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.h>
18 #include <linux/file.h>
19 #include <linux/signal.h>
20 #include <linux/errno.h>
22 #include <linux/slab.h>
23 #include <linux/poll.h>
24 #include <linux/string.h>
25 #include <linux/list.h>
26 #include <linux/hash.h>
27 #include <linux/spinlock.h>
28 #include <linux/syscalls.h>
29 #include <linux/rbtree.h>
30 #include <linux/wait.h>
31 #include <linux/eventpoll.h>
32 #include <linux/mount.h>
33 #include <linux/bitops.h>
34 #include <linux/mutex.h>
35 #include <linux/anon_inodes.h>
36 #include <asm/uaccess.h>
37 #include <asm/system.h>
40 #include <linux/atomic.h>
44 * There are three level of locking required by epoll :
48 * 3) ep->lock (spinlock)
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a spinlock (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 (EPOLLONESHOT | EPOLLET)
93 /* Maximum number of nesting allowed inside epoll sets */
94 #define EP_MAX_NESTS 4
96 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
98 #define EP_UNACTIVE_PTR ((void *) -1L)
100 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
102 struct epoll_filefd {
108 * Structure used to track possible nested calls, for too deep recursions
111 struct nested_call_node {
112 struct list_head llink;
118 * This structure is used as collector for nested calls, to check for
119 * maximum recursion dept and loop cycles.
121 struct nested_calls {
122 struct list_head tasks_call_list;
127 * Each file descriptor added to the eventpoll interface will
128 * have an entry of this type linked to the "rbr" RB tree.
131 /* RB tree node used to link this structure to the eventpoll RB tree */
134 /* List header used to link this structure to the eventpoll ready list */
135 struct list_head rdllink;
138 * Works together "struct eventpoll"->ovflist in keeping the
139 * single linked chain of items.
143 /* The file descriptor information this item refers to */
144 struct epoll_filefd ffd;
146 /* Number of active wait queue attached to poll operations */
149 /* List containing poll wait queues */
150 struct list_head pwqlist;
152 /* The "container" of this item */
153 struct eventpoll *ep;
155 /* List header used to link this item to the "struct file" items list */
156 struct list_head fllink;
158 /* The structure that describe the interested events and the source fd */
159 struct epoll_event event;
163 * This structure is stored inside the "private_data" member of the file
164 * structure and represents the main data structure for the eventpoll
168 /* Protect the access to this structure */
172 * This mutex is used to ensure that files are not removed
173 * while epoll is using them. This is held during the event
174 * collection loop, the file cleanup path, the epoll file exit
175 * code and the ctl operations.
179 /* Wait queue used by sys_epoll_wait() */
180 wait_queue_head_t wq;
182 /* Wait queue used by file->poll() */
183 wait_queue_head_t poll_wait;
185 /* List of ready file descriptors */
186 struct list_head rdllist;
188 /* RB tree root used to store monitored fd structs */
192 * This is a single linked list that chains all the "struct epitem" that
193 * happened while transferring ready events to userspace w/out
196 struct epitem *ovflist;
198 /* The user that created the eventpoll descriptor */
199 struct user_struct *user;
203 /* used to optimize loop detection check */
205 struct list_head visited_list_link;
208 /* Wait structure used by the poll hooks */
209 struct eppoll_entry {
210 /* List header used to link this structure to the "struct epitem" */
211 struct list_head llink;
213 /* The "base" pointer is set to the container "struct epitem" */
217 * Wait queue item that will be linked to the target file wait
222 /* The wait queue head that linked the "wait" wait queue item */
223 wait_queue_head_t *whead;
226 /* Wrapper struct used by poll queueing */
232 /* Used by the ep_send_events() function as callback private data */
233 struct ep_send_events_data {
235 struct epoll_event __user *events;
239 * Configuration options available inside /proc/sys/fs/epoll/
241 /* Maximum number of epoll watched descriptors, per user */
242 static long max_user_watches __read_mostly;
245 * This mutex is used to serialize ep_free() and eventpoll_release_file().
247 static DEFINE_MUTEX(epmutex);
249 /* Used to check for epoll file descriptor inclusion loops */
250 static struct nested_calls poll_loop_ncalls;
252 /* Used for safe wake up implementation */
253 static struct nested_calls poll_safewake_ncalls;
255 /* Used to call file's f_op->poll() under the nested calls boundaries */
256 static struct nested_calls poll_readywalk_ncalls;
258 /* Slab cache used to allocate "struct epitem" */
259 static struct kmem_cache *epi_cache __read_mostly;
261 /* Slab cache used to allocate "struct eppoll_entry" */
262 static struct kmem_cache *pwq_cache __read_mostly;
264 /* Visited nodes during ep_loop_check(), so we can unset them when we finish */
265 static LIST_HEAD(visited_list);
268 * List of files with newly added links, where we may need to limit the number
269 * of emanating paths. Protected by the epmutex.
271 static LIST_HEAD(tfile_check_list);
275 #include <linux/sysctl.h>
278 static long long_max = LONG_MAX;
280 ctl_table epoll_table[] = {
282 .procname = "max_user_watches",
283 .data = &max_user_watches,
284 .maxlen = sizeof(max_user_watches),
286 .proc_handler = proc_doulongvec_minmax,
292 #endif /* CONFIG_SYSCTL */
294 static const struct file_operations eventpoll_fops;
296 static inline int is_file_epoll(struct file *f)
298 return f->f_op == &eventpoll_fops;
301 /* Setup the structure that is used as key for the RB tree */
302 static inline void ep_set_ffd(struct epoll_filefd *ffd,
303 struct file *file, int fd)
309 /* Compare RB tree keys */
310 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
311 struct epoll_filefd *p2)
313 return (p1->file > p2->file ? +1:
314 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
317 /* Tells us if the item is currently linked */
318 static inline int ep_is_linked(struct list_head *p)
320 return !list_empty(p);
323 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
325 return container_of(p, struct eppoll_entry, wait);
328 /* Get the "struct epitem" from a wait queue pointer */
329 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
331 return container_of(p, struct eppoll_entry, wait)->base;
334 /* Get the "struct epitem" from an epoll queue wrapper */
335 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
337 return container_of(p, struct ep_pqueue, pt)->epi;
340 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
341 static inline int ep_op_has_event(int op)
343 return op != EPOLL_CTL_DEL;
346 /* Initialize the poll safe wake up structure */
347 static void ep_nested_calls_init(struct nested_calls *ncalls)
349 INIT_LIST_HEAD(&ncalls->tasks_call_list);
350 spin_lock_init(&ncalls->lock);
354 * ep_events_available - Checks if ready events might be available.
356 * @ep: Pointer to the eventpoll context.
358 * Returns: Returns a value different than zero if ready events are available,
361 static inline int ep_events_available(struct eventpoll *ep)
363 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
367 * ep_call_nested - Perform a bound (possibly) nested call, by checking
368 * that the recursion limit is not exceeded, and that
369 * the same nested call (by the meaning of same cookie) is
372 * @ncalls: Pointer to the nested_calls structure to be used for this call.
373 * @max_nests: Maximum number of allowed nesting calls.
374 * @nproc: Nested call core function pointer.
375 * @priv: Opaque data to be passed to the @nproc callback.
376 * @cookie: Cookie to be used to identify this nested call.
377 * @ctx: This instance context.
379 * Returns: Returns the code returned by the @nproc callback, or -1 if
380 * the maximum recursion limit has been exceeded.
382 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
383 int (*nproc)(void *, void *, int), void *priv,
384 void *cookie, void *ctx)
386 int error, call_nests = 0;
388 struct list_head *lsthead = &ncalls->tasks_call_list;
389 struct nested_call_node *tncur;
390 struct nested_call_node tnode;
392 spin_lock_irqsave(&ncalls->lock, flags);
395 * Try to see if the current task is already inside this wakeup call.
396 * We use a list here, since the population inside this set is always
399 list_for_each_entry(tncur, lsthead, llink) {
400 if (tncur->ctx == ctx &&
401 (tncur->cookie == cookie || ++call_nests > max_nests)) {
403 * Ops ... loop detected or maximum nest level reached.
404 * We abort this wake by breaking the cycle itself.
411 /* Add the current task and cookie to the list */
413 tnode.cookie = cookie;
414 list_add(&tnode.llink, lsthead);
416 spin_unlock_irqrestore(&ncalls->lock, flags);
418 /* Call the nested function */
419 error = (*nproc)(priv, cookie, call_nests);
421 /* Remove the current task from the list */
422 spin_lock_irqsave(&ncalls->lock, flags);
423 list_del(&tnode.llink);
425 spin_unlock_irqrestore(&ncalls->lock, flags);
430 #ifdef CONFIG_DEBUG_LOCK_ALLOC
431 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
432 unsigned long events, int subclass)
436 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
437 wake_up_locked_poll(wqueue, events);
438 spin_unlock_irqrestore(&wqueue->lock, flags);
441 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
442 unsigned long events, int subclass)
444 wake_up_poll(wqueue, events);
448 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
450 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
456 * Perform a safe wake up of the poll wait list. The problem is that
457 * with the new callback'd wake up system, it is possible that the
458 * poll callback is reentered from inside the call to wake_up() done
459 * on the poll wait queue head. The rule is that we cannot reenter the
460 * wake up code from the same task more than EP_MAX_NESTS times,
461 * and we cannot reenter the same wait queue head at all. This will
462 * enable to have a hierarchy of epoll file descriptor of no more than
465 static void ep_poll_safewake(wait_queue_head_t *wq)
467 int this_cpu = get_cpu();
469 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
470 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
475 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
477 wait_queue_head_t *whead;
481 * If it is cleared by POLLFREE, it should be rcu-safe.
482 * If we read NULL we need a barrier paired with
483 * smp_store_release() in ep_poll_callback(), otherwise
484 * we rely on whead->lock.
486 whead = ACCESS_ONCE(pwq->whead);
489 remove_wait_queue(whead, &pwq->wait);
494 * This function unregisters poll callbacks from the associated file
495 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
498 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
500 struct list_head *lsthead = &epi->pwqlist;
501 struct eppoll_entry *pwq;
503 while (!list_empty(lsthead)) {
504 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
506 list_del(&pwq->llink);
507 ep_remove_wait_queue(pwq);
508 kmem_cache_free(pwq_cache, pwq);
513 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
514 * the scan code, to call f_op->poll(). Also allows for
515 * O(NumReady) performance.
517 * @ep: Pointer to the epoll private data structure.
518 * @sproc: Pointer to the scan callback.
519 * @priv: Private opaque data passed to the @sproc callback.
520 * @depth: The current depth of recursive f_op->poll calls.
522 * Returns: The same integer error code returned by the @sproc callback.
524 static int ep_scan_ready_list(struct eventpoll *ep,
525 int (*sproc)(struct eventpoll *,
526 struct list_head *, void *),
530 int error, pwake = 0;
532 struct epitem *epi, *nepi;
536 * We need to lock this because we could be hit by
537 * eventpoll_release_file() and epoll_ctl().
539 mutex_lock_nested(&ep->mtx, depth);
542 * Steal the ready list, and re-init the original one to the
543 * empty list. Also, set ep->ovflist to NULL so that events
544 * happening while looping w/out locks, are not lost. We cannot
545 * have the poll callback to queue directly on ep->rdllist,
546 * because we want the "sproc" callback to be able to do it
549 spin_lock_irqsave(&ep->lock, flags);
550 list_splice_init(&ep->rdllist, &txlist);
552 spin_unlock_irqrestore(&ep->lock, flags);
555 * Now call the callback function.
557 error = (*sproc)(ep, &txlist, priv);
559 spin_lock_irqsave(&ep->lock, flags);
561 * During the time we spent inside the "sproc" callback, some
562 * other events might have been queued by the poll callback.
563 * We re-insert them inside the main ready-list here.
565 for (nepi = ep->ovflist; (epi = nepi) != NULL;
566 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
568 * We need to check if the item is already in the list.
569 * During the "sproc" callback execution time, items are
570 * queued into ->ovflist but the "txlist" might already
571 * contain them, and the list_splice() below takes care of them.
573 if (!ep_is_linked(&epi->rdllink))
574 list_add_tail(&epi->rdllink, &ep->rdllist);
577 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
578 * releasing the lock, events will be queued in the normal way inside
581 ep->ovflist = EP_UNACTIVE_PTR;
584 * Quickly re-inject items left on "txlist".
586 list_splice(&txlist, &ep->rdllist);
588 if (!list_empty(&ep->rdllist)) {
590 * Wake up (if active) both the eventpoll wait list and
591 * the ->poll() wait list (delayed after we release the lock).
593 if (waitqueue_active(&ep->wq))
594 wake_up_locked(&ep->wq);
595 if (waitqueue_active(&ep->poll_wait))
598 spin_unlock_irqrestore(&ep->lock, flags);
600 mutex_unlock(&ep->mtx);
602 /* We have to call this outside the lock */
604 ep_poll_safewake(&ep->poll_wait);
610 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
611 * all the associated resources. Must be called with "mtx" held.
613 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
616 struct file *file = epi->ffd.file;
619 * Removes poll wait queue hooks. We _have_ to do this without holding
620 * the "ep->lock" otherwise a deadlock might occur. This because of the
621 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
622 * queue head lock when unregistering the wait queue. The wakeup callback
623 * will run by holding the wait queue head lock and will call our callback
624 * that will try to get "ep->lock".
626 ep_unregister_pollwait(ep, epi);
628 /* Remove the current item from the list of epoll hooks */
629 spin_lock(&file->f_lock);
630 if (ep_is_linked(&epi->fllink))
631 list_del_init(&epi->fllink);
632 spin_unlock(&file->f_lock);
634 rb_erase(&epi->rbn, &ep->rbr);
636 spin_lock_irqsave(&ep->lock, flags);
637 if (ep_is_linked(&epi->rdllink))
638 list_del_init(&epi->rdllink);
639 spin_unlock_irqrestore(&ep->lock, flags);
641 /* At this point it is safe to free the eventpoll item */
642 kmem_cache_free(epi_cache, epi);
644 atomic_long_dec(&ep->user->epoll_watches);
649 static void ep_free(struct eventpoll *ep)
654 /* We need to release all tasks waiting for these file */
655 if (waitqueue_active(&ep->poll_wait))
656 ep_poll_safewake(&ep->poll_wait);
659 * We need to lock this because we could be hit by
660 * eventpoll_release_file() while we're freeing the "struct eventpoll".
661 * We do not need to hold "ep->mtx" here because the epoll file
662 * is on the way to be removed and no one has references to it
663 * anymore. The only hit might come from eventpoll_release_file() but
664 * holding "epmutex" is sufficient here.
666 mutex_lock(&epmutex);
669 * Walks through the whole tree by unregistering poll callbacks.
671 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
672 epi = rb_entry(rbp, struct epitem, rbn);
674 ep_unregister_pollwait(ep, epi);
678 * Walks through the whole tree by freeing each "struct epitem". At this
679 * point we are sure no poll callbacks will be lingering around, and also by
680 * holding "epmutex" we can be sure that no file cleanup code will hit
681 * us during this operation. So we can avoid the lock on "ep->lock".
683 while ((rbp = rb_first(&ep->rbr)) != NULL) {
684 epi = rb_entry(rbp, struct epitem, rbn);
688 mutex_unlock(&epmutex);
689 mutex_destroy(&ep->mtx);
694 static int ep_eventpoll_release(struct inode *inode, struct file *file)
696 struct eventpoll *ep = file->private_data;
704 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
707 struct epitem *epi, *tmp;
709 list_for_each_entry_safe(epi, tmp, head, rdllink) {
710 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
712 return POLLIN | POLLRDNORM;
715 * Item has been dropped into the ready list by the poll
716 * callback, but it's not actually ready, as far as
717 * caller requested events goes. We can remove it here.
719 list_del_init(&epi->rdllink);
726 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
728 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
731 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
734 struct eventpoll *ep = file->private_data;
736 /* Insert inside our poll wait queue */
737 poll_wait(file, &ep->poll_wait, wait);
740 * Proceed to find out if wanted events are really available inside
741 * the ready list. This need to be done under ep_call_nested()
742 * supervision, since the call to f_op->poll() done on listed files
743 * could re-enter here.
745 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
746 ep_poll_readyevents_proc, ep, ep, current);
748 return pollflags != -1 ? pollflags : 0;
751 /* File callbacks that implement the eventpoll file behaviour */
752 static const struct file_operations eventpoll_fops = {
753 .release = ep_eventpoll_release,
754 .poll = ep_eventpoll_poll,
755 .llseek = noop_llseek,
759 * This is called from eventpoll_release() to unlink files from the eventpoll
760 * interface. We need to have this facility to cleanup correctly files that are
761 * closed without being removed from the eventpoll interface.
763 void eventpoll_release_file(struct file *file)
765 struct list_head *lsthead = &file->f_ep_links;
766 struct eventpoll *ep;
770 * We don't want to get "file->f_lock" because it is not
771 * necessary. It is not necessary because we're in the "struct file"
772 * cleanup path, and this means that no one is using this file anymore.
773 * So, for example, epoll_ctl() cannot hit here since if we reach this
774 * point, the file counter already went to zero and fget() would fail.
775 * The only hit might come from ep_free() but by holding the mutex
776 * will correctly serialize the operation. We do need to acquire
777 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
778 * from anywhere but ep_free().
780 * Besides, ep_remove() acquires the lock, so we can't hold it here.
782 mutex_lock(&epmutex);
784 while (!list_empty(lsthead)) {
785 epi = list_first_entry(lsthead, struct epitem, fllink);
788 list_del_init(&epi->fllink);
789 mutex_lock_nested(&ep->mtx, 0);
791 mutex_unlock(&ep->mtx);
794 mutex_unlock(&epmutex);
797 static int ep_alloc(struct eventpoll **pep)
800 struct user_struct *user;
801 struct eventpoll *ep;
803 user = get_current_user();
805 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
809 spin_lock_init(&ep->lock);
810 mutex_init(&ep->mtx);
811 init_waitqueue_head(&ep->wq);
812 init_waitqueue_head(&ep->poll_wait);
813 INIT_LIST_HEAD(&ep->rdllist);
815 ep->ovflist = EP_UNACTIVE_PTR;
828 * Search the file inside the eventpoll tree. The RB tree operations
829 * are protected by the "mtx" mutex, and ep_find() must be called with
832 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
836 struct epitem *epi, *epir = NULL;
837 struct epoll_filefd ffd;
839 ep_set_ffd(&ffd, file, fd);
840 for (rbp = ep->rbr.rb_node; rbp; ) {
841 epi = rb_entry(rbp, struct epitem, rbn);
842 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
857 * This is the callback that is passed to the wait queue wakeup
858 * mechanism. It is called by the stored file descriptors when they
859 * have events to report.
861 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
865 struct epitem *epi = ep_item_from_wait(wait);
866 struct eventpoll *ep = epi->ep;
868 spin_lock_irqsave(&ep->lock, flags);
871 * If the event mask does not contain any poll(2) event, we consider the
872 * descriptor to be disabled. This condition is likely the effect of the
873 * EPOLLONESHOT bit that disables the descriptor when an event is received,
874 * until the next EPOLL_CTL_MOD will be issued.
876 if (!(epi->event.events & ~EP_PRIVATE_BITS))
880 * Check the events coming with the callback. At this stage, not
881 * every device reports the events in the "key" parameter of the
882 * callback. We need to be able to handle both cases here, hence the
883 * test for "key" != NULL before the event match test.
885 if (key && !((unsigned long) key & epi->event.events))
889 * If we are transferring events to userspace, we can hold no locks
890 * (because we're accessing user memory, and because of linux f_op->poll()
891 * semantics). All the events that happen during that period of time are
892 * chained in ep->ovflist and requeued later on.
894 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
895 if (epi->next == EP_UNACTIVE_PTR) {
896 epi->next = ep->ovflist;
902 /* If this file is already in the ready list we exit soon */
903 if (!ep_is_linked(&epi->rdllink))
904 list_add_tail(&epi->rdllink, &ep->rdllist);
907 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
910 if (waitqueue_active(&ep->wq))
911 wake_up_locked(&ep->wq);
912 if (waitqueue_active(&ep->poll_wait))
916 spin_unlock_irqrestore(&ep->lock, flags);
918 /* We have to call this outside the lock */
920 ep_poll_safewake(&ep->poll_wait);
922 if ((unsigned long)key & POLLFREE) {
924 * If we race with ep_remove_wait_queue() it can miss
925 * ->whead = NULL and do another remove_wait_queue() after
926 * us, so we can't use __remove_wait_queue().
928 list_del_init(&wait->task_list);
930 * ->whead != NULL protects us from the race with ep_free()
931 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
932 * held by the caller. Once we nullify it, nothing protects
933 * ep/epi or even wait.
936 ACCESS_ONCE(ep_pwq_from_wait(wait)->whead) = NULL;
943 * This is the callback that is used to add our wait queue to the
944 * target file wakeup lists.
946 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
949 struct epitem *epi = ep_item_from_epqueue(pt);
950 struct eppoll_entry *pwq;
952 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
953 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
956 add_wait_queue(whead, &pwq->wait);
957 list_add_tail(&pwq->llink, &epi->pwqlist);
960 /* We have to signal that an error occurred */
965 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
968 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
973 epic = rb_entry(parent, struct epitem, rbn);
974 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
976 p = &parent->rb_right;
978 p = &parent->rb_left;
980 rb_link_node(&epi->rbn, parent, p);
981 rb_insert_color(&epi->rbn, &ep->rbr);
986 #define PATH_ARR_SIZE 5
988 * These are the number paths of length 1 to 5, that we are allowing to emanate
989 * from a single file of interest. For example, we allow 1000 paths of length
990 * 1, to emanate from each file of interest. This essentially represents the
991 * potential wakeup paths, which need to be limited in order to avoid massive
992 * uncontrolled wakeup storms. The common use case should be a single ep which
993 * is connected to n file sources. In this case each file source has 1 path
994 * of length 1. Thus, the numbers below should be more than sufficient. These
995 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
996 * and delete can't add additional paths. Protected by the epmutex.
998 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
999 static int path_count[PATH_ARR_SIZE];
1001 static int path_count_inc(int nests)
1003 /* Allow an arbitrary number of depth 1 paths */
1007 if (++path_count[nests] > path_limits[nests])
1012 static void path_count_init(void)
1016 for (i = 0; i < PATH_ARR_SIZE; i++)
1020 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1023 struct file *file = priv;
1024 struct file *child_file;
1027 list_for_each_entry(epi, &file->f_ep_links, fllink) {
1028 child_file = epi->ep->file;
1029 if (is_file_epoll(child_file)) {
1030 if (list_empty(&child_file->f_ep_links)) {
1031 if (path_count_inc(call_nests)) {
1036 error = ep_call_nested(&poll_loop_ncalls,
1038 reverse_path_check_proc,
1039 child_file, child_file,
1045 printk(KERN_ERR "reverse_path_check_proc: "
1046 "file is not an ep!\n");
1053 * reverse_path_check - The tfile_check_list is list of file *, which have
1054 * links that are proposed to be newly added. We need to
1055 * make sure that those added links don't add too many
1056 * paths such that we will spend all our time waking up
1057 * eventpoll objects.
1059 * Returns: Returns zero if the proposed links don't create too many paths,
1062 static int reverse_path_check(void)
1066 struct file *current_file;
1068 /* let's call this for all tfiles */
1069 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1072 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1073 reverse_path_check_proc, current_file,
1074 current_file, current);
1082 * Must be called with "mtx" held.
1084 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1085 struct file *tfile, int fd)
1087 int error, revents, pwake = 0;
1088 unsigned long flags;
1091 struct ep_pqueue epq;
1093 user_watches = atomic_long_read(&ep->user->epoll_watches);
1094 if (unlikely(user_watches >= max_user_watches))
1096 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1099 /* Item initialization follow here ... */
1100 INIT_LIST_HEAD(&epi->rdllink);
1101 INIT_LIST_HEAD(&epi->fllink);
1102 INIT_LIST_HEAD(&epi->pwqlist);
1104 ep_set_ffd(&epi->ffd, tfile, fd);
1105 epi->event = *event;
1107 epi->next = EP_UNACTIVE_PTR;
1109 /* Initialize the poll table using the queue callback */
1111 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1114 * Attach the item to the poll hooks and get current event bits.
1115 * We can safely use the file* here because its usage count has
1116 * been increased by the caller of this function. Note that after
1117 * this operation completes, the poll callback can start hitting
1120 revents = tfile->f_op->poll(tfile, &epq.pt);
1123 * We have to check if something went wrong during the poll wait queue
1124 * install process. Namely an allocation for a wait queue failed due
1125 * high memory pressure.
1129 goto error_unregister;
1131 /* Add the current item to the list of active epoll hook for this file */
1132 spin_lock(&tfile->f_lock);
1133 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1134 spin_unlock(&tfile->f_lock);
1137 * Add the current item to the RB tree. All RB tree operations are
1138 * protected by "mtx", and ep_insert() is called with "mtx" held.
1140 ep_rbtree_insert(ep, epi);
1142 /* now check if we've created too many backpaths */
1144 if (reverse_path_check())
1145 goto error_remove_epi;
1147 /* We have to drop the new item inside our item list to keep track of it */
1148 spin_lock_irqsave(&ep->lock, flags);
1150 /* If the file is already "ready" we drop it inside the ready list */
1151 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1152 list_add_tail(&epi->rdllink, &ep->rdllist);
1154 /* Notify waiting tasks that events are available */
1155 if (waitqueue_active(&ep->wq))
1156 wake_up_locked(&ep->wq);
1157 if (waitqueue_active(&ep->poll_wait))
1161 spin_unlock_irqrestore(&ep->lock, flags);
1163 atomic_long_inc(&ep->user->epoll_watches);
1165 /* We have to call this outside the lock */
1167 ep_poll_safewake(&ep->poll_wait);
1172 spin_lock(&tfile->f_lock);
1173 if (ep_is_linked(&epi->fllink))
1174 list_del_init(&epi->fllink);
1175 spin_unlock(&tfile->f_lock);
1177 rb_erase(&epi->rbn, &ep->rbr);
1180 ep_unregister_pollwait(ep, epi);
1183 * We need to do this because an event could have been arrived on some
1184 * allocated wait queue. Note that we don't care about the ep->ovflist
1185 * list, since that is used/cleaned only inside a section bound by "mtx".
1186 * And ep_insert() is called with "mtx" held.
1188 spin_lock_irqsave(&ep->lock, flags);
1189 if (ep_is_linked(&epi->rdllink))
1190 list_del_init(&epi->rdllink);
1191 spin_unlock_irqrestore(&ep->lock, flags);
1193 kmem_cache_free(epi_cache, epi);
1199 * Modify the interest event mask by dropping an event if the new mask
1200 * has a match in the current file status. Must be called with "mtx" held.
1202 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1205 unsigned int revents;
1208 * Set the new event interest mask before calling f_op->poll();
1209 * otherwise we might miss an event that happens between the
1210 * f_op->poll() call and the new event set registering.
1212 epi->event.events = event->events; /* need barrier below */
1213 epi->event.data = event->data; /* protected by mtx */
1216 * The following barrier has two effects:
1218 * 1) Flush epi changes above to other CPUs. This ensures
1219 * we do not miss events from ep_poll_callback if an
1220 * event occurs immediately after we call f_op->poll().
1221 * We need this because we did not take ep->lock while
1222 * changing epi above (but ep_poll_callback does take
1225 * 2) We also need to ensure we do not miss _past_ events
1226 * when calling f_op->poll(). This barrier also
1227 * pairs with the barrier in wq_has_sleeper (see
1228 * comments for wq_has_sleeper).
1230 * This barrier will now guarantee ep_poll_callback or f_op->poll
1231 * (or both) will notice the readiness of an item.
1236 * Get current event bits. We can safely use the file* here because
1237 * its usage count has been increased by the caller of this function.
1239 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1242 * If the item is "hot" and it is not registered inside the ready
1243 * list, push it inside.
1245 if (revents & event->events) {
1246 spin_lock_irq(&ep->lock);
1247 if (!ep_is_linked(&epi->rdllink)) {
1248 list_add_tail(&epi->rdllink, &ep->rdllist);
1250 /* Notify waiting tasks that events are available */
1251 if (waitqueue_active(&ep->wq))
1252 wake_up_locked(&ep->wq);
1253 if (waitqueue_active(&ep->poll_wait))
1256 spin_unlock_irq(&ep->lock);
1259 /* We have to call this outside the lock */
1261 ep_poll_safewake(&ep->poll_wait);
1266 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1269 struct ep_send_events_data *esed = priv;
1271 unsigned int revents;
1273 struct epoll_event __user *uevent;
1276 * We can loop without lock because we are passed a task private list.
1277 * Items cannot vanish during the loop because ep_scan_ready_list() is
1278 * holding "mtx" during this call.
1280 for (eventcnt = 0, uevent = esed->events;
1281 !list_empty(head) && eventcnt < esed->maxevents;) {
1282 epi = list_first_entry(head, struct epitem, rdllink);
1284 list_del_init(&epi->rdllink);
1286 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1290 * If the event mask intersect the caller-requested one,
1291 * deliver the event to userspace. Again, ep_scan_ready_list()
1292 * is holding "mtx", so no operations coming from userspace
1293 * can change the item.
1296 if (__put_user(revents, &uevent->events) ||
1297 __put_user(epi->event.data, &uevent->data)) {
1298 list_add(&epi->rdllink, head);
1299 return eventcnt ? eventcnt : -EFAULT;
1303 if (epi->event.events & EPOLLONESHOT)
1304 epi->event.events &= EP_PRIVATE_BITS;
1305 else if (!(epi->event.events & EPOLLET)) {
1307 * If this file has been added with Level
1308 * Trigger mode, we need to insert back inside
1309 * the ready list, so that the next call to
1310 * epoll_wait() will check again the events
1311 * availability. At this point, no one can insert
1312 * into ep->rdllist besides us. The epoll_ctl()
1313 * callers are locked out by
1314 * ep_scan_ready_list() holding "mtx" and the
1315 * poll callback will queue them in ep->ovflist.
1317 list_add_tail(&epi->rdllink, &ep->rdllist);
1325 static int ep_send_events(struct eventpoll *ep,
1326 struct epoll_event __user *events, int maxevents)
1328 struct ep_send_events_data esed;
1330 esed.maxevents = maxevents;
1331 esed.events = events;
1333 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
1336 static inline struct timespec ep_set_mstimeout(long ms)
1338 struct timespec now, ts = {
1339 .tv_sec = ms / MSEC_PER_SEC,
1340 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1344 return timespec_add_safe(now, ts);
1348 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1351 * @ep: Pointer to the eventpoll context.
1352 * @events: Pointer to the userspace buffer where the ready events should be
1354 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1355 * @timeout: Maximum timeout for the ready events fetch operation, in
1356 * milliseconds. If the @timeout is zero, the function will not block,
1357 * while if the @timeout is less than zero, the function will block
1358 * until at least one event has been retrieved (or an error
1361 * Returns: Returns the number of ready events which have been fetched, or an
1362 * error code, in case of error.
1364 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1365 int maxevents, long timeout)
1367 int res = 0, eavail, timed_out = 0;
1368 unsigned long flags;
1371 ktime_t expires, *to = NULL;
1374 struct timespec end_time = ep_set_mstimeout(timeout);
1376 slack = select_estimate_accuracy(&end_time);
1378 *to = timespec_to_ktime(end_time);
1379 } else if (timeout == 0) {
1381 * Avoid the unnecessary trip to the wait queue loop, if the
1382 * caller specified a non blocking operation.
1385 spin_lock_irqsave(&ep->lock, flags);
1390 spin_lock_irqsave(&ep->lock, flags);
1392 if (!ep_events_available(ep)) {
1394 * We don't have any available event to return to the caller.
1395 * We need to sleep here, and we will be wake up by
1396 * ep_poll_callback() when events will become available.
1398 init_waitqueue_entry(&wait, current);
1399 __add_wait_queue_exclusive(&ep->wq, &wait);
1403 * We don't want to sleep if the ep_poll_callback() sends us
1404 * a wakeup in between. That's why we set the task state
1405 * to TASK_INTERRUPTIBLE before doing the checks.
1407 set_current_state(TASK_INTERRUPTIBLE);
1408 if (ep_events_available(ep) || timed_out)
1410 if (signal_pending(current)) {
1415 spin_unlock_irqrestore(&ep->lock, flags);
1416 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1419 spin_lock_irqsave(&ep->lock, flags);
1421 __remove_wait_queue(&ep->wq, &wait);
1423 set_current_state(TASK_RUNNING);
1426 /* Is it worth to try to dig for events ? */
1427 eavail = ep_events_available(ep);
1429 spin_unlock_irqrestore(&ep->lock, flags);
1432 * Try to transfer events to user space. In case we get 0 events and
1433 * there's still timeout left over, we go trying again in search of
1436 if (!res && eavail &&
1437 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1444 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1445 * API, to verify that adding an epoll file inside another
1446 * epoll structure, does not violate the constraints, in
1447 * terms of closed loops, or too deep chains (which can
1448 * result in excessive stack usage).
1450 * @priv: Pointer to the epoll file to be currently checked.
1451 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1452 * data structure pointer.
1453 * @call_nests: Current dept of the @ep_call_nested() call stack.
1455 * Returns: Returns zero if adding the epoll @file inside current epoll
1456 * structure @ep does not violate the constraints, or -1 otherwise.
1458 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1461 struct file *file = priv;
1462 struct eventpoll *ep = file->private_data;
1463 struct eventpoll *ep_tovisit;
1464 struct rb_node *rbp;
1467 mutex_lock_nested(&ep->mtx, call_nests + 1);
1469 list_add(&ep->visited_list_link, &visited_list);
1470 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1471 epi = rb_entry(rbp, struct epitem, rbn);
1472 if (unlikely(is_file_epoll(epi->ffd.file))) {
1473 ep_tovisit = epi->ffd.file->private_data;
1474 if (ep_tovisit->visited)
1476 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1477 ep_loop_check_proc, epi->ffd.file,
1478 ep_tovisit, current);
1483 * If we've reached a file that is not associated with
1484 * an ep, then we need to check if the newly added
1485 * links are going to add too many wakeup paths. We do
1486 * this by adding it to the tfile_check_list, if it's
1487 * not already there, and calling reverse_path_check()
1488 * during ep_insert().
1490 if (list_empty(&epi->ffd.file->f_tfile_llink))
1491 list_add(&epi->ffd.file->f_tfile_llink,
1495 mutex_unlock(&ep->mtx);
1501 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1502 * another epoll file (represented by @ep) does not create
1503 * closed loops or too deep chains.
1505 * @ep: Pointer to the epoll private data structure.
1506 * @file: Pointer to the epoll file to be checked.
1508 * Returns: Returns zero if adding the epoll @file inside current epoll
1509 * structure @ep does not violate the constraints, or -1 otherwise.
1511 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1514 struct eventpoll *ep_cur, *ep_next;
1516 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1517 ep_loop_check_proc, file, ep, current);
1518 /* clear visited list */
1519 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1520 visited_list_link) {
1521 ep_cur->visited = 0;
1522 list_del(&ep_cur->visited_list_link);
1527 static void clear_tfile_check_list(void)
1531 /* first clear the tfile_check_list */
1532 while (!list_empty(&tfile_check_list)) {
1533 file = list_first_entry(&tfile_check_list, struct file,
1535 list_del_init(&file->f_tfile_llink);
1537 INIT_LIST_HEAD(&tfile_check_list);
1541 * Open an eventpoll file descriptor.
1543 SYSCALL_DEFINE1(epoll_create1, int, flags)
1546 struct eventpoll *ep = NULL;
1549 /* Check the EPOLL_* constant for consistency. */
1550 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1552 if (flags & ~EPOLL_CLOEXEC)
1555 * Create the internal data structure ("struct eventpoll").
1557 error = ep_alloc(&ep);
1561 * Creates all the items needed to setup an eventpoll file. That is,
1562 * a file structure and a free file descriptor.
1564 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1569 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1570 O_RDWR | (flags & O_CLOEXEC));
1572 error = PTR_ERR(file);
1575 fd_install(fd, file);
1586 SYSCALL_DEFINE1(epoll_create, int, size)
1591 return sys_epoll_create1(0);
1595 * The following function implements the controller interface for
1596 * the eventpoll file that enables the insertion/removal/change of
1597 * file descriptors inside the interest set.
1599 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1600 struct epoll_event __user *, event)
1603 int did_lock_epmutex = 0;
1604 struct file *file, *tfile;
1605 struct eventpoll *ep;
1607 struct epoll_event epds;
1610 if (ep_op_has_event(op) &&
1611 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1614 /* Get the "struct file *" for the eventpoll file */
1620 /* Get the "struct file *" for the target file */
1625 /* The target file descriptor must support poll */
1627 if (!tfile->f_op || !tfile->f_op->poll)
1628 goto error_tgt_fput;
1631 * We have to check that the file structure underneath the file descriptor
1632 * the user passed to us _is_ an eventpoll file. And also we do not permit
1633 * adding an epoll file descriptor inside itself.
1636 if (file == tfile || !is_file_epoll(file))
1637 goto error_tgt_fput;
1640 * At this point it is safe to assume that the "private_data" contains
1641 * our own data structure.
1643 ep = file->private_data;
1646 * When we insert an epoll file descriptor, inside another epoll file
1647 * descriptor, there is the change of creating closed loops, which are
1648 * better be handled here, than in more critical paths. While we are
1649 * checking for loops we also determine the list of files reachable
1650 * and hang them on the tfile_check_list, so we can check that we
1651 * haven't created too many possible wakeup paths.
1653 * We need to hold the epmutex across both ep_insert and ep_remove
1654 * b/c we want to make sure we are looking at a coherent view of
1657 if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
1658 mutex_lock(&epmutex);
1659 did_lock_epmutex = 1;
1661 if (op == EPOLL_CTL_ADD) {
1662 if (is_file_epoll(tfile)) {
1664 if (ep_loop_check(ep, tfile) != 0) {
1665 clear_tfile_check_list();
1666 goto error_tgt_fput;
1669 list_add(&tfile->f_tfile_llink, &tfile_check_list);
1672 mutex_lock_nested(&ep->mtx, 0);
1675 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1676 * above, we can be sure to be able to use the item looked up by
1677 * ep_find() till we release the mutex.
1679 epi = ep_find(ep, tfile, fd);
1685 epds.events |= POLLERR | POLLHUP;
1686 error = ep_insert(ep, &epds, tfile, fd);
1689 clear_tfile_check_list();
1693 error = ep_remove(ep, epi);
1699 epds.events |= POLLERR | POLLHUP;
1700 error = ep_modify(ep, epi, &epds);
1705 mutex_unlock(&ep->mtx);
1708 if (did_lock_epmutex)
1709 mutex_unlock(&epmutex);
1720 * Implement the event wait interface for the eventpoll file. It is the kernel
1721 * part of the user space epoll_wait(2).
1723 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1724 int, maxevents, int, timeout)
1728 struct eventpoll *ep;
1730 /* The maximum number of event must be greater than zero */
1731 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1734 /* Verify that the area passed by the user is writeable */
1735 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1740 /* Get the "struct file *" for the eventpoll file */
1747 * We have to check that the file structure underneath the fd
1748 * the user passed to us _is_ an eventpoll file.
1751 if (!is_file_epoll(file))
1755 * At this point it is safe to assume that the "private_data" contains
1756 * our own data structure.
1758 ep = file->private_data;
1760 /* Time to fish for events ... */
1761 error = ep_poll(ep, events, maxevents, timeout);
1770 #ifdef HAVE_SET_RESTORE_SIGMASK
1773 * Implement the event wait interface for the eventpoll file. It is the kernel
1774 * part of the user space epoll_pwait(2).
1776 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1777 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1781 sigset_t ksigmask, sigsaved;
1784 * If the caller wants a certain signal mask to be set during the wait,
1788 if (sigsetsize != sizeof(sigset_t))
1790 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1792 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1793 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1796 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1799 * If we changed the signal mask, we need to restore the original one.
1800 * In case we've got a signal while waiting, we do not restore the
1801 * signal mask yet, and we allow do_signal() to deliver the signal on
1802 * the way back to userspace, before the signal mask is restored.
1805 if (error == -EINTR) {
1806 memcpy(¤t->saved_sigmask, &sigsaved,
1808 set_restore_sigmask();
1810 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1816 #endif /* HAVE_SET_RESTORE_SIGMASK */
1818 static int __init eventpoll_init(void)
1824 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1826 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1828 BUG_ON(max_user_watches < 0);
1831 * Initialize the structure used to perform epoll file descriptor
1832 * inclusion loops checks.
1834 ep_nested_calls_init(&poll_loop_ncalls);
1836 /* Initialize the structure used to perform safe poll wait head wake ups */
1837 ep_nested_calls_init(&poll_safewake_ncalls);
1839 /* Initialize the structure used to perform file's f_op->poll() calls */
1840 ep_nested_calls_init(&poll_readywalk_ncalls);
1842 /* Allocates slab cache used to allocate "struct epitem" items */
1843 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1844 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1846 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1847 pwq_cache = kmem_cache_create("eventpoll_pwq",
1848 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1852 fs_initcall(eventpoll_init);