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 <asm/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 toepoll_ctl(EPOLL_CTL_DEL).
66 * It is possible to drop the "ep->mtx" and to use the global
67 * mutex "epmutex" (together with "ep->lock") to have it working,
68 * but having "ep->mtx" will make the interface more scalable.
69 * Events that require holding "epmutex" are very rare, while for
70 * normal operations the epoll private "ep->mtx" will guarantee
71 * a better scalability.
74 /* Epoll private bits inside the event mask */
75 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
77 /* Maximum number of nesting allowed inside epoll sets */
78 #define EP_MAX_NESTS 4
80 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
82 #define EP_UNACTIVE_PTR ((void *) -1L)
84 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
92 * Structure used to track possible nested calls, for too deep recursions
95 struct nested_call_node {
96 struct list_head llink;
102 * This structure is used as collector for nested calls, to check for
103 * maximum recursion dept and loop cycles.
105 struct nested_calls {
106 struct list_head tasks_call_list;
111 * Each file descriptor added to the eventpoll interface will
112 * have an entry of this type linked to the "rbr" RB tree.
115 /* RB tree node used to link this structure to the eventpoll RB tree */
118 /* List header used to link this structure to the eventpoll ready list */
119 struct list_head rdllink;
122 * Works together "struct eventpoll"->ovflist in keeping the
123 * single linked chain of items.
127 /* The file descriptor information this item refers to */
128 struct epoll_filefd ffd;
130 /* Number of active wait queue attached to poll operations */
133 /* List containing poll wait queues */
134 struct list_head pwqlist;
136 /* The "container" of this item */
137 struct eventpoll *ep;
139 /* List header used to link this item to the "struct file" items list */
140 struct list_head fllink;
142 /* The structure that describe the interested events and the source fd */
143 struct epoll_event event;
147 * This structure is stored inside the "private_data" member of the file
148 * structure and rapresent the main data sructure for the eventpoll
152 /* Protect the this structure access */
156 * This mutex is used to ensure that files are not removed
157 * while epoll is using them. This is held during the event
158 * collection loop, the file cleanup path, the epoll file exit
159 * code and the ctl operations.
163 /* Wait queue used by sys_epoll_wait() */
164 wait_queue_head_t wq;
166 /* Wait queue used by file->poll() */
167 wait_queue_head_t poll_wait;
169 /* List of ready file descriptors */
170 struct list_head rdllist;
172 /* RB tree root used to store monitored fd structs */
176 * This is a single linked list that chains all the "struct epitem" that
177 * happened while transfering ready events to userspace w/out
180 struct epitem *ovflist;
182 /* The user that created the eventpoll descriptor */
183 struct user_struct *user;
186 /* Wait structure used by the poll hooks */
187 struct eppoll_entry {
188 /* List header used to link this structure to the "struct epitem" */
189 struct list_head llink;
191 /* The "base" pointer is set to the container "struct epitem" */
195 * Wait queue item that will be linked to the target file wait
200 /* The wait queue head that linked the "wait" wait queue item */
201 wait_queue_head_t *whead;
204 /* Wrapper struct used by poll queueing */
210 /* Used by the ep_send_events() function as callback private data */
211 struct ep_send_events_data {
213 struct epoll_event __user *events;
217 * Configuration options available inside /proc/sys/fs/epoll/
219 /* Maximum number of epoll watched descriptors, per user */
220 static long max_user_watches __read_mostly;
223 * This mutex is used to serialize ep_free() and eventpoll_release_file().
225 static DEFINE_MUTEX(epmutex);
227 /* Used for safe wake up implementation */
228 static struct nested_calls poll_safewake_ncalls;
230 /* Used to call file's f_op->poll() under the nested calls boundaries */
231 static struct nested_calls poll_readywalk_ncalls;
233 /* Slab cache used to allocate "struct epitem" */
234 static struct kmem_cache *epi_cache __read_mostly;
236 /* Slab cache used to allocate "struct eppoll_entry" */
237 static struct kmem_cache *pwq_cache __read_mostly;
241 #include <linux/sysctl.h>
244 static long long_max = LONG_MAX;
246 ctl_table epoll_table[] = {
248 .procname = "max_user_watches",
249 .data = &max_user_watches,
250 .maxlen = sizeof(max_user_watches),
252 .proc_handler = proc_doulongvec_minmax,
258 #endif /* CONFIG_SYSCTL */
261 /* Setup the structure that is used as key for the RB tree */
262 static inline void ep_set_ffd(struct epoll_filefd *ffd,
263 struct file *file, int fd)
269 /* Compare RB tree keys */
270 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
271 struct epoll_filefd *p2)
273 return (p1->file > p2->file ? +1:
274 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
277 /* Tells us if the item is currently linked */
278 static inline int ep_is_linked(struct list_head *p)
280 return !list_empty(p);
283 /* Get the "struct epitem" from a wait queue pointer */
284 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
286 return container_of(p, struct eppoll_entry, wait)->base;
289 /* Get the "struct epitem" from an epoll queue wrapper */
290 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
292 return container_of(p, struct ep_pqueue, pt)->epi;
295 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
296 static inline int ep_op_has_event(int op)
298 return op != EPOLL_CTL_DEL;
301 /* Initialize the poll safe wake up structure */
302 static void ep_nested_calls_init(struct nested_calls *ncalls)
304 INIT_LIST_HEAD(&ncalls->tasks_call_list);
305 spin_lock_init(&ncalls->lock);
309 * ep_call_nested - Perform a bound (possibly) nested call, by checking
310 * that the recursion limit is not exceeded, and that
311 * the same nested call (by the meaning of same cookie) is
314 * @ncalls: Pointer to the nested_calls structure to be used for this call.
315 * @max_nests: Maximum number of allowed nesting calls.
316 * @nproc: Nested call core function pointer.
317 * @priv: Opaque data to be passed to the @nproc callback.
318 * @cookie: Cookie to be used to identify this nested call.
319 * @ctx: This instance context.
321 * Returns: Returns the code returned by the @nproc callback, or -1 if
322 * the maximum recursion limit has been exceeded.
324 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
325 int (*nproc)(void *, void *, int), void *priv,
326 void *cookie, void *ctx)
328 int error, call_nests = 0;
330 struct list_head *lsthead = &ncalls->tasks_call_list;
331 struct nested_call_node *tncur;
332 struct nested_call_node tnode;
334 spin_lock_irqsave(&ncalls->lock, flags);
337 * Try to see if the current task is already inside this wakeup call.
338 * We use a list here, since the population inside this set is always
341 list_for_each_entry(tncur, lsthead, llink) {
342 if (tncur->ctx == ctx &&
343 (tncur->cookie == cookie || ++call_nests > max_nests)) {
345 * Ops ... loop detected or maximum nest level reached.
346 * We abort this wake by breaking the cycle itself.
353 /* Add the current task and cookie to the list */
355 tnode.cookie = cookie;
356 list_add(&tnode.llink, lsthead);
358 spin_unlock_irqrestore(&ncalls->lock, flags);
360 /* Call the nested function */
361 error = (*nproc)(priv, cookie, call_nests);
363 /* Remove the current task from the list */
364 spin_lock_irqsave(&ncalls->lock, flags);
365 list_del(&tnode.llink);
367 spin_unlock_irqrestore(&ncalls->lock, flags);
372 #ifdef CONFIG_DEBUG_LOCK_ALLOC
373 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
374 unsigned long events, int subclass)
378 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
379 wake_up_locked_poll(wqueue, events);
380 spin_unlock_irqrestore(&wqueue->lock, flags);
383 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
384 unsigned long events, int subclass)
386 wake_up_poll(wqueue, events);
390 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
392 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
398 * Perform a safe wake up of the poll wait list. The problem is that
399 * with the new callback'd wake up system, it is possible that the
400 * poll callback is reentered from inside the call to wake_up() done
401 * on the poll wait queue head. The rule is that we cannot reenter the
402 * wake up code from the same task more than EP_MAX_NESTS times,
403 * and we cannot reenter the same wait queue head at all. This will
404 * enable to have a hierarchy of epoll file descriptor of no more than
407 static void ep_poll_safewake(wait_queue_head_t *wq)
409 int this_cpu = get_cpu();
411 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
412 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
418 * This function unregisters poll callbacks from the associated file
419 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
422 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
424 struct list_head *lsthead = &epi->pwqlist;
425 struct eppoll_entry *pwq;
427 while (!list_empty(lsthead)) {
428 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
430 list_del(&pwq->llink);
431 remove_wait_queue(pwq->whead, &pwq->wait);
432 kmem_cache_free(pwq_cache, pwq);
437 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
438 * the scan code, to call f_op->poll(). Also allows for
439 * O(NumReady) performance.
441 * @ep: Pointer to the epoll private data structure.
442 * @sproc: Pointer to the scan callback.
443 * @priv: Private opaque data passed to the @sproc callback.
445 * Returns: The same integer error code returned by the @sproc callback.
447 static int ep_scan_ready_list(struct eventpoll *ep,
448 int (*sproc)(struct eventpoll *,
449 struct list_head *, void *),
452 int error, pwake = 0;
454 struct epitem *epi, *nepi;
458 * We need to lock this because we could be hit by
459 * eventpoll_release_file() and epoll_ctl().
461 mutex_lock(&ep->mtx);
464 * Steal the ready list, and re-init the original one to the
465 * empty list. Also, set ep->ovflist to NULL so that events
466 * happening while looping w/out locks, are not lost. We cannot
467 * have the poll callback to queue directly on ep->rdllist,
468 * because we want the "sproc" callback to be able to do it
471 spin_lock_irqsave(&ep->lock, flags);
472 list_splice_init(&ep->rdllist, &txlist);
474 spin_unlock_irqrestore(&ep->lock, flags);
477 * Now call the callback function.
479 error = (*sproc)(ep, &txlist, priv);
481 spin_lock_irqsave(&ep->lock, flags);
483 * During the time we spent inside the "sproc" callback, some
484 * other events might have been queued by the poll callback.
485 * We re-insert them inside the main ready-list here.
487 for (nepi = ep->ovflist; (epi = nepi) != NULL;
488 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
490 * We need to check if the item is already in the list.
491 * During the "sproc" callback execution time, items are
492 * queued into ->ovflist but the "txlist" might already
493 * contain them, and the list_splice() below takes care of them.
495 if (!ep_is_linked(&epi->rdllink))
496 list_add_tail(&epi->rdllink, &ep->rdllist);
499 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
500 * releasing the lock, events will be queued in the normal way inside
503 ep->ovflist = EP_UNACTIVE_PTR;
506 * Quickly re-inject items left on "txlist".
508 list_splice(&txlist, &ep->rdllist);
510 if (!list_empty(&ep->rdllist)) {
512 * Wake up (if active) both the eventpoll wait list and
513 * the ->poll() wait list (delayed after we release the lock).
515 if (waitqueue_active(&ep->wq))
516 wake_up_locked(&ep->wq);
517 if (waitqueue_active(&ep->poll_wait))
520 spin_unlock_irqrestore(&ep->lock, flags);
522 mutex_unlock(&ep->mtx);
524 /* We have to call this outside the lock */
526 ep_poll_safewake(&ep->poll_wait);
532 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
533 * all the associated resources. Must be called with "mtx" held.
535 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
538 struct file *file = epi->ffd.file;
541 * Removes poll wait queue hooks. We _have_ to do this without holding
542 * the "ep->lock" otherwise a deadlock might occur. This because of the
543 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
544 * queue head lock when unregistering the wait queue. The wakeup callback
545 * will run by holding the wait queue head lock and will call our callback
546 * that will try to get "ep->lock".
548 ep_unregister_pollwait(ep, epi);
550 /* Remove the current item from the list of epoll hooks */
551 spin_lock(&file->f_lock);
552 if (ep_is_linked(&epi->fllink))
553 list_del_init(&epi->fllink);
554 spin_unlock(&file->f_lock);
556 rb_erase(&epi->rbn, &ep->rbr);
558 spin_lock_irqsave(&ep->lock, flags);
559 if (ep_is_linked(&epi->rdllink))
560 list_del_init(&epi->rdllink);
561 spin_unlock_irqrestore(&ep->lock, flags);
563 /* At this point it is safe to free the eventpoll item */
564 kmem_cache_free(epi_cache, epi);
566 atomic_long_dec(&ep->user->epoll_watches);
571 static void ep_free(struct eventpoll *ep)
576 /* We need to release all tasks waiting for these file */
577 if (waitqueue_active(&ep->poll_wait))
578 ep_poll_safewake(&ep->poll_wait);
581 * We need to lock this because we could be hit by
582 * eventpoll_release_file() while we're freeing the "struct eventpoll".
583 * We do not need to hold "ep->mtx" here because the epoll file
584 * is on the way to be removed and no one has references to it
585 * anymore. The only hit might come from eventpoll_release_file() but
586 * holding "epmutex" is sufficent here.
588 mutex_lock(&epmutex);
591 * Walks through the whole tree by unregistering poll callbacks.
593 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
594 epi = rb_entry(rbp, struct epitem, rbn);
596 ep_unregister_pollwait(ep, epi);
600 * Walks through the whole tree by freeing each "struct epitem". At this
601 * point we are sure no poll callbacks will be lingering around, and also by
602 * holding "epmutex" we can be sure that no file cleanup code will hit
603 * us during this operation. So we can avoid the lock on "ep->lock".
605 while ((rbp = rb_first(&ep->rbr)) != NULL) {
606 epi = rb_entry(rbp, struct epitem, rbn);
610 mutex_unlock(&epmutex);
611 mutex_destroy(&ep->mtx);
616 static int ep_eventpoll_release(struct inode *inode, struct file *file)
618 struct eventpoll *ep = file->private_data;
626 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
629 struct epitem *epi, *tmp;
631 list_for_each_entry_safe(epi, tmp, head, rdllink) {
632 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
634 return POLLIN | POLLRDNORM;
637 * Item has been dropped into the ready list by the poll
638 * callback, but it's not actually ready, as far as
639 * caller requested events goes. We can remove it here.
641 list_del_init(&epi->rdllink);
648 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
650 return ep_scan_ready_list(priv, ep_read_events_proc, NULL);
653 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
656 struct eventpoll *ep = file->private_data;
658 /* Insert inside our poll wait queue */
659 poll_wait(file, &ep->poll_wait, wait);
662 * Proceed to find out if wanted events are really available inside
663 * the ready list. This need to be done under ep_call_nested()
664 * supervision, since the call to f_op->poll() done on listed files
665 * could re-enter here.
667 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
668 ep_poll_readyevents_proc, ep, ep, current);
670 return pollflags != -1 ? pollflags : 0;
673 /* File callbacks that implement the eventpoll file behaviour */
674 static const struct file_operations eventpoll_fops = {
675 .release = ep_eventpoll_release,
676 .poll = ep_eventpoll_poll,
677 .llseek = noop_llseek,
680 /* Fast test to see if the file is an evenpoll file */
681 static inline int is_file_epoll(struct file *f)
683 return f->f_op == &eventpoll_fops;
687 * This is called from eventpoll_release() to unlink files from the eventpoll
688 * interface. We need to have this facility to cleanup correctly files that are
689 * closed without being removed from the eventpoll interface.
691 void eventpoll_release_file(struct file *file)
693 struct list_head *lsthead = &file->f_ep_links;
694 struct eventpoll *ep;
698 * We don't want to get "file->f_lock" because it is not
699 * necessary. It is not necessary because we're in the "struct file"
700 * cleanup path, and this means that noone is using this file anymore.
701 * So, for example, epoll_ctl() cannot hit here since if we reach this
702 * point, the file counter already went to zero and fget() would fail.
703 * The only hit might come from ep_free() but by holding the mutex
704 * will correctly serialize the operation. We do need to acquire
705 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
706 * from anywhere but ep_free().
708 * Besides, ep_remove() acquires the lock, so we can't hold it here.
710 mutex_lock(&epmutex);
712 while (!list_empty(lsthead)) {
713 epi = list_first_entry(lsthead, struct epitem, fllink);
716 list_del_init(&epi->fllink);
717 mutex_lock(&ep->mtx);
719 mutex_unlock(&ep->mtx);
722 mutex_unlock(&epmutex);
725 static int ep_alloc(struct eventpoll **pep)
728 struct user_struct *user;
729 struct eventpoll *ep;
731 user = get_current_user();
733 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
737 spin_lock_init(&ep->lock);
738 mutex_init(&ep->mtx);
739 init_waitqueue_head(&ep->wq);
740 init_waitqueue_head(&ep->poll_wait);
741 INIT_LIST_HEAD(&ep->rdllist);
743 ep->ovflist = EP_UNACTIVE_PTR;
756 * Search the file inside the eventpoll tree. The RB tree operations
757 * are protected by the "mtx" mutex, and ep_find() must be called with
760 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
764 struct epitem *epi, *epir = NULL;
765 struct epoll_filefd ffd;
767 ep_set_ffd(&ffd, file, fd);
768 for (rbp = ep->rbr.rb_node; rbp; ) {
769 epi = rb_entry(rbp, struct epitem, rbn);
770 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
785 * This is the callback that is passed to the wait queue wakeup
786 * machanism. It is called by the stored file descriptors when they
787 * have events to report.
789 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
793 struct epitem *epi = ep_item_from_wait(wait);
794 struct eventpoll *ep = epi->ep;
796 spin_lock_irqsave(&ep->lock, flags);
799 * If the event mask does not contain any poll(2) event, we consider the
800 * descriptor to be disabled. This condition is likely the effect of the
801 * EPOLLONESHOT bit that disables the descriptor when an event is received,
802 * until the next EPOLL_CTL_MOD will be issued.
804 if (!(epi->event.events & ~EP_PRIVATE_BITS))
808 * Check the events coming with the callback. At this stage, not
809 * every device reports the events in the "key" parameter of the
810 * callback. We need to be able to handle both cases here, hence the
811 * test for "key" != NULL before the event match test.
813 if (key && !((unsigned long) key & epi->event.events))
817 * If we are trasfering events to userspace, we can hold no locks
818 * (because we're accessing user memory, and because of linux f_op->poll()
819 * semantics). All the events that happens during that period of time are
820 * chained in ep->ovflist and requeued later on.
822 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
823 if (epi->next == EP_UNACTIVE_PTR) {
824 epi->next = ep->ovflist;
830 /* If this file is already in the ready list we exit soon */
831 if (!ep_is_linked(&epi->rdllink))
832 list_add_tail(&epi->rdllink, &ep->rdllist);
835 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
838 if (waitqueue_active(&ep->wq))
839 wake_up_locked(&ep->wq);
840 if (waitqueue_active(&ep->poll_wait))
844 spin_unlock_irqrestore(&ep->lock, flags);
846 /* We have to call this outside the lock */
848 ep_poll_safewake(&ep->poll_wait);
854 * This is the callback that is used to add our wait queue to the
855 * target file wakeup lists.
857 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
860 struct epitem *epi = ep_item_from_epqueue(pt);
861 struct eppoll_entry *pwq;
863 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
864 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
867 add_wait_queue(whead, &pwq->wait);
868 list_add_tail(&pwq->llink, &epi->pwqlist);
871 /* We have to signal that an error occurred */
876 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
879 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
884 epic = rb_entry(parent, struct epitem, rbn);
885 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
887 p = &parent->rb_right;
889 p = &parent->rb_left;
891 rb_link_node(&epi->rbn, parent, p);
892 rb_insert_color(&epi->rbn, &ep->rbr);
896 * Must be called with "mtx" held.
898 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
899 struct file *tfile, int fd)
901 int error, revents, pwake = 0;
905 struct ep_pqueue epq;
907 user_watches = atomic_long_read(&ep->user->epoll_watches);
908 if (unlikely(user_watches >= max_user_watches))
910 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
913 /* Item initialization follow here ... */
914 INIT_LIST_HEAD(&epi->rdllink);
915 INIT_LIST_HEAD(&epi->fllink);
916 INIT_LIST_HEAD(&epi->pwqlist);
918 ep_set_ffd(&epi->ffd, tfile, fd);
921 epi->next = EP_UNACTIVE_PTR;
923 /* Initialize the poll table using the queue callback */
925 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
928 * Attach the item to the poll hooks and get current event bits.
929 * We can safely use the file* here because its usage count has
930 * been increased by the caller of this function. Note that after
931 * this operation completes, the poll callback can start hitting
934 revents = tfile->f_op->poll(tfile, &epq.pt);
937 * We have to check if something went wrong during the poll wait queue
938 * install process. Namely an allocation for a wait queue failed due
939 * high memory pressure.
943 goto error_unregister;
945 /* Add the current item to the list of active epoll hook for this file */
946 spin_lock(&tfile->f_lock);
947 list_add_tail(&epi->fllink, &tfile->f_ep_links);
948 spin_unlock(&tfile->f_lock);
951 * Add the current item to the RB tree. All RB tree operations are
952 * protected by "mtx", and ep_insert() is called with "mtx" held.
954 ep_rbtree_insert(ep, epi);
956 /* We have to drop the new item inside our item list to keep track of it */
957 spin_lock_irqsave(&ep->lock, flags);
959 /* If the file is already "ready" we drop it inside the ready list */
960 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
961 list_add_tail(&epi->rdllink, &ep->rdllist);
963 /* Notify waiting tasks that events are available */
964 if (waitqueue_active(&ep->wq))
965 wake_up_locked(&ep->wq);
966 if (waitqueue_active(&ep->poll_wait))
970 spin_unlock_irqrestore(&ep->lock, flags);
972 atomic_long_inc(&ep->user->epoll_watches);
974 /* We have to call this outside the lock */
976 ep_poll_safewake(&ep->poll_wait);
981 ep_unregister_pollwait(ep, epi);
984 * We need to do this because an event could have been arrived on some
985 * allocated wait queue. Note that we don't care about the ep->ovflist
986 * list, since that is used/cleaned only inside a section bound by "mtx".
987 * And ep_insert() is called with "mtx" held.
989 spin_lock_irqsave(&ep->lock, flags);
990 if (ep_is_linked(&epi->rdllink))
991 list_del_init(&epi->rdllink);
992 spin_unlock_irqrestore(&ep->lock, flags);
994 kmem_cache_free(epi_cache, epi);
1000 * Modify the interest event mask by dropping an event if the new mask
1001 * has a match in the current file status. Must be called with "mtx" held.
1003 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1006 unsigned int revents;
1009 * Set the new event interest mask before calling f_op->poll();
1010 * otherwise we might miss an event that happens between the
1011 * f_op->poll() call and the new event set registering.
1013 epi->event.events = event->events;
1014 epi->event.data = event->data; /* protected by mtx */
1017 * Get current event bits. We can safely use the file* here because
1018 * its usage count has been increased by the caller of this function.
1020 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1023 * If the item is "hot" and it is not registered inside the ready
1024 * list, push it inside.
1026 if (revents & event->events) {
1027 spin_lock_irq(&ep->lock);
1028 if (!ep_is_linked(&epi->rdllink)) {
1029 list_add_tail(&epi->rdllink, &ep->rdllist);
1031 /* Notify waiting tasks that events are available */
1032 if (waitqueue_active(&ep->wq))
1033 wake_up_locked(&ep->wq);
1034 if (waitqueue_active(&ep->poll_wait))
1037 spin_unlock_irq(&ep->lock);
1040 /* We have to call this outside the lock */
1042 ep_poll_safewake(&ep->poll_wait);
1047 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1050 struct ep_send_events_data *esed = priv;
1052 unsigned int revents;
1054 struct epoll_event __user *uevent;
1057 * We can loop without lock because we are passed a task private list.
1058 * Items cannot vanish during the loop because ep_scan_ready_list() is
1059 * holding "mtx" during this call.
1061 for (eventcnt = 0, uevent = esed->events;
1062 !list_empty(head) && eventcnt < esed->maxevents;) {
1063 epi = list_first_entry(head, struct epitem, rdllink);
1065 list_del_init(&epi->rdllink);
1067 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1071 * If the event mask intersect the caller-requested one,
1072 * deliver the event to userspace. Again, ep_scan_ready_list()
1073 * is holding "mtx", so no operations coming from userspace
1074 * can change the item.
1077 if (__put_user(revents, &uevent->events) ||
1078 __put_user(epi->event.data, &uevent->data)) {
1079 list_add(&epi->rdllink, head);
1080 return eventcnt ? eventcnt : -EFAULT;
1084 if (epi->event.events & EPOLLONESHOT)
1085 epi->event.events &= EP_PRIVATE_BITS;
1086 else if (!(epi->event.events & EPOLLET)) {
1088 * If this file has been added with Level
1089 * Trigger mode, we need to insert back inside
1090 * the ready list, so that the next call to
1091 * epoll_wait() will check again the events
1092 * availability. At this point, noone can insert
1093 * into ep->rdllist besides us. The epoll_ctl()
1094 * callers are locked out by
1095 * ep_scan_ready_list() holding "mtx" and the
1096 * poll callback will queue them in ep->ovflist.
1098 list_add_tail(&epi->rdllink, &ep->rdllist);
1106 static int ep_send_events(struct eventpoll *ep,
1107 struct epoll_event __user *events, int maxevents)
1109 struct ep_send_events_data esed;
1111 esed.maxevents = maxevents;
1112 esed.events = events;
1114 return ep_scan_ready_list(ep, ep_send_events_proc, &esed);
1117 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1118 int maxevents, long timeout)
1120 int res, eavail, timed_out = 0;
1121 unsigned long flags;
1124 struct timespec end_time;
1125 ktime_t expires, *to = NULL;
1128 ktime_get_ts(&end_time);
1129 timespec_add_ns(&end_time, (u64)timeout * NSEC_PER_MSEC);
1130 slack = select_estimate_accuracy(&end_time);
1132 *to = timespec_to_ktime(end_time);
1133 } else if (timeout == 0) {
1138 spin_lock_irqsave(&ep->lock, flags);
1141 if (list_empty(&ep->rdllist)) {
1143 * We don't have any available event to return to the caller.
1144 * We need to sleep here, and we will be wake up by
1145 * ep_poll_callback() when events will become available.
1147 init_waitqueue_entry(&wait, current);
1148 __add_wait_queue_exclusive(&ep->wq, &wait);
1152 * We don't want to sleep if the ep_poll_callback() sends us
1153 * a wakeup in between. That's why we set the task state
1154 * to TASK_INTERRUPTIBLE before doing the checks.
1156 set_current_state(TASK_INTERRUPTIBLE);
1157 if (!list_empty(&ep->rdllist) || timed_out)
1159 if (signal_pending(current)) {
1164 spin_unlock_irqrestore(&ep->lock, flags);
1165 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1168 spin_lock_irqsave(&ep->lock, flags);
1170 __remove_wait_queue(&ep->wq, &wait);
1172 set_current_state(TASK_RUNNING);
1174 /* Is it worth to try to dig for events ? */
1175 eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
1177 spin_unlock_irqrestore(&ep->lock, flags);
1180 * Try to transfer events to user space. In case we get 0 events and
1181 * there's still timeout left over, we go trying again in search of
1184 if (!res && eavail &&
1185 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1192 * Open an eventpoll file descriptor.
1194 SYSCALL_DEFINE1(epoll_create1, int, flags)
1197 struct eventpoll *ep = NULL;
1199 /* Check the EPOLL_* constant for consistency. */
1200 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1202 if (flags & ~EPOLL_CLOEXEC)
1205 * Create the internal data structure ("struct eventpoll").
1207 error = ep_alloc(&ep);
1211 * Creates all the items needed to setup an eventpoll file. That is,
1212 * a file structure and a free file descriptor.
1214 error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1215 O_RDWR | (flags & O_CLOEXEC));
1222 SYSCALL_DEFINE1(epoll_create, int, size)
1227 return sys_epoll_create1(0);
1231 * The following function implements the controller interface for
1232 * the eventpoll file that enables the insertion/removal/change of
1233 * file descriptors inside the interest set.
1235 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1236 struct epoll_event __user *, event)
1239 struct file *file, *tfile;
1240 struct eventpoll *ep;
1242 struct epoll_event epds;
1245 if (ep_op_has_event(op) &&
1246 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1249 /* Get the "struct file *" for the eventpoll file */
1255 /* Get the "struct file *" for the target file */
1260 /* The target file descriptor must support poll */
1262 if (!tfile->f_op || !tfile->f_op->poll)
1263 goto error_tgt_fput;
1266 * We have to check that the file structure underneath the file descriptor
1267 * the user passed to us _is_ an eventpoll file. And also we do not permit
1268 * adding an epoll file descriptor inside itself.
1271 if (file == tfile || !is_file_epoll(file))
1272 goto error_tgt_fput;
1275 * At this point it is safe to assume that the "private_data" contains
1276 * our own data structure.
1278 ep = file->private_data;
1280 mutex_lock(&ep->mtx);
1283 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1284 * above, we can be sure to be able to use the item looked up by
1285 * ep_find() till we release the mutex.
1287 epi = ep_find(ep, tfile, fd);
1293 epds.events |= POLLERR | POLLHUP;
1294 error = ep_insert(ep, &epds, tfile, fd);
1300 error = ep_remove(ep, epi);
1306 epds.events |= POLLERR | POLLHUP;
1307 error = ep_modify(ep, epi, &epds);
1312 mutex_unlock(&ep->mtx);
1324 * Implement the event wait interface for the eventpoll file. It is the kernel
1325 * part of the user space epoll_wait(2).
1327 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1328 int, maxevents, int, timeout)
1332 struct eventpoll *ep;
1334 /* The maximum number of event must be greater than zero */
1335 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1338 /* Verify that the area passed by the user is writeable */
1339 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1344 /* Get the "struct file *" for the eventpoll file */
1351 * We have to check that the file structure underneath the fd
1352 * the user passed to us _is_ an eventpoll file.
1355 if (!is_file_epoll(file))
1359 * At this point it is safe to assume that the "private_data" contains
1360 * our own data structure.
1362 ep = file->private_data;
1364 /* Time to fish for events ... */
1365 error = ep_poll(ep, events, maxevents, timeout);
1374 #ifdef HAVE_SET_RESTORE_SIGMASK
1377 * Implement the event wait interface for the eventpoll file. It is the kernel
1378 * part of the user space epoll_pwait(2).
1380 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1381 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1385 sigset_t ksigmask, sigsaved;
1388 * If the caller wants a certain signal mask to be set during the wait,
1392 if (sigsetsize != sizeof(sigset_t))
1394 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1396 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1397 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1400 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1403 * If we changed the signal mask, we need to restore the original one.
1404 * In case we've got a signal while waiting, we do not restore the
1405 * signal mask yet, and we allow do_signal() to deliver the signal on
1406 * the way back to userspace, before the signal mask is restored.
1409 if (error == -EINTR) {
1410 memcpy(¤t->saved_sigmask, &sigsaved,
1412 set_restore_sigmask();
1414 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1420 #endif /* HAVE_SET_RESTORE_SIGMASK */
1422 static int __init eventpoll_init(void)
1428 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1430 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1432 BUG_ON(max_user_watches < 0);
1434 /* Initialize the structure used to perform safe poll wait head wake ups */
1435 ep_nested_calls_init(&poll_safewake_ncalls);
1437 /* Initialize the structure used to perform file's f_op->poll() calls */
1438 ep_nested_calls_init(&poll_readywalk_ncalls);
1440 /* Allocates slab cache used to allocate "struct epitem" items */
1441 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1442 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1444 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1445 pwq_cache = kmem_cache_create("eventpoll_pwq",
1446 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1450 fs_initcall(eventpoll_init);