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
, false,
404 static inline void ep_reset_busy_poll_napi_id(struct eventpoll
*ep
)
411 * Set epoll busy poll NAPI ID from sk.
413 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
415 struct eventpoll
*ep
;
416 unsigned int napi_id
;
421 if (!net_busy_loop_on())
424 sock
= sock_from_file(epi
->ffd
.file
, &err
);
432 napi_id
= READ_ONCE(sk
->sk_napi_id
);
435 /* Non-NAPI IDs can be rejected
437 * Nothing to do if we already have this ID
439 if (napi_id
< MIN_NAPI_ID
|| napi_id
== ep
->napi_id
)
442 /* record NAPI ID for use in next busy poll */
443 ep
->napi_id
= napi_id
;
448 static inline void ep_busy_loop(struct eventpoll
*ep
, int nonblock
)
452 static inline void ep_reset_busy_poll_napi_id(struct eventpoll
*ep
)
456 static inline void ep_set_busy_poll_napi_id(struct epitem
*epi
)
460 #endif /* CONFIG_NET_RX_BUSY_POLL */
463 * ep_call_nested - Perform a bound (possibly) nested call, by checking
464 * that the recursion limit is not exceeded, and that
465 * the same nested call (by the meaning of same cookie) is
468 * @ncalls: Pointer to the nested_calls structure to be used for this call.
469 * @nproc: Nested call core function pointer.
470 * @priv: Opaque data to be passed to the @nproc callback.
471 * @cookie: Cookie to be used to identify this nested call.
472 * @ctx: This instance context.
474 * Returns: Returns the code returned by the @nproc callback, or -1 if
475 * the maximum recursion limit has been exceeded.
477 static int ep_call_nested(struct nested_calls
*ncalls
,
478 int (*nproc
)(void *, void *, int), void *priv
,
479 void *cookie
, void *ctx
)
481 int error
, call_nests
= 0;
483 struct list_head
*lsthead
= &ncalls
->tasks_call_list
;
484 struct nested_call_node
*tncur
;
485 struct nested_call_node tnode
;
487 spin_lock_irqsave(&ncalls
->lock
, flags
);
490 * Try to see if the current task is already inside this wakeup call.
491 * We use a list here, since the population inside this set is always
494 list_for_each_entry(tncur
, lsthead
, llink
) {
495 if (tncur
->ctx
== ctx
&&
496 (tncur
->cookie
== cookie
|| ++call_nests
> EP_MAX_NESTS
)) {
498 * Ops ... loop detected or maximum nest level reached.
499 * We abort this wake by breaking the cycle itself.
506 /* Add the current task and cookie to the list */
508 tnode
.cookie
= cookie
;
509 list_add(&tnode
.llink
, lsthead
);
511 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
513 /* Call the nested function */
514 error
= (*nproc
)(priv
, cookie
, call_nests
);
516 /* Remove the current task from the list */
517 spin_lock_irqsave(&ncalls
->lock
, flags
);
518 list_del(&tnode
.llink
);
520 spin_unlock_irqrestore(&ncalls
->lock
, flags
);
526 * As described in commit 0ccf831cb lockdep: annotate epoll
527 * the use of wait queues used by epoll is done in a very controlled
528 * manner. Wake ups can nest inside each other, but are never done
529 * with the same locking. For example:
532 * efd1 = epoll_create();
533 * efd2 = epoll_create();
534 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
535 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
537 * When a packet arrives to the device underneath "dfd", the net code will
538 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
539 * callback wakeup entry on that queue, and the wake_up() performed by the
540 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
541 * (efd1) notices that it may have some event ready, so it needs to wake up
542 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
543 * that ends up in another wake_up(), after having checked about the
544 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
545 * avoid stack blasting.
547 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
548 * this special case of epoll.
550 #ifdef CONFIG_DEBUG_LOCK_ALLOC
552 static void ep_poll_safewake(struct eventpoll
*ep
, struct epitem
*epi
)
554 struct eventpoll
*ep_src
;
559 * To set the subclass or nesting level for spin_lock_irqsave_nested()
560 * it might be natural to create a per-cpu nest count. However, since
561 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
562 * schedule() in the -rt kernel, the per-cpu variable are no longer
563 * protected. Thus, we are introducing a per eventpoll nest field.
564 * If we are not being call from ep_poll_callback(), epi is NULL and
565 * we are at the first level of nesting, 0. Otherwise, we are being
566 * called from ep_poll_callback() and if a previous wakeup source is
567 * not an epoll file itself, we are at depth 1 since the wakeup source
568 * is depth 0. If the wakeup source is a previous epoll file in the
569 * wakeup chain then we use its nests value and record ours as
570 * nests + 1. The previous epoll file nests value is stable since its
571 * already holding its own poll_wait.lock.
574 if ((is_file_epoll(epi
->ffd
.file
))) {
575 ep_src
= epi
->ffd
.file
->private_data
;
576 nests
= ep_src
->nests
;
581 spin_lock_irqsave_nested(&ep
->poll_wait
.lock
, flags
, nests
);
582 ep
->nests
= nests
+ 1;
583 wake_up_locked_poll(&ep
->poll_wait
, EPOLLIN
);
585 spin_unlock_irqrestore(&ep
->poll_wait
.lock
, flags
);
590 static void ep_poll_safewake(struct eventpoll
*ep
, struct epitem
*epi
)
592 wake_up_poll(&ep
->poll_wait
, EPOLLIN
);
597 static void ep_remove_wait_queue(struct eppoll_entry
*pwq
)
599 wait_queue_head_t
*whead
;
603 * If it is cleared by POLLFREE, it should be rcu-safe.
604 * If we read NULL we need a barrier paired with
605 * smp_store_release() in ep_poll_callback(), otherwise
606 * we rely on whead->lock.
608 whead
= smp_load_acquire(&pwq
->whead
);
610 remove_wait_queue(whead
, &pwq
->wait
);
615 * This function unregisters poll callbacks from the associated file
616 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
619 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
621 struct list_head
*lsthead
= &epi
->pwqlist
;
622 struct eppoll_entry
*pwq
;
624 while (!list_empty(lsthead
)) {
625 pwq
= list_first_entry(lsthead
, struct eppoll_entry
, llink
);
627 list_del(&pwq
->llink
);
628 ep_remove_wait_queue(pwq
);
629 kmem_cache_free(pwq_cache
, pwq
);
633 /* call only when ep->mtx is held */
634 static inline struct wakeup_source
*ep_wakeup_source(struct epitem
*epi
)
636 return rcu_dereference_check(epi
->ws
, lockdep_is_held(&epi
->ep
->mtx
));
639 /* call only when ep->mtx is held */
640 static inline void ep_pm_stay_awake(struct epitem
*epi
)
642 struct wakeup_source
*ws
= ep_wakeup_source(epi
);
648 static inline bool ep_has_wakeup_source(struct epitem
*epi
)
650 return rcu_access_pointer(epi
->ws
) ? true : false;
653 /* call when ep->mtx cannot be held (ep_poll_callback) */
654 static inline void ep_pm_stay_awake_rcu(struct epitem
*epi
)
656 struct wakeup_source
*ws
;
659 ws
= rcu_dereference(epi
->ws
);
666 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
667 * the scan code, to call f_op->poll(). Also allows for
668 * O(NumReady) performance.
670 * @ep: Pointer to the epoll private data structure.
671 * @sproc: Pointer to the scan callback.
672 * @priv: Private opaque data passed to the @sproc callback.
673 * @depth: The current depth of recursive f_op->poll calls.
674 * @ep_locked: caller already holds ep->mtx
676 * Returns: The same integer error code returned by the @sproc callback.
678 static __poll_t
ep_scan_ready_list(struct eventpoll
*ep
,
679 __poll_t (*sproc
)(struct eventpoll
*,
680 struct list_head
*, void *),
681 void *priv
, int depth
, bool ep_locked
)
684 struct epitem
*epi
, *nepi
;
687 lockdep_assert_irqs_enabled();
690 * We need to lock this because we could be hit by
691 * eventpoll_release_file() and epoll_ctl().
695 mutex_lock_nested(&ep
->mtx
, depth
);
698 * Steal the ready list, and re-init the original one to the
699 * empty list. Also, set ep->ovflist to NULL so that events
700 * happening while looping w/out locks, are not lost. We cannot
701 * have the poll callback to queue directly on ep->rdllist,
702 * because we want the "sproc" callback to be able to do it
705 write_lock_irq(&ep
->lock
);
706 list_splice_init(&ep
->rdllist
, &txlist
);
707 WRITE_ONCE(ep
->ovflist
, NULL
);
708 write_unlock_irq(&ep
->lock
);
711 * Now call the callback function.
713 res
= (*sproc
)(ep
, &txlist
, priv
);
715 write_lock_irq(&ep
->lock
);
717 * During the time we spent inside the "sproc" callback, some
718 * other events might have been queued by the poll callback.
719 * We re-insert them inside the main ready-list here.
721 for (nepi
= READ_ONCE(ep
->ovflist
); (epi
= nepi
) != NULL
;
722 nepi
= epi
->next
, epi
->next
= EP_UNACTIVE_PTR
) {
724 * We need to check if the item is already in the list.
725 * During the "sproc" callback execution time, items are
726 * queued into ->ovflist but the "txlist" might already
727 * contain them, and the list_splice() below takes care of them.
729 if (!ep_is_linked(epi
)) {
731 * ->ovflist is LIFO, so we have to reverse it in order
734 list_add(&epi
->rdllink
, &ep
->rdllist
);
735 ep_pm_stay_awake(epi
);
739 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
740 * releasing the lock, events will be queued in the normal way inside
743 WRITE_ONCE(ep
->ovflist
, EP_UNACTIVE_PTR
);
746 * Quickly re-inject items left on "txlist".
748 list_splice(&txlist
, &ep
->rdllist
);
750 write_unlock_irq(&ep
->lock
);
753 mutex_unlock(&ep
->mtx
);
758 static void epi_rcu_free(struct rcu_head
*head
)
760 struct epitem
*epi
= container_of(head
, struct epitem
, rcu
);
761 kmem_cache_free(epi_cache
, epi
);
765 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
766 * all the associated resources. Must be called with "mtx" held.
768 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
770 struct file
*file
= epi
->ffd
.file
;
772 lockdep_assert_irqs_enabled();
775 * Removes poll wait queue hooks.
777 ep_unregister_pollwait(ep
, epi
);
779 /* Remove the current item from the list of epoll hooks */
780 spin_lock(&file
->f_lock
);
781 list_del_rcu(&epi
->fllink
);
782 spin_unlock(&file
->f_lock
);
784 rb_erase_cached(&epi
->rbn
, &ep
->rbr
);
786 write_lock_irq(&ep
->lock
);
787 if (ep_is_linked(epi
))
788 list_del_init(&epi
->rdllink
);
789 write_unlock_irq(&ep
->lock
);
791 wakeup_source_unregister(ep_wakeup_source(epi
));
793 * At this point it is safe to free the eventpoll item. Use the union
794 * field epi->rcu, since we are trying to minimize the size of
795 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
796 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
797 * use of the rbn field.
799 call_rcu(&epi
->rcu
, epi_rcu_free
);
801 atomic_long_dec(&ep
->user
->epoll_watches
);
806 static void ep_free(struct eventpoll
*ep
)
811 /* We need to release all tasks waiting for these file */
812 if (waitqueue_active(&ep
->poll_wait
))
813 ep_poll_safewake(ep
, NULL
);
816 * We need to lock this because we could be hit by
817 * eventpoll_release_file() while we're freeing the "struct eventpoll".
818 * We do not need to hold "ep->mtx" here because the epoll file
819 * is on the way to be removed and no one has references to it
820 * anymore. The only hit might come from eventpoll_release_file() but
821 * holding "epmutex" is sufficient here.
823 mutex_lock(&epmutex
);
826 * Walks through the whole tree by unregistering poll callbacks.
828 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
829 epi
= rb_entry(rbp
, struct epitem
, rbn
);
831 ep_unregister_pollwait(ep
, epi
);
836 * Walks through the whole tree by freeing each "struct epitem". At this
837 * point we are sure no poll callbacks will be lingering around, and also by
838 * holding "epmutex" we can be sure that no file cleanup code will hit
839 * us during this operation. So we can avoid the lock on "ep->lock".
840 * We do not need to lock ep->mtx, either, we only do it to prevent
843 mutex_lock(&ep
->mtx
);
844 while ((rbp
= rb_first_cached(&ep
->rbr
)) != NULL
) {
845 epi
= rb_entry(rbp
, struct epitem
, rbn
);
849 mutex_unlock(&ep
->mtx
);
851 mutex_unlock(&epmutex
);
852 mutex_destroy(&ep
->mtx
);
854 wakeup_source_unregister(ep
->ws
);
858 static int ep_eventpoll_release(struct inode
*inode
, struct file
*file
)
860 struct eventpoll
*ep
= file
->private_data
;
868 static __poll_t
ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
870 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
874 * Differs from ep_eventpoll_poll() in that internal callers already have
875 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
876 * is correctly annotated.
878 static __poll_t
ep_item_poll(const struct epitem
*epi
, poll_table
*pt
,
881 struct eventpoll
*ep
;
884 pt
->_key
= epi
->event
.events
;
885 if (!is_file_epoll(epi
->ffd
.file
))
886 return vfs_poll(epi
->ffd
.file
, pt
) & epi
->event
.events
;
888 ep
= epi
->ffd
.file
->private_data
;
889 poll_wait(epi
->ffd
.file
, &ep
->poll_wait
, pt
);
890 locked
= pt
&& (pt
->_qproc
== ep_ptable_queue_proc
);
892 return ep_scan_ready_list(epi
->ffd
.file
->private_data
,
893 ep_read_events_proc
, &depth
, depth
,
894 locked
) & epi
->event
.events
;
897 static __poll_t
ep_read_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
900 struct epitem
*epi
, *tmp
;
902 int depth
= *(int *)priv
;
904 init_poll_funcptr(&pt
, NULL
);
907 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
908 if (ep_item_poll(epi
, &pt
, depth
)) {
909 return EPOLLIN
| EPOLLRDNORM
;
912 * Item has been dropped into the ready list by the poll
913 * callback, but it's not actually ready, as far as
914 * caller requested events goes. We can remove it here.
916 __pm_relax(ep_wakeup_source(epi
));
917 list_del_init(&epi
->rdllink
);
924 static __poll_t
ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
926 struct eventpoll
*ep
= file
->private_data
;
929 /* Insert inside our poll wait queue */
930 poll_wait(file
, &ep
->poll_wait
, wait
);
933 * Proceed to find out if wanted events are really available inside
936 return ep_scan_ready_list(ep
, ep_read_events_proc
,
937 &depth
, depth
, false);
940 #ifdef CONFIG_PROC_FS
941 static void ep_show_fdinfo(struct seq_file
*m
, struct file
*f
)
943 struct eventpoll
*ep
= f
->private_data
;
946 mutex_lock(&ep
->mtx
);
947 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
948 struct epitem
*epi
= rb_entry(rbp
, struct epitem
, rbn
);
949 struct inode
*inode
= file_inode(epi
->ffd
.file
);
951 seq_printf(m
, "tfd: %8d events: %8x data: %16llx "
952 " pos:%lli ino:%lx sdev:%x\n",
953 epi
->ffd
.fd
, epi
->event
.events
,
954 (long long)epi
->event
.data
,
955 (long long)epi
->ffd
.file
->f_pos
,
956 inode
->i_ino
, inode
->i_sb
->s_dev
);
957 if (seq_has_overflowed(m
))
960 mutex_unlock(&ep
->mtx
);
964 /* File callbacks that implement the eventpoll file behaviour */
965 static const struct file_operations eventpoll_fops
= {
966 #ifdef CONFIG_PROC_FS
967 .show_fdinfo
= ep_show_fdinfo
,
969 .release
= ep_eventpoll_release
,
970 .poll
= ep_eventpoll_poll
,
971 .llseek
= noop_llseek
,
975 * This is called from eventpoll_release() to unlink files from the eventpoll
976 * interface. We need to have this facility to cleanup correctly files that are
977 * closed without being removed from the eventpoll interface.
979 void eventpoll_release_file(struct file
*file
)
981 struct eventpoll
*ep
;
982 struct epitem
*epi
, *next
;
985 * We don't want to get "file->f_lock" because it is not
986 * necessary. It is not necessary because we're in the "struct file"
987 * cleanup path, and this means that no one is using this file anymore.
988 * So, for example, epoll_ctl() cannot hit here since if we reach this
989 * point, the file counter already went to zero and fget() would fail.
990 * The only hit might come from ep_free() but by holding the mutex
991 * will correctly serialize the operation. We do need to acquire
992 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
993 * from anywhere but ep_free().
995 * Besides, ep_remove() acquires the lock, so we can't hold it here.
997 mutex_lock(&epmutex
);
998 list_for_each_entry_safe(epi
, next
, &file
->f_ep_links
, fllink
) {
1000 mutex_lock_nested(&ep
->mtx
, 0);
1002 mutex_unlock(&ep
->mtx
);
1004 mutex_unlock(&epmutex
);
1007 static int ep_alloc(struct eventpoll
**pep
)
1010 struct user_struct
*user
;
1011 struct eventpoll
*ep
;
1013 user
= get_current_user();
1015 ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
1019 mutex_init(&ep
->mtx
);
1020 rwlock_init(&ep
->lock
);
1021 init_waitqueue_head(&ep
->wq
);
1022 init_waitqueue_head(&ep
->poll_wait
);
1023 INIT_LIST_HEAD(&ep
->rdllist
);
1024 ep
->rbr
= RB_ROOT_CACHED
;
1025 ep
->ovflist
= EP_UNACTIVE_PTR
;
1038 * Search the file inside the eventpoll tree. The RB tree operations
1039 * are protected by the "mtx" mutex, and ep_find() must be called with
1042 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
1045 struct rb_node
*rbp
;
1046 struct epitem
*epi
, *epir
= NULL
;
1047 struct epoll_filefd ffd
;
1049 ep_set_ffd(&ffd
, file
, fd
);
1050 for (rbp
= ep
->rbr
.rb_root
.rb_node
; rbp
; ) {
1051 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1052 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
1054 rbp
= rbp
->rb_right
;
1066 #ifdef CONFIG_CHECKPOINT_RESTORE
1067 static struct epitem
*ep_find_tfd(struct eventpoll
*ep
, int tfd
, unsigned long toff
)
1069 struct rb_node
*rbp
;
1072 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1073 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1074 if (epi
->ffd
.fd
== tfd
) {
1086 struct file
*get_epoll_tfile_raw_ptr(struct file
*file
, int tfd
,
1089 struct file
*file_raw
;
1090 struct eventpoll
*ep
;
1093 if (!is_file_epoll(file
))
1094 return ERR_PTR(-EINVAL
);
1096 ep
= file
->private_data
;
1098 mutex_lock(&ep
->mtx
);
1099 epi
= ep_find_tfd(ep
, tfd
, toff
);
1101 file_raw
= epi
->ffd
.file
;
1103 file_raw
= ERR_PTR(-ENOENT
);
1104 mutex_unlock(&ep
->mtx
);
1108 #endif /* CONFIG_CHECKPOINT_RESTORE */
1111 * Adds a new entry to the tail of the list in a lockless way, i.e.
1112 * multiple CPUs are allowed to call this function concurrently.
1114 * Beware: it is necessary to prevent any other modifications of the
1115 * existing list until all changes are completed, in other words
1116 * concurrent list_add_tail_lockless() calls should be protected
1117 * with a read lock, where write lock acts as a barrier which
1118 * makes sure all list_add_tail_lockless() calls are fully
1121 * Also an element can be locklessly added to the list only in one
1122 * direction i.e. either to the tail either to the head, otherwise
1123 * concurrent access will corrupt the list.
1125 * Returns %false if element has been already added to the list, %true
1128 static inline bool list_add_tail_lockless(struct list_head
*new,
1129 struct list_head
*head
)
1131 struct list_head
*prev
;
1134 * This is simple 'new->next = head' operation, but cmpxchg()
1135 * is used in order to detect that same element has been just
1136 * added to the list from another CPU: the winner observes
1139 if (cmpxchg(&new->next
, new, head
) != new)
1143 * Initially ->next of a new element must be updated with the head
1144 * (we are inserting to the tail) and only then pointers are atomically
1145 * exchanged. XCHG guarantees memory ordering, thus ->next should be
1146 * updated before pointers are actually swapped and pointers are
1147 * swapped before prev->next is updated.
1150 prev
= xchg(&head
->prev
, new);
1153 * It is safe to modify prev->next and new->prev, because a new element
1154 * is added only to the tail and new->next is updated before XCHG.
1164 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,
1165 * i.e. multiple CPUs are allowed to call this function concurrently.
1167 * Returns %false if epi element has been already chained, %true otherwise.
1169 static inline bool chain_epi_lockless(struct epitem
*epi
)
1171 struct eventpoll
*ep
= epi
->ep
;
1173 /* Fast preliminary check */
1174 if (epi
->next
!= EP_UNACTIVE_PTR
)
1177 /* Check that the same epi has not been just chained from another CPU */
1178 if (cmpxchg(&epi
->next
, EP_UNACTIVE_PTR
, NULL
) != EP_UNACTIVE_PTR
)
1181 /* Atomically exchange tail */
1182 epi
->next
= xchg(&ep
->ovflist
, epi
);
1188 * This is the callback that is passed to the wait queue wakeup
1189 * mechanism. It is called by the stored file descriptors when they
1190 * have events to report.
1192 * This callback takes a read lock in order not to content with concurrent
1193 * events from another file descriptors, thus all modifications to ->rdllist
1194 * or ->ovflist are lockless. Read lock is paired with the write lock from
1195 * ep_scan_ready_list(), which stops all list modifications and guarantees
1196 * that lists state is seen correctly.
1198 * Another thing worth to mention is that ep_poll_callback() can be called
1199 * concurrently for the same @epi from different CPUs if poll table was inited
1200 * with several wait queues entries. Plural wakeup from different CPUs of a
1201 * single wait queue is serialized by wq.lock, but the case when multiple wait
1202 * queues are used should be detected accordingly. This is detected using
1203 * cmpxchg() operation.
1205 static int ep_poll_callback(wait_queue_entry_t
*wait
, unsigned mode
, int sync
, void *key
)
1208 struct epitem
*epi
= ep_item_from_wait(wait
);
1209 struct eventpoll
*ep
= epi
->ep
;
1210 __poll_t pollflags
= key_to_poll(key
);
1211 unsigned long flags
;
1214 read_lock_irqsave(&ep
->lock
, flags
);
1216 ep_set_busy_poll_napi_id(epi
);
1219 * If the event mask does not contain any poll(2) event, we consider the
1220 * descriptor to be disabled. This condition is likely the effect of the
1221 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1222 * until the next EPOLL_CTL_MOD will be issued.
1224 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1228 * Check the events coming with the callback. At this stage, not
1229 * every device reports the events in the "key" parameter of the
1230 * callback. We need to be able to handle both cases here, hence the
1231 * test for "key" != NULL before the event match test.
1233 if (pollflags
&& !(pollflags
& epi
->event
.events
))
1237 * If we are transferring events to userspace, we can hold no locks
1238 * (because we're accessing user memory, and because of linux f_op->poll()
1239 * semantics). All the events that happen during that period of time are
1240 * chained in ep->ovflist and requeued later on.
1242 if (READ_ONCE(ep
->ovflist
) != EP_UNACTIVE_PTR
) {
1243 if (chain_epi_lockless(epi
))
1244 ep_pm_stay_awake_rcu(epi
);
1245 } else if (!ep_is_linked(epi
)) {
1246 /* In the usual case, add event to ready list. */
1247 if (list_add_tail_lockless(&epi
->rdllink
, &ep
->rdllist
))
1248 ep_pm_stay_awake_rcu(epi
);
1252 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1255 if (waitqueue_active(&ep
->wq
)) {
1256 if ((epi
->event
.events
& EPOLLEXCLUSIVE
) &&
1257 !(pollflags
& POLLFREE
)) {
1258 switch (pollflags
& EPOLLINOUT_BITS
) {
1260 if (epi
->event
.events
& EPOLLIN
)
1264 if (epi
->event
.events
& EPOLLOUT
)
1274 if (waitqueue_active(&ep
->poll_wait
))
1278 read_unlock_irqrestore(&ep
->lock
, flags
);
1280 /* We have to call this outside the lock */
1282 ep_poll_safewake(ep
, epi
);
1284 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
))
1287 if (pollflags
& POLLFREE
) {
1289 * If we race with ep_remove_wait_queue() it can miss
1290 * ->whead = NULL and do another remove_wait_queue() after
1291 * us, so we can't use __remove_wait_queue().
1293 list_del_init(&wait
->entry
);
1295 * ->whead != NULL protects us from the race with ep_free()
1296 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1297 * held by the caller. Once we nullify it, nothing protects
1298 * ep/epi or even wait.
1300 smp_store_release(&ep_pwq_from_wait(wait
)->whead
, NULL
);
1307 * This is the callback that is used to add our wait queue to the
1308 * target file wakeup lists.
1310 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
1313 struct epitem
*epi
= ep_item_from_epqueue(pt
);
1314 struct eppoll_entry
*pwq
;
1316 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
1317 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
1320 if (epi
->event
.events
& EPOLLEXCLUSIVE
)
1321 add_wait_queue_exclusive(whead
, &pwq
->wait
);
1323 add_wait_queue(whead
, &pwq
->wait
);
1324 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
1327 /* We have to signal that an error occurred */
1332 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
1335 struct rb_node
**p
= &ep
->rbr
.rb_root
.rb_node
, *parent
= NULL
;
1336 struct epitem
*epic
;
1337 bool leftmost
= true;
1341 epic
= rb_entry(parent
, struct epitem
, rbn
);
1342 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
1344 p
= &parent
->rb_right
;
1347 p
= &parent
->rb_left
;
1349 rb_link_node(&epi
->rbn
, parent
, p
);
1350 rb_insert_color_cached(&epi
->rbn
, &ep
->rbr
, leftmost
);
1355 #define PATH_ARR_SIZE 5
1357 * These are the number paths of length 1 to 5, that we are allowing to emanate
1358 * from a single file of interest. For example, we allow 1000 paths of length
1359 * 1, to emanate from each file of interest. This essentially represents the
1360 * potential wakeup paths, which need to be limited in order to avoid massive
1361 * uncontrolled wakeup storms. The common use case should be a single ep which
1362 * is connected to n file sources. In this case each file source has 1 path
1363 * of length 1. Thus, the numbers below should be more than sufficient. These
1364 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1365 * and delete can't add additional paths. Protected by the epmutex.
1367 static const int path_limits
[PATH_ARR_SIZE
] = { 1000, 500, 100, 50, 10 };
1368 static int path_count
[PATH_ARR_SIZE
];
1370 static int path_count_inc(int nests
)
1372 /* Allow an arbitrary number of depth 1 paths */
1376 if (++path_count
[nests
] > path_limits
[nests
])
1381 static void path_count_init(void)
1385 for (i
= 0; i
< PATH_ARR_SIZE
; i
++)
1389 static int reverse_path_check_proc(void *priv
, void *cookie
, int call_nests
)
1392 struct file
*file
= priv
;
1393 struct file
*child_file
;
1396 /* CTL_DEL can remove links here, but that can't increase our count */
1398 list_for_each_entry_rcu(epi
, &file
->f_ep_links
, fllink
) {
1399 child_file
= epi
->ep
->file
;
1400 if (is_file_epoll(child_file
)) {
1401 if (list_empty(&child_file
->f_ep_links
)) {
1402 if (path_count_inc(call_nests
)) {
1407 error
= ep_call_nested(&poll_loop_ncalls
,
1408 reverse_path_check_proc
,
1409 child_file
, child_file
,
1415 printk(KERN_ERR
"reverse_path_check_proc: "
1416 "file is not an ep!\n");
1424 * reverse_path_check - The tfile_check_list is list of file *, which have
1425 * links that are proposed to be newly added. We need to
1426 * make sure that those added links don't add too many
1427 * paths such that we will spend all our time waking up
1428 * eventpoll objects.
1430 * Returns: Returns zero if the proposed links don't create too many paths,
1433 static int reverse_path_check(void)
1436 struct file
*current_file
;
1438 /* let's call this for all tfiles */
1439 list_for_each_entry(current_file
, &tfile_check_list
, f_tfile_llink
) {
1441 error
= ep_call_nested(&poll_loop_ncalls
,
1442 reverse_path_check_proc
, current_file
,
1443 current_file
, current
);
1450 static int ep_create_wakeup_source(struct epitem
*epi
)
1452 struct name_snapshot n
;
1453 struct wakeup_source
*ws
;
1456 epi
->ep
->ws
= wakeup_source_register(NULL
, "eventpoll");
1461 take_dentry_name_snapshot(&n
, epi
->ffd
.file
->f_path
.dentry
);
1462 ws
= wakeup_source_register(NULL
, n
.name
.name
);
1463 release_dentry_name_snapshot(&n
);
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 /* Add the current item to the list of active epoll hook for this file */
1526 spin_lock(&tfile
->f_lock
);
1527 list_add_tail_rcu(&epi
->fllink
, &tfile
->f_ep_links
);
1528 spin_unlock(&tfile
->f_lock
);
1531 * Add the current item to the RB tree. All RB tree operations are
1532 * protected by "mtx", and ep_insert() is called with "mtx" held.
1534 ep_rbtree_insert(ep
, epi
);
1536 /* now check if we've created too many backpaths */
1538 if (full_check
&& reverse_path_check())
1539 goto error_remove_epi
;
1541 /* Initialize the poll table using the queue callback */
1543 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1546 * Attach the item to the poll hooks and get current event bits.
1547 * We can safely use the file* here because its usage count has
1548 * been increased by the caller of this function. Note that after
1549 * this operation completes, the poll callback can start hitting
1552 revents
= ep_item_poll(epi
, &epq
.pt
, 1);
1555 * We have to check if something went wrong during the poll wait queue
1556 * install process. Namely an allocation for a wait queue failed due
1557 * high memory pressure.
1561 goto error_unregister
;
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 ep_unregister_pollwait(ep
, epi
);
1594 spin_lock(&tfile
->f_lock
);
1595 list_del_rcu(&epi
->fllink
);
1596 spin_unlock(&tfile
->f_lock
);
1598 rb_erase_cached(&epi
->rbn
, &ep
->rbr
);
1601 * We need to do this because an event could have been arrived on some
1602 * allocated wait queue. Note that we don't care about the ep->ovflist
1603 * list, since that is used/cleaned only inside a section bound by "mtx".
1604 * And ep_insert() is called with "mtx" held.
1606 write_lock_irq(&ep
->lock
);
1607 if (ep_is_linked(epi
))
1608 list_del_init(&epi
->rdllink
);
1609 write_unlock_irq(&ep
->lock
);
1611 wakeup_source_unregister(ep_wakeup_source(epi
));
1613 error_create_wakeup_source
:
1614 kmem_cache_free(epi_cache
, epi
);
1620 * Modify the interest event mask by dropping an event if the new mask
1621 * has a match in the current file status. Must be called with "mtx" held.
1623 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
,
1624 const struct epoll_event
*event
)
1629 lockdep_assert_irqs_enabled();
1631 init_poll_funcptr(&pt
, NULL
);
1634 * Set the new event interest mask before calling f_op->poll();
1635 * otherwise we might miss an event that happens between the
1636 * f_op->poll() call and the new event set registering.
1638 epi
->event
.events
= event
->events
; /* need barrier below */
1639 epi
->event
.data
= event
->data
; /* protected by mtx */
1640 if (epi
->event
.events
& EPOLLWAKEUP
) {
1641 if (!ep_has_wakeup_source(epi
))
1642 ep_create_wakeup_source(epi
);
1643 } else if (ep_has_wakeup_source(epi
)) {
1644 ep_destroy_wakeup_source(epi
);
1648 * The following barrier has two effects:
1650 * 1) Flush epi changes above to other CPUs. This ensures
1651 * we do not miss events from ep_poll_callback if an
1652 * event occurs immediately after we call f_op->poll().
1653 * We need this because we did not take ep->lock while
1654 * changing epi above (but ep_poll_callback does take
1657 * 2) We also need to ensure we do not miss _past_ events
1658 * when calling f_op->poll(). This barrier also
1659 * pairs with the barrier in wq_has_sleeper (see
1660 * comments for wq_has_sleeper).
1662 * This barrier will now guarantee ep_poll_callback or f_op->poll
1663 * (or both) will notice the readiness of an item.
1668 * Get current event bits. We can safely use the file* here because
1669 * its usage count has been increased by the caller of this function.
1670 * If the item is "hot" and it is not registered inside the ready
1671 * list, push it inside.
1673 if (ep_item_poll(epi
, &pt
, 1)) {
1674 write_lock_irq(&ep
->lock
);
1675 if (!ep_is_linked(epi
)) {
1676 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1677 ep_pm_stay_awake(epi
);
1679 /* Notify waiting tasks that events are available */
1680 if (waitqueue_active(&ep
->wq
))
1682 if (waitqueue_active(&ep
->poll_wait
))
1685 write_unlock_irq(&ep
->lock
);
1688 /* We have to call this outside the lock */
1690 ep_poll_safewake(ep
, NULL
);
1695 static __poll_t
ep_send_events_proc(struct eventpoll
*ep
, struct list_head
*head
,
1698 struct ep_send_events_data
*esed
= priv
;
1700 struct epitem
*epi
, *tmp
;
1701 struct epoll_event __user
*uevent
= esed
->events
;
1702 struct wakeup_source
*ws
;
1705 init_poll_funcptr(&pt
, NULL
);
1709 * We can loop without lock because we are passed a task private list.
1710 * Items cannot vanish during the loop because ep_scan_ready_list() is
1711 * holding "mtx" during this call.
1713 lockdep_assert_held(&ep
->mtx
);
1715 list_for_each_entry_safe(epi
, tmp
, head
, rdllink
) {
1716 if (esed
->res
>= esed
->maxevents
)
1720 * Activate ep->ws before deactivating epi->ws to prevent
1721 * triggering auto-suspend here (in case we reactive epi->ws
1724 * This could be rearranged to delay the deactivation of epi->ws
1725 * instead, but then epi->ws would temporarily be out of sync
1726 * with ep_is_linked().
1728 ws
= ep_wakeup_source(epi
);
1731 __pm_stay_awake(ep
->ws
);
1735 list_del_init(&epi
->rdllink
);
1738 * If the event mask intersect the caller-requested one,
1739 * deliver the event to userspace. Again, ep_scan_ready_list()
1740 * is holding ep->mtx, so no operations coming from userspace
1741 * can change the item.
1743 revents
= ep_item_poll(epi
, &pt
, 1);
1747 if (__put_user(revents
, &uevent
->events
) ||
1748 __put_user(epi
->event
.data
, &uevent
->data
)) {
1749 list_add(&epi
->rdllink
, head
);
1750 ep_pm_stay_awake(epi
);
1752 esed
->res
= -EFAULT
;
1757 if (epi
->event
.events
& EPOLLONESHOT
)
1758 epi
->event
.events
&= EP_PRIVATE_BITS
;
1759 else if (!(epi
->event
.events
& EPOLLET
)) {
1761 * If this file has been added with Level
1762 * Trigger mode, we need to insert back inside
1763 * the ready list, so that the next call to
1764 * epoll_wait() will check again the events
1765 * availability. At this point, no one can insert
1766 * into ep->rdllist besides us. The epoll_ctl()
1767 * callers are locked out by
1768 * ep_scan_ready_list() holding "mtx" and the
1769 * poll callback will queue them in ep->ovflist.
1771 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1772 ep_pm_stay_awake(epi
);
1779 static int ep_send_events(struct eventpoll
*ep
,
1780 struct epoll_event __user
*events
, int maxevents
)
1782 struct ep_send_events_data esed
;
1784 esed
.maxevents
= maxevents
;
1785 esed
.events
= events
;
1787 ep_scan_ready_list(ep
, ep_send_events_proc
, &esed
, 0, false);
1791 static inline struct timespec64
ep_set_mstimeout(long ms
)
1793 struct timespec64 now
, ts
= {
1794 .tv_sec
= ms
/ MSEC_PER_SEC
,
1795 .tv_nsec
= NSEC_PER_MSEC
* (ms
% MSEC_PER_SEC
),
1798 ktime_get_ts64(&now
);
1799 return timespec64_add_safe(now
, ts
);
1803 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1806 * @ep: Pointer to the eventpoll context.
1807 * @events: Pointer to the userspace buffer where the ready events should be
1809 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1810 * @timeout: Maximum timeout for the ready events fetch operation, in
1811 * milliseconds. If the @timeout is zero, the function will not block,
1812 * while if the @timeout is less than zero, the function will block
1813 * until at least one event has been retrieved (or an error
1816 * Returns: Returns the number of ready events which have been fetched, or an
1817 * error code, in case of error.
1819 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1820 int maxevents
, long timeout
)
1822 int res
= 0, eavail
, timed_out
= 0;
1824 wait_queue_entry_t wait
;
1825 ktime_t expires
, *to
= NULL
;
1827 lockdep_assert_irqs_enabled();
1830 struct timespec64 end_time
= ep_set_mstimeout(timeout
);
1832 slack
= select_estimate_accuracy(&end_time
);
1834 *to
= timespec64_to_ktime(end_time
);
1835 } else if (timeout
== 0) {
1837 * Avoid the unnecessary trip to the wait queue loop, if the
1838 * caller specified a non blocking operation. We still need
1839 * lock because we could race and not see an epi being added
1840 * to the ready list while in irq callback. Thus incorrectly
1841 * returning 0 back to userspace.
1845 write_lock_irq(&ep
->lock
);
1846 eavail
= ep_events_available(ep
);
1847 write_unlock_irq(&ep
->lock
);
1854 if (!ep_events_available(ep
))
1855 ep_busy_loop(ep
, timed_out
);
1857 eavail
= ep_events_available(ep
);
1862 * Busy poll timed out. Drop NAPI ID for now, we can add
1863 * it back in when we have moved a socket with a valid NAPI
1864 * ID onto the ready list.
1866 ep_reset_busy_poll_napi_id(ep
);
1870 * Internally init_wait() uses autoremove_wake_function(),
1871 * thus wait entry is removed from the wait queue on each
1872 * wakeup. Why it is important? In case of several waiters
1873 * each new wakeup will hit the next waiter, giving it the
1874 * chance to harvest new event. Otherwise wakeup can be
1875 * lost. This is also good performance-wise, because on
1876 * normal wakeup path no need to call __remove_wait_queue()
1877 * explicitly, thus ep->lock is not taken, which halts the
1882 write_lock_irq(&ep
->lock
);
1884 * Barrierless variant, waitqueue_active() is called under
1885 * the same lock on wakeup ep_poll_callback() side, so it
1886 * is safe to avoid an explicit barrier.
1888 __set_current_state(TASK_INTERRUPTIBLE
);
1891 * Do the final check under the lock. ep_scan_ready_list()
1892 * plays with two lists (->rdllist and ->ovflist) and there
1893 * is always a race when both lists are empty for short
1894 * period of time although events are pending, so lock is
1897 eavail
= ep_events_available(ep
);
1899 if (signal_pending(current
))
1902 __add_wait_queue_exclusive(&ep
->wq
, &wait
);
1904 write_unlock_irq(&ep
->lock
);
1909 if (!schedule_hrtimeout_range(to
, slack
, HRTIMER_MODE_ABS
)) {
1914 /* We were woken up, thus go and try to harvest some events */
1919 __set_current_state(TASK_RUNNING
);
1921 if (!list_empty_careful(&wait
.entry
)) {
1922 write_lock_irq(&ep
->lock
);
1923 __remove_wait_queue(&ep
->wq
, &wait
);
1924 write_unlock_irq(&ep
->lock
);
1928 if (fatal_signal_pending(current
)) {
1930 * Always short-circuit for fatal signals to allow
1931 * threads to make a timely exit without the chance of
1932 * finding more events available and fetching
1938 * Try to transfer events to user space. In case we get 0 events and
1939 * there's still timeout left over, we go trying again in search of
1942 if (!res
&& eavail
&&
1943 !(res
= ep_send_events(ep
, events
, maxevents
)) && !timed_out
)
1950 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1951 * API, to verify that adding an epoll file inside another
1952 * epoll structure, does not violate the constraints, in
1953 * terms of closed loops, or too deep chains (which can
1954 * result in excessive stack usage).
1956 * @priv: Pointer to the epoll file to be currently checked.
1957 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1958 * data structure pointer.
1959 * @call_nests: Current dept of the @ep_call_nested() call stack.
1961 * Returns: Returns zero if adding the epoll @file inside current epoll
1962 * structure @ep does not violate the constraints, or -1 otherwise.
1964 static int ep_loop_check_proc(void *priv
, void *cookie
, int call_nests
)
1967 struct file
*file
= priv
;
1968 struct eventpoll
*ep
= file
->private_data
;
1969 struct eventpoll
*ep_tovisit
;
1970 struct rb_node
*rbp
;
1973 mutex_lock_nested(&ep
->mtx
, call_nests
+ 1);
1974 ep
->gen
= loop_check_gen
;
1975 for (rbp
= rb_first_cached(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
1976 epi
= rb_entry(rbp
, struct epitem
, rbn
);
1977 if (unlikely(is_file_epoll(epi
->ffd
.file
))) {
1978 ep_tovisit
= epi
->ffd
.file
->private_data
;
1979 if (ep_tovisit
->gen
== loop_check_gen
)
1981 error
= ep_call_nested(&poll_loop_ncalls
,
1982 ep_loop_check_proc
, epi
->ffd
.file
,
1983 ep_tovisit
, current
);
1988 * If we've reached a file that is not associated with
1989 * an ep, then we need to check if the newly added
1990 * links are going to add too many wakeup paths. We do
1991 * this by adding it to the tfile_check_list, if it's
1992 * not already there, and calling reverse_path_check()
1993 * during ep_insert().
1995 if (list_empty(&epi
->ffd
.file
->f_tfile_llink
)) {
1996 if (get_file_rcu(epi
->ffd
.file
))
1997 list_add(&epi
->ffd
.file
->f_tfile_llink
,
2002 mutex_unlock(&ep
->mtx
);
2008 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
2009 * another epoll file (represented by @ep) does not create
2010 * closed loops or too deep chains.
2012 * @ep: Pointer to the epoll private data structure.
2013 * @file: Pointer to the epoll file to be checked.
2015 * Returns: Returns zero if adding the epoll @file inside current epoll
2016 * structure @ep does not violate the constraints, or -1 otherwise.
2018 static int ep_loop_check(struct eventpoll
*ep
, struct file
*file
)
2020 return ep_call_nested(&poll_loop_ncalls
,
2021 ep_loop_check_proc
, file
, ep
, current
);
2024 static void clear_tfile_check_list(void)
2028 /* first clear the tfile_check_list */
2029 while (!list_empty(&tfile_check_list
)) {
2030 file
= list_first_entry(&tfile_check_list
, struct file
,
2032 list_del_init(&file
->f_tfile_llink
);
2035 INIT_LIST_HEAD(&tfile_check_list
);
2039 * Open an eventpoll file descriptor.
2041 static int do_epoll_create(int flags
)
2044 struct eventpoll
*ep
= NULL
;
2047 /* Check the EPOLL_* constant for consistency. */
2048 BUILD_BUG_ON(EPOLL_CLOEXEC
!= O_CLOEXEC
);
2050 if (flags
& ~EPOLL_CLOEXEC
)
2053 * Create the internal data structure ("struct eventpoll").
2055 error
= ep_alloc(&ep
);
2059 * Creates all the items needed to setup an eventpoll file. That is,
2060 * a file structure and a free file descriptor.
2062 fd
= get_unused_fd_flags(O_RDWR
| (flags
& O_CLOEXEC
));
2067 file
= anon_inode_getfile("[eventpoll]", &eventpoll_fops
, ep
,
2068 O_RDWR
| (flags
& O_CLOEXEC
));
2070 error
= PTR_ERR(file
);
2074 fd_install(fd
, file
);
2084 SYSCALL_DEFINE1(epoll_create1
, int, flags
)
2086 return do_epoll_create(flags
);
2089 SYSCALL_DEFINE1(epoll_create
, int, size
)
2094 return do_epoll_create(0);
2097 static inline int epoll_mutex_lock(struct mutex
*mutex
, int depth
,
2101 mutex_lock_nested(mutex
, depth
);
2104 if (mutex_trylock(mutex
))
2109 int do_epoll_ctl(int epfd
, int op
, int fd
, struct epoll_event
*epds
,
2115 struct eventpoll
*ep
;
2117 struct eventpoll
*tep
= NULL
;
2124 /* Get the "struct file *" for the target file */
2129 /* The target file descriptor must support poll */
2131 if (!file_can_poll(tf
.file
))
2132 goto error_tgt_fput
;
2134 /* Check if EPOLLWAKEUP is allowed */
2135 if (ep_op_has_event(op
))
2136 ep_take_care_of_epollwakeup(epds
);
2139 * We have to check that the file structure underneath the file descriptor
2140 * the user passed to us _is_ an eventpoll file. And also we do not permit
2141 * adding an epoll file descriptor inside itself.
2144 if (f
.file
== tf
.file
|| !is_file_epoll(f
.file
))
2145 goto error_tgt_fput
;
2148 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2149 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2150 * Also, we do not currently supported nested exclusive wakeups.
2152 if (ep_op_has_event(op
) && (epds
->events
& EPOLLEXCLUSIVE
)) {
2153 if (op
== EPOLL_CTL_MOD
)
2154 goto error_tgt_fput
;
2155 if (op
== EPOLL_CTL_ADD
&& (is_file_epoll(tf
.file
) ||
2156 (epds
->events
& ~EPOLLEXCLUSIVE_OK_BITS
)))
2157 goto error_tgt_fput
;
2161 * At this point it is safe to assume that the "private_data" contains
2162 * our own data structure.
2164 ep
= f
.file
->private_data
;
2167 * When we insert an epoll file descriptor, inside another epoll file
2168 * descriptor, there is the change of creating closed loops, which are
2169 * better be handled here, than in more critical paths. While we are
2170 * checking for loops we also determine the list of files reachable
2171 * and hang them on the tfile_check_list, so we can check that we
2172 * haven't created too many possible wakeup paths.
2174 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2175 * the epoll file descriptor is attaching directly to a wakeup source,
2176 * unless the epoll file descriptor is nested. The purpose of taking the
2177 * 'epmutex' on add is to prevent complex toplogies such as loops and
2178 * deep wakeup paths from forming in parallel through multiple
2179 * EPOLL_CTL_ADD operations.
2181 error
= epoll_mutex_lock(&ep
->mtx
, 0, nonblock
);
2183 goto error_tgt_fput
;
2184 if (op
== EPOLL_CTL_ADD
) {
2185 if (!list_empty(&f
.file
->f_ep_links
) ||
2186 ep
->gen
== loop_check_gen
||
2187 is_file_epoll(tf
.file
)) {
2188 mutex_unlock(&ep
->mtx
);
2189 error
= epoll_mutex_lock(&epmutex
, 0, nonblock
);
2191 goto error_tgt_fput
;
2194 if (is_file_epoll(tf
.file
)) {
2196 if (ep_loop_check(ep
, tf
.file
) != 0)
2197 goto error_tgt_fput
;
2200 list_add(&tf
.file
->f_tfile_llink
,
2203 error
= epoll_mutex_lock(&ep
->mtx
, 0, nonblock
);
2205 goto error_tgt_fput
;
2206 if (is_file_epoll(tf
.file
)) {
2207 tep
= tf
.file
->private_data
;
2208 error
= epoll_mutex_lock(&tep
->mtx
, 1, nonblock
);
2210 mutex_unlock(&ep
->mtx
);
2211 goto error_tgt_fput
;
2218 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2219 * above, we can be sure to be able to use the item looked up by
2220 * ep_find() till we release the mutex.
2222 epi
= ep_find(ep
, tf
.file
, fd
);
2228 epds
->events
|= EPOLLERR
| EPOLLHUP
;
2229 error
= ep_insert(ep
, epds
, tf
.file
, fd
, full_check
);
2235 error
= ep_remove(ep
, epi
);
2241 if (!(epi
->event
.events
& EPOLLEXCLUSIVE
)) {
2242 epds
->events
|= EPOLLERR
| EPOLLHUP
;
2243 error
= ep_modify(ep
, epi
, epds
);
2250 mutex_unlock(&tep
->mtx
);
2251 mutex_unlock(&ep
->mtx
);
2255 clear_tfile_check_list();
2257 mutex_unlock(&epmutex
);
2269 * The following function implements the controller interface for
2270 * the eventpoll file that enables the insertion/removal/change of
2271 * file descriptors inside the interest set.
2273 SYSCALL_DEFINE4(epoll_ctl
, int, epfd
, int, op
, int, fd
,
2274 struct epoll_event __user
*, event
)
2276 struct epoll_event epds
;
2278 if (ep_op_has_event(op
) &&
2279 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
2282 return do_epoll_ctl(epfd
, op
, fd
, &epds
, false);
2286 * Implement the event wait interface for the eventpoll file. It is the kernel
2287 * part of the user space epoll_wait(2).
2289 static int do_epoll_wait(int epfd
, struct epoll_event __user
*events
,
2290 int maxevents
, int timeout
)
2294 struct eventpoll
*ep
;
2296 /* The maximum number of event must be greater than zero */
2297 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
2300 /* Verify that the area passed by the user is writeable */
2301 if (!access_ok(events
, maxevents
* sizeof(struct epoll_event
)))
2304 /* Get the "struct file *" for the eventpoll file */
2310 * We have to check that the file structure underneath the fd
2311 * the user passed to us _is_ an eventpoll file.
2314 if (!is_file_epoll(f
.file
))
2318 * At this point it is safe to assume that the "private_data" contains
2319 * our own data structure.
2321 ep
= f
.file
->private_data
;
2323 /* Time to fish for events ... */
2324 error
= ep_poll(ep
, events
, maxevents
, timeout
);
2331 SYSCALL_DEFINE4(epoll_wait
, int, epfd
, struct epoll_event __user
*, events
,
2332 int, maxevents
, int, timeout
)
2334 return do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2338 * Implement the event wait interface for the eventpoll file. It is the kernel
2339 * part of the user space epoll_pwait(2).
2341 SYSCALL_DEFINE6(epoll_pwait
, int, epfd
, struct epoll_event __user
*, events
,
2342 int, maxevents
, int, timeout
, const sigset_t __user
*, sigmask
,
2348 * If the caller wants a certain signal mask to be set during the wait,
2351 error
= set_user_sigmask(sigmask
, sigsetsize
);
2355 error
= do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2356 restore_saved_sigmask_unless(error
== -EINTR
);
2361 #ifdef CONFIG_COMPAT
2362 COMPAT_SYSCALL_DEFINE6(epoll_pwait
, int, epfd
,
2363 struct epoll_event __user
*, events
,
2364 int, maxevents
, int, timeout
,
2365 const compat_sigset_t __user
*, sigmask
,
2366 compat_size_t
, sigsetsize
)
2371 * If the caller wants a certain signal mask to be set during the wait,
2374 err
= set_compat_user_sigmask(sigmask
, sigsetsize
);
2378 err
= do_epoll_wait(epfd
, events
, maxevents
, timeout
);
2379 restore_saved_sigmask_unless(err
== -EINTR
);
2385 static int __init
eventpoll_init(void)
2391 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2393 max_user_watches
= (((si
.totalram
- si
.totalhigh
) / 25) << PAGE_SHIFT
) /
2395 BUG_ON(max_user_watches
< 0);
2398 * Initialize the structure used to perform epoll file descriptor
2399 * inclusion loops checks.
2401 ep_nested_calls_init(&poll_loop_ncalls
);
2404 * We can have many thousands of epitems, so prevent this from
2405 * using an extra cache line on 64-bit (and smaller) CPUs
2407 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem
) > 128);
2409 /* Allocates slab cache used to allocate "struct epitem" items */
2410 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
2411 0, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2413 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2414 pwq_cache
= kmem_cache_create("eventpoll_pwq",
2415 sizeof(struct eppoll_entry
), 0, SLAB_PANIC
|SLAB_ACCOUNT
, NULL
);
2419 fs_initcall(eventpoll_init
);