xfs: fix type usage
[linux/fpc-iii.git] / fs / eventpoll.c
blob2fabd19cdeea76e32f51f3c84348a22b9fe22843
1 /*
2 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * Davide Libenzi <davidel@xmailserver.org>
14 #include <linux/init.h>
15 #include <linux/kernel.h>
16 #include <linux/sched/signal.h>
17 #include <linux/fs.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
21 #include <linux/mm.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/device.h>
37 #include <linux/uaccess.h>
38 #include <asm/io.h>
39 #include <asm/mman.h>
40 #include <linux/atomic.h>
41 #include <linux/proc_fs.h>
42 #include <linux/seq_file.h>
43 #include <linux/compat.h>
44 #include <linux/rculist.h>
45 #include <net/busy_poll.h>
48 * LOCKING:
49 * There are three level of locking required by epoll :
51 * 1) epmutex (mutex)
52 * 2) ep->mtx (mutex)
53 * 3) ep->lock (spinlock)
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
63 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
66 * and ep_free().
67 * This mutex is acquired by ep_free() during the epoll file
68 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
77 * going to.
78 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
86 * the lockdep subkey.
87 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
90 * Events that require holding "epmutex" are very rare, while for
91 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
95 /* Epoll private bits inside the event mask */
96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
98 #define EPOLLINOUT_BITS (POLLIN | POLLOUT)
100 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | POLLERR | POLLHUP | \
101 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
103 /* Maximum number of nesting allowed inside epoll sets */
104 #define EP_MAX_NESTS 4
106 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
108 #define EP_UNACTIVE_PTR ((void *) -1L)
110 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
112 struct epoll_filefd {
113 struct file *file;
114 int fd;
115 } __packed;
118 * Structure used to track possible nested calls, for too deep recursions
119 * and loop cycles.
121 struct nested_call_node {
122 struct list_head llink;
123 void *cookie;
124 void *ctx;
128 * This structure is used as collector for nested calls, to check for
129 * maximum recursion dept and loop cycles.
131 struct nested_calls {
132 struct list_head tasks_call_list;
133 spinlock_t lock;
137 * Each file descriptor added to the eventpoll interface will
138 * have an entry of this type linked to the "rbr" RB tree.
139 * Avoid increasing the size of this struct, there can be many thousands
140 * of these on a server and we do not want this to take another cache line.
142 struct epitem {
143 union {
144 /* RB tree node links this structure to the eventpoll RB tree */
145 struct rb_node rbn;
146 /* Used to free the struct epitem */
147 struct rcu_head rcu;
150 /* List header used to link this structure to the eventpoll ready list */
151 struct list_head rdllink;
154 * Works together "struct eventpoll"->ovflist in keeping the
155 * single linked chain of items.
157 struct epitem *next;
159 /* The file descriptor information this item refers to */
160 struct epoll_filefd ffd;
162 /* Number of active wait queue attached to poll operations */
163 int nwait;
165 /* List containing poll wait queues */
166 struct list_head pwqlist;
168 /* The "container" of this item */
169 struct eventpoll *ep;
171 /* List header used to link this item to the "struct file" items list */
172 struct list_head fllink;
174 /* wakeup_source used when EPOLLWAKEUP is set */
175 struct wakeup_source __rcu *ws;
177 /* The structure that describe the interested events and the source fd */
178 struct epoll_event event;
182 * This structure is stored inside the "private_data" member of the file
183 * structure and represents the main data structure for the eventpoll
184 * interface.
186 struct eventpoll {
187 /* Protect the access to this structure */
188 spinlock_t lock;
191 * This mutex is used to ensure that files are not removed
192 * while epoll is using them. This is held during the event
193 * collection loop, the file cleanup path, the epoll file exit
194 * code and the ctl operations.
196 struct mutex mtx;
198 /* Wait queue used by sys_epoll_wait() */
199 wait_queue_head_t wq;
201 /* Wait queue used by file->poll() */
202 wait_queue_head_t poll_wait;
204 /* List of ready file descriptors */
205 struct list_head rdllist;
207 /* RB tree root used to store monitored fd structs */
208 struct rb_root_cached rbr;
211 * This is a single linked list that chains all the "struct epitem" that
212 * happened while transferring ready events to userspace w/out
213 * holding ->lock.
215 struct epitem *ovflist;
217 /* wakeup_source used when ep_scan_ready_list is running */
218 struct wakeup_source *ws;
220 /* The user that created the eventpoll descriptor */
221 struct user_struct *user;
223 struct file *file;
225 /* used to optimize loop detection check */
226 int visited;
227 struct list_head visited_list_link;
229 #ifdef CONFIG_NET_RX_BUSY_POLL
230 /* used to track busy poll napi_id */
231 unsigned int napi_id;
232 #endif
235 /* Wait structure used by the poll hooks */
236 struct eppoll_entry {
237 /* List header used to link this structure to the "struct epitem" */
238 struct list_head llink;
240 /* The "base" pointer is set to the container "struct epitem" */
241 struct epitem *base;
244 * Wait queue item that will be linked to the target file wait
245 * queue head.
247 wait_queue_entry_t wait;
249 /* The wait queue head that linked the "wait" wait queue item */
250 wait_queue_head_t *whead;
253 /* Wrapper struct used by poll queueing */
254 struct ep_pqueue {
255 poll_table pt;
256 struct epitem *epi;
259 /* Used by the ep_send_events() function as callback private data */
260 struct ep_send_events_data {
261 int maxevents;
262 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 /* Used to check for epoll file descriptor inclusion loops */
277 static struct nested_calls poll_loop_ncalls;
279 /* Used for safe wake up implementation */
280 static struct nested_calls poll_safewake_ncalls;
282 /* Used to call file's f_op->poll() under the nested calls boundaries */
283 static struct nested_calls poll_readywalk_ncalls;
285 /* Slab cache used to allocate "struct epitem" */
286 static struct kmem_cache *epi_cache __read_mostly;
288 /* Slab cache used to allocate "struct eppoll_entry" */
289 static struct kmem_cache *pwq_cache __read_mostly;
291 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
292 static LIST_HEAD(visited_list);
295 * List of files with newly added links, where we may need to limit the number
296 * of emanating paths. Protected by the epmutex.
298 static LIST_HEAD(tfile_check_list);
300 #ifdef CONFIG_SYSCTL
302 #include <linux/sysctl.h>
304 static long zero;
305 static long long_max = LONG_MAX;
307 struct ctl_table epoll_table[] = {
309 .procname = "max_user_watches",
310 .data = &max_user_watches,
311 .maxlen = sizeof(max_user_watches),
312 .mode = 0644,
313 .proc_handler = proc_doulongvec_minmax,
314 .extra1 = &zero,
315 .extra2 = &long_max,
319 #endif /* CONFIG_SYSCTL */
321 static const struct file_operations eventpoll_fops;
323 static inline int is_file_epoll(struct file *f)
325 return f->f_op == &eventpoll_fops;
328 /* Setup the structure that is used as key for the RB tree */
329 static inline void ep_set_ffd(struct epoll_filefd *ffd,
330 struct file *file, int fd)
332 ffd->file = file;
333 ffd->fd = fd;
336 /* Compare RB tree keys */
337 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
338 struct epoll_filefd *p2)
340 return (p1->file > p2->file ? +1:
341 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
344 /* Tells us if the item is currently linked */
345 static inline int ep_is_linked(struct list_head *p)
347 return !list_empty(p);
350 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
352 return container_of(p, struct eppoll_entry, wait);
355 /* Get the "struct epitem" from a wait queue pointer */
356 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
358 return container_of(p, struct eppoll_entry, wait)->base;
361 /* Get the "struct epitem" from an epoll queue wrapper */
362 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
364 return container_of(p, struct ep_pqueue, pt)->epi;
367 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
368 static inline int ep_op_has_event(int op)
370 return op != EPOLL_CTL_DEL;
373 /* Initialize the poll safe wake up structure */
374 static void ep_nested_calls_init(struct nested_calls *ncalls)
376 INIT_LIST_HEAD(&ncalls->tasks_call_list);
377 spin_lock_init(&ncalls->lock);
381 * ep_events_available - Checks if ready events might be available.
383 * @ep: Pointer to the eventpoll context.
385 * Returns: Returns a value different than zero if ready events are available,
386 * or zero otherwise.
388 static inline int ep_events_available(struct eventpoll *ep)
390 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
393 #ifdef CONFIG_NET_RX_BUSY_POLL
394 static bool ep_busy_loop_end(void *p, unsigned long start_time)
396 struct eventpoll *ep = p;
398 return ep_events_available(ep) || busy_loop_timeout(start_time);
400 #endif /* CONFIG_NET_RX_BUSY_POLL */
403 * Busy poll if globally on and supporting sockets found && no events,
404 * busy loop will return if need_resched or ep_events_available.
406 * we must do our busy polling with irqs enabled
408 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
410 #ifdef CONFIG_NET_RX_BUSY_POLL
411 unsigned int napi_id = READ_ONCE(ep->napi_id);
413 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
414 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
415 #endif
418 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
420 #ifdef CONFIG_NET_RX_BUSY_POLL
421 if (ep->napi_id)
422 ep->napi_id = 0;
423 #endif
427 * Set epoll busy poll NAPI ID from sk.
429 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
431 #ifdef CONFIG_NET_RX_BUSY_POLL
432 struct eventpoll *ep;
433 unsigned int napi_id;
434 struct socket *sock;
435 struct sock *sk;
436 int err;
438 if (!net_busy_loop_on())
439 return;
441 sock = sock_from_file(epi->ffd.file, &err);
442 if (!sock)
443 return;
445 sk = sock->sk;
446 if (!sk)
447 return;
449 napi_id = READ_ONCE(sk->sk_napi_id);
450 ep = epi->ep;
452 /* Non-NAPI IDs can be rejected
453 * or
454 * Nothing to do if we already have this ID
456 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
457 return;
459 /* record NAPI ID for use in next busy poll */
460 ep->napi_id = napi_id;
461 #endif
465 * ep_call_nested - Perform a bound (possibly) nested call, by checking
466 * that the recursion limit is not exceeded, and that
467 * the same nested call (by the meaning of same cookie) is
468 * no re-entered.
470 * @ncalls: Pointer to the nested_calls structure to be used for this call.
471 * @max_nests: Maximum number of allowed nesting calls.
472 * @nproc: Nested call core function pointer.
473 * @priv: Opaque data to be passed to the @nproc callback.
474 * @cookie: Cookie to be used to identify this nested call.
475 * @ctx: This instance context.
477 * Returns: Returns the code returned by the @nproc callback, or -1 if
478 * the maximum recursion limit has been exceeded.
480 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
481 int (*nproc)(void *, void *, int), void *priv,
482 void *cookie, void *ctx)
484 int error, call_nests = 0;
485 unsigned long flags;
486 struct list_head *lsthead = &ncalls->tasks_call_list;
487 struct nested_call_node *tncur;
488 struct nested_call_node tnode;
490 spin_lock_irqsave(&ncalls->lock, flags);
493 * Try to see if the current task is already inside this wakeup call.
494 * We use a list here, since the population inside this set is always
495 * very much limited.
497 list_for_each_entry(tncur, lsthead, llink) {
498 if (tncur->ctx == ctx &&
499 (tncur->cookie == cookie || ++call_nests > max_nests)) {
501 * Ops ... loop detected or maximum nest level reached.
502 * We abort this wake by breaking the cycle itself.
504 error = -1;
505 goto out_unlock;
509 /* Add the current task and cookie to the list */
510 tnode.ctx = ctx;
511 tnode.cookie = cookie;
512 list_add(&tnode.llink, lsthead);
514 spin_unlock_irqrestore(&ncalls->lock, flags);
516 /* Call the nested function */
517 error = (*nproc)(priv, cookie, call_nests);
519 /* Remove the current task from the list */
520 spin_lock_irqsave(&ncalls->lock, flags);
521 list_del(&tnode.llink);
522 out_unlock:
523 spin_unlock_irqrestore(&ncalls->lock, flags);
525 return error;
529 * As described in commit 0ccf831cb lockdep: annotate epoll
530 * the use of wait queues used by epoll is done in a very controlled
531 * manner. Wake ups can nest inside each other, but are never done
532 * with the same locking. For example:
534 * dfd = socket(...);
535 * efd1 = epoll_create();
536 * efd2 = epoll_create();
537 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
538 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
540 * When a packet arrives to the device underneath "dfd", the net code will
541 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
542 * callback wakeup entry on that queue, and the wake_up() performed by the
543 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
544 * (efd1) notices that it may have some event ready, so it needs to wake up
545 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
546 * that ends up in another wake_up(), after having checked about the
547 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
548 * avoid stack blasting.
550 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
551 * this special case of epoll.
553 #ifdef CONFIG_DEBUG_LOCK_ALLOC
554 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
555 unsigned long events, int subclass)
557 unsigned long flags;
559 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
560 wake_up_locked_poll(wqueue, events);
561 spin_unlock_irqrestore(&wqueue->lock, flags);
563 #else
564 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
565 unsigned long events, int subclass)
567 wake_up_poll(wqueue, events);
569 #endif
571 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
573 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
574 1 + call_nests);
575 return 0;
579 * Perform a safe wake up of the poll wait list. The problem is that
580 * with the new callback'd wake up system, it is possible that the
581 * poll callback is reentered from inside the call to wake_up() done
582 * on the poll wait queue head. The rule is that we cannot reenter the
583 * wake up code from the same task more than EP_MAX_NESTS times,
584 * and we cannot reenter the same wait queue head at all. This will
585 * enable to have a hierarchy of epoll file descriptor of no more than
586 * EP_MAX_NESTS deep.
588 static void ep_poll_safewake(wait_queue_head_t *wq)
590 int this_cpu = get_cpu();
592 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
593 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
595 put_cpu();
598 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
600 wait_queue_head_t *whead;
602 rcu_read_lock();
604 * If it is cleared by POLLFREE, it should be rcu-safe.
605 * If we read NULL we need a barrier paired with
606 * smp_store_release() in ep_poll_callback(), otherwise
607 * we rely on whead->lock.
609 whead = smp_load_acquire(&pwq->whead);
610 if (whead)
611 remove_wait_queue(whead, &pwq->wait);
612 rcu_read_unlock();
616 * This function unregisters poll callbacks from the associated file
617 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
618 * ep_free).
620 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
622 struct list_head *lsthead = &epi->pwqlist;
623 struct eppoll_entry *pwq;
625 while (!list_empty(lsthead)) {
626 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
628 list_del(&pwq->llink);
629 ep_remove_wait_queue(pwq);
630 kmem_cache_free(pwq_cache, pwq);
634 /* call only when ep->mtx is held */
635 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
637 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
640 /* call only when ep->mtx is held */
641 static inline void ep_pm_stay_awake(struct epitem *epi)
643 struct wakeup_source *ws = ep_wakeup_source(epi);
645 if (ws)
646 __pm_stay_awake(ws);
649 static inline bool ep_has_wakeup_source(struct epitem *epi)
651 return rcu_access_pointer(epi->ws) ? true : false;
654 /* call when ep->mtx cannot be held (ep_poll_callback) */
655 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
657 struct wakeup_source *ws;
659 rcu_read_lock();
660 ws = rcu_dereference(epi->ws);
661 if (ws)
662 __pm_stay_awake(ws);
663 rcu_read_unlock();
667 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
668 * the scan code, to call f_op->poll(). Also allows for
669 * O(NumReady) performance.
671 * @ep: Pointer to the epoll private data structure.
672 * @sproc: Pointer to the scan callback.
673 * @priv: Private opaque data passed to the @sproc callback.
674 * @depth: The current depth of recursive f_op->poll calls.
675 * @ep_locked: caller already holds ep->mtx
677 * Returns: The same integer error code returned by the @sproc callback.
679 static int ep_scan_ready_list(struct eventpoll *ep,
680 int (*sproc)(struct eventpoll *,
681 struct list_head *, void *),
682 void *priv, int depth, bool ep_locked)
684 int error, pwake = 0;
685 unsigned long flags;
686 struct epitem *epi, *nepi;
687 LIST_HEAD(txlist);
690 * We need to lock this because we could be hit by
691 * eventpoll_release_file() and epoll_ctl().
694 if (!ep_locked)
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
703 * in a lockless way.
705 spin_lock_irqsave(&ep->lock, flags);
706 list_splice_init(&ep->rdllist, &txlist);
707 ep->ovflist = NULL;
708 spin_unlock_irqrestore(&ep->lock, flags);
711 * Now call the callback function.
713 error = (*sproc)(ep, &txlist, priv);
715 spin_lock_irqsave(&ep->lock, flags);
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 = 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->rdllink)) {
730 list_add_tail(&epi->rdllink, &ep->rdllist);
731 ep_pm_stay_awake(epi);
735 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
736 * releasing the lock, events will be queued in the normal way inside
737 * ep->rdllist.
739 ep->ovflist = EP_UNACTIVE_PTR;
742 * Quickly re-inject items left on "txlist".
744 list_splice(&txlist, &ep->rdllist);
745 __pm_relax(ep->ws);
747 if (!list_empty(&ep->rdllist)) {
749 * Wake up (if active) both the eventpoll wait list and
750 * the ->poll() wait list (delayed after we release the lock).
752 if (waitqueue_active(&ep->wq))
753 wake_up_locked(&ep->wq);
754 if (waitqueue_active(&ep->poll_wait))
755 pwake++;
757 spin_unlock_irqrestore(&ep->lock, flags);
759 if (!ep_locked)
760 mutex_unlock(&ep->mtx);
762 /* We have to call this outside the lock */
763 if (pwake)
764 ep_poll_safewake(&ep->poll_wait);
766 return error;
769 static void epi_rcu_free(struct rcu_head *head)
771 struct epitem *epi = container_of(head, struct epitem, rcu);
772 kmem_cache_free(epi_cache, epi);
776 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
777 * all the associated resources. Must be called with "mtx" held.
779 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
781 unsigned long flags;
782 struct file *file = epi->ffd.file;
785 * Removes poll wait queue hooks. We _have_ to do this without holding
786 * the "ep->lock" otherwise a deadlock might occur. This because of the
787 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
788 * queue head lock when unregistering the wait queue. The wakeup callback
789 * will run by holding the wait queue head lock and will call our callback
790 * that will try to get "ep->lock".
792 ep_unregister_pollwait(ep, epi);
794 /* Remove the current item from the list of epoll hooks */
795 spin_lock(&file->f_lock);
796 list_del_rcu(&epi->fllink);
797 spin_unlock(&file->f_lock);
799 rb_erase_cached(&epi->rbn, &ep->rbr);
801 spin_lock_irqsave(&ep->lock, flags);
802 if (ep_is_linked(&epi->rdllink))
803 list_del_init(&epi->rdllink);
804 spin_unlock_irqrestore(&ep->lock, flags);
806 wakeup_source_unregister(ep_wakeup_source(epi));
808 * At this point it is safe to free the eventpoll item. Use the union
809 * field epi->rcu, since we are trying to minimize the size of
810 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
811 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
812 * use of the rbn field.
814 call_rcu(&epi->rcu, epi_rcu_free);
816 atomic_long_dec(&ep->user->epoll_watches);
818 return 0;
821 static void ep_free(struct eventpoll *ep)
823 struct rb_node *rbp;
824 struct epitem *epi;
826 /* We need to release all tasks waiting for these file */
827 if (waitqueue_active(&ep->poll_wait))
828 ep_poll_safewake(&ep->poll_wait);
831 * We need to lock this because we could be hit by
832 * eventpoll_release_file() while we're freeing the "struct eventpoll".
833 * We do not need to hold "ep->mtx" here because the epoll file
834 * is on the way to be removed and no one has references to it
835 * anymore. The only hit might come from eventpoll_release_file() but
836 * holding "epmutex" is sufficient here.
838 mutex_lock(&epmutex);
841 * Walks through the whole tree by unregistering poll callbacks.
843 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
844 epi = rb_entry(rbp, struct epitem, rbn);
846 ep_unregister_pollwait(ep, epi);
847 cond_resched();
851 * Walks through the whole tree by freeing each "struct epitem". At this
852 * point we are sure no poll callbacks will be lingering around, and also by
853 * holding "epmutex" we can be sure that no file cleanup code will hit
854 * us during this operation. So we can avoid the lock on "ep->lock".
855 * We do not need to lock ep->mtx, either, we only do it to prevent
856 * a lockdep warning.
858 mutex_lock(&ep->mtx);
859 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
860 epi = rb_entry(rbp, struct epitem, rbn);
861 ep_remove(ep, epi);
862 cond_resched();
864 mutex_unlock(&ep->mtx);
866 mutex_unlock(&epmutex);
867 mutex_destroy(&ep->mtx);
868 free_uid(ep->user);
869 wakeup_source_unregister(ep->ws);
870 kfree(ep);
873 static int ep_eventpoll_release(struct inode *inode, struct file *file)
875 struct eventpoll *ep = file->private_data;
877 if (ep)
878 ep_free(ep);
880 return 0;
883 static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
885 pt->_key = epi->event.events;
887 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
890 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
891 void *priv)
893 struct epitem *epi, *tmp;
894 poll_table pt;
896 init_poll_funcptr(&pt, NULL);
898 list_for_each_entry_safe(epi, tmp, head, rdllink) {
899 if (ep_item_poll(epi, &pt))
900 return POLLIN | POLLRDNORM;
901 else {
903 * Item has been dropped into the ready list by the poll
904 * callback, but it's not actually ready, as far as
905 * caller requested events goes. We can remove it here.
907 __pm_relax(ep_wakeup_source(epi));
908 list_del_init(&epi->rdllink);
912 return 0;
915 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
916 poll_table *pt);
918 struct readyevents_arg {
919 struct eventpoll *ep;
920 bool locked;
923 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
925 struct readyevents_arg *arg = priv;
927 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
928 call_nests + 1, arg->locked);
931 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
933 int pollflags;
934 struct eventpoll *ep = file->private_data;
935 struct readyevents_arg arg;
938 * During ep_insert() we already hold the ep->mtx for the tfile.
939 * Prevent re-aquisition.
941 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
942 arg.ep = ep;
944 /* Insert inside our poll wait queue */
945 poll_wait(file, &ep->poll_wait, wait);
948 * Proceed to find out if wanted events are really available inside
949 * the ready list. This need to be done under ep_call_nested()
950 * supervision, since the call to f_op->poll() done on listed files
951 * could re-enter here.
953 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
954 ep_poll_readyevents_proc, &arg, ep, current);
956 return pollflags != -1 ? pollflags : 0;
959 #ifdef CONFIG_PROC_FS
960 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
962 struct eventpoll *ep = f->private_data;
963 struct rb_node *rbp;
965 mutex_lock(&ep->mtx);
966 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
967 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
968 struct inode *inode = file_inode(epi->ffd.file);
970 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
971 " pos:%lli ino:%lx sdev:%x\n",
972 epi->ffd.fd, epi->event.events,
973 (long long)epi->event.data,
974 (long long)epi->ffd.file->f_pos,
975 inode->i_ino, inode->i_sb->s_dev);
976 if (seq_has_overflowed(m))
977 break;
979 mutex_unlock(&ep->mtx);
981 #endif
983 /* File callbacks that implement the eventpoll file behaviour */
984 static const struct file_operations eventpoll_fops = {
985 #ifdef CONFIG_PROC_FS
986 .show_fdinfo = ep_show_fdinfo,
987 #endif
988 .release = ep_eventpoll_release,
989 .poll = ep_eventpoll_poll,
990 .llseek = noop_llseek,
994 * This is called from eventpoll_release() to unlink files from the eventpoll
995 * interface. We need to have this facility to cleanup correctly files that are
996 * closed without being removed from the eventpoll interface.
998 void eventpoll_release_file(struct file *file)
1000 struct eventpoll *ep;
1001 struct epitem *epi, *next;
1004 * We don't want to get "file->f_lock" because it is not
1005 * necessary. It is not necessary because we're in the "struct file"
1006 * cleanup path, and this means that no one is using this file anymore.
1007 * So, for example, epoll_ctl() cannot hit here since if we reach this
1008 * point, the file counter already went to zero and fget() would fail.
1009 * The only hit might come from ep_free() but by holding the mutex
1010 * will correctly serialize the operation. We do need to acquire
1011 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
1012 * from anywhere but ep_free().
1014 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1016 mutex_lock(&epmutex);
1017 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1018 ep = epi->ep;
1019 mutex_lock_nested(&ep->mtx, 0);
1020 ep_remove(ep, epi);
1021 mutex_unlock(&ep->mtx);
1023 mutex_unlock(&epmutex);
1026 static int ep_alloc(struct eventpoll **pep)
1028 int error;
1029 struct user_struct *user;
1030 struct eventpoll *ep;
1032 user = get_current_user();
1033 error = -ENOMEM;
1034 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1035 if (unlikely(!ep))
1036 goto free_uid;
1038 spin_lock_init(&ep->lock);
1039 mutex_init(&ep->mtx);
1040 init_waitqueue_head(&ep->wq);
1041 init_waitqueue_head(&ep->poll_wait);
1042 INIT_LIST_HEAD(&ep->rdllist);
1043 ep->rbr = RB_ROOT_CACHED;
1044 ep->ovflist = EP_UNACTIVE_PTR;
1045 ep->user = user;
1047 *pep = ep;
1049 return 0;
1051 free_uid:
1052 free_uid(user);
1053 return error;
1057 * Search the file inside the eventpoll tree. The RB tree operations
1058 * are protected by the "mtx" mutex, and ep_find() must be called with
1059 * "mtx" held.
1061 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1063 int kcmp;
1064 struct rb_node *rbp;
1065 struct epitem *epi, *epir = NULL;
1066 struct epoll_filefd ffd;
1068 ep_set_ffd(&ffd, file, fd);
1069 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1070 epi = rb_entry(rbp, struct epitem, rbn);
1071 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1072 if (kcmp > 0)
1073 rbp = rbp->rb_right;
1074 else if (kcmp < 0)
1075 rbp = rbp->rb_left;
1076 else {
1077 epir = epi;
1078 break;
1082 return epir;
1085 #ifdef CONFIG_CHECKPOINT_RESTORE
1086 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1088 struct rb_node *rbp;
1089 struct epitem *epi;
1091 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1092 epi = rb_entry(rbp, struct epitem, rbn);
1093 if (epi->ffd.fd == tfd) {
1094 if (toff == 0)
1095 return epi;
1096 else
1097 toff--;
1099 cond_resched();
1102 return NULL;
1105 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1106 unsigned long toff)
1108 struct file *file_raw;
1109 struct eventpoll *ep;
1110 struct epitem *epi;
1112 if (!is_file_epoll(file))
1113 return ERR_PTR(-EINVAL);
1115 ep = file->private_data;
1117 mutex_lock(&ep->mtx);
1118 epi = ep_find_tfd(ep, tfd, toff);
1119 if (epi)
1120 file_raw = epi->ffd.file;
1121 else
1122 file_raw = ERR_PTR(-ENOENT);
1123 mutex_unlock(&ep->mtx);
1125 return file_raw;
1127 #endif /* CONFIG_CHECKPOINT_RESTORE */
1130 * This is the callback that is passed to the wait queue wakeup
1131 * mechanism. It is called by the stored file descriptors when they
1132 * have events to report.
1134 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1136 int pwake = 0;
1137 unsigned long flags;
1138 struct epitem *epi = ep_item_from_wait(wait);
1139 struct eventpoll *ep = epi->ep;
1140 int ewake = 0;
1142 spin_lock_irqsave(&ep->lock, flags);
1144 ep_set_busy_poll_napi_id(epi);
1147 * If the event mask does not contain any poll(2) event, we consider the
1148 * descriptor to be disabled. This condition is likely the effect of the
1149 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1150 * until the next EPOLL_CTL_MOD will be issued.
1152 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1153 goto out_unlock;
1156 * Check the events coming with the callback. At this stage, not
1157 * every device reports the events in the "key" parameter of the
1158 * callback. We need to be able to handle both cases here, hence the
1159 * test for "key" != NULL before the event match test.
1161 if (key && !((unsigned long) key & epi->event.events))
1162 goto out_unlock;
1165 * If we are transferring events to userspace, we can hold no locks
1166 * (because we're accessing user memory, and because of linux f_op->poll()
1167 * semantics). All the events that happen during that period of time are
1168 * chained in ep->ovflist and requeued later on.
1170 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1171 if (epi->next == EP_UNACTIVE_PTR) {
1172 epi->next = ep->ovflist;
1173 ep->ovflist = epi;
1174 if (epi->ws) {
1176 * Activate ep->ws since epi->ws may get
1177 * deactivated at any time.
1179 __pm_stay_awake(ep->ws);
1183 goto out_unlock;
1186 /* If this file is already in the ready list we exit soon */
1187 if (!ep_is_linked(&epi->rdllink)) {
1188 list_add_tail(&epi->rdllink, &ep->rdllist);
1189 ep_pm_stay_awake_rcu(epi);
1193 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1194 * wait list.
1196 if (waitqueue_active(&ep->wq)) {
1197 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1198 !((unsigned long)key & POLLFREE)) {
1199 switch ((unsigned long)key & EPOLLINOUT_BITS) {
1200 case POLLIN:
1201 if (epi->event.events & POLLIN)
1202 ewake = 1;
1203 break;
1204 case POLLOUT:
1205 if (epi->event.events & POLLOUT)
1206 ewake = 1;
1207 break;
1208 case 0:
1209 ewake = 1;
1210 break;
1213 wake_up_locked(&ep->wq);
1215 if (waitqueue_active(&ep->poll_wait))
1216 pwake++;
1218 out_unlock:
1219 spin_unlock_irqrestore(&ep->lock, flags);
1221 /* We have to call this outside the lock */
1222 if (pwake)
1223 ep_poll_safewake(&ep->poll_wait);
1225 if (!(epi->event.events & EPOLLEXCLUSIVE))
1226 ewake = 1;
1228 if ((unsigned long)key & POLLFREE) {
1230 * If we race with ep_remove_wait_queue() it can miss
1231 * ->whead = NULL and do another remove_wait_queue() after
1232 * us, so we can't use __remove_wait_queue().
1234 list_del_init(&wait->entry);
1236 * ->whead != NULL protects us from the race with ep_free()
1237 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1238 * held by the caller. Once we nullify it, nothing protects
1239 * ep/epi or even wait.
1241 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1244 return ewake;
1248 * This is the callback that is used to add our wait queue to the
1249 * target file wakeup lists.
1251 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1252 poll_table *pt)
1254 struct epitem *epi = ep_item_from_epqueue(pt);
1255 struct eppoll_entry *pwq;
1257 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1258 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1259 pwq->whead = whead;
1260 pwq->base = epi;
1261 if (epi->event.events & EPOLLEXCLUSIVE)
1262 add_wait_queue_exclusive(whead, &pwq->wait);
1263 else
1264 add_wait_queue(whead, &pwq->wait);
1265 list_add_tail(&pwq->llink, &epi->pwqlist);
1266 epi->nwait++;
1267 } else {
1268 /* We have to signal that an error occurred */
1269 epi->nwait = -1;
1273 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1275 int kcmp;
1276 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1277 struct epitem *epic;
1278 bool leftmost = true;
1280 while (*p) {
1281 parent = *p;
1282 epic = rb_entry(parent, struct epitem, rbn);
1283 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1284 if (kcmp > 0) {
1285 p = &parent->rb_right;
1286 leftmost = false;
1287 } else
1288 p = &parent->rb_left;
1290 rb_link_node(&epi->rbn, parent, p);
1291 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1296 #define PATH_ARR_SIZE 5
1298 * These are the number paths of length 1 to 5, that we are allowing to emanate
1299 * from a single file of interest. For example, we allow 1000 paths of length
1300 * 1, to emanate from each file of interest. This essentially represents the
1301 * potential wakeup paths, which need to be limited in order to avoid massive
1302 * uncontrolled wakeup storms. The common use case should be a single ep which
1303 * is connected to n file sources. In this case each file source has 1 path
1304 * of length 1. Thus, the numbers below should be more than sufficient. These
1305 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1306 * and delete can't add additional paths. Protected by the epmutex.
1308 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1309 static int path_count[PATH_ARR_SIZE];
1311 static int path_count_inc(int nests)
1313 /* Allow an arbitrary number of depth 1 paths */
1314 if (nests == 0)
1315 return 0;
1317 if (++path_count[nests] > path_limits[nests])
1318 return -1;
1319 return 0;
1322 static void path_count_init(void)
1324 int i;
1326 for (i = 0; i < PATH_ARR_SIZE; i++)
1327 path_count[i] = 0;
1330 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1332 int error = 0;
1333 struct file *file = priv;
1334 struct file *child_file;
1335 struct epitem *epi;
1337 /* CTL_DEL can remove links here, but that can't increase our count */
1338 rcu_read_lock();
1339 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1340 child_file = epi->ep->file;
1341 if (is_file_epoll(child_file)) {
1342 if (list_empty(&child_file->f_ep_links)) {
1343 if (path_count_inc(call_nests)) {
1344 error = -1;
1345 break;
1347 } else {
1348 error = ep_call_nested(&poll_loop_ncalls,
1349 EP_MAX_NESTS,
1350 reverse_path_check_proc,
1351 child_file, child_file,
1352 current);
1354 if (error != 0)
1355 break;
1356 } else {
1357 printk(KERN_ERR "reverse_path_check_proc: "
1358 "file is not an ep!\n");
1361 rcu_read_unlock();
1362 return error;
1366 * reverse_path_check - The tfile_check_list is list of file *, which have
1367 * links that are proposed to be newly added. We need to
1368 * make sure that those added links don't add too many
1369 * paths such that we will spend all our time waking up
1370 * eventpoll objects.
1372 * Returns: Returns zero if the proposed links don't create too many paths,
1373 * -1 otherwise.
1375 static int reverse_path_check(void)
1377 int error = 0;
1378 struct file *current_file;
1380 /* let's call this for all tfiles */
1381 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1382 path_count_init();
1383 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1384 reverse_path_check_proc, current_file,
1385 current_file, current);
1386 if (error)
1387 break;
1389 return error;
1392 static int ep_create_wakeup_source(struct epitem *epi)
1394 const char *name;
1395 struct wakeup_source *ws;
1397 if (!epi->ep->ws) {
1398 epi->ep->ws = wakeup_source_register("eventpoll");
1399 if (!epi->ep->ws)
1400 return -ENOMEM;
1403 name = epi->ffd.file->f_path.dentry->d_name.name;
1404 ws = wakeup_source_register(name);
1406 if (!ws)
1407 return -ENOMEM;
1408 rcu_assign_pointer(epi->ws, ws);
1410 return 0;
1413 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1414 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1416 struct wakeup_source *ws = ep_wakeup_source(epi);
1418 RCU_INIT_POINTER(epi->ws, NULL);
1421 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1422 * used internally by wakeup_source_remove, too (called by
1423 * wakeup_source_unregister), so we cannot use call_rcu
1425 synchronize_rcu();
1426 wakeup_source_unregister(ws);
1430 * Must be called with "mtx" held.
1432 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1433 struct file *tfile, int fd, int full_check)
1435 int error, revents, pwake = 0;
1436 unsigned long flags;
1437 long user_watches;
1438 struct epitem *epi;
1439 struct ep_pqueue epq;
1441 user_watches = atomic_long_read(&ep->user->epoll_watches);
1442 if (unlikely(user_watches >= max_user_watches))
1443 return -ENOSPC;
1444 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1445 return -ENOMEM;
1447 /* Item initialization follow here ... */
1448 INIT_LIST_HEAD(&epi->rdllink);
1449 INIT_LIST_HEAD(&epi->fllink);
1450 INIT_LIST_HEAD(&epi->pwqlist);
1451 epi->ep = ep;
1452 ep_set_ffd(&epi->ffd, tfile, fd);
1453 epi->event = *event;
1454 epi->nwait = 0;
1455 epi->next = EP_UNACTIVE_PTR;
1456 if (epi->event.events & EPOLLWAKEUP) {
1457 error = ep_create_wakeup_source(epi);
1458 if (error)
1459 goto error_create_wakeup_source;
1460 } else {
1461 RCU_INIT_POINTER(epi->ws, NULL);
1464 /* Initialize the poll table using the queue callback */
1465 epq.epi = epi;
1466 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1469 * Attach the item to the poll hooks and get current event bits.
1470 * We can safely use the file* here because its usage count has
1471 * been increased by the caller of this function. Note that after
1472 * this operation completes, the poll callback can start hitting
1473 * the new item.
1475 revents = ep_item_poll(epi, &epq.pt);
1478 * We have to check if something went wrong during the poll wait queue
1479 * install process. Namely an allocation for a wait queue failed due
1480 * high memory pressure.
1482 error = -ENOMEM;
1483 if (epi->nwait < 0)
1484 goto error_unregister;
1486 /* Add the current item to the list of active epoll hook for this file */
1487 spin_lock(&tfile->f_lock);
1488 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1489 spin_unlock(&tfile->f_lock);
1492 * Add the current item to the RB tree. All RB tree operations are
1493 * protected by "mtx", and ep_insert() is called with "mtx" held.
1495 ep_rbtree_insert(ep, epi);
1497 /* now check if we've created too many backpaths */
1498 error = -EINVAL;
1499 if (full_check && reverse_path_check())
1500 goto error_remove_epi;
1502 /* We have to drop the new item inside our item list to keep track of it */
1503 spin_lock_irqsave(&ep->lock, flags);
1505 /* record NAPI ID of new item if present */
1506 ep_set_busy_poll_napi_id(epi);
1508 /* If the file is already "ready" we drop it inside the ready list */
1509 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1510 list_add_tail(&epi->rdllink, &ep->rdllist);
1511 ep_pm_stay_awake(epi);
1513 /* Notify waiting tasks that events are available */
1514 if (waitqueue_active(&ep->wq))
1515 wake_up_locked(&ep->wq);
1516 if (waitqueue_active(&ep->poll_wait))
1517 pwake++;
1520 spin_unlock_irqrestore(&ep->lock, flags);
1522 atomic_long_inc(&ep->user->epoll_watches);
1524 /* We have to call this outside the lock */
1525 if (pwake)
1526 ep_poll_safewake(&ep->poll_wait);
1528 return 0;
1530 error_remove_epi:
1531 spin_lock(&tfile->f_lock);
1532 list_del_rcu(&epi->fllink);
1533 spin_unlock(&tfile->f_lock);
1535 rb_erase_cached(&epi->rbn, &ep->rbr);
1537 error_unregister:
1538 ep_unregister_pollwait(ep, epi);
1541 * We need to do this because an event could have been arrived on some
1542 * allocated wait queue. Note that we don't care about the ep->ovflist
1543 * list, since that is used/cleaned only inside a section bound by "mtx".
1544 * And ep_insert() is called with "mtx" held.
1546 spin_lock_irqsave(&ep->lock, flags);
1547 if (ep_is_linked(&epi->rdllink))
1548 list_del_init(&epi->rdllink);
1549 spin_unlock_irqrestore(&ep->lock, flags);
1551 wakeup_source_unregister(ep_wakeup_source(epi));
1553 error_create_wakeup_source:
1554 kmem_cache_free(epi_cache, epi);
1556 return error;
1560 * Modify the interest event mask by dropping an event if the new mask
1561 * has a match in the current file status. Must be called with "mtx" held.
1563 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1565 int pwake = 0;
1566 unsigned int revents;
1567 poll_table pt;
1569 init_poll_funcptr(&pt, NULL);
1572 * Set the new event interest mask before calling f_op->poll();
1573 * otherwise we might miss an event that happens between the
1574 * f_op->poll() call and the new event set registering.
1576 epi->event.events = event->events; /* need barrier below */
1577 epi->event.data = event->data; /* protected by mtx */
1578 if (epi->event.events & EPOLLWAKEUP) {
1579 if (!ep_has_wakeup_source(epi))
1580 ep_create_wakeup_source(epi);
1581 } else if (ep_has_wakeup_source(epi)) {
1582 ep_destroy_wakeup_source(epi);
1586 * The following barrier has two effects:
1588 * 1) Flush epi changes above to other CPUs. This ensures
1589 * we do not miss events from ep_poll_callback if an
1590 * event occurs immediately after we call f_op->poll().
1591 * We need this because we did not take ep->lock while
1592 * changing epi above (but ep_poll_callback does take
1593 * ep->lock).
1595 * 2) We also need to ensure we do not miss _past_ events
1596 * when calling f_op->poll(). This barrier also
1597 * pairs with the barrier in wq_has_sleeper (see
1598 * comments for wq_has_sleeper).
1600 * This barrier will now guarantee ep_poll_callback or f_op->poll
1601 * (or both) will notice the readiness of an item.
1603 smp_mb();
1606 * Get current event bits. We can safely use the file* here because
1607 * its usage count has been increased by the caller of this function.
1609 revents = ep_item_poll(epi, &pt);
1612 * If the item is "hot" and it is not registered inside the ready
1613 * list, push it inside.
1615 if (revents & event->events) {
1616 spin_lock_irq(&ep->lock);
1617 if (!ep_is_linked(&epi->rdllink)) {
1618 list_add_tail(&epi->rdllink, &ep->rdllist);
1619 ep_pm_stay_awake(epi);
1621 /* Notify waiting tasks that events are available */
1622 if (waitqueue_active(&ep->wq))
1623 wake_up_locked(&ep->wq);
1624 if (waitqueue_active(&ep->poll_wait))
1625 pwake++;
1627 spin_unlock_irq(&ep->lock);
1630 /* We have to call this outside the lock */
1631 if (pwake)
1632 ep_poll_safewake(&ep->poll_wait);
1634 return 0;
1637 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1638 void *priv)
1640 struct ep_send_events_data *esed = priv;
1641 int eventcnt;
1642 unsigned int revents;
1643 struct epitem *epi;
1644 struct epoll_event __user *uevent;
1645 struct wakeup_source *ws;
1646 poll_table pt;
1648 init_poll_funcptr(&pt, NULL);
1651 * We can loop without lock because we are passed a task private list.
1652 * Items cannot vanish during the loop because ep_scan_ready_list() is
1653 * holding "mtx" during this call.
1655 for (eventcnt = 0, uevent = esed->events;
1656 !list_empty(head) && eventcnt < esed->maxevents;) {
1657 epi = list_first_entry(head, struct epitem, rdllink);
1660 * Activate ep->ws before deactivating epi->ws to prevent
1661 * triggering auto-suspend here (in case we reactive epi->ws
1662 * below).
1664 * This could be rearranged to delay the deactivation of epi->ws
1665 * instead, but then epi->ws would temporarily be out of sync
1666 * with ep_is_linked().
1668 ws = ep_wakeup_source(epi);
1669 if (ws) {
1670 if (ws->active)
1671 __pm_stay_awake(ep->ws);
1672 __pm_relax(ws);
1675 list_del_init(&epi->rdllink);
1677 revents = ep_item_poll(epi, &pt);
1680 * If the event mask intersect the caller-requested one,
1681 * deliver the event to userspace. Again, ep_scan_ready_list()
1682 * is holding "mtx", so no operations coming from userspace
1683 * can change the item.
1685 if (revents) {
1686 if (__put_user(revents, &uevent->events) ||
1687 __put_user(epi->event.data, &uevent->data)) {
1688 list_add(&epi->rdllink, head);
1689 ep_pm_stay_awake(epi);
1690 return eventcnt ? eventcnt : -EFAULT;
1692 eventcnt++;
1693 uevent++;
1694 if (epi->event.events & EPOLLONESHOT)
1695 epi->event.events &= EP_PRIVATE_BITS;
1696 else if (!(epi->event.events & EPOLLET)) {
1698 * If this file has been added with Level
1699 * Trigger mode, we need to insert back inside
1700 * the ready list, so that the next call to
1701 * epoll_wait() will check again the events
1702 * availability. At this point, no one can insert
1703 * into ep->rdllist besides us. The epoll_ctl()
1704 * callers are locked out by
1705 * ep_scan_ready_list() holding "mtx" and the
1706 * poll callback will queue them in ep->ovflist.
1708 list_add_tail(&epi->rdllink, &ep->rdllist);
1709 ep_pm_stay_awake(epi);
1714 return eventcnt;
1717 static int ep_send_events(struct eventpoll *ep,
1718 struct epoll_event __user *events, int maxevents)
1720 struct ep_send_events_data esed;
1722 esed.maxevents = maxevents;
1723 esed.events = events;
1725 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1728 static inline struct timespec64 ep_set_mstimeout(long ms)
1730 struct timespec64 now, ts = {
1731 .tv_sec = ms / MSEC_PER_SEC,
1732 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1735 ktime_get_ts64(&now);
1736 return timespec64_add_safe(now, ts);
1740 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1741 * event buffer.
1743 * @ep: Pointer to the eventpoll context.
1744 * @events: Pointer to the userspace buffer where the ready events should be
1745 * stored.
1746 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1747 * @timeout: Maximum timeout for the ready events fetch operation, in
1748 * milliseconds. If the @timeout is zero, the function will not block,
1749 * while if the @timeout is less than zero, the function will block
1750 * until at least one event has been retrieved (or an error
1751 * occurred).
1753 * Returns: Returns the number of ready events which have been fetched, or an
1754 * error code, in case of error.
1756 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1757 int maxevents, long timeout)
1759 int res = 0, eavail, timed_out = 0;
1760 unsigned long flags;
1761 u64 slack = 0;
1762 wait_queue_entry_t wait;
1763 ktime_t expires, *to = NULL;
1765 if (timeout > 0) {
1766 struct timespec64 end_time = ep_set_mstimeout(timeout);
1768 slack = select_estimate_accuracy(&end_time);
1769 to = &expires;
1770 *to = timespec64_to_ktime(end_time);
1771 } else if (timeout == 0) {
1773 * Avoid the unnecessary trip to the wait queue loop, if the
1774 * caller specified a non blocking operation.
1776 timed_out = 1;
1777 spin_lock_irqsave(&ep->lock, flags);
1778 goto check_events;
1781 fetch_events:
1783 if (!ep_events_available(ep))
1784 ep_busy_loop(ep, timed_out);
1786 spin_lock_irqsave(&ep->lock, flags);
1788 if (!ep_events_available(ep)) {
1790 * Busy poll timed out. Drop NAPI ID for now, we can add
1791 * it back in when we have moved a socket with a valid NAPI
1792 * ID onto the ready list.
1794 ep_reset_busy_poll_napi_id(ep);
1797 * We don't have any available event to return to the caller.
1798 * We need to sleep here, and we will be wake up by
1799 * ep_poll_callback() when events will become available.
1801 init_waitqueue_entry(&wait, current);
1802 __add_wait_queue_exclusive(&ep->wq, &wait);
1804 for (;;) {
1806 * We don't want to sleep if the ep_poll_callback() sends us
1807 * a wakeup in between. That's why we set the task state
1808 * to TASK_INTERRUPTIBLE before doing the checks.
1810 set_current_state(TASK_INTERRUPTIBLE);
1812 * Always short-circuit for fatal signals to allow
1813 * threads to make a timely exit without the chance of
1814 * finding more events available and fetching
1815 * repeatedly.
1817 if (fatal_signal_pending(current)) {
1818 res = -EINTR;
1819 break;
1821 if (ep_events_available(ep) || timed_out)
1822 break;
1823 if (signal_pending(current)) {
1824 res = -EINTR;
1825 break;
1828 spin_unlock_irqrestore(&ep->lock, flags);
1829 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1830 timed_out = 1;
1832 spin_lock_irqsave(&ep->lock, flags);
1835 __remove_wait_queue(&ep->wq, &wait);
1836 __set_current_state(TASK_RUNNING);
1838 check_events:
1839 /* Is it worth to try to dig for events ? */
1840 eavail = ep_events_available(ep);
1842 spin_unlock_irqrestore(&ep->lock, flags);
1845 * Try to transfer events to user space. In case we get 0 events and
1846 * there's still timeout left over, we go trying again in search of
1847 * more luck.
1849 if (!res && eavail &&
1850 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1851 goto fetch_events;
1853 return res;
1857 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1858 * API, to verify that adding an epoll file inside another
1859 * epoll structure, does not violate the constraints, in
1860 * terms of closed loops, or too deep chains (which can
1861 * result in excessive stack usage).
1863 * @priv: Pointer to the epoll file to be currently checked.
1864 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1865 * data structure pointer.
1866 * @call_nests: Current dept of the @ep_call_nested() call stack.
1868 * Returns: Returns zero if adding the epoll @file inside current epoll
1869 * structure @ep does not violate the constraints, or -1 otherwise.
1871 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1873 int error = 0;
1874 struct file *file = priv;
1875 struct eventpoll *ep = file->private_data;
1876 struct eventpoll *ep_tovisit;
1877 struct rb_node *rbp;
1878 struct epitem *epi;
1880 mutex_lock_nested(&ep->mtx, call_nests + 1);
1881 ep->visited = 1;
1882 list_add(&ep->visited_list_link, &visited_list);
1883 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1884 epi = rb_entry(rbp, struct epitem, rbn);
1885 if (unlikely(is_file_epoll(epi->ffd.file))) {
1886 ep_tovisit = epi->ffd.file->private_data;
1887 if (ep_tovisit->visited)
1888 continue;
1889 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1890 ep_loop_check_proc, epi->ffd.file,
1891 ep_tovisit, current);
1892 if (error != 0)
1893 break;
1894 } else {
1896 * If we've reached a file that is not associated with
1897 * an ep, then we need to check if the newly added
1898 * links are going to add too many wakeup paths. We do
1899 * this by adding it to the tfile_check_list, if it's
1900 * not already there, and calling reverse_path_check()
1901 * during ep_insert().
1903 if (list_empty(&epi->ffd.file->f_tfile_llink))
1904 list_add(&epi->ffd.file->f_tfile_llink,
1905 &tfile_check_list);
1908 mutex_unlock(&ep->mtx);
1910 return error;
1914 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1915 * another epoll file (represented by @ep) does not create
1916 * closed loops or too deep chains.
1918 * @ep: Pointer to the epoll private data structure.
1919 * @file: Pointer to the epoll file to be checked.
1921 * Returns: Returns zero if adding the epoll @file inside current epoll
1922 * structure @ep does not violate the constraints, or -1 otherwise.
1924 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1926 int ret;
1927 struct eventpoll *ep_cur, *ep_next;
1929 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1930 ep_loop_check_proc, file, ep, current);
1931 /* clear visited list */
1932 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1933 visited_list_link) {
1934 ep_cur->visited = 0;
1935 list_del(&ep_cur->visited_list_link);
1937 return ret;
1940 static void clear_tfile_check_list(void)
1942 struct file *file;
1944 /* first clear the tfile_check_list */
1945 while (!list_empty(&tfile_check_list)) {
1946 file = list_first_entry(&tfile_check_list, struct file,
1947 f_tfile_llink);
1948 list_del_init(&file->f_tfile_llink);
1950 INIT_LIST_HEAD(&tfile_check_list);
1954 * Open an eventpoll file descriptor.
1956 SYSCALL_DEFINE1(epoll_create1, int, flags)
1958 int error, fd;
1959 struct eventpoll *ep = NULL;
1960 struct file *file;
1962 /* Check the EPOLL_* constant for consistency. */
1963 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1965 if (flags & ~EPOLL_CLOEXEC)
1966 return -EINVAL;
1968 * Create the internal data structure ("struct eventpoll").
1970 error = ep_alloc(&ep);
1971 if (error < 0)
1972 return error;
1974 * Creates all the items needed to setup an eventpoll file. That is,
1975 * a file structure and a free file descriptor.
1977 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1978 if (fd < 0) {
1979 error = fd;
1980 goto out_free_ep;
1982 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1983 O_RDWR | (flags & O_CLOEXEC));
1984 if (IS_ERR(file)) {
1985 error = PTR_ERR(file);
1986 goto out_free_fd;
1988 ep->file = file;
1989 fd_install(fd, file);
1990 return fd;
1992 out_free_fd:
1993 put_unused_fd(fd);
1994 out_free_ep:
1995 ep_free(ep);
1996 return error;
1999 SYSCALL_DEFINE1(epoll_create, int, size)
2001 if (size <= 0)
2002 return -EINVAL;
2004 return sys_epoll_create1(0);
2008 * The following function implements the controller interface for
2009 * the eventpoll file that enables the insertion/removal/change of
2010 * file descriptors inside the interest set.
2012 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
2013 struct epoll_event __user *, event)
2015 int error;
2016 int full_check = 0;
2017 struct fd f, tf;
2018 struct eventpoll *ep;
2019 struct epitem *epi;
2020 struct epoll_event epds;
2021 struct eventpoll *tep = NULL;
2023 error = -EFAULT;
2024 if (ep_op_has_event(op) &&
2025 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2026 goto error_return;
2028 error = -EBADF;
2029 f = fdget(epfd);
2030 if (!f.file)
2031 goto error_return;
2033 /* Get the "struct file *" for the target file */
2034 tf = fdget(fd);
2035 if (!tf.file)
2036 goto error_fput;
2038 /* The target file descriptor must support poll */
2039 error = -EPERM;
2040 if (!tf.file->f_op->poll)
2041 goto error_tgt_fput;
2043 /* Check if EPOLLWAKEUP is allowed */
2044 if (ep_op_has_event(op))
2045 ep_take_care_of_epollwakeup(&epds);
2048 * We have to check that the file structure underneath the file descriptor
2049 * the user passed to us _is_ an eventpoll file. And also we do not permit
2050 * adding an epoll file descriptor inside itself.
2052 error = -EINVAL;
2053 if (f.file == tf.file || !is_file_epoll(f.file))
2054 goto error_tgt_fput;
2057 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2058 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2059 * Also, we do not currently supported nested exclusive wakeups.
2061 if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2062 if (op == EPOLL_CTL_MOD)
2063 goto error_tgt_fput;
2064 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2065 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2066 goto error_tgt_fput;
2070 * At this point it is safe to assume that the "private_data" contains
2071 * our own data structure.
2073 ep = f.file->private_data;
2076 * When we insert an epoll file descriptor, inside another epoll file
2077 * descriptor, there is the change of creating closed loops, which are
2078 * better be handled here, than in more critical paths. While we are
2079 * checking for loops we also determine the list of files reachable
2080 * and hang them on the tfile_check_list, so we can check that we
2081 * haven't created too many possible wakeup paths.
2083 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2084 * the epoll file descriptor is attaching directly to a wakeup source,
2085 * unless the epoll file descriptor is nested. The purpose of taking the
2086 * 'epmutex' on add is to prevent complex toplogies such as loops and
2087 * deep wakeup paths from forming in parallel through multiple
2088 * EPOLL_CTL_ADD operations.
2090 mutex_lock_nested(&ep->mtx, 0);
2091 if (op == EPOLL_CTL_ADD) {
2092 if (!list_empty(&f.file->f_ep_links) ||
2093 is_file_epoll(tf.file)) {
2094 full_check = 1;
2095 mutex_unlock(&ep->mtx);
2096 mutex_lock(&epmutex);
2097 if (is_file_epoll(tf.file)) {
2098 error = -ELOOP;
2099 if (ep_loop_check(ep, tf.file) != 0) {
2100 clear_tfile_check_list();
2101 goto error_tgt_fput;
2103 } else
2104 list_add(&tf.file->f_tfile_llink,
2105 &tfile_check_list);
2106 mutex_lock_nested(&ep->mtx, 0);
2107 if (is_file_epoll(tf.file)) {
2108 tep = tf.file->private_data;
2109 mutex_lock_nested(&tep->mtx, 1);
2115 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2116 * above, we can be sure to be able to use the item looked up by
2117 * ep_find() till we release the mutex.
2119 epi = ep_find(ep, tf.file, fd);
2121 error = -EINVAL;
2122 switch (op) {
2123 case EPOLL_CTL_ADD:
2124 if (!epi) {
2125 epds.events |= POLLERR | POLLHUP;
2126 error = ep_insert(ep, &epds, tf.file, fd, full_check);
2127 } else
2128 error = -EEXIST;
2129 if (full_check)
2130 clear_tfile_check_list();
2131 break;
2132 case EPOLL_CTL_DEL:
2133 if (epi)
2134 error = ep_remove(ep, epi);
2135 else
2136 error = -ENOENT;
2137 break;
2138 case EPOLL_CTL_MOD:
2139 if (epi) {
2140 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2141 epds.events |= POLLERR | POLLHUP;
2142 error = ep_modify(ep, epi, &epds);
2144 } else
2145 error = -ENOENT;
2146 break;
2148 if (tep != NULL)
2149 mutex_unlock(&tep->mtx);
2150 mutex_unlock(&ep->mtx);
2152 error_tgt_fput:
2153 if (full_check)
2154 mutex_unlock(&epmutex);
2156 fdput(tf);
2157 error_fput:
2158 fdput(f);
2159 error_return:
2161 return error;
2165 * Implement the event wait interface for the eventpoll file. It is the kernel
2166 * part of the user space epoll_wait(2).
2168 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2169 int, maxevents, int, timeout)
2171 int error;
2172 struct fd f;
2173 struct eventpoll *ep;
2175 /* The maximum number of event must be greater than zero */
2176 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2177 return -EINVAL;
2179 /* Verify that the area passed by the user is writeable */
2180 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2181 return -EFAULT;
2183 /* Get the "struct file *" for the eventpoll file */
2184 f = fdget(epfd);
2185 if (!f.file)
2186 return -EBADF;
2189 * We have to check that the file structure underneath the fd
2190 * the user passed to us _is_ an eventpoll file.
2192 error = -EINVAL;
2193 if (!is_file_epoll(f.file))
2194 goto error_fput;
2197 * At this point it is safe to assume that the "private_data" contains
2198 * our own data structure.
2200 ep = f.file->private_data;
2202 /* Time to fish for events ... */
2203 error = ep_poll(ep, events, maxevents, timeout);
2205 error_fput:
2206 fdput(f);
2207 return error;
2211 * Implement the event wait interface for the eventpoll file. It is the kernel
2212 * part of the user space epoll_pwait(2).
2214 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2215 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2216 size_t, sigsetsize)
2218 int error;
2219 sigset_t ksigmask, sigsaved;
2222 * If the caller wants a certain signal mask to be set during the wait,
2223 * we apply it here.
2225 if (sigmask) {
2226 if (sigsetsize != sizeof(sigset_t))
2227 return -EINVAL;
2228 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2229 return -EFAULT;
2230 sigsaved = current->blocked;
2231 set_current_blocked(&ksigmask);
2234 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2237 * If we changed the signal mask, we need to restore the original one.
2238 * In case we've got a signal while waiting, we do not restore the
2239 * signal mask yet, and we allow do_signal() to deliver the signal on
2240 * the way back to userspace, before the signal mask is restored.
2242 if (sigmask) {
2243 if (error == -EINTR) {
2244 memcpy(&current->saved_sigmask, &sigsaved,
2245 sizeof(sigsaved));
2246 set_restore_sigmask();
2247 } else
2248 set_current_blocked(&sigsaved);
2251 return error;
2254 #ifdef CONFIG_COMPAT
2255 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2256 struct epoll_event __user *, events,
2257 int, maxevents, int, timeout,
2258 const compat_sigset_t __user *, sigmask,
2259 compat_size_t, sigsetsize)
2261 long err;
2262 compat_sigset_t csigmask;
2263 sigset_t ksigmask, sigsaved;
2266 * If the caller wants a certain signal mask to be set during the wait,
2267 * we apply it here.
2269 if (sigmask) {
2270 if (sigsetsize != sizeof(compat_sigset_t))
2271 return -EINVAL;
2272 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2273 return -EFAULT;
2274 sigset_from_compat(&ksigmask, &csigmask);
2275 sigsaved = current->blocked;
2276 set_current_blocked(&ksigmask);
2279 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2282 * If we changed the signal mask, we need to restore the original one.
2283 * In case we've got a signal while waiting, we do not restore the
2284 * signal mask yet, and we allow do_signal() to deliver the signal on
2285 * the way back to userspace, before the signal mask is restored.
2287 if (sigmask) {
2288 if (err == -EINTR) {
2289 memcpy(&current->saved_sigmask, &sigsaved,
2290 sizeof(sigsaved));
2291 set_restore_sigmask();
2292 } else
2293 set_current_blocked(&sigsaved);
2296 return err;
2298 #endif
2300 static int __init eventpoll_init(void)
2302 struct sysinfo si;
2304 si_meminfo(&si);
2306 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2308 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2309 EP_ITEM_COST;
2310 BUG_ON(max_user_watches < 0);
2313 * Initialize the structure used to perform epoll file descriptor
2314 * inclusion loops checks.
2316 ep_nested_calls_init(&poll_loop_ncalls);
2318 /* Initialize the structure used to perform safe poll wait head wake ups */
2319 ep_nested_calls_init(&poll_safewake_ncalls);
2321 /* Initialize the structure used to perform file's f_op->poll() calls */
2322 ep_nested_calls_init(&poll_readywalk_ncalls);
2325 * We can have many thousands of epitems, so prevent this from
2326 * using an extra cache line on 64-bit (and smaller) CPUs
2328 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2330 /* Allocates slab cache used to allocate "struct epitem" items */
2331 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2332 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2334 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2335 pwq_cache = kmem_cache_create("eventpoll_pwq",
2336 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2338 return 0;
2340 fs_initcall(eventpoll_init);