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 */
223 #ifdef CONFIG_NET_RX_BUSY_POLL
224 /* used to track busy poll napi_id */
225 unsigned int napi_id;
228 #ifdef CONFIG_DEBUG_LOCK_ALLOC
229 /* tracks wakeup nests for lockdep validation */
234 /* Wait structure used by the poll hooks */
235 struct eppoll_entry {
236 /* List header used to link this structure to the "struct epitem" */
237 struct list_head llink;
239 /* The "base" pointer is set to the container "struct epitem" */
243 * Wait queue item that will be linked to the target file wait
246 wait_queue_entry_t wait;
248 /* The wait queue head that linked the "wait" wait queue item */
249 wait_queue_head_t *whead;
252 /* Wrapper struct used by poll queueing */
258 /* Used by the ep_send_events() function as callback private data */
259 struct ep_send_events_data {
261 struct epoll_event __user *events;
266 * Configuration options available inside /proc/sys/fs/epoll/
268 /* Maximum number of epoll watched descriptors, per user */
269 static long max_user_watches __read_mostly;
272 * This mutex is used to serialize ep_free() and eventpoll_release_file().
274 static DEFINE_MUTEX(epmutex);
276 static u64 loop_check_gen = 0;
278 /* Used to check for epoll file descriptor inclusion loops */
279 static struct nested_calls poll_loop_ncalls;
281 /* Slab cache used to allocate "struct epitem" */
282 static struct kmem_cache *epi_cache __read_mostly;
284 /* Slab cache used to allocate "struct eppoll_entry" */
285 static struct kmem_cache *pwq_cache __read_mostly;
288 * List of files with newly added links, where we may need to limit the number
289 * of emanating paths. Protected by the epmutex.
291 static LIST_HEAD(tfile_check_list);
295 #include <linux/sysctl.h>
297 static long long_zero;
298 static long long_max = LONG_MAX;
300 struct ctl_table epoll_table[] = {
302 .procname = "max_user_watches",
303 .data = &max_user_watches,
304 .maxlen = sizeof(max_user_watches),
306 .proc_handler = proc_doulongvec_minmax,
307 .extra1 = &long_zero,
312 #endif /* CONFIG_SYSCTL */
314 static const struct file_operations eventpoll_fops;
316 static inline int is_file_epoll(struct file *f)
318 return f->f_op == &eventpoll_fops;
321 /* Setup the structure that is used as key for the RB tree */
322 static inline void ep_set_ffd(struct epoll_filefd *ffd,
323 struct file *file, int fd)
329 /* Compare RB tree keys */
330 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
331 struct epoll_filefd *p2)
333 return (p1->file > p2->file ? +1:
334 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
337 /* Tells us if the item is currently linked */
338 static inline int ep_is_linked(struct epitem *epi)
340 return !list_empty(&epi->rdllink);
343 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
345 return container_of(p, struct eppoll_entry, wait);
348 /* Get the "struct epitem" from a wait queue pointer */
349 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
351 return container_of(p, struct eppoll_entry, wait)->base;
354 /* Get the "struct epitem" from an epoll queue wrapper */
355 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
357 return container_of(p, struct ep_pqueue, pt)->epi;
360 /* Initialize the poll safe wake up structure */
361 static void ep_nested_calls_init(struct nested_calls *ncalls)
363 INIT_LIST_HEAD(&ncalls->tasks_call_list);
364 spin_lock_init(&ncalls->lock);
368 * ep_events_available - Checks if ready events might be available.
370 * @ep: Pointer to the eventpoll context.
372 * Returns: Returns a value different than zero if ready events are available,
375 static inline int ep_events_available(struct eventpoll *ep)
377 return !list_empty_careful(&ep->rdllist) ||
378 READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
381 #ifdef CONFIG_NET_RX_BUSY_POLL
382 static bool ep_busy_loop_end(void *p, unsigned long start_time)
384 struct eventpoll *ep = p;
386 return ep_events_available(ep) || busy_loop_timeout(start_time);
390 * Busy poll if globally on and supporting sockets found && no events,
391 * busy loop will return if need_resched or ep_events_available.
393 * we must do our busy polling with irqs enabled
395 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
397 unsigned int napi_id = READ_ONCE(ep->napi_id);
399 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
400 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
403 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
410 * Set epoll busy poll NAPI ID from sk.
412 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
414 struct eventpoll *ep;
415 unsigned int napi_id;
420 if (!net_busy_loop_on())
423 sock = sock_from_file(epi->ffd.file, &err);
431 napi_id = READ_ONCE(sk->sk_napi_id);
434 /* Non-NAPI IDs can be rejected
436 * Nothing to do if we already have this ID
438 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
441 /* record NAPI ID for use in next busy poll */
442 ep->napi_id = napi_id;
447 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
451 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
455 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
459 #endif /* CONFIG_NET_RX_BUSY_POLL */
462 * ep_call_nested - Perform a bound (possibly) nested call, by checking
463 * that the recursion limit is not exceeded, and that
464 * the same nested call (by the meaning of same cookie) is
467 * @ncalls: Pointer to the nested_calls structure to be used for this call.
468 * @nproc: Nested call core function pointer.
469 * @priv: Opaque data to be passed to the @nproc callback.
470 * @cookie: Cookie to be used to identify this nested call.
471 * @ctx: This instance context.
473 * Returns: Returns the code returned by the @nproc callback, or -1 if
474 * the maximum recursion limit has been exceeded.
476 static int ep_call_nested(struct nested_calls *ncalls,
477 int (*nproc)(void *, void *, int), void *priv,
478 void *cookie, void *ctx)
480 int error, call_nests = 0;
482 struct list_head *lsthead = &ncalls->tasks_call_list;
483 struct nested_call_node *tncur;
484 struct nested_call_node tnode;
486 spin_lock_irqsave(&ncalls->lock, flags);
489 * Try to see if the current task is already inside this wakeup call.
490 * We use a list here, since the population inside this set is always
493 list_for_each_entry(tncur, lsthead, llink) {
494 if (tncur->ctx == ctx &&
495 (tncur->cookie == cookie || ++call_nests > EP_MAX_NESTS)) {
497 * Ops ... loop detected or maximum nest level reached.
498 * We abort this wake by breaking the cycle itself.
505 /* Add the current task and cookie to the list */
507 tnode.cookie = cookie;
508 list_add(&tnode.llink, lsthead);
510 spin_unlock_irqrestore(&ncalls->lock, flags);
512 /* Call the nested function */
513 error = (*nproc)(priv, cookie, call_nests);
515 /* Remove the current task from the list */
516 spin_lock_irqsave(&ncalls->lock, flags);
517 list_del(&tnode.llink);
519 spin_unlock_irqrestore(&ncalls->lock, flags);
525 * As described in commit 0ccf831cb lockdep: annotate epoll
526 * the use of wait queues used by epoll is done in a very controlled
527 * manner. Wake ups can nest inside each other, but are never done
528 * with the same locking. For example:
531 * efd1 = epoll_create();
532 * efd2 = epoll_create();
533 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
534 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
536 * When a packet arrives to the device underneath "dfd", the net code will
537 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
538 * callback wakeup entry on that queue, and the wake_up() performed by the
539 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
540 * (efd1) notices that it may have some event ready, so it needs to wake up
541 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
542 * that ends up in another wake_up(), after having checked about the
543 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
544 * avoid stack blasting.
546 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
547 * this special case of epoll.
549 #ifdef CONFIG_DEBUG_LOCK_ALLOC
551 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
553 struct eventpoll *ep_src;
558 * To set the subclass or nesting level for spin_lock_irqsave_nested()
559 * it might be natural to create a per-cpu nest count. However, since
560 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
561 * schedule() in the -rt kernel, the per-cpu variable are no longer
562 * protected. Thus, we are introducing a per eventpoll nest field.
563 * If we are not being call from ep_poll_callback(), epi is NULL and
564 * we are at the first level of nesting, 0. Otherwise, we are being
565 * called from ep_poll_callback() and if a previous wakeup source is
566 * not an epoll file itself, we are at depth 1 since the wakeup source
567 * is depth 0. If the wakeup source is a previous epoll file in the
568 * wakeup chain then we use its nests value and record ours as
569 * nests + 1. The previous epoll file nests value is stable since its
570 * already holding its own poll_wait.lock.
573 if ((is_file_epoll(epi->ffd.file))) {
574 ep_src = epi->ffd.file->private_data;
575 nests = ep_src->nests;
580 spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
581 ep->nests = nests + 1;
582 wake_up_locked_poll(&ep->poll_wait, EPOLLIN);
584 spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
589 static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi)
591 wake_up_poll(&ep->poll_wait, EPOLLIN);
596 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
598 wait_queue_head_t *whead;
602 * If it is cleared by POLLFREE, it should be rcu-safe.
603 * If we read NULL we need a barrier paired with
604 * smp_store_release() in ep_poll_callback(), otherwise
605 * we rely on whead->lock.
607 whead = smp_load_acquire(&pwq->whead);
609 remove_wait_queue(whead, &pwq->wait);
614 * This function unregisters poll callbacks from the associated file
615 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
618 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
620 struct list_head *lsthead = &epi->pwqlist;
621 struct eppoll_entry *pwq;
623 while (!list_empty(lsthead)) {
624 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
626 list_del(&pwq->llink);
627 ep_remove_wait_queue(pwq);
628 kmem_cache_free(pwq_cache, pwq);
632 /* call only when ep->mtx is held */
633 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
635 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
638 /* call only when ep->mtx is held */
639 static inline void ep_pm_stay_awake(struct epitem *epi)
641 struct wakeup_source *ws = ep_wakeup_source(epi);
647 static inline bool ep_has_wakeup_source(struct epitem *epi)
649 return rcu_access_pointer(epi->ws) ? true : false;
652 /* call when ep->mtx cannot be held (ep_poll_callback) */
653 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
655 struct wakeup_source *ws;
658 ws = rcu_dereference(epi->ws);
665 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
666 * the scan code, to call f_op->poll(). Also allows for
667 * O(NumReady) performance.
669 * @ep: Pointer to the epoll private data structure.
670 * @sproc: Pointer to the scan callback.
671 * @priv: Private opaque data passed to the @sproc callback.
672 * @depth: The current depth of recursive f_op->poll calls.
673 * @ep_locked: caller already holds ep->mtx
675 * Returns: The same integer error code returned by the @sproc callback.
677 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
678 __poll_t (*sproc)(struct eventpoll *,
679 struct list_head *, void *),
680 void *priv, int depth, bool ep_locked)
683 struct epitem *epi, *nepi;
686 lockdep_assert_irqs_enabled();
689 * We need to lock this because we could be hit by
690 * eventpoll_release_file() and epoll_ctl().
694 mutex_lock_nested(&ep->mtx, depth);
697 * Steal the ready list, and re-init the original one to the
698 * empty list. Also, set ep->ovflist to NULL so that events
699 * happening while looping w/out locks, are not lost. We cannot
700 * have the poll callback to queue directly on ep->rdllist,
701 * because we want the "sproc" callback to be able to do it
704 write_lock_irq(&ep->lock);
705 list_splice_init(&ep->rdllist, &txlist);
706 WRITE_ONCE(ep->ovflist, NULL);
707 write_unlock_irq(&ep->lock);
710 * Now call the callback function.
712 res = (*sproc)(ep, &txlist, priv);
714 write_lock_irq(&ep->lock);
716 * During the time we spent inside the "sproc" callback, some
717 * other events might have been queued by the poll callback.
718 * We re-insert them inside the main ready-list here.
720 for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;
721 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
723 * We need to check if the item is already in the list.
724 * During the "sproc" callback execution time, items are
725 * queued into ->ovflist but the "txlist" might already
726 * contain them, and the list_splice() below takes care of them.
728 if (!ep_is_linked(epi)) {
730 * ->ovflist is LIFO, so we have to reverse it in order
733 list_add(&epi->rdllink, &ep->rdllist);
734 ep_pm_stay_awake(epi);
738 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
739 * releasing the lock, events will be queued in the normal way inside
742 WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
745 * Quickly re-inject items left on "txlist".
747 list_splice(&txlist, &ep->rdllist);
750 if (!list_empty(&ep->rdllist)) {
751 if (waitqueue_active(&ep->wq))
755 write_unlock_irq(&ep->lock);
758 mutex_unlock(&ep->mtx);
763 static void epi_rcu_free(struct rcu_head *head)
765 struct epitem *epi = container_of(head, struct epitem, rcu);
766 kmem_cache_free(epi_cache, epi);
770 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
771 * all the associated resources. Must be called with "mtx" held.
773 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
775 struct file *file = epi->ffd.file;
777 lockdep_assert_irqs_enabled();
780 * Removes poll wait queue hooks.
782 ep_unregister_pollwait(ep, epi);
784 /* Remove the current item from the list of epoll hooks */
785 spin_lock(&file->f_lock);
786 list_del_rcu(&epi->fllink);
787 spin_unlock(&file->f_lock);
789 rb_erase_cached(&epi->rbn, &ep->rbr);
791 write_lock_irq(&ep->lock);
792 if (ep_is_linked(epi))
793 list_del_init(&epi->rdllink);
794 write_unlock_irq(&ep->lock);
796 wakeup_source_unregister(ep_wakeup_source(epi));
798 * At this point it is safe to free the eventpoll item. Use the union
799 * field epi->rcu, since we are trying to minimize the size of
800 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
801 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
802 * use of the rbn field.
804 call_rcu(&epi->rcu, epi_rcu_free);
806 atomic_long_dec(&ep->user->epoll_watches);
811 static void ep_free(struct eventpoll *ep)
816 /* We need to release all tasks waiting for these file */
817 if (waitqueue_active(&ep->poll_wait))
818 ep_poll_safewake(ep, NULL);
821 * We need to lock this because we could be hit by
822 * eventpoll_release_file() while we're freeing the "struct eventpoll".
823 * We do not need to hold "ep->mtx" here because the epoll file
824 * is on the way to be removed and no one has references to it
825 * anymore. The only hit might come from eventpoll_release_file() but
826 * holding "epmutex" is sufficient here.
828 mutex_lock(&epmutex);
831 * Walks through the whole tree by unregistering poll callbacks.
833 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
834 epi = rb_entry(rbp, struct epitem, rbn);
836 ep_unregister_pollwait(ep, epi);
841 * Walks through the whole tree by freeing each "struct epitem". At this
842 * point we are sure no poll callbacks will be lingering around, and also by
843 * holding "epmutex" we can be sure that no file cleanup code will hit
844 * us during this operation. So we can avoid the lock on "ep->lock".
845 * We do not need to lock ep->mtx, either, we only do it to prevent
848 mutex_lock(&ep->mtx);
849 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
850 epi = rb_entry(rbp, struct epitem, rbn);
854 mutex_unlock(&ep->mtx);
856 mutex_unlock(&epmutex);
857 mutex_destroy(&ep->mtx);
859 wakeup_source_unregister(ep->ws);
863 static int ep_eventpoll_release(struct inode *inode, struct file *file)
865 struct eventpoll *ep = file->private_data;
873 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
875 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
879 * Differs from ep_eventpoll_poll() in that internal callers already have
880 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
881 * is correctly annotated.
883 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
886 struct eventpoll *ep;
889 pt->_key = epi->event.events;
890 if (!is_file_epoll(epi->ffd.file))
891 return vfs_poll(epi->ffd.file, pt) & epi->event.events;
893 ep = epi->ffd.file->private_data;
894 poll_wait(epi->ffd.file, &ep->poll_wait, pt);
895 locked = pt && (pt->_qproc == ep_ptable_queue_proc);
897 return ep_scan_ready_list(epi->ffd.file->private_data,
898 ep_read_events_proc, &depth, depth,
899 locked) & epi->event.events;
902 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
905 struct epitem *epi, *tmp;
907 int depth = *(int *)priv;
909 init_poll_funcptr(&pt, NULL);
912 list_for_each_entry_safe(epi, tmp, head, rdllink) {
913 if (ep_item_poll(epi, &pt, depth)) {
914 return EPOLLIN | EPOLLRDNORM;
917 * Item has been dropped into the ready list by the poll
918 * callback, but it's not actually ready, as far as
919 * caller requested events goes. We can remove it here.
921 __pm_relax(ep_wakeup_source(epi));
922 list_del_init(&epi->rdllink);
929 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
931 struct eventpoll *ep = file->private_data;
934 /* Insert inside our poll wait queue */
935 poll_wait(file, &ep->poll_wait, wait);
938 * Proceed to find out if wanted events are really available inside
941 return ep_scan_ready_list(ep, ep_read_events_proc,
942 &depth, depth, false);
945 #ifdef CONFIG_PROC_FS
946 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
948 struct eventpoll *ep = f->private_data;
951 mutex_lock(&ep->mtx);
952 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
953 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
954 struct inode *inode = file_inode(epi->ffd.file);
956 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
957 " pos:%lli ino:%lx sdev:%x\n",
958 epi->ffd.fd, epi->event.events,
959 (long long)epi->event.data,
960 (long long)epi->ffd.file->f_pos,
961 inode->i_ino, inode->i_sb->s_dev);
962 if (seq_has_overflowed(m))
965 mutex_unlock(&ep->mtx);
969 /* File callbacks that implement the eventpoll file behaviour */
970 static const struct file_operations eventpoll_fops = {
971 #ifdef CONFIG_PROC_FS
972 .show_fdinfo = ep_show_fdinfo,
974 .release = ep_eventpoll_release,
975 .poll = ep_eventpoll_poll,
976 .llseek = noop_llseek,
980 * This is called from eventpoll_release() to unlink files from the eventpoll
981 * interface. We need to have this facility to cleanup correctly files that are
982 * closed without being removed from the eventpoll interface.
984 void eventpoll_release_file(struct file *file)
986 struct eventpoll *ep;
987 struct epitem *epi, *next;
990 * We don't want to get "file->f_lock" because it is not
991 * necessary. It is not necessary because we're in the "struct file"
992 * cleanup path, and this means that no one is using this file anymore.
993 * So, for example, epoll_ctl() cannot hit here since if we reach this
994 * point, the file counter already went to zero and fget() would fail.
995 * The only hit might come from ep_free() but by holding the mutex
996 * will correctly serialize the operation. We do need to acquire
997 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
998 * from anywhere but ep_free().
1000 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1002 mutex_lock(&epmutex);
1003 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1005 mutex_lock_nested(&ep->mtx, 0);
1007 mutex_unlock(&ep->mtx);
1009 mutex_unlock(&epmutex);
1012 static int ep_alloc(struct eventpoll **pep)
1015 struct user_struct *user;
1016 struct eventpoll *ep;
1018 user = get_current_user();
1020 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1024 mutex_init(&ep->mtx);
1025 rwlock_init(&ep->lock);
1026 init_waitqueue_head(&ep->wq);
1027 init_waitqueue_head(&ep->poll_wait);
1028 INIT_LIST_HEAD(&ep->rdllist);
1029 ep->rbr = RB_ROOT_CACHED;
1030 ep->ovflist = EP_UNACTIVE_PTR;
1043 * Search the file inside the eventpoll tree. The RB tree operations
1044 * are protected by the "mtx" mutex, and ep_find() must be called with
1047 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1050 struct rb_node *rbp;
1051 struct epitem *epi, *epir = NULL;
1052 struct epoll_filefd ffd;
1054 ep_set_ffd(&ffd, file, fd);
1055 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1056 epi = rb_entry(rbp, struct epitem, rbn);
1057 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1059 rbp = rbp->rb_right;
1072 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1074 struct rb_node *rbp;
1077 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1078 epi = rb_entry(rbp, struct epitem, rbn);
1079 if (epi->ffd.fd == tfd) {
1091 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1094 struct file *file_raw;
1095 struct eventpoll *ep;
1098 if (!is_file_epoll(file))
1099 return ERR_PTR(-EINVAL);
1101 ep = file->private_data;
1103 mutex_lock(&ep->mtx);
1104 epi = ep_find_tfd(ep, tfd, toff);
1106 file_raw = epi->ffd.file;
1108 file_raw = ERR_PTR(-ENOENT);
1109 mutex_unlock(&ep->mtx);
1113 #endif /* CONFIG_KCMP */
1116 * Adds a new entry to the tail of the list in a lockless way, i.e.
1117 * multiple CPUs are allowed to call this function concurrently.
1119 * Beware: it is necessary to prevent any other modifications of the
1120 * existing list until all changes are completed, in other words
1121 * concurrent list_add_tail_lockless() calls should be protected
1122 * with a read lock, where write lock acts as a barrier which
1123 * makes sure all list_add_tail_lockless() calls are fully
1126 * Also an element can be locklessly added to the list only in one
1127 * direction i.e. either to the tail either to the head, otherwise
1128 * concurrent access will corrupt the list.
1130 * Returns %false if element has been already added to the list, %true
1133 static inline bool list_add_tail_lockless(struct list_head *new,
1134 struct list_head *head)
1136 struct list_head *prev;
1139 * This is simple 'new->next = head' operation, but cmpxchg()
1140 * is used in order to detect that same element has been just
1141 * added to the list from another CPU: the winner observes
1144 if (cmpxchg(&new->next, new, head) != new)
1148 * Initially ->next of a new element must be updated with the head
1149 * (we are inserting to the tail) and only then pointers are atomically
1150 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1151 * updated before pointers are actually swapped and pointers are
1152 * swapped before prev->next is updated.
1155 prev = xchg(&head->prev, new);
1158 * It is safe to modify prev->next and new->prev, because a new element
1159 * is added only to the tail and new->next is updated before XCHG.
1169 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1170 * i.e. multiple CPUs are allowed to call this function concurrently.
1172 * Returns %false if epi element has been already chained, %true otherwise.
1174 static inline bool chain_epi_lockless(struct epitem *epi)
1176 struct eventpoll *ep = epi->ep;
1178 /* Fast preliminary check */
1179 if (epi->next != EP_UNACTIVE_PTR)
1182 /* Check that the same epi has not been just chained from another CPU */
1183 if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)
1186 /* Atomically exchange tail */
1187 epi->next = xchg(&ep->ovflist, epi);
1193 * This is the callback that is passed to the wait queue wakeup
1194 * mechanism. It is called by the stored file descriptors when they
1195 * have events to report.
1197 * This callback takes a read lock in order not to content with concurrent
1198 * events from another file descriptors, thus all modifications to ->rdllist
1199 * or ->ovflist are lockless. Read lock is paired with the write lock from
1200 * ep_scan_ready_list(), which stops all list modifications and guarantees
1201 * that lists state is seen correctly.
1203 * Another thing worth to mention is that ep_poll_callback() can be called
1204 * concurrently for the same @epi from different CPUs if poll table was inited
1205 * with several wait queues entries. Plural wakeup from different CPUs of a
1206 * single wait queue is serialized by wq.lock, but the case when multiple wait
1207 * queues are used should be detected accordingly. This is detected using
1208 * cmpxchg() operation.
1210 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1213 struct epitem *epi = ep_item_from_wait(wait);
1214 struct eventpoll *ep = epi->ep;
1215 __poll_t pollflags = key_to_poll(key);
1216 unsigned long flags;
1219 read_lock_irqsave(&ep->lock, flags);
1221 ep_set_busy_poll_napi_id(epi);
1224 * If the event mask does not contain any poll(2) event, we consider the
1225 * descriptor to be disabled. This condition is likely the effect of the
1226 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1227 * until the next EPOLL_CTL_MOD will be issued.
1229 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1233 * Check the events coming with the callback. At this stage, not
1234 * every device reports the events in the "key" parameter of the
1235 * callback. We need to be able to handle both cases here, hence the
1236 * test for "key" != NULL before the event match test.
1238 if (pollflags && !(pollflags & epi->event.events))
1242 * If we are transferring events to userspace, we can hold no locks
1243 * (because we're accessing user memory, and because of linux f_op->poll()
1244 * semantics). All the events that happen during that period of time are
1245 * chained in ep->ovflist and requeued later on.
1247 if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {
1248 if (chain_epi_lockless(epi))
1249 ep_pm_stay_awake_rcu(epi);
1250 } else if (!ep_is_linked(epi)) {
1251 /* In the usual case, add event to ready list. */
1252 if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))
1253 ep_pm_stay_awake_rcu(epi);
1257 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1260 if (waitqueue_active(&ep->wq)) {
1261 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1262 !(pollflags & POLLFREE)) {
1263 switch (pollflags & EPOLLINOUT_BITS) {
1265 if (epi->event.events & EPOLLIN)
1269 if (epi->event.events & EPOLLOUT)
1279 if (waitqueue_active(&ep->poll_wait))
1283 read_unlock_irqrestore(&ep->lock, flags);
1285 /* We have to call this outside the lock */
1287 ep_poll_safewake(ep, epi);
1289 if (!(epi->event.events & EPOLLEXCLUSIVE))
1292 if (pollflags & POLLFREE) {
1294 * If we race with ep_remove_wait_queue() it can miss
1295 * ->whead = NULL and do another remove_wait_queue() after
1296 * us, so we can't use __remove_wait_queue().
1298 list_del_init(&wait->entry);
1300 * ->whead != NULL protects us from the race with ep_free()
1301 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1302 * held by the caller. Once we nullify it, nothing protects
1303 * ep/epi or even wait.
1305 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1312 * This is the callback that is used to add our wait queue to the
1313 * target file wakeup lists.
1315 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1318 struct epitem *epi = ep_item_from_epqueue(pt);
1319 struct eppoll_entry *pwq;
1321 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1322 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1325 if (epi->event.events & EPOLLEXCLUSIVE)
1326 add_wait_queue_exclusive(whead, &pwq->wait);
1328 add_wait_queue(whead, &pwq->wait);
1329 list_add_tail(&pwq->llink, &epi->pwqlist);
1332 /* We have to signal that an error occurred */
1337 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1340 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1341 struct epitem *epic;
1342 bool leftmost = true;
1346 epic = rb_entry(parent, struct epitem, rbn);
1347 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1349 p = &parent->rb_right;
1352 p = &parent->rb_left;
1354 rb_link_node(&epi->rbn, parent, p);
1355 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1360 #define PATH_ARR_SIZE 5
1362 * These are the number paths of length 1 to 5, that we are allowing to emanate
1363 * from a single file of interest. For example, we allow 1000 paths of length
1364 * 1, to emanate from each file of interest. This essentially represents the
1365 * potential wakeup paths, which need to be limited in order to avoid massive
1366 * uncontrolled wakeup storms. The common use case should be a single ep which
1367 * is connected to n file sources. In this case each file source has 1 path
1368 * of length 1. Thus, the numbers below should be more than sufficient. These
1369 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1370 * and delete can't add additional paths. Protected by the epmutex.
1372 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1373 static int path_count[PATH_ARR_SIZE];
1375 static int path_count_inc(int nests)
1377 /* Allow an arbitrary number of depth 1 paths */
1381 if (++path_count[nests] > path_limits[nests])
1386 static void path_count_init(void)
1390 for (i = 0; i < PATH_ARR_SIZE; i++)
1394 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1397 struct file *file = priv;
1398 struct file *child_file;
1401 /* CTL_DEL can remove links here, but that can't increase our count */
1403 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1404 child_file = epi->ep->file;
1405 if (is_file_epoll(child_file)) {
1406 if (list_empty(&child_file->f_ep_links)) {
1407 if (path_count_inc(call_nests)) {
1412 error = ep_call_nested(&poll_loop_ncalls,
1413 reverse_path_check_proc,
1414 child_file, child_file,
1420 printk(KERN_ERR "reverse_path_check_proc: "
1421 "file is not an ep!\n");
1429 * reverse_path_check - The tfile_check_list is list of file *, which have
1430 * links that are proposed to be newly added. We need to
1431 * make sure that those added links don't add too many
1432 * paths such that we will spend all our time waking up
1433 * eventpoll objects.
1435 * Returns: Returns zero if the proposed links don't create too many paths,
1438 static int reverse_path_check(void)
1441 struct file *current_file;
1443 /* let's call this for all tfiles */
1444 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1446 error = ep_call_nested(&poll_loop_ncalls,
1447 reverse_path_check_proc, current_file,
1448 current_file, current);
1455 static int ep_create_wakeup_source(struct epitem *epi)
1457 struct name_snapshot n;
1458 struct wakeup_source *ws;
1461 epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
1466 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1467 ws = wakeup_source_register(NULL, n.name.name);
1468 release_dentry_name_snapshot(&n);
1472 rcu_assign_pointer(epi->ws, ws);
1477 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1478 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1480 struct wakeup_source *ws = ep_wakeup_source(epi);
1482 RCU_INIT_POINTER(epi->ws, NULL);
1485 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1486 * used internally by wakeup_source_remove, too (called by
1487 * wakeup_source_unregister), so we cannot use call_rcu
1490 wakeup_source_unregister(ws);
1494 * Must be called with "mtx" held.
1496 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1497 struct file *tfile, int fd, int full_check)
1499 int error, pwake = 0;
1503 struct ep_pqueue epq;
1505 lockdep_assert_irqs_enabled();
1507 user_watches = atomic_long_read(&ep->user->epoll_watches);
1508 if (unlikely(user_watches >= max_user_watches))
1510 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1513 /* Item initialization follow here ... */
1514 INIT_LIST_HEAD(&epi->rdllink);
1515 INIT_LIST_HEAD(&epi->fllink);
1516 INIT_LIST_HEAD(&epi->pwqlist);
1518 ep_set_ffd(&epi->ffd, tfile, fd);
1519 epi->event = *event;
1521 epi->next = EP_UNACTIVE_PTR;
1522 if (epi->event.events & EPOLLWAKEUP) {
1523 error = ep_create_wakeup_source(epi);
1525 goto error_create_wakeup_source;
1527 RCU_INIT_POINTER(epi->ws, NULL);
1530 /* Add the current item to the list of active epoll hook for this file */
1531 spin_lock(&tfile->f_lock);
1532 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1533 spin_unlock(&tfile->f_lock);
1536 * Add the current item to the RB tree. All RB tree operations are
1537 * protected by "mtx", and ep_insert() is called with "mtx" held.
1539 ep_rbtree_insert(ep, epi);
1541 /* now check if we've created too many backpaths */
1543 if (full_check && reverse_path_check())
1544 goto error_remove_epi;
1546 /* Initialize the poll table using the queue callback */
1548 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1551 * Attach the item to the poll hooks and get current event bits.
1552 * We can safely use the file* here because its usage count has
1553 * been increased by the caller of this function. Note that after
1554 * this operation completes, the poll callback can start hitting
1557 revents = ep_item_poll(epi, &epq.pt, 1);
1560 * We have to check if something went wrong during the poll wait queue
1561 * install process. Namely an allocation for a wait queue failed due
1562 * high memory pressure.
1566 goto error_unregister;
1568 /* We have to drop the new item inside our item list to keep track of it */
1569 write_lock_irq(&ep->lock);
1571 /* record NAPI ID of new item if present */
1572 ep_set_busy_poll_napi_id(epi);
1574 /* If the file is already "ready" we drop it inside the ready list */
1575 if (revents && !ep_is_linked(epi)) {
1576 list_add_tail(&epi->rdllink, &ep->rdllist);
1577 ep_pm_stay_awake(epi);
1579 /* Notify waiting tasks that events are available */
1580 if (waitqueue_active(&ep->wq))
1582 if (waitqueue_active(&ep->poll_wait))
1586 write_unlock_irq(&ep->lock);
1588 atomic_long_inc(&ep->user->epoll_watches);
1590 /* We have to call this outside the lock */
1592 ep_poll_safewake(ep, NULL);
1597 ep_unregister_pollwait(ep, epi);
1599 spin_lock(&tfile->f_lock);
1600 list_del_rcu(&epi->fllink);
1601 spin_unlock(&tfile->f_lock);
1603 rb_erase_cached(&epi->rbn, &ep->rbr);
1606 * We need to do this because an event could have been arrived on some
1607 * allocated wait queue. Note that we don't care about the ep->ovflist
1608 * list, since that is used/cleaned only inside a section bound by "mtx".
1609 * And ep_insert() is called with "mtx" held.
1611 write_lock_irq(&ep->lock);
1612 if (ep_is_linked(epi))
1613 list_del_init(&epi->rdllink);
1614 write_unlock_irq(&ep->lock);
1616 wakeup_source_unregister(ep_wakeup_source(epi));
1618 error_create_wakeup_source:
1619 kmem_cache_free(epi_cache, epi);
1625 * Modify the interest event mask by dropping an event if the new mask
1626 * has a match in the current file status. Must be called with "mtx" held.
1628 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1629 const struct epoll_event *event)
1634 lockdep_assert_irqs_enabled();
1636 init_poll_funcptr(&pt, NULL);
1639 * Set the new event interest mask before calling f_op->poll();
1640 * otherwise we might miss an event that happens between the
1641 * f_op->poll() call and the new event set registering.
1643 epi->event.events = event->events; /* need barrier below */
1644 epi->event.data = event->data; /* protected by mtx */
1645 if (epi->event.events & EPOLLWAKEUP) {
1646 if (!ep_has_wakeup_source(epi))
1647 ep_create_wakeup_source(epi);
1648 } else if (ep_has_wakeup_source(epi)) {
1649 ep_destroy_wakeup_source(epi);
1653 * The following barrier has two effects:
1655 * 1) Flush epi changes above to other CPUs. This ensures
1656 * we do not miss events from ep_poll_callback if an
1657 * event occurs immediately after we call f_op->poll().
1658 * We need this because we did not take ep->lock while
1659 * changing epi above (but ep_poll_callback does take
1662 * 2) We also need to ensure we do not miss _past_ events
1663 * when calling f_op->poll(). This barrier also
1664 * pairs with the barrier in wq_has_sleeper (see
1665 * comments for wq_has_sleeper).
1667 * This barrier will now guarantee ep_poll_callback or f_op->poll
1668 * (or both) will notice the readiness of an item.
1673 * Get current event bits. We can safely use the file* here because
1674 * its usage count has been increased by the caller of this function.
1675 * If the item is "hot" and it is not registered inside the ready
1676 * list, push it inside.
1678 if (ep_item_poll(epi, &pt, 1)) {
1679 write_lock_irq(&ep->lock);
1680 if (!ep_is_linked(epi)) {
1681 list_add_tail(&epi->rdllink, &ep->rdllist);
1682 ep_pm_stay_awake(epi);
1684 /* Notify waiting tasks that events are available */
1685 if (waitqueue_active(&ep->wq))
1687 if (waitqueue_active(&ep->poll_wait))
1690 write_unlock_irq(&ep->lock);
1693 /* We have to call this outside the lock */
1695 ep_poll_safewake(ep, NULL);
1700 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1703 struct ep_send_events_data *esed = priv;
1705 struct epitem *epi, *tmp;
1706 struct epoll_event __user *uevent = esed->events;
1707 struct wakeup_source *ws;
1710 init_poll_funcptr(&pt, NULL);
1714 * We can loop without lock because we are passed a task private list.
1715 * Items cannot vanish during the loop because ep_scan_ready_list() is
1716 * holding "mtx" during this call.
1718 lockdep_assert_held(&ep->mtx);
1720 list_for_each_entry_safe(epi, tmp, head, rdllink) {
1721 if (esed->res >= esed->maxevents)
1725 * Activate ep->ws before deactivating epi->ws to prevent
1726 * triggering auto-suspend here (in case we reactive epi->ws
1729 * This could be rearranged to delay the deactivation of epi->ws
1730 * instead, but then epi->ws would temporarily be out of sync
1731 * with ep_is_linked().
1733 ws = ep_wakeup_source(epi);
1736 __pm_stay_awake(ep->ws);
1740 list_del_init(&epi->rdllink);
1743 * If the event mask intersect the caller-requested one,
1744 * deliver the event to userspace. Again, ep_scan_ready_list()
1745 * is holding ep->mtx, so no operations coming from userspace
1746 * can change the item.
1748 revents = ep_item_poll(epi, &pt, 1);
1752 if (__put_user(revents, &uevent->events) ||
1753 __put_user(epi->event.data, &uevent->data)) {
1754 list_add(&epi->rdllink, head);
1755 ep_pm_stay_awake(epi);
1757 esed->res = -EFAULT;
1762 if (epi->event.events & EPOLLONESHOT)
1763 epi->event.events &= EP_PRIVATE_BITS;
1764 else if (!(epi->event.events & EPOLLET)) {
1766 * If this file has been added with Level
1767 * Trigger mode, we need to insert back inside
1768 * the ready list, so that the next call to
1769 * epoll_wait() will check again the events
1770 * availability. At this point, no one can insert
1771 * into ep->rdllist besides us. The epoll_ctl()
1772 * callers are locked out by
1773 * ep_scan_ready_list() holding "mtx" and the
1774 * poll callback will queue them in ep->ovflist.
1776 list_add_tail(&epi->rdllink, &ep->rdllist);
1777 ep_pm_stay_awake(epi);
1784 static int ep_send_events(struct eventpoll *ep,
1785 struct epoll_event __user *events, int maxevents)
1787 struct ep_send_events_data esed;
1789 esed.maxevents = maxevents;
1790 esed.events = events;
1792 ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1796 static inline struct timespec64 ep_set_mstimeout(long ms)
1798 struct timespec64 now, ts = {
1799 .tv_sec = ms / MSEC_PER_SEC,
1800 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1803 ktime_get_ts64(&now);
1804 return timespec64_add_safe(now, ts);
1808 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1811 * @ep: Pointer to the eventpoll context.
1812 * @events: Pointer to the userspace buffer where the ready events should be
1814 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1815 * @timeout: Maximum timeout for the ready events fetch operation, in
1816 * milliseconds. If the @timeout is zero, the function will not block,
1817 * while if the @timeout is less than zero, the function will block
1818 * until at least one event has been retrieved (or an error
1821 * Returns: Returns the number of ready events which have been fetched, or an
1822 * error code, in case of error.
1824 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1825 int maxevents, long timeout)
1827 int res = 0, eavail, timed_out = 0;
1829 wait_queue_entry_t wait;
1830 ktime_t expires, *to = NULL;
1832 lockdep_assert_irqs_enabled();
1835 struct timespec64 end_time = ep_set_mstimeout(timeout);
1837 slack = select_estimate_accuracy(&end_time);
1839 *to = timespec64_to_ktime(end_time);
1840 } else if (timeout == 0) {
1842 * Avoid the unnecessary trip to the wait queue loop, if the
1843 * caller specified a non blocking operation. We still need
1844 * lock because we could race and not see an epi being added
1845 * to the ready list while in irq callback. Thus incorrectly
1846 * returning 0 back to userspace.
1850 write_lock_irq(&ep->lock);
1851 eavail = ep_events_available(ep);
1852 write_unlock_irq(&ep->lock);
1859 if (!ep_events_available(ep))
1860 ep_busy_loop(ep, timed_out);
1862 eavail = ep_events_available(ep);
1867 * Busy poll timed out. Drop NAPI ID for now, we can add
1868 * it back in when we have moved a socket with a valid NAPI
1869 * ID onto the ready list.
1871 ep_reset_busy_poll_napi_id(ep);
1875 * Internally init_wait() uses autoremove_wake_function(),
1876 * thus wait entry is removed from the wait queue on each
1877 * wakeup. Why it is important? In case of several waiters
1878 * each new wakeup will hit the next waiter, giving it the
1879 * chance to harvest new event. Otherwise wakeup can be
1880 * lost. This is also good performance-wise, because on
1881 * normal wakeup path no need to call __remove_wait_queue()
1882 * explicitly, thus ep->lock is not taken, which halts the
1887 write_lock_irq(&ep->lock);
1889 * Barrierless variant, waitqueue_active() is called under
1890 * the same lock on wakeup ep_poll_callback() side, so it
1891 * is safe to avoid an explicit barrier.
1893 __set_current_state(TASK_INTERRUPTIBLE);
1896 * Do the final check under the lock. ep_scan_ready_list()
1897 * plays with two lists (->rdllist and ->ovflist) and there
1898 * is always a race when both lists are empty for short
1899 * period of time although events are pending, so lock is
1902 eavail = ep_events_available(ep);
1904 if (signal_pending(current))
1907 __add_wait_queue_exclusive(&ep->wq, &wait);
1909 write_unlock_irq(&ep->lock);
1911 if (!eavail && !res)
1912 timed_out = !schedule_hrtimeout_range(to, slack,
1916 * We were woken up, thus go and try to harvest some events.
1917 * If timed out and still on the wait queue, recheck eavail
1918 * carefully under lock, below.
1923 __set_current_state(TASK_RUNNING);
1925 if (!list_empty_careful(&wait.entry)) {
1926 write_lock_irq(&ep->lock);
1928 * If the thread timed out and is not on the wait queue, it
1929 * means that the thread was woken up after its timeout expired
1930 * before it could reacquire the lock. Thus, when wait.entry is
1931 * empty, it needs to harvest events.
1934 eavail = list_empty(&wait.entry);
1935 __remove_wait_queue(&ep->wq, &wait);
1936 write_unlock_irq(&ep->lock);
1940 if (fatal_signal_pending(current)) {
1942 * Always short-circuit for fatal signals to allow
1943 * threads to make a timely exit without the chance of
1944 * finding more events available and fetching
1950 * Try to transfer events to user space. In case we get 0 events and
1951 * there's still timeout left over, we go trying again in search of
1954 if (!res && eavail &&
1955 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1962 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1963 * API, to verify that adding an epoll file inside another
1964 * epoll structure, does not violate the constraints, in
1965 * terms of closed loops, or too deep chains (which can
1966 * result in excessive stack usage).
1968 * @priv: Pointer to the epoll file to be currently checked.
1969 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1970 * data structure pointer.
1971 * @call_nests: Current dept of the @ep_call_nested() call stack.
1973 * Returns: Returns zero if adding the epoll @file inside current epoll
1974 * structure @ep does not violate the constraints, or -1 otherwise.
1976 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1979 struct file *file = priv;
1980 struct eventpoll *ep = file->private_data;
1981 struct eventpoll *ep_tovisit;
1982 struct rb_node *rbp;
1985 mutex_lock_nested(&ep->mtx, call_nests + 1);
1986 ep->gen = loop_check_gen;
1987 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1988 epi = rb_entry(rbp, struct epitem, rbn);
1989 if (unlikely(is_file_epoll(epi->ffd.file))) {
1990 ep_tovisit = epi->ffd.file->private_data;
1991 if (ep_tovisit->gen == loop_check_gen)
1993 error = ep_call_nested(&poll_loop_ncalls,
1994 ep_loop_check_proc, epi->ffd.file,
1995 ep_tovisit, current);
2000 * If we've reached a file that is not associated with
2001 * an ep, then we need to check if the newly added
2002 * links are going to add too many wakeup paths. We do
2003 * this by adding it to the tfile_check_list, if it's
2004 * not already there, and calling reverse_path_check()
2005 * during ep_insert().
2007 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
2008 if (get_file_rcu(epi->ffd.file))
2009 list_add(&epi->ffd.file->f_tfile_llink,
2014 mutex_unlock(&ep->mtx);
2020 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2021 * another epoll file (represented by @ep) does not create
2022 * closed loops or too deep chains.
2024 * @ep: Pointer to the epoll private data structure.
2025 * @file: Pointer to the epoll file to be checked.
2027 * Returns: Returns zero if adding the epoll @file inside current epoll
2028 * structure @ep does not violate the constraints, or -1 otherwise.
2030 static int ep_loop_check(struct eventpoll *ep, struct file *file)
2032 return ep_call_nested(&poll_loop_ncalls,
2033 ep_loop_check_proc, file, ep, current);
2036 static void clear_tfile_check_list(void)
2040 /* first clear the tfile_check_list */
2041 while (!list_empty(&tfile_check_list)) {
2042 file = list_first_entry(&tfile_check_list, struct file,
2044 list_del_init(&file->f_tfile_llink);
2047 INIT_LIST_HEAD(&tfile_check_list);
2051 * Open an eventpoll file descriptor.
2053 static int do_epoll_create(int flags)
2056 struct eventpoll *ep = NULL;
2059 /* Check the EPOLL_* constant for consistency. */
2060 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
2062 if (flags & ~EPOLL_CLOEXEC)
2065 * Create the internal data structure ("struct eventpoll").
2067 error = ep_alloc(&ep);
2071 * Creates all the items needed to setup an eventpoll file. That is,
2072 * a file structure and a free file descriptor.
2074 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
2079 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
2080 O_RDWR | (flags & O_CLOEXEC));
2082 error = PTR_ERR(file);
2086 fd_install(fd, file);
2096 SYSCALL_DEFINE1(epoll_create1, int, flags)
2098 return do_epoll_create(flags);
2101 SYSCALL_DEFINE1(epoll_create, int, size)
2106 return do_epoll_create(0);
2109 static inline int epoll_mutex_lock(struct mutex *mutex, int depth,
2113 mutex_lock_nested(mutex, depth);
2116 if (mutex_trylock(mutex))
2121 int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
2127 struct eventpoll *ep;
2129 struct eventpoll *tep = NULL;
2136 /* Get the "struct file *" for the target file */
2141 /* The target file descriptor must support poll */
2143 if (!file_can_poll(tf.file))
2144 goto error_tgt_fput;
2146 /* Check if EPOLLWAKEUP is allowed */
2147 if (ep_op_has_event(op))
2148 ep_take_care_of_epollwakeup(epds);
2151 * We have to check that the file structure underneath the file descriptor
2152 * the user passed to us _is_ an eventpoll file. And also we do not permit
2153 * adding an epoll file descriptor inside itself.
2156 if (f.file == tf.file || !is_file_epoll(f.file))
2157 goto error_tgt_fput;
2160 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2161 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2162 * Also, we do not currently supported nested exclusive wakeups.
2164 if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
2165 if (op == EPOLL_CTL_MOD)
2166 goto error_tgt_fput;
2167 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2168 (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
2169 goto error_tgt_fput;
2173 * At this point it is safe to assume that the "private_data" contains
2174 * our own data structure.
2176 ep = f.file->private_data;
2179 * When we insert an epoll file descriptor, inside another epoll file
2180 * descriptor, there is the change of creating closed loops, which are
2181 * better be handled here, than in more critical paths. While we are
2182 * checking for loops we also determine the list of files reachable
2183 * and hang them on the tfile_check_list, so we can check that we
2184 * haven't created too many possible wakeup paths.
2186 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2187 * the epoll file descriptor is attaching directly to a wakeup source,
2188 * unless the epoll file descriptor is nested. The purpose of taking the
2189 * 'epmutex' on add is to prevent complex toplogies such as loops and
2190 * deep wakeup paths from forming in parallel through multiple
2191 * EPOLL_CTL_ADD operations.
2193 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2195 goto error_tgt_fput;
2196 if (op == EPOLL_CTL_ADD) {
2197 if (!list_empty(&f.file->f_ep_links) ||
2198 ep->gen == loop_check_gen ||
2199 is_file_epoll(tf.file)) {
2200 mutex_unlock(&ep->mtx);
2201 error = epoll_mutex_lock(&epmutex, 0, nonblock);
2203 goto error_tgt_fput;
2206 if (is_file_epoll(tf.file)) {
2208 if (ep_loop_check(ep, tf.file) != 0)
2209 goto error_tgt_fput;
2212 list_add(&tf.file->f_tfile_llink,
2215 error = epoll_mutex_lock(&ep->mtx, 0, nonblock);
2217 goto error_tgt_fput;
2218 if (is_file_epoll(tf.file)) {
2219 tep = tf.file->private_data;
2220 error = epoll_mutex_lock(&tep->mtx, 1, nonblock);
2222 mutex_unlock(&ep->mtx);
2223 goto error_tgt_fput;
2230 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2231 * above, we can be sure to be able to use the item looked up by
2232 * ep_find() till we release the mutex.
2234 epi = ep_find(ep, tf.file, fd);
2240 epds->events |= EPOLLERR | EPOLLHUP;
2241 error = ep_insert(ep, epds, tf.file, fd, full_check);
2247 error = ep_remove(ep, epi);
2253 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2254 epds->events |= EPOLLERR | EPOLLHUP;
2255 error = ep_modify(ep, epi, epds);
2262 mutex_unlock(&tep->mtx);
2263 mutex_unlock(&ep->mtx);
2267 clear_tfile_check_list();
2269 mutex_unlock(&epmutex);
2281 * The following function implements the controller interface for
2282 * the eventpoll file that enables the insertion/removal/change of
2283 * file descriptors inside the interest set.
2285 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2286 struct epoll_event __user *, event)
2288 struct epoll_event epds;
2290 if (ep_op_has_event(op) &&
2291 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2294 return do_epoll_ctl(epfd, op, fd, &epds, false);
2298 * Implement the event wait interface for the eventpoll file. It is the kernel
2299 * part of the user space epoll_wait(2).
2301 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2302 int maxevents, int timeout)
2306 struct eventpoll *ep;
2308 /* The maximum number of event must be greater than zero */
2309 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2312 /* Verify that the area passed by the user is writeable */
2313 if (!access_ok(events, maxevents * sizeof(struct epoll_event)))
2316 /* Get the "struct file *" for the eventpoll file */
2322 * We have to check that the file structure underneath the fd
2323 * the user passed to us _is_ an eventpoll file.
2326 if (!is_file_epoll(f.file))
2330 * At this point it is safe to assume that the "private_data" contains
2331 * our own data structure.
2333 ep = f.file->private_data;
2335 /* Time to fish for events ... */
2336 error = ep_poll(ep, events, maxevents, timeout);
2343 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2344 int, maxevents, int, timeout)
2346 return do_epoll_wait(epfd, events, maxevents, timeout);
2350 * Implement the event wait interface for the eventpoll file. It is the kernel
2351 * part of the user space epoll_pwait(2).
2353 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2354 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2360 * If the caller wants a certain signal mask to be set during the wait,
2363 error = set_user_sigmask(sigmask, sigsetsize);
2367 error = do_epoll_wait(epfd, events, maxevents, timeout);
2368 restore_saved_sigmask_unless(error == -EINTR);
2373 #ifdef CONFIG_COMPAT
2374 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2375 struct epoll_event __user *, events,
2376 int, maxevents, int, timeout,
2377 const compat_sigset_t __user *, sigmask,
2378 compat_size_t, sigsetsize)
2383 * If the caller wants a certain signal mask to be set during the wait,
2386 err = set_compat_user_sigmask(sigmask, sigsetsize);
2390 err = do_epoll_wait(epfd, events, maxevents, timeout);
2391 restore_saved_sigmask_unless(err == -EINTR);
2397 static int __init eventpoll_init(void)
2403 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2405 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2407 BUG_ON(max_user_watches < 0);
2410 * Initialize the structure used to perform epoll file descriptor
2411 * inclusion loops checks.
2413 ep_nested_calls_init(&poll_loop_ncalls);
2416 * We can have many thousands of epitems, so prevent this from
2417 * using an extra cache line on 64-bit (and smaller) CPUs
2419 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2421 /* Allocates slab cache used to allocate "struct epitem" items */
2422 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2423 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2425 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2426 pwq_cache = kmem_cache_create("eventpoll_pwq",
2427 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2431 fs_initcall(eventpoll_init);