1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * fs/eventpoll.c (Efficient event retrieval implementation)
4 * Copyright (C) 2001,...,2009 Davide Libenzi
6 * Davide Libenzi <davidel@xmailserver.org>
9 #include <linux/init.h>
10 #include <linux/kernel.h>
11 #include <linux/sched/signal.h>
13 #include <linux/file.h>
14 #include <linux/signal.h>
15 #include <linux/errno.h>
17 #include <linux/slab.h>
18 #include <linux/poll.h>
19 #include <linux/string.h>
20 #include <linux/list.h>
21 #include <linux/hash.h>
22 #include <linux/spinlock.h>
23 #include <linux/syscalls.h>
24 #include <linux/rbtree.h>
25 #include <linux/wait.h>
26 #include <linux/eventpoll.h>
27 #include <linux/mount.h>
28 #include <linux/bitops.h>
29 #include <linux/mutex.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/device.h>
32 #include <linux/uaccess.h>
35 #include <linux/atomic.h>
36 #include <linux/proc_fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/compat.h>
39 #include <linux/rculist.h>
40 #include <net/busy_poll.h>
44 * There are three level of locking required by epoll :
48 * 3) ep->lock (rwlock)
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a rwlock (ep->lock) because we manipulate objects
52 * from inside the poll callback, that might be triggered from
53 * a wake_up() that in turn might be called from IRQ context.
54 * So we can't sleep inside the poll callback and hence we need
55 * a spinlock. During the event transfer loop (from kernel to
56 * user space) we could end up sleeping due a copy_to_user(), so
57 * we need a lock that will allow us to sleep. This lock is a
58 * mutex (ep->mtx). It is acquired during the event transfer loop,
59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * This mutex is acquired by ep_free() during the epoll file
63 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
66 * It is also acquired when inserting an epoll fd onto another epoll
67 * fd. We do this so that we walk the epoll tree and ensure that this
68 * insertion does not create a cycle of epoll file descriptors, which
69 * could lead to deadlock. We need a global mutex to prevent two
70 * simultaneous inserts (A into B and B into A) from racing and
71 * constructing a cycle without either insert observing that it is
73 * It is necessary to acquire multiple "ep->mtx"es at once in the
74 * case when one epoll fd is added to another. In this case, we
75 * always acquire the locks in the order of nesting (i.e. after
76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77 * before e2->mtx). Since we disallow cycles of epoll file
78 * descriptors, this ensures that the mutexes are well-ordered. In
79 * order to communicate this nesting to lockdep, when walking a tree
80 * of epoll file descriptors, we use the current recursion depth as
82 * It is possible to drop the "ep->mtx" and to use the global
83 * mutex "epmutex" (together with "ep->lock") to have it working,
84 * but having "ep->mtx" will make the interface more scalable.
85 * Events that require holding "epmutex" are very rare, while for
86 * normal operations the epoll private "ep->mtx" will guarantee
87 * a better scalability.
90 /* Epoll private bits inside the event mask */
91 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
93 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
95 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
96 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EP_MAX_NESTS 4
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
103 #define EP_UNACTIVE_PTR ((void *) -1L)
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
107 struct epoll_filefd {
113 * Structure used to track possible nested calls, for too deep recursions
116 struct nested_call_node {
117 struct list_head llink;
123 * This structure is used as collector for nested calls, to check for
124 * maximum recursion dept and loop cycles.
126 struct nested_calls {
127 struct list_head tasks_call_list;
132 * Each file descriptor added to the eventpoll interface will
133 * have an entry of this type linked to the "rbr" RB tree.
134 * Avoid increasing the size of this struct, there can be many thousands
135 * of these on a server and we do not want this to take another cache line.
139 /* RB tree node links this structure to the eventpoll RB tree */
141 /* Used to free the struct epitem */
145 /* List header used to link this structure to the eventpoll ready list */
146 struct list_head rdllink;
149 * Works together "struct eventpoll"->ovflist in keeping the
150 * single linked chain of items.
154 /* The file descriptor information this item refers to */
155 struct epoll_filefd ffd;
157 /* Number of active wait queue attached to poll operations */
160 /* List containing poll wait queues */
161 struct list_head pwqlist;
163 /* The "container" of this item */
164 struct eventpoll *ep;
166 /* List header used to link this item to the "struct file" items list */
167 struct list_head fllink;
169 /* wakeup_source used when EPOLLWAKEUP is set */
170 struct wakeup_source __rcu *ws;
172 /* The structure that describe the interested events and the source fd */
173 struct epoll_event event;
177 * This structure is stored inside the "private_data" member of the file
178 * structure and represents the main data structure for the eventpoll
183 * This mutex is used to ensure that files are not removed
184 * while epoll is using them. This is held during the event
185 * collection loop, the file cleanup path, the epoll file exit
186 * code and the ctl operations.
190 /* Wait queue used by sys_epoll_wait() */
191 wait_queue_head_t wq;
193 /* Wait queue used by file->poll() */
194 wait_queue_head_t poll_wait;
196 /* List of ready file descriptors */
197 struct list_head rdllist;
199 /* Lock which protects rdllist and ovflist */
202 /* RB tree root used to store monitored fd structs */
203 struct rb_root_cached rbr;
206 * This is a single linked list that chains all the "struct epitem" that
207 * happened while transferring ready events to userspace w/out
210 struct epitem *ovflist;
212 /* wakeup_source used when ep_scan_ready_list is running */
213 struct wakeup_source *ws;
215 /* The user that created the eventpoll descriptor */
216 struct user_struct *user;
220 /* used to optimize loop detection check */
221 struct list_head visited_list_link;
224 #ifdef CONFIG_NET_RX_BUSY_POLL
225 /* used to track busy poll napi_id */
226 unsigned int napi_id;
229 #ifdef CONFIG_DEBUG_LOCK_ALLOC
230 /* tracks wakeup nests for lockdep validation */
235 /* Wait structure used by the poll hooks */
236 struct eppoll_entry {
237 /* List header used to link this structure to the "struct epitem" */
238 struct list_head llink;
240 /* The "base" pointer is set to the container "struct epitem" */
244 * Wait queue item that will be linked to the target file wait
247 wait_queue_entry_t wait;
249 /* The wait queue head that linked the "wait" wait queue item */
250 wait_queue_head_t *whead;
253 /* Wrapper struct used by poll queueing */
259 /* Used by the ep_send_events() function as callback private data */
260 struct ep_send_events_data {
262 struct epoll_event __user *events;
267 * Configuration options available inside /proc/sys/fs/epoll/
269 /* Maximum number of epoll watched descriptors, per user */
270 static long max_user_watches __read_mostly;
273 * This mutex is used to serialize ep_free() and eventpoll_release_file().
275 static DEFINE_MUTEX(epmutex);
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;
286 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
287 static LIST_HEAD(visited_list);
290 * List of files with newly added links, where we may need to limit the number
291 * of emanating paths. Protected by the epmutex.
293 static LIST_HEAD(tfile_check_list);
297 #include <linux/sysctl.h>
299 static long long_zero;
300 static long long_max = LONG_MAX;
302 struct ctl_table epoll_table[] = {
304 .procname = "max_user_watches",
305 .data = &max_user_watches,
306 .maxlen = sizeof(max_user_watches),
308 .proc_handler = proc_doulongvec_minmax,
309 .extra1 = &long_zero,
314 #endif /* CONFIG_SYSCTL */
316 static const struct file_operations eventpoll_fops;
318 static inline int is_file_epoll(struct file *f)
320 return f->f_op == &eventpoll_fops;
323 /* Setup the structure that is used as key for the RB tree */
324 static inline void ep_set_ffd(struct epoll_filefd *ffd,
325 struct file *file, int fd)
331 /* Compare RB tree keys */
332 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
333 struct epoll_filefd *p2)
335 return (p1->file > p2->file ? +1:
336 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
339 /* Tells us if the item is currently linked */
340 static inline int ep_is_linked(struct epitem *epi)
342 return !list_empty(&epi->rdllink);
345 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
347 return container_of(p, struct eppoll_entry, wait);
350 /* Get the "struct epitem" from a wait queue pointer */
351 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
353 return container_of(p, struct eppoll_entry, wait)->base;
356 /* Get the "struct epitem" from an epoll queue wrapper */
357 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
359 return container_of(p, struct ep_pqueue, pt)->epi;
362 /* Initialize the poll safe wake up structure */
363 static void ep_nested_calls_init(struct nested_calls *ncalls)
365 INIT_LIST_HEAD(&ncalls->tasks_call_list);
366 spin_lock_init(&ncalls->lock);
370 * ep_events_available - Checks if ready events might be available.
372 * @ep: Pointer to the eventpoll context.
374 * Returns: Returns a value different than zero if ready events are available,
377 static inline int ep_events_available(struct eventpoll *ep)
379 return !list_empty_careful(&ep->rdllist) ||
380 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
383 #ifdef CONFIG_NET_RX_BUSY_POLL
384 static bool ep_busy_loop_end(void *p, unsigned long start_time)
386 struct eventpoll *ep = p;
388 return ep_events_available(ep) || busy_loop_timeout(start_time);
392 * Busy poll if globally on and supporting sockets found && no events,
393 * busy loop will return if need_resched or ep_events_available.
395 * we must do our busy polling with irqs enabled
397 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
399 unsigned int napi_id = READ_ONCE(ep->napi_id);
401 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
402 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
405 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
412 * Set epoll busy poll NAPI ID from sk.
414 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
416 struct eventpoll *ep;
417 unsigned int napi_id;
422 if (!net_busy_loop_on())
425 sock = sock_from_file(epi->ffd.file, &err);
433 napi_id = READ_ONCE(sk->sk_napi_id);
436 /* Non-NAPI IDs can be rejected
438 * Nothing to do if we already have this ID
440 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
443 /* record NAPI ID for use in next busy poll */
444 ep->napi_id = napi_id;
449 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
453 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
457 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
461 #endif /* CONFIG_NET_RX_BUSY_POLL */
464 * ep_call_nested - Perform a bound (possibly) nested call, by checking
465 * that the recursion limit is not exceeded, and that
466 * the same nested call (by the meaning of same cookie) is
469 * @ncalls: Pointer to the nested_calls structure to be used for this call.
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,
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 > EP_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
553 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
555 struct eventpoll *ep_src;
560 * To set the subclass or nesting level for spin_lock_irqsave_nested()
561 * it might be natural to create a per-cpu nest count. However, since
562 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
563 * schedule() in the -rt kernel, the per-cpu variable are no longer
564 * protected. Thus, we are introducing a per eventpoll nest field.
565 * If we are not being call from ep_poll_callback(), epi is NULL and
566 * we are at the first level of nesting, 0. Otherwise, we are being
567 * called from ep_poll_callback() and if a previous wakeup source is
568 * not an epoll file itself, we are at depth 1 since the wakeup source
569 * is depth 0. If the wakeup source is a previous epoll file in the
570 * wakeup chain then we use its nests value and record ours as
571 * nests + 1. The previous epoll file nests value is stable since its
572 * already holding its own poll_wait.lock.
575 if ((is_file_epoll(epi->ffd.file))) {
576 ep_src = epi->ffd.file->private_data;
577 nests = ep_src->nests;
582 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
583 ep->nests = nests + 1;
584 wake_up_locked_poll(&ep->poll_wait, EPOLLIN);
586 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
591 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
593 wake_up_poll(&ep->poll_wait, EPOLLIN);
598 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
600 wait_queue_head_t *whead;
604 * If it is cleared by POLLFREE, it should be rcu-safe.
605 * If we read NULL we need a barrier paired with
606 * smp_store_release() in ep_poll_callback(), otherwise
607 * we rely on whead->lock.
609 whead = smp_load_acquire(&pwq->whead);
611 remove_wait_queue(whead, &pwq->wait);
616 * This function unregisters poll callbacks from the associated file
617 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
620 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
622 struct list_head *lsthead = &epi->pwqlist;
623 struct eppoll_entry *pwq;
625 while (!list_empty(lsthead)) {
626 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
628 list_del(&pwq->llink);
629 ep_remove_wait_queue(pwq);
630 kmem_cache_free(pwq_cache, pwq);
634 /* call only when ep->mtx is held */
635 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
637 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
640 /* call only when ep->mtx is held */
641 static inline void ep_pm_stay_awake(struct epitem *epi)
643 struct wakeup_source *ws = ep_wakeup_source(epi);
649 static inline bool ep_has_wakeup_source(struct epitem *epi)
651 return rcu_access_pointer(epi->ws) ? true : false;
654 /* call when ep->mtx cannot be held (ep_poll_callback) */
655 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
657 struct wakeup_source *ws;
660 ws = rcu_dereference(epi->ws);
667 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
668 * the scan code, to call f_op->poll(). Also allows for
669 * O(NumReady) performance.
671 * @ep: Pointer to the epoll private data structure.
672 * @sproc: Pointer to the scan callback.
673 * @priv: Private opaque data passed to the @sproc callback.
674 * @depth: The current depth of recursive f_op->poll calls.
675 * @ep_locked: caller already holds ep->mtx
677 * Returns: The same integer error code returned by the @sproc callback.
679 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
680 __poll_t (*sproc)(struct eventpoll *,
681 struct list_head *, void *),
682 void *priv, int depth, bool ep_locked)
685 struct epitem *epi, *nepi;
688 lockdep_assert_irqs_enabled();
691 * We need to lock this because we could be hit by
692 * eventpoll_release_file() and epoll_ctl().
696 mutex_lock_nested(&ep->mtx, depth);
699 * Steal the ready list, and re-init the original one to the
700 * empty list. Also, set ep->ovflist to NULL so that events
701 * happening while looping w/out locks, are not lost. We cannot
702 * have the poll callback to queue directly on ep->rdllist,
703 * because we want the "sproc" callback to be able to do it
706 write_lock_irq(&ep->lock);
707 list_splice_init(&ep->rdllist, &txlist);
708 WRITE_ONCE(ep->ovflist, NULL);
709 write_unlock_irq(&ep->lock);
712 * Now call the callback function.
714 res = (*sproc)(ep, &txlist, priv);
716 write_lock_irq(&ep->lock);
718 * During the time we spent inside the "sproc" callback, some
719 * other events might have been queued by the poll callback.
720 * We re-insert them inside the main ready-list here.
722 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
723 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
725 * We need to check if the item is already in the list.
726 * During the "sproc" callback execution time, items are
727 * queued into ->ovflist but the "txlist" might already
728 * contain them, and the list_splice() below takes care of them.
730 if (!ep_is_linked(epi)) {
732 * ->ovflist is LIFO, so we have to reverse it in order
735 list_add(&epi->rdllink, &ep->rdllist);
736 ep_pm_stay_awake(epi);
740 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
741 * releasing the lock, events will be queued in the normal way inside
744 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
747 * Quickly re-inject items left on "txlist".
749 list_splice(&txlist, &ep->rdllist);
751 write_unlock_irq(&ep->lock);
754 mutex_unlock(&ep->mtx);
759 static void epi_rcu_free(struct rcu_head *head)
761 struct epitem *epi = container_of(head, struct epitem, rcu);
762 kmem_cache_free(epi_cache, epi);
766 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
767 * all the associated resources. Must be called with "mtx" held.
769 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
771 struct file *file = epi->ffd.file;
773 lockdep_assert_irqs_enabled();
776 * Removes poll wait queue hooks.
778 ep_unregister_pollwait(ep, epi);
780 /* Remove the current item from the list of epoll hooks */
781 spin_lock(&file->f_lock);
782 list_del_rcu(&epi->fllink);
783 spin_unlock(&file->f_lock);
785 rb_erase_cached(&epi->rbn, &ep->rbr);
787 write_lock_irq(&ep->lock);
788 if (ep_is_linked(epi))
789 list_del_init(&epi->rdllink);
790 write_unlock_irq(&ep->lock);
792 wakeup_source_unregister(ep_wakeup_source(epi));
794 * At this point it is safe to free the eventpoll item. Use the union
795 * field epi->rcu, since we are trying to minimize the size of
796 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
797 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
798 * use of the rbn field.
800 call_rcu(&epi->rcu, epi_rcu_free);
802 atomic_long_dec(&ep->user->epoll_watches);
807 static void ep_free(struct eventpoll *ep)
812 /* We need to release all tasks waiting for these file */
813 if (waitqueue_active(&ep->poll_wait))
814 ep_poll_safewake(ep, NULL);
817 * We need to lock this because we could be hit by
818 * eventpoll_release_file() while we're freeing the "struct eventpoll".
819 * We do not need to hold "ep->mtx" here because the epoll file
820 * is on the way to be removed and no one has references to it
821 * anymore. The only hit might come from eventpoll_release_file() but
822 * holding "epmutex" is sufficient here.
824 mutex_lock(&epmutex);
827 * Walks through the whole tree by unregistering poll callbacks.
829 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
830 epi = rb_entry(rbp, struct epitem, rbn);
832 ep_unregister_pollwait(ep, epi);
837 * Walks through the whole tree by freeing each "struct epitem". At this
838 * point we are sure no poll callbacks will be lingering around, and also by
839 * holding "epmutex" we can be sure that no file cleanup code will hit
840 * us during this operation. So we can avoid the lock on "ep->lock".
841 * We do not need to lock ep->mtx, either, we only do it to prevent
844 mutex_lock(&ep->mtx);
845 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
846 epi = rb_entry(rbp, struct epitem, rbn);
850 mutex_unlock(&ep->mtx);
852 mutex_unlock(&epmutex);
853 mutex_destroy(&ep->mtx);
855 wakeup_source_unregister(ep->ws);
859 static int ep_eventpoll_release(struct inode *inode, struct file *file)
861 struct eventpoll *ep = file->private_data;
869 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
871 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
875 * Differs from ep_eventpoll_poll() in that internal callers already have
876 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
877 * is correctly annotated.
879 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
882 struct eventpoll *ep;
885 pt->_key = epi->event.events;
886 if (!is_file_epoll(epi->ffd.file))
887 return vfs_poll(epi->ffd.file, pt) & epi->event.events;
889 ep = epi->ffd.file->private_data;
890 poll_wait(epi->ffd.file, &ep->poll_wait, pt);
891 locked = pt && (pt->_qproc == ep_ptable_queue_proc);
893 return ep_scan_ready_list(epi->ffd.file->private_data,
894 ep_read_events_proc, &depth, depth,
895 locked) & epi->event.events;
898 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
901 struct epitem *epi, *tmp;
903 int depth = *(int *)priv;
905 init_poll_funcptr(&pt, NULL);
908 list_for_each_entry_safe(epi, tmp, head, rdllink) {
909 if (ep_item_poll(epi, &pt, depth)) {
910 return EPOLLIN | EPOLLRDNORM;
913 * Item has been dropped into the ready list by the poll
914 * callback, but it's not actually ready, as far as
915 * caller requested events goes. We can remove it here.
917 __pm_relax(ep_wakeup_source(epi));
918 list_del_init(&epi->rdllink);
925 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
927 struct eventpoll *ep = file->private_data;
930 /* Insert inside our poll wait queue */
931 poll_wait(file, &ep->poll_wait, wait);
934 * Proceed to find out if wanted events are really available inside
937 return ep_scan_ready_list(ep, ep_read_events_proc,
938 &depth, depth, false);
941 #ifdef CONFIG_PROC_FS
942 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
944 struct eventpoll *ep = f->private_data;
947 mutex_lock(&ep->mtx);
948 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
949 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
950 struct inode *inode = file_inode(epi->ffd.file);
952 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
953 " pos:%lli ino:%lx sdev:%x\n",
954 epi->ffd.fd, epi->event.events,
955 (long long)epi->event.data,
956 (long long)epi->ffd.file->f_pos,
957 inode->i_ino, inode->i_sb->s_dev);
958 if (seq_has_overflowed(m))
961 mutex_unlock(&ep->mtx);
965 /* File callbacks that implement the eventpoll file behaviour */
966 static const struct file_operations eventpoll_fops = {
967 #ifdef CONFIG_PROC_FS
968 .show_fdinfo = ep_show_fdinfo,
970 .release = ep_eventpoll_release,
971 .poll = ep_eventpoll_poll,
972 .llseek = noop_llseek,
976 * This is called from eventpoll_release() to unlink files from the eventpoll
977 * interface. We need to have this facility to cleanup correctly files that are
978 * closed without being removed from the eventpoll interface.
980 void eventpoll_release_file(struct file *file)
982 struct eventpoll *ep;
983 struct epitem *epi, *next;
986 * We don't want to get "file->f_lock" because it is not
987 * necessary. It is not necessary because we're in the "struct file"
988 * cleanup path, and this means that no one is using this file anymore.
989 * So, for example, epoll_ctl() cannot hit here since if we reach this
990 * point, the file counter already went to zero and fget() would fail.
991 * The only hit might come from ep_free() but by holding the mutex
992 * will correctly serialize the operation. We do need to acquire
993 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
994 * from anywhere but ep_free().
996 * Besides, ep_remove() acquires the lock, so we can't hold it here.
998 mutex_lock(&epmutex);
999 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1001 mutex_lock_nested(&ep->mtx, 0);
1003 mutex_unlock(&ep->mtx);
1005 mutex_unlock(&epmutex);
1008 static int ep_alloc(struct eventpoll **pep)
1011 struct user_struct *user;
1012 struct eventpoll *ep;
1014 user = get_current_user();
1016 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1020 mutex_init(&ep->mtx);
1021 rwlock_init(&ep->lock);
1022 init_waitqueue_head(&ep->wq);
1023 init_waitqueue_head(&ep->poll_wait);
1024 INIT_LIST_HEAD(&ep->rdllist);
1025 ep->rbr = RB_ROOT_CACHED;
1026 ep->ovflist = EP_UNACTIVE_PTR;
1039 * Search the file inside the eventpoll tree. The RB tree operations
1040 * are protected by the "mtx" mutex, and ep_find() must be called with
1043 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1046 struct rb_node *rbp;
1047 struct epitem *epi, *epir = NULL;
1048 struct epoll_filefd ffd;
1050 ep_set_ffd(&ffd, file, fd);
1051 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1052 epi = rb_entry(rbp, struct epitem, rbn);
1053 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1055 rbp = rbp->rb_right;
1067 #ifdef CONFIG_CHECKPOINT_RESTORE
1068 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1070 struct rb_node *rbp;
1073 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1074 epi = rb_entry(rbp, struct epitem, rbn);
1075 if (epi->ffd.fd == tfd) {
1087 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1090 struct file *file_raw;
1091 struct eventpoll *ep;
1094 if (!is_file_epoll(file))
1095 return ERR_PTR(-EINVAL);
1097 ep = file->private_data;
1099 mutex_lock(&ep->mtx);
1100 epi = ep_find_tfd(ep, tfd, toff);
1102 file_raw = epi->ffd.file;
1104 file_raw = ERR_PTR(-ENOENT);
1105 mutex_unlock(&ep->mtx);
1109 #endif /* CONFIG_CHECKPOINT_RESTORE */
1112 * Adds a new entry to the tail of the list in a lockless way, i.e.
1113 * multiple CPUs are allowed to call this function concurrently.
1115 * Beware: it is necessary to prevent any other modifications of the
1116 * existing list until all changes are completed, in other words
1117 * concurrent list_add_tail_lockless() calls should be protected
1118 * with a read lock, where write lock acts as a barrier which
1119 * makes sure all list_add_tail_lockless() calls are fully
1122 * Also an element can be locklessly added to the list only in one
1123 * direction i.e. either to the tail either to the head, otherwise
1124 * concurrent access will corrupt the list.
1126 * Returns %false if element has been already added to the list, %true
1129 static inline bool list_add_tail_lockless(struct list_head *new,
1130 struct list_head *head)
1132 struct list_head *prev;
1135 * This is simple 'new->next = head' operation, but cmpxchg()
1136 * is used in order to detect that same element has been just
1137 * added to the list from another CPU: the winner observes
1140 if (cmpxchg(&new->next, new, head) != new)
1144 * Initially ->next of a new element must be updated with the head
1145 * (we are inserting to the tail) and only then pointers are atomically
1146 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1147 * updated before pointers are actually swapped and pointers are
1148 * swapped before prev->next is updated.
1151 prev = xchg(&head->prev, new);
1154 * It is safe to modify prev->next and new->prev, because a new element
1155 * is added only to the tail and new->next is updated before XCHG.
1165 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1166 * i.e. multiple CPUs are allowed to call this function concurrently.
1168 * Returns %false if epi element has been already chained, %true otherwise.
1170 static inline bool chain_epi_lockless(struct epitem *epi)
1172 struct eventpoll *ep = epi->ep;
1174 /* Check that the same epi has not been just chained from another CPU */
1175 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1178 /* Atomically exchange tail */
1179 epi->next = xchg(&ep->ovflist, epi);
1185 * This is the callback that is passed to the wait queue wakeup
1186 * mechanism. It is called by the stored file descriptors when they
1187 * have events to report.
1189 * This callback takes a read lock in order not to content with concurrent
1190 * events from another file descriptors, thus all modifications to ->rdllist
1191 * or ->ovflist are lockless. Read lock is paired with the write lock from
1192 * ep_scan_ready_list(), which stops all list modifications and guarantees
1193 * that lists state is seen correctly.
1195 * Another thing worth to mention is that ep_poll_callback() can be called
1196 * concurrently for the same @epi from different CPUs if poll table was inited
1197 * with several wait queues entries. Plural wakeup from different CPUs of a
1198 * single wait queue is serialized by wq.lock, but the case when multiple wait
1199 * queues are used should be detected accordingly. This is detected using
1200 * cmpxchg() operation.
1202 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1205 struct epitem *epi = ep_item_from_wait(wait);
1206 struct eventpoll *ep = epi->ep;
1207 __poll_t pollflags = key_to_poll(key);
1208 unsigned long flags;
1211 read_lock_irqsave(&ep->lock, flags);
1213 ep_set_busy_poll_napi_id(epi);
1216 * If the event mask does not contain any poll(2) event, we consider the
1217 * descriptor to be disabled. This condition is likely the effect of the
1218 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1219 * until the next EPOLL_CTL_MOD will be issued.
1221 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1225 * Check the events coming with the callback. At this stage, not
1226 * every device reports the events in the "key" parameter of the
1227 * callback. We need to be able to handle both cases here, hence the
1228 * test for "key" != NULL before the event match test.
1230 if (pollflags && !(pollflags & epi->event.events))
1234 * If we are transferring events to userspace, we can hold no locks
1235 * (because we're accessing user memory, and because of linux f_op->poll()
1236 * semantics). All the events that happen during that period of time are
1237 * chained in ep->ovflist and requeued later on.
1239 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1240 if (epi->next == EP_UNACTIVE_PTR &&
1241 chain_epi_lockless(epi))
1242 ep_pm_stay_awake_rcu(epi);
1246 /* If this file is already in the ready list we exit soon */
1247 if (!ep_is_linked(epi) &&
1248 list_add_tail_lockless(&epi->rdllink, &ep->rdllist)) {
1249 ep_pm_stay_awake_rcu(epi);
1253 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1256 if (waitqueue_active(&ep->wq)) {
1257 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1258 !(pollflags & POLLFREE)) {
1259 switch (pollflags & EPOLLINOUT_BITS) {
1261 if (epi->event.events & EPOLLIN)
1265 if (epi->event.events & EPOLLOUT)
1275 if (waitqueue_active(&ep->poll_wait))
1279 read_unlock_irqrestore(&ep->lock, flags);
1281 /* We have to call this outside the lock */
1283 ep_poll_safewake(ep, epi);
1285 if (!(epi->event.events & EPOLLEXCLUSIVE))
1288 if (pollflags & POLLFREE) {
1290 * If we race with ep_remove_wait_queue() it can miss
1291 * ->whead = NULL and do another remove_wait_queue() after
1292 * us, so we can't use __remove_wait_queue().
1294 list_del_init(&wait->entry);
1296 * ->whead != NULL protects us from the race with ep_free()
1297 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1298 * held by the caller. Once we nullify it, nothing protects
1299 * ep/epi or even wait.
1301 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1308 * This is the callback that is used to add our wait queue to the
1309 * target file wakeup lists.
1311 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1314 struct epitem *epi = ep_item_from_epqueue(pt);
1315 struct eppoll_entry *pwq;
1317 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1318 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1321 if (epi->event.events & EPOLLEXCLUSIVE)
1322 add_wait_queue_exclusive(whead, &pwq->wait);
1324 add_wait_queue(whead, &pwq->wait);
1325 list_add_tail(&pwq->llink, &epi->pwqlist);
1328 /* We have to signal that an error occurred */
1333 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1336 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1337 struct epitem *epic;
1338 bool leftmost = true;
1342 epic = rb_entry(parent, struct epitem, rbn);
1343 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1345 p = &parent->rb_right;
1348 p = &parent->rb_left;
1350 rb_link_node(&epi->rbn, parent, p);
1351 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1356 #define PATH_ARR_SIZE 5
1358 * These are the number paths of length 1 to 5, that we are allowing to emanate
1359 * from a single file of interest. For example, we allow 1000 paths of length
1360 * 1, to emanate from each file of interest. This essentially represents the
1361 * potential wakeup paths, which need to be limited in order to avoid massive
1362 * uncontrolled wakeup storms. The common use case should be a single ep which
1363 * is connected to n file sources. In this case each file source has 1 path
1364 * of length 1. Thus, the numbers below should be more than sufficient. These
1365 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1366 * and delete can't add additional paths. Protected by the epmutex.
1368 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1369 static int path_count[PATH_ARR_SIZE];
1371 static int path_count_inc(int nests)
1373 /* Allow an arbitrary number of depth 1 paths */
1377 if (++path_count[nests] > path_limits[nests])
1382 static void path_count_init(void)
1386 for (i = 0; i < PATH_ARR_SIZE; i++)
1390 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1393 struct file *file = priv;
1394 struct file *child_file;
1397 /* CTL_DEL can remove links here, but that can't increase our count */
1399 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1400 child_file = epi->ep->file;
1401 if (is_file_epoll(child_file)) {
1402 if (list_empty(&child_file->f_ep_links)) {
1403 if (path_count_inc(call_nests)) {
1408 error = ep_call_nested(&poll_loop_ncalls,
1409 reverse_path_check_proc,
1410 child_file, child_file,
1416 printk(KERN_ERR "reverse_path_check_proc: "
1417 "file is not an ep!\n");
1425 * reverse_path_check - The tfile_check_list is list of file *, which have
1426 * links that are proposed to be newly added. We need to
1427 * make sure that those added links don't add too many
1428 * paths such that we will spend all our time waking up
1429 * eventpoll objects.
1431 * Returns: Returns zero if the proposed links don't create too many paths,
1434 static int reverse_path_check(void)
1437 struct file *current_file;
1439 /* let's call this for all tfiles */
1440 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1442 error = ep_call_nested(&poll_loop_ncalls,
1443 reverse_path_check_proc, current_file,
1444 current_file, current);
1451 static int ep_create_wakeup_source(struct epitem *epi)
1454 struct wakeup_source *ws;
1457 epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1462 name = epi->ffd.file->f_path.dentry->d_name.name;
1463 ws = wakeup_source_register(NULL, name);
1467 rcu_assign_pointer(epi->ws, ws);
1472 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1473 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1475 struct wakeup_source *ws = ep_wakeup_source(epi);
1477 RCU_INIT_POINTER(epi->ws, NULL);
1480 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1481 * used internally by wakeup_source_remove, too (called by
1482 * wakeup_source_unregister), so we cannot use call_rcu
1485 wakeup_source_unregister(ws);
1489 * Must be called with "mtx" held.
1491 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1492 struct file *tfile, int fd, int full_check)
1494 int error, pwake = 0;
1498 struct ep_pqueue epq;
1500 lockdep_assert_irqs_enabled();
1502 user_watches = atomic_long_read(&ep->user->epoll_watches);
1503 if (unlikely(user_watches >= max_user_watches))
1505 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1508 /* Item initialization follow here ... */
1509 INIT_LIST_HEAD(&epi->rdllink);
1510 INIT_LIST_HEAD(&epi->fllink);
1511 INIT_LIST_HEAD(&epi->pwqlist);
1513 ep_set_ffd(&epi->ffd, tfile, fd);
1514 epi->event = *event;
1516 epi->next = EP_UNACTIVE_PTR;
1517 if (epi->event.events & EPOLLWAKEUP) {
1518 error = ep_create_wakeup_source(epi);
1520 goto error_create_wakeup_source;
1522 RCU_INIT_POINTER(epi->ws, NULL);
1525 /* Initialize the poll table using the queue callback */
1527 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1530 * Attach the item to the poll hooks and get current event bits.
1531 * We can safely use the file* here because its usage count has
1532 * been increased by the caller of this function. Note that after
1533 * this operation completes, the poll callback can start hitting
1536 revents = ep_item_poll(epi, &epq.pt, 1);
1539 * We have to check if something went wrong during the poll wait queue
1540 * install process. Namely an allocation for a wait queue failed due
1541 * high memory pressure.
1545 goto error_unregister;
1547 /* Add the current item to the list of active epoll hook for this file */
1548 spin_lock(&tfile->f_lock);
1549 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1550 spin_unlock(&tfile->f_lock);
1553 * Add the current item to the RB tree. All RB tree operations are
1554 * protected by "mtx", and ep_insert() is called with "mtx" held.
1556 ep_rbtree_insert(ep, epi);
1558 /* now check if we've created too many backpaths */
1560 if (full_check && reverse_path_check())
1561 goto error_remove_epi;
1563 /* We have to drop the new item inside our item list to keep track of it */
1564 write_lock_irq(&ep->lock);
1566 /* record NAPI ID of new item if present */
1567 ep_set_busy_poll_napi_id(epi);
1569 /* If the file is already "ready" we drop it inside the ready list */
1570 if (revents && !ep_is_linked(epi)) {
1571 list_add_tail(&epi->rdllink, &ep->rdllist);
1572 ep_pm_stay_awake(epi);
1574 /* Notify waiting tasks that events are available */
1575 if (waitqueue_active(&ep->wq))
1577 if (waitqueue_active(&ep->poll_wait))
1581 write_unlock_irq(&ep->lock);
1583 atomic_long_inc(&ep->user->epoll_watches);
1585 /* We have to call this outside the lock */
1587 ep_poll_safewake(ep, NULL);
1592 spin_lock(&tfile->f_lock);
1593 list_del_rcu(&epi->fllink);
1594 spin_unlock(&tfile->f_lock);
1596 rb_erase_cached(&epi->rbn, &ep->rbr);
1599 ep_unregister_pollwait(ep, epi);
1602 * We need to do this because an event could have been arrived on some
1603 * allocated wait queue. Note that we don't care about the ep->ovflist
1604 * list, since that is used/cleaned only inside a section bound by "mtx".
1605 * And ep_insert() is called with "mtx" held.
1607 write_lock_irq(&ep->lock);
1608 if (ep_is_linked(epi))
1609 list_del_init(&epi->rdllink);
1610 write_unlock_irq(&ep->lock);
1612 wakeup_source_unregister(ep_wakeup_source(epi));
1614 error_create_wakeup_source:
1615 kmem_cache_free(epi_cache, epi);
1621 * Modify the interest event mask by dropping an event if the new mask
1622 * has a match in the current file status. Must be called with "mtx" held.
1624 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1625 const struct epoll_event *event)
1630 lockdep_assert_irqs_enabled();
1632 init_poll_funcptr(&pt, NULL);
1635 * Set the new event interest mask before calling f_op->poll();
1636 * otherwise we might miss an event that happens between the
1637 * f_op->poll() call and the new event set registering.
1639 epi->event.events = event->events; /* need barrier below */
1640 epi->event.data = event->data; /* protected by mtx */
1641 if (epi->event.events & EPOLLWAKEUP) {
1642 if (!ep_has_wakeup_source(epi))
1643 ep_create_wakeup_source(epi);
1644 } else if (ep_has_wakeup_source(epi)) {
1645 ep_destroy_wakeup_source(epi);
1649 * The following barrier has two effects:
1651 * 1) Flush epi changes above to other CPUs. This ensures
1652 * we do not miss events from ep_poll_callback if an
1653 * event occurs immediately after we call f_op->poll().
1654 * We need this because we did not take ep->lock while
1655 * changing epi above (but ep_poll_callback does take
1658 * 2) We also need to ensure we do not miss _past_ events
1659 * when calling f_op->poll(). This barrier also
1660 * pairs with the barrier in wq_has_sleeper (see
1661 * comments for wq_has_sleeper).
1663 * This barrier will now guarantee ep_poll_callback or f_op->poll
1664 * (or both) will notice the readiness of an item.
1669 * Get current event bits. We can safely use the file* here because
1670 * its usage count has been increased by the caller of this function.
1671 * If the item is "hot" and it is not registered inside the ready
1672 * list, push it inside.
1674 if (ep_item_poll(epi, &pt, 1)) {
1675 write_lock_irq(&ep->lock);
1676 if (!ep_is_linked(epi)) {
1677 list_add_tail(&epi->rdllink, &ep->rdllist);
1678 ep_pm_stay_awake(epi);
1680 /* Notify waiting tasks that events are available */
1681 if (waitqueue_active(&ep->wq))
1683 if (waitqueue_active(&ep->poll_wait))
1686 write_unlock_irq(&ep->lock);
1689 /* We have to call this outside the lock */
1691 ep_poll_safewake(ep, NULL);
1696 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1699 struct ep_send_events_data *esed = priv;
1701 struct epitem *epi, *tmp;
1702 struct epoll_event __user *uevent = esed->events;
1703 struct wakeup_source *ws;
1706 init_poll_funcptr(&pt, NULL);
1710 * We can loop without lock because we are passed a task private list.
1711 * Items cannot vanish during the loop because ep_scan_ready_list() is
1712 * holding "mtx" during this call.
1714 lockdep_assert_held(&ep->mtx);
1716 list_for_each_entry_safe(epi, tmp, head, rdllink) {
1717 if (esed->res >= esed->maxevents)
1721 * Activate ep->ws before deactivating epi->ws to prevent
1722 * triggering auto-suspend here (in case we reactive epi->ws
1725 * This could be rearranged to delay the deactivation of epi->ws
1726 * instead, but then epi->ws would temporarily be out of sync
1727 * with ep_is_linked().
1729 ws = ep_wakeup_source(epi);
1732 __pm_stay_awake(ep->ws);
1736 list_del_init(&epi->rdllink);
1739 * If the event mask intersect the caller-requested one,
1740 * deliver the event to userspace. Again, ep_scan_ready_list()
1741 * is holding ep->mtx, so no operations coming from userspace
1742 * can change the item.
1744 revents = ep_item_poll(epi, &pt, 1);
1748 if (__put_user(revents, &uevent->events) ||
1749 __put_user(epi->event.data, &uevent->data)) {
1750 list_add(&epi->rdllink, head);
1751 ep_pm_stay_awake(epi);
1753 esed->res = -EFAULT;
1758 if (epi->event.events & EPOLLONESHOT)
1759 epi->event.events &= EP_PRIVATE_BITS;
1760 else if (!(epi->event.events & EPOLLET)) {
1762 * If this file has been added with Level
1763 * Trigger mode, we need to insert back inside
1764 * the ready list, so that the next call to
1765 * epoll_wait() will check again the events
1766 * availability. At this point, no one can insert
1767 * into ep->rdllist besides us. The epoll_ctl()
1768 * callers are locked out by
1769 * ep_scan_ready_list() holding "mtx" and the
1770 * poll callback will queue them in ep->ovflist.
1772 list_add_tail(&epi->rdllink, &ep->rdllist);
1773 ep_pm_stay_awake(epi);
1780 static int ep_send_events(struct eventpoll *ep,
1781 struct epoll_event __user *events, int maxevents)
1783 struct ep_send_events_data esed;
1785 esed.maxevents = maxevents;
1786 esed.events = events;
1788 ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1792 static inline struct timespec64 ep_set_mstimeout(long ms)
1794 struct timespec64 now, ts = {
1795 .tv_sec = ms / MSEC_PER_SEC,
1796 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1799 ktime_get_ts64(&now);
1800 return timespec64_add_safe(now, ts);
1804 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1807 * @ep: Pointer to the eventpoll context.
1808 * @events: Pointer to the userspace buffer where the ready events should be
1810 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1811 * @timeout: Maximum timeout for the ready events fetch operation, in
1812 * milliseconds. If the @timeout is zero, the function will not block,
1813 * while if the @timeout is less than zero, the function will block
1814 * until at least one event has been retrieved (or an error
1817 * Returns: Returns the number of ready events which have been fetched, or an
1818 * error code, in case of error.
1820 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1821 int maxevents, long timeout)
1823 int res = 0, eavail, timed_out = 0;
1825 bool waiter = false;
1826 wait_queue_entry_t wait;
1827 ktime_t expires, *to = NULL;
1829 lockdep_assert_irqs_enabled();
1832 struct timespec64 end_time = ep_set_mstimeout(timeout);
1834 slack = select_estimate_accuracy(&end_time);
1836 *to = timespec64_to_ktime(end_time);
1837 } else if (timeout == 0) {
1839 * Avoid the unnecessary trip to the wait queue loop, if the
1840 * caller specified a non blocking operation. We still need
1841 * lock because we could race and not see an epi being added
1842 * to the ready list while in irq callback. Thus incorrectly
1843 * returning 0 back to userspace.
1847 write_lock_irq(&ep->lock);
1848 eavail = ep_events_available(ep);
1849 write_unlock_irq(&ep->lock);
1856 if (!ep_events_available(ep))
1857 ep_busy_loop(ep, timed_out);
1859 eavail = ep_events_available(ep);
1864 * Busy poll timed out. Drop NAPI ID for now, we can add
1865 * it back in when we have moved a socket with a valid NAPI
1866 * ID onto the ready list.
1868 ep_reset_busy_poll_napi_id(ep);
1871 * We don't have any available event to return to the caller. We need
1872 * to sleep here, and we will be woken by ep_poll_callback() when events
1877 init_waitqueue_entry(&wait, current);
1879 write_lock_irq(&ep->lock);
1880 __add_wait_queue_exclusive(&ep->wq, &wait);
1881 write_unlock_irq(&ep->lock);
1886 * We don't want to sleep if the ep_poll_callback() sends us
1887 * a wakeup in between. That's why we set the task state
1888 * to TASK_INTERRUPTIBLE before doing the checks.
1890 set_current_state(TASK_INTERRUPTIBLE);
1892 * Always short-circuit for fatal signals to allow
1893 * threads to make a timely exit without the chance of
1894 * finding more events available and fetching
1897 if (fatal_signal_pending(current)) {
1902 eavail = ep_events_available(ep);
1905 if (signal_pending(current)) {
1910 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS)) {
1916 __set_current_state(TASK_RUNNING);
1920 * Try to transfer events to user space. In case we get 0 events and
1921 * there's still timeout left over, we go trying again in search of
1924 if (!res && eavail &&
1925 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1929 write_lock_irq(&ep->lock);
1930 __remove_wait_queue(&ep->wq, &wait);
1931 write_unlock_irq(&ep->lock);
1938 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1939 * API, to verify that adding an epoll file inside another
1940 * epoll structure, does not violate the constraints, in
1941 * terms of closed loops, or too deep chains (which can
1942 * result in excessive stack usage).
1944 * @priv: Pointer to the epoll file to be currently checked.
1945 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1946 * data structure pointer.
1947 * @call_nests: Current dept of the @ep_call_nested() call stack.
1949 * Returns: Returns zero if adding the epoll @file inside current epoll
1950 * structure @ep does not violate the constraints, or -1 otherwise.
1952 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1955 struct file *file = priv;
1956 struct eventpoll *ep = file->private_data;
1957 struct eventpoll *ep_tovisit;
1958 struct rb_node *rbp;
1961 mutex_lock_nested(&ep->mtx, call_nests + 1);
1963 list_add(&ep->visited_list_link, &visited_list);
1964 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1965 epi = rb_entry(rbp, struct epitem, rbn);
1966 if (unlikely(is_file_epoll(epi->ffd.file))) {
1967 ep_tovisit = epi->ffd.file->private_data;
1968 if (ep_tovisit->visited)
1970 error = ep_call_nested(&poll_loop_ncalls,
1971 ep_loop_check_proc, epi->ffd.file,
1972 ep_tovisit, current);
1977 * If we've reached a file that is not associated with
1978 * an ep, then we need to check if the newly added
1979 * links are going to add too many wakeup paths. We do
1980 * this by adding it to the tfile_check_list, if it's
1981 * not already there, and calling reverse_path_check()
1982 * during ep_insert().
1984 if (list_empty(&epi->ffd.file->f_tfile_llink))
1985 list_add(&epi->ffd.file->f_tfile_llink,
1989 mutex_unlock(&ep->mtx);
1995 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1996 * another epoll file (represented by @ep) does not create
1997 * closed loops or too deep chains.
1999 * @ep: Pointer to the epoll private data structure.
2000 * @file: Pointer to the epoll file to be checked.
2002 * Returns: Returns zero if adding the epoll @file inside current epoll
2003 * structure @ep does not violate the constraints, or -1 otherwise.
2005 static int ep_loop_check(struct eventpoll *ep, struct file *file)
2008 struct eventpoll *ep_cur, *ep_next;
2010 ret = ep_call_nested(&poll_loop_ncalls,
2011 ep_loop_check_proc, file, ep, current);
2012 /* clear visited list */
2013 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
2014 visited_list_link) {
2015 ep_cur->visited = 0;
2016 list_del(&ep_cur->visited_list_link);
2021 static void clear_tfile_check_list(void)
2025 /* first clear the tfile_check_list */
2026 while (!list_empty(&tfile_check_list)) {
2027 file = list_first_entry(&tfile_check_list, struct file,
2029 list_del_init(&file->f_tfile_llink);
2031 INIT_LIST_HEAD(&tfile_check_list);
2035 * Open an eventpoll file descriptor.
2037 static int do_epoll_create(int flags)
2040 struct eventpoll *ep = NULL;
2043 /* Check the EPOLL_* constant for consistency. */
2044 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2046 if (flags & ~EPOLL_CLOEXEC)
2049 * Create the internal data structure ("struct eventpoll").
2051 error = ep_alloc(&ep);
2055 * Creates all the items needed to setup an eventpoll file. That is,
2056 * a file structure and a free file descriptor.
2058 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2063 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2064 O_RDWR | (flags & O_CLOEXEC));
2066 error = PTR_ERR(file);
2070 fd_install(fd, file);
2080 SYSCALL_DEFINE1(epoll_create1, int, flags)
2082 return do_epoll_create(flags);
2085 SYSCALL_DEFINE1(epoll_create, int, size)
2090 return do_epoll_create(0);
2093 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2097 mutex_lock_nested(mutex, depth);
2100 if (mutex_trylock(mutex))
2105 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2111 struct eventpoll *ep;
2113 struct eventpoll *tep = NULL;
2120 /* Get the "struct file *" for the target file */
2125 /* The target file descriptor must support poll */
2127 if (!file_can_poll(tf.file))
2128 goto error_tgt_fput;
2130 /* Check if EPOLLWAKEUP is allowed */
2131 if (ep_op_has_event(op))
2132 ep_take_care_of_epollwakeup(epds);
2135 * We have to check that the file structure underneath the file descriptor
2136 * the user passed to us _is_ an eventpoll file. And also we do not permit
2137 * adding an epoll file descriptor inside itself.
2140 if (f.file == tf.file || !is_file_epoll(f.file))
2141 goto error_tgt_fput;
2144 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2145 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2146 * Also, we do not currently supported nested exclusive wakeups.
2148 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2149 if (op == EPOLL_CTL_MOD)
2150 goto error_tgt_fput;
2151 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2152 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2153 goto error_tgt_fput;
2157 * At this point it is safe to assume that the "private_data" contains
2158 * our own data structure.
2160 ep = f.file->private_data;
2163 * When we insert an epoll file descriptor, inside another epoll file
2164 * descriptor, there is the change of creating closed loops, which are
2165 * better be handled here, than in more critical paths. While we are
2166 * checking for loops we also determine the list of files reachable
2167 * and hang them on the tfile_check_list, so we can check that we
2168 * haven't created too many possible wakeup paths.
2170 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2171 * the epoll file descriptor is attaching directly to a wakeup source,
2172 * unless the epoll file descriptor is nested. The purpose of taking the
2173 * 'epmutex' on add is to prevent complex toplogies such as loops and
2174 * deep wakeup paths from forming in parallel through multiple
2175 * EPOLL_CTL_ADD operations.
2177 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2179 goto error_tgt_fput;
2180 if (op == EPOLL_CTL_ADD) {
2181 if (!list_empty(&f.file->f_ep_links) ||
2182 is_file_epoll(tf.file)) {
2183 mutex_unlock(&ep->mtx);
2184 error = epoll_mutex_lock(&epmutex, 0, nonblock);
2186 goto error_tgt_fput;
2188 if (is_file_epoll(tf.file)) {
2190 if (ep_loop_check(ep, tf.file) != 0) {
2191 clear_tfile_check_list();
2192 goto error_tgt_fput;
2195 list_add(&tf.file->f_tfile_llink,
2197 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2200 list_del(&tf.file->f_tfile_llink);
2201 goto error_tgt_fput;
2203 if (is_file_epoll(tf.file)) {
2204 tep = tf.file->private_data;
2205 error = epoll_mutex_lock(&tep->mtx, 1, nonblock);
2207 mutex_unlock(&ep->mtx);
2215 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2216 * above, we can be sure to be able to use the item looked up by
2217 * ep_find() till we release the mutex.
2219 epi = ep_find(ep, tf.file, fd);
2225 epds->events |= EPOLLERR | EPOLLHUP;
2226 error = ep_insert(ep, epds, tf.file, fd, full_check);
2230 clear_tfile_check_list();
2234 error = ep_remove(ep, epi);
2240 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2241 epds->events |= EPOLLERR | EPOLLHUP;
2242 error = ep_modify(ep, epi, epds);
2249 mutex_unlock(&tep->mtx);
2250 mutex_unlock(&ep->mtx);
2254 mutex_unlock(&epmutex);
2265 * The following function implements the controller interface for
2266 * the eventpoll file that enables the insertion/removal/change of
2267 * file descriptors inside the interest set.
2269 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2270 struct epoll_event __user *, event)
2272 struct epoll_event epds;
2274 if (ep_op_has_event(op) &&
2275 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2278 return do_epoll_ctl(epfd, op, fd, &epds, false);
2282 * Implement the event wait interface for the eventpoll file. It is the kernel
2283 * part of the user space epoll_wait(2).
2285 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2286 int maxevents, int timeout)
2290 struct eventpoll *ep;
2292 /* The maximum number of event must be greater than zero */
2293 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2296 /* Verify that the area passed by the user is writeable */
2297 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2300 /* Get the "struct file *" for the eventpoll file */
2306 * We have to check that the file structure underneath the fd
2307 * the user passed to us _is_ an eventpoll file.
2310 if (!is_file_epoll(f.file))
2314 * At this point it is safe to assume that the "private_data" contains
2315 * our own data structure.
2317 ep = f.file->private_data;
2319 /* Time to fish for events ... */
2320 error = ep_poll(ep, events, maxevents, timeout);
2327 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2328 int, maxevents, int, timeout)
2330 return do_epoll_wait(epfd, events, maxevents, timeout);
2334 * Implement the event wait interface for the eventpoll file. It is the kernel
2335 * part of the user space epoll_pwait(2).
2337 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2338 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2344 * If the caller wants a certain signal mask to be set during the wait,
2347 error = set_user_sigmask(sigmask, sigsetsize);
2351 error = do_epoll_wait(epfd, events, maxevents, timeout);
2352 restore_saved_sigmask_unless(error == -EINTR);
2357 #ifdef CONFIG_COMPAT
2358 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2359 struct epoll_event __user *, events,
2360 int, maxevents, int, timeout,
2361 const compat_sigset_t __user *, sigmask,
2362 compat_size_t, sigsetsize)
2367 * If the caller wants a certain signal mask to be set during the wait,
2370 err = set_compat_user_sigmask(sigmask, sigsetsize);
2374 err = do_epoll_wait(epfd, events, maxevents, timeout);
2375 restore_saved_sigmask_unless(err == -EINTR);
2381 static int __init eventpoll_init(void)
2387 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2389 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2391 BUG_ON(max_user_watches < 0);
2394 * Initialize the structure used to perform epoll file descriptor
2395 * inclusion loops checks.
2397 ep_nested_calls_init(&poll_loop_ncalls);
2400 * We can have many thousands of epitems, so prevent this from
2401 * using an extra cache line on 64-bit (and smaller) CPUs
2403 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2405 /* Allocates slab cache used to allocate "struct epitem" items */
2406 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2407 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2409 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2410 pwq_cache = kmem_cache_create("eventpoll_pwq",
2411 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2415 fs_initcall(eventpoll_init);