1942 /* Check the EPOLL_* constant for consistency. */
1943 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1945 if (flags & ~EPOLL_CLOEXEC)
1948 * Create the internal data structure ("struct eventpoll").
1950 error = ep_alloc(&ep);
1954 * Creates all the items needed to setup an eventpoll file. That is,
1955 * a file structure and a free file descriptor.
1957 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1962 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1963 O_RDWR | (flags & O_CLOEXEC));
1965 error = PTR_ERR(file);
1969 fd_install(fd, file);
1979 SYSCALL_DEFINE1(epoll_create1, int, flags)
1981 return do_epoll_create(flags);
1984 SYSCALL_DEFINE1(epoll_create, int, size)
1989 return do_epoll_create(0);
1993 * The following function implements the controller interface for
1994 * the eventpoll file that enables the insertion/removal/change of
1995 * file descriptors inside the interest set.
1997 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1998 struct epoll_event __user *, event)
2003 struct eventpoll *ep;
2005 struct epoll_event epds;
2006 struct eventpoll *tep = NULL;
2009 if (ep_op_has_event(op) &&
2010 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2018 /* Get the "struct file *" for the target file */
2023 /* The target file descriptor must support poll */
2025 if (!file_can_poll(tf.file))
2026 goto error_tgt_fput;
2028 /* Check if EPOLLWAKEUP is allowed */
2029 if (ep_op_has_event(op))
2030 ep_take_care_of_epollwakeup(&epds);
2033 * We have to check that the file structure underneath the file descriptor
2034 * the user passed to us _is_ an eventpoll file. And also we do not permit
2035 * adding an epoll file descriptor inside itself.
2038 if (f.file == tf.file || !is_file_epoll(f.file))
2039 goto error_tgt_fput;
2042 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2043 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2044 * Also, we do not currently supported nested exclusive wakeups.
2046 if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2047 if (op == EPOLL_CTL_MOD)
2048 goto error_tgt_fput;
2049 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2050 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2051 goto error_tgt_fput;
2055 * At this point it is safe to assume that the "private_data" contains
2056 * our own data structure.
2058 ep = f.file->private_data;
2061 * When we insert an epoll file descriptor, inside another epoll file
2062 * descriptor, there is the change of creating closed loops, which are
2063 * better be handled here, than in more critical paths. While we are
2064 * checking for loops we also determine the list of files reachable
2065 * and hang them on the tfile_check_list, so we can check that we
2066 * haven't created too many possible wakeup paths.
2068 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2069 * the epoll file descriptor is attaching directly to a wakeup source,
2070 * unless the epoll file descriptor is nested. The purpose of taking the
2071 * 'epmutex' on add is to prevent complex toplogies such as loops and
2072 * deep wakeup paths from forming in parallel through multiple
2073 * EPOLL_CTL_ADD operations.
2075 mutex_lock_nested(&ep->mtx, 0);
2076 if (op == EPOLL_CTL_ADD) {
2077 if (!list_empty(&f.file->f_ep_links) ||
2078 ep->gen == loop_check_gen ||
2079 is_file_epoll(tf.file)) {
2081 mutex_unlock(&ep->mtx);
2082 mutex_lock(&epmutex);
2083 if (is_file_epoll(tf.file)) {
2085 if (ep_loop_check(ep, tf.file) != 0)
2086 goto error_tgt_fput;
2089 list_add(&tf.file->f_tfile_llink,
2092 mutex_lock_nested(&ep->mtx, 0);
2093 if (is_file_epoll(tf.file)) {
2094 tep = tf.file->private_data;
2095 mutex_lock_nested(&tep->mtx, 1);
2101 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2102 * above, we can be sure to be able to use the item looked up by
2103 * ep_find() till we release the mutex.
2105 epi = ep_find(ep, tf.file, fd);
2111 epds.events |= EPOLLERR | EPOLLHUP;
2112 error = ep_insert(ep, &epds, tf.file, fd, full_check);
2118 error = ep_remove(ep, epi);
2124 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2125 epds.events |= EPOLLERR | EPOLLHUP;
2126 error = ep_modify(ep, epi, &epds);
2133 mutex_unlock(&tep->mtx);
2134 mutex_unlock(&ep->mtx);
2138 clear_tfile_check_list();
2140 mutex_unlock(&epmutex);
2152 * Implement the event wait interface for the eventpoll file. It is the kernel
2153 * part of the user space epoll_wait(2).
2155 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2156 int maxevents, int timeout)
2160 struct eventpoll *ep;
2162 /* The maximum number of event must be greater than zero */
2163 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2166 /* Verify that the area passed by the user is writeable */
2167 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2170 /* Get the "struct file *" for the eventpoll file */
2176 * We have to check that the file structure underneath the fd
2177 * the user passed to us _is_ an eventpoll file.
2180 if (!is_file_epoll(f.file))
2184 * At this point it is safe to assume that the "private_data" contains
2185 * our own data structure.
2187 ep = f.file->private_data;
2189 /* Time to fish for events ... */
2190 error = ep_poll(ep, events, maxevents, timeout);
2197 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2198 int, maxevents, int, timeout)
2200 return do_epoll_wait(epfd, events, maxevents, timeout);
2204 * Implement the event wait interface for the eventpoll file. It is the kernel
2205 * part of the user space epoll_pwait(2).
2207 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2208 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2212 sigset_t ksigmask, sigsaved;
2215 * If the caller wants a certain signal mask to be set during the wait,
2219 if (sigsetsize != sizeof(sigset_t))
2221 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2223 sigsaved = current->blocked;
2224 set_current_blocked(&ksigmask);
2227 error = do_epoll_wait(epfd, events, maxevents, timeout);
2230 * If we changed the signal mask, we need to restore the original one.
2231 * In case we've got a signal while waiting, we do not restore the
2232 * signal mask yet, and we allow do_signal() to deliver the signal on
2233 * the way back to userspace, before the signal mask is restored.
2236 if (error == -EINTR) {
2237 memcpy(¤t->saved_sigmask, &sigsaved,
2239 set_restore_sigmask();
2241 set_current_blocked(&sigsaved);
2247 #ifdef CONFIG_COMPAT
2248 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2249 struct epoll_event __user *, events,
2250 int, maxevents, int, timeout,
2251 const compat_sigset_t __user *, sigmask,
2252 compat_size_t, sigsetsize)
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 (get_compat_sigset(&ksigmask, sigmask))
2266 sigsaved = current->blocked;
2267 set_current_blocked(&ksigmask);
2270 err = do_epoll_wait(epfd, events, maxevents, timeout);
2273 * If we changed the signal mask, we need to restore the original one.
2274 * In case we've got a signal while waiting, we do not restore the
2275 * signal mask yet, and we allow do_signal() to deliver the signal on
2276 * the way back to userspace, before the signal mask is restored.
2279 if (err == -EINTR) {
2280 memcpy(¤t->saved_sigmask, &sigsaved,
2282 set_restore_sigmask();
2284 set_current_blocked(&sigsaved);
2291 static int __init eventpoll_init(void)
2297 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2299 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2301 BUG_ON(max_user_watches < 0);
2304 * Initialize the structure used to perform epoll file descriptor
2305 * inclusion loops checks.
2307 ep_nested_calls_init(&poll_loop_ncalls);
2309 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2310 /* Initialize the structure used to perform safe poll wait head wake ups */
2311 ep_nested_calls_init(&poll_safewake_ncalls);
2315 * We can have many thousands of epitems, so prevent this from
2316 * using an extra cache line on 64-bit (and smaller) CPUs
2318 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2320 /* Allocates slab cache used to allocate "struct epitem" items */
2321 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2322 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2324 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2325 pwq_cache = kmem_cache_create("eventpoll_pwq",
2326 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2330 fs_initcall(eventpoll_init);