ACPI: Introduce acpi_get_pci_dev()
[linux-2.6/linux-acpi-2.6.git] / fs / eventpoll.c
blob5458e80fc55841d136227f7ce0d7ee97f8ba0f38
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.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 <asm/uaccess.h>
37 #include <asm/system.h>
38 #include <asm/io.h>
39 #include <asm/mman.h>
40 #include <asm/atomic.h>
43 * LOCKING:
44 * There are three level of locking required by epoll :
46 * 1) epmutex (mutex)
47 * 2) ep->mtx (mutex)
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()
61 * and ep_free().
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 /* Maximum msec timeout value storeable in a long int */
81 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
83 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
85 #define EP_UNACTIVE_PTR ((void *) -1L)
87 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
89 struct epoll_filefd {
90 struct file *file;
91 int fd;
95 * Structure used to track possible nested calls, for too deep recursions
96 * and loop cycles.
98 struct nested_call_node {
99 struct list_head llink;
100 void *cookie;
101 int cpu;
105 * This structure is used as collector for nested calls, to check for
106 * maximum recursion dept and loop cycles.
108 struct nested_calls {
109 struct list_head tasks_call_list;
110 spinlock_t lock;
114 * Each file descriptor added to the eventpoll interface will
115 * have an entry of this type linked to the "rbr" RB tree.
117 struct epitem {
118 /* RB tree node used to link this structure to the eventpoll RB tree */
119 struct rb_node rbn;
121 /* List header used to link this structure to the eventpoll ready list */
122 struct list_head rdllink;
125 * Works together "struct eventpoll"->ovflist in keeping the
126 * single linked chain of items.
128 struct epitem *next;
130 /* The file descriptor information this item refers to */
131 struct epoll_filefd ffd;
133 /* Number of active wait queue attached to poll operations */
134 int nwait;
136 /* List containing poll wait queues */
137 struct list_head pwqlist;
139 /* The "container" of this item */
140 struct eventpoll *ep;
142 /* List header used to link this item to the "struct file" items list */
143 struct list_head fllink;
145 /* The structure that describe the interested events and the source fd */
146 struct epoll_event event;
150 * This structure is stored inside the "private_data" member of the file
151 * structure and rapresent the main data sructure for the eventpoll
152 * interface.
154 struct eventpoll {
155 /* Protect the this structure access */
156 spinlock_t lock;
159 * This mutex is used to ensure that files are not removed
160 * while epoll is using them. This is held during the event
161 * collection loop, the file cleanup path, the epoll file exit
162 * code and the ctl operations.
164 struct mutex mtx;
166 /* Wait queue used by sys_epoll_wait() */
167 wait_queue_head_t wq;
169 /* Wait queue used by file->poll() */
170 wait_queue_head_t poll_wait;
172 /* List of ready file descriptors */
173 struct list_head rdllist;
175 /* RB tree root used to store monitored fd structs */
176 struct rb_root rbr;
179 * This is a single linked list that chains all the "struct epitem" that
180 * happened while transfering ready events to userspace w/out
181 * holding ->lock.
183 struct epitem *ovflist;
185 /* The user that created the eventpoll descriptor */
186 struct user_struct *user;
189 /* Wait structure used by the poll hooks */
190 struct eppoll_entry {
191 /* List header used to link this structure to the "struct epitem" */
192 struct list_head llink;
194 /* The "base" pointer is set to the container "struct epitem" */
195 struct epitem *base;
198 * Wait queue item that will be linked to the target file wait
199 * queue head.
201 wait_queue_t wait;
203 /* The wait queue head that linked the "wait" wait queue item */
204 wait_queue_head_t *whead;
207 /* Wrapper struct used by poll queueing */
208 struct ep_pqueue {
209 poll_table pt;
210 struct epitem *epi;
213 /* Used by the ep_send_events() function as callback private data */
214 struct ep_send_events_data {
215 int maxevents;
216 struct epoll_event __user *events;
220 * Configuration options available inside /proc/sys/fs/epoll/
222 /* Maximum number of epoll watched descriptors, per user */
223 static int max_user_watches __read_mostly;
226 * This mutex is used to serialize ep_free() and eventpoll_release_file().
228 static DEFINE_MUTEX(epmutex);
230 /* Used for safe wake up implementation */
231 static struct nested_calls poll_safewake_ncalls;
233 /* Used to call file's f_op->poll() under the nested calls boundaries */
234 static struct nested_calls poll_readywalk_ncalls;
236 /* Slab cache used to allocate "struct epitem" */
237 static struct kmem_cache *epi_cache __read_mostly;
239 /* Slab cache used to allocate "struct eppoll_entry" */
240 static struct kmem_cache *pwq_cache __read_mostly;
242 #ifdef CONFIG_SYSCTL
244 #include <linux/sysctl.h>
246 static int zero;
248 ctl_table epoll_table[] = {
250 .procname = "max_user_watches",
251 .data = &max_user_watches,
252 .maxlen = sizeof(int),
253 .mode = 0644,
254 .proc_handler = &proc_dointvec_minmax,
255 .extra1 = &zero,
257 { .ctl_name = 0 }
259 #endif /* CONFIG_SYSCTL */
262 /* Setup the structure that is used as key for the RB tree */
263 static inline void ep_set_ffd(struct epoll_filefd *ffd,
264 struct file *file, int fd)
266 ffd->file = file;
267 ffd->fd = fd;
270 /* Compare RB tree keys */
271 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
272 struct epoll_filefd *p2)
274 return (p1->file > p2->file ? +1:
275 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
278 /* Tells us if the item is currently linked */
279 static inline int ep_is_linked(struct list_head *p)
281 return !list_empty(p);
284 /* Get the "struct epitem" from a wait queue pointer */
285 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
287 return container_of(p, struct eppoll_entry, wait)->base;
290 /* Get the "struct epitem" from an epoll queue wrapper */
291 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
293 return container_of(p, struct ep_pqueue, pt)->epi;
296 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
297 static inline int ep_op_has_event(int op)
299 return op != EPOLL_CTL_DEL;
302 /* Initialize the poll safe wake up structure */
303 static void ep_nested_calls_init(struct nested_calls *ncalls)
305 INIT_LIST_HEAD(&ncalls->tasks_call_list);
306 spin_lock_init(&ncalls->lock);
310 * ep_call_nested - Perform a bound (possibly) nested call, by checking
311 * that the recursion limit is not exceeded, and that
312 * the same nested call (by the meaning of same cookie) is
313 * no re-entered.
315 * @ncalls: Pointer to the nested_calls structure to be used for this call.
316 * @max_nests: Maximum number of allowed nesting calls.
317 * @nproc: Nested call core function pointer.
318 * @priv: Opaque data to be passed to the @nproc callback.
319 * @cookie: Cookie to be used to identify this nested call.
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)
328 int error, call_nests = 0;
329 unsigned long flags;
330 int this_cpu = get_cpu();
331 struct list_head *lsthead = &ncalls->tasks_call_list;
332 struct nested_call_node *tncur;
333 struct nested_call_node tnode;
335 spin_lock_irqsave(&ncalls->lock, flags);
338 * Try to see if the current task is already inside this wakeup call.
339 * We use a list here, since the population inside this set is always
340 * very much limited.
342 list_for_each_entry(tncur, lsthead, llink) {
343 if (tncur->cpu == this_cpu &&
344 (tncur->cookie == cookie || ++call_nests > max_nests)) {
346 * Ops ... loop detected or maximum nest level reached.
347 * We abort this wake by breaking the cycle itself.
349 error = -1;
350 goto out_unlock;
354 /* Add the current task and cookie to the list */
355 tnode.cpu = this_cpu;
356 tnode.cookie = cookie;
357 list_add(&tnode.llink, lsthead);
359 spin_unlock_irqrestore(&ncalls->lock, flags);
361 /* Call the nested function */
362 error = (*nproc)(priv, cookie, call_nests);
364 /* Remove the current task from the list */
365 spin_lock_irqsave(&ncalls->lock, flags);
366 list_del(&tnode.llink);
367 out_unlock:
368 spin_unlock_irqrestore(&ncalls->lock, flags);
370 put_cpu();
371 return error;
374 #ifdef CONFIG_DEBUG_LOCK_ALLOC
375 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
376 unsigned long events, int subclass)
378 unsigned long flags;
380 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
381 wake_up_locked_poll(wqueue, events);
382 spin_unlock_irqrestore(&wqueue->lock, flags);
384 #else
385 static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
386 unsigned long events, int subclass)
388 wake_up_poll(wqueue, events);
390 #endif
392 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
394 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
395 1 + call_nests);
396 return 0;
400 * Perform a safe wake up of the poll wait list. The problem is that
401 * with the new callback'd wake up system, it is possible that the
402 * poll callback is reentered from inside the call to wake_up() done
403 * on the poll wait queue head. The rule is that we cannot reenter the
404 * wake up code from the same task more than EP_MAX_NESTS times,
405 * and we cannot reenter the same wait queue head at all. This will
406 * enable to have a hierarchy of epoll file descriptor of no more than
407 * EP_MAX_NESTS deep.
409 static void ep_poll_safewake(wait_queue_head_t *wq)
411 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
412 ep_poll_wakeup_proc, NULL, wq);
416 * This function unregisters poll callbacks from the associated file
417 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
418 * ep_free).
420 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
422 struct list_head *lsthead = &epi->pwqlist;
423 struct eppoll_entry *pwq;
425 while (!list_empty(lsthead)) {
426 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
428 list_del(&pwq->llink);
429 remove_wait_queue(pwq->whead, &pwq->wait);
430 kmem_cache_free(pwq_cache, pwq);
435 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
436 * the scan code, to call f_op->poll(). Also allows for
437 * O(NumReady) performance.
439 * @ep: Pointer to the epoll private data structure.
440 * @sproc: Pointer to the scan callback.
441 * @priv: Private opaque data passed to the @sproc callback.
443 * Returns: The same integer error code returned by the @sproc callback.
445 static int ep_scan_ready_list(struct eventpoll *ep,
446 int (*sproc)(struct eventpoll *,
447 struct list_head *, void *),
448 void *priv)
450 int error, pwake = 0;
451 unsigned long flags;
452 struct epitem *epi, *nepi;
453 LIST_HEAD(txlist);
456 * We need to lock this because we could be hit by
457 * eventpoll_release_file() and epoll_ctl().
459 mutex_lock(&ep->mtx);
462 * Steal the ready list, and re-init the original one to the
463 * empty list. Also, set ep->ovflist to NULL so that events
464 * happening while looping w/out locks, are not lost. We cannot
465 * have the poll callback to queue directly on ep->rdllist,
466 * because we want the "sproc" callback to be able to do it
467 * in a lockless way.
469 spin_lock_irqsave(&ep->lock, flags);
470 list_splice_init(&ep->rdllist, &txlist);
471 ep->ovflist = NULL;
472 spin_unlock_irqrestore(&ep->lock, flags);
475 * Now call the callback function.
477 error = (*sproc)(ep, &txlist, priv);
479 spin_lock_irqsave(&ep->lock, flags);
481 * During the time we spent inside the "sproc" callback, some
482 * other events might have been queued by the poll callback.
483 * We re-insert them inside the main ready-list here.
485 for (nepi = ep->ovflist; (epi = nepi) != NULL;
486 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
488 * We need to check if the item is already in the list.
489 * During the "sproc" callback execution time, items are
490 * queued into ->ovflist but the "txlist" might already
491 * contain them, and the list_splice() below takes care of them.
493 if (!ep_is_linked(&epi->rdllink))
494 list_add_tail(&epi->rdllink, &ep->rdllist);
497 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
498 * releasing the lock, events will be queued in the normal way inside
499 * ep->rdllist.
501 ep->ovflist = EP_UNACTIVE_PTR;
504 * Quickly re-inject items left on "txlist".
506 list_splice(&txlist, &ep->rdllist);
508 if (!list_empty(&ep->rdllist)) {
510 * Wake up (if active) both the eventpoll wait list and
511 * the ->poll() wait list (delayed after we release the lock).
513 if (waitqueue_active(&ep->wq))
514 wake_up_locked(&ep->wq);
515 if (waitqueue_active(&ep->poll_wait))
516 pwake++;
518 spin_unlock_irqrestore(&ep->lock, flags);
520 mutex_unlock(&ep->mtx);
522 /* We have to call this outside the lock */
523 if (pwake)
524 ep_poll_safewake(&ep->poll_wait);
526 return error;
530 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
531 * all the associated resources. Must be called with "mtx" held.
533 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
535 unsigned long flags;
536 struct file *file = epi->ffd.file;
539 * Removes poll wait queue hooks. We _have_ to do this without holding
540 * the "ep->lock" otherwise a deadlock might occur. This because of the
541 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
542 * queue head lock when unregistering the wait queue. The wakeup callback
543 * will run by holding the wait queue head lock and will call our callback
544 * that will try to get "ep->lock".
546 ep_unregister_pollwait(ep, epi);
548 /* Remove the current item from the list of epoll hooks */
549 spin_lock(&file->f_lock);
550 if (ep_is_linked(&epi->fllink))
551 list_del_init(&epi->fllink);
552 spin_unlock(&file->f_lock);
554 rb_erase(&epi->rbn, &ep->rbr);
556 spin_lock_irqsave(&ep->lock, flags);
557 if (ep_is_linked(&epi->rdllink))
558 list_del_init(&epi->rdllink);
559 spin_unlock_irqrestore(&ep->lock, flags);
561 /* At this point it is safe to free the eventpoll item */
562 kmem_cache_free(epi_cache, epi);
564 atomic_dec(&ep->user->epoll_watches);
566 return 0;
569 static void ep_free(struct eventpoll *ep)
571 struct rb_node *rbp;
572 struct epitem *epi;
574 /* We need to release all tasks waiting for these file */
575 if (waitqueue_active(&ep->poll_wait))
576 ep_poll_safewake(&ep->poll_wait);
579 * We need to lock this because we could be hit by
580 * eventpoll_release_file() while we're freeing the "struct eventpoll".
581 * We do not need to hold "ep->mtx" here because the epoll file
582 * is on the way to be removed and no one has references to it
583 * anymore. The only hit might come from eventpoll_release_file() but
584 * holding "epmutex" is sufficent here.
586 mutex_lock(&epmutex);
589 * Walks through the whole tree by unregistering poll callbacks.
591 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
592 epi = rb_entry(rbp, struct epitem, rbn);
594 ep_unregister_pollwait(ep, epi);
598 * Walks through the whole tree by freeing each "struct epitem". At this
599 * point we are sure no poll callbacks will be lingering around, and also by
600 * holding "epmutex" we can be sure that no file cleanup code will hit
601 * us during this operation. So we can avoid the lock on "ep->lock".
603 while ((rbp = rb_first(&ep->rbr)) != NULL) {
604 epi = rb_entry(rbp, struct epitem, rbn);
605 ep_remove(ep, epi);
608 mutex_unlock(&epmutex);
609 mutex_destroy(&ep->mtx);
610 free_uid(ep->user);
611 kfree(ep);
614 static int ep_eventpoll_release(struct inode *inode, struct file *file)
616 struct eventpoll *ep = file->private_data;
618 if (ep)
619 ep_free(ep);
621 return 0;
624 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
625 void *priv)
627 struct epitem *epi, *tmp;
629 list_for_each_entry_safe(epi, tmp, head, rdllink) {
630 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
631 epi->event.events)
632 return POLLIN | POLLRDNORM;
633 else {
635 * Item has been dropped into the ready list by the poll
636 * callback, but it's not actually ready, as far as
637 * caller requested events goes. We can remove it here.
639 list_del_init(&epi->rdllink);
643 return 0;
646 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
648 return ep_scan_ready_list(priv, ep_read_events_proc, NULL);
651 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
653 int pollflags;
654 struct eventpoll *ep = file->private_data;
656 /* Insert inside our poll wait queue */
657 poll_wait(file, &ep->poll_wait, wait);
660 * Proceed to find out if wanted events are really available inside
661 * the ready list. This need to be done under ep_call_nested()
662 * supervision, since the call to f_op->poll() done on listed files
663 * could re-enter here.
665 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
666 ep_poll_readyevents_proc, ep, ep);
668 return pollflags != -1 ? pollflags : 0;
671 /* File callbacks that implement the eventpoll file behaviour */
672 static const struct file_operations eventpoll_fops = {
673 .release = ep_eventpoll_release,
674 .poll = ep_eventpoll_poll
677 /* Fast test to see if the file is an evenpoll file */
678 static inline int is_file_epoll(struct file *f)
680 return f->f_op == &eventpoll_fops;
684 * This is called from eventpoll_release() to unlink files from the eventpoll
685 * interface. We need to have this facility to cleanup correctly files that are
686 * closed without being removed from the eventpoll interface.
688 void eventpoll_release_file(struct file *file)
690 struct list_head *lsthead = &file->f_ep_links;
691 struct eventpoll *ep;
692 struct epitem *epi;
695 * We don't want to get "file->f_lock" because it is not
696 * necessary. It is not necessary because we're in the "struct file"
697 * cleanup path, and this means that noone is using this file anymore.
698 * So, for example, epoll_ctl() cannot hit here since if we reach this
699 * point, the file counter already went to zero and fget() would fail.
700 * The only hit might come from ep_free() but by holding the mutex
701 * will correctly serialize the operation. We do need to acquire
702 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
703 * from anywhere but ep_free().
705 * Besides, ep_remove() acquires the lock, so we can't hold it here.
707 mutex_lock(&epmutex);
709 while (!list_empty(lsthead)) {
710 epi = list_first_entry(lsthead, struct epitem, fllink);
712 ep = epi->ep;
713 list_del_init(&epi->fllink);
714 mutex_lock(&ep->mtx);
715 ep_remove(ep, epi);
716 mutex_unlock(&ep->mtx);
719 mutex_unlock(&epmutex);
722 static int ep_alloc(struct eventpoll **pep)
724 int error;
725 struct user_struct *user;
726 struct eventpoll *ep;
728 user = get_current_user();
729 error = -ENOMEM;
730 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
731 if (unlikely(!ep))
732 goto free_uid;
734 spin_lock_init(&ep->lock);
735 mutex_init(&ep->mtx);
736 init_waitqueue_head(&ep->wq);
737 init_waitqueue_head(&ep->poll_wait);
738 INIT_LIST_HEAD(&ep->rdllist);
739 ep->rbr = RB_ROOT;
740 ep->ovflist = EP_UNACTIVE_PTR;
741 ep->user = user;
743 *pep = ep;
745 return 0;
747 free_uid:
748 free_uid(user);
749 return error;
753 * Search the file inside the eventpoll tree. The RB tree operations
754 * are protected by the "mtx" mutex, and ep_find() must be called with
755 * "mtx" held.
757 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
759 int kcmp;
760 struct rb_node *rbp;
761 struct epitem *epi, *epir = NULL;
762 struct epoll_filefd ffd;
764 ep_set_ffd(&ffd, file, fd);
765 for (rbp = ep->rbr.rb_node; rbp; ) {
766 epi = rb_entry(rbp, struct epitem, rbn);
767 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
768 if (kcmp > 0)
769 rbp = rbp->rb_right;
770 else if (kcmp < 0)
771 rbp = rbp->rb_left;
772 else {
773 epir = epi;
774 break;
778 return epir;
782 * This is the callback that is passed to the wait queue wakeup
783 * machanism. It is called by the stored file descriptors when they
784 * have events to report.
786 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
788 int pwake = 0;
789 unsigned long flags;
790 struct epitem *epi = ep_item_from_wait(wait);
791 struct eventpoll *ep = epi->ep;
793 spin_lock_irqsave(&ep->lock, flags);
796 * If the event mask does not contain any poll(2) event, we consider the
797 * descriptor to be disabled. This condition is likely the effect of the
798 * EPOLLONESHOT bit that disables the descriptor when an event is received,
799 * until the next EPOLL_CTL_MOD will be issued.
801 if (!(epi->event.events & ~EP_PRIVATE_BITS))
802 goto out_unlock;
805 * Check the events coming with the callback. At this stage, not
806 * every device reports the events in the "key" parameter of the
807 * callback. We need to be able to handle both cases here, hence the
808 * test for "key" != NULL before the event match test.
810 if (key && !((unsigned long) key & epi->event.events))
811 goto out_unlock;
814 * If we are trasfering events to userspace, we can hold no locks
815 * (because we're accessing user memory, and because of linux f_op->poll()
816 * semantics). All the events that happens during that period of time are
817 * chained in ep->ovflist and requeued later on.
819 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
820 if (epi->next == EP_UNACTIVE_PTR) {
821 epi->next = ep->ovflist;
822 ep->ovflist = epi;
824 goto out_unlock;
827 /* If this file is already in the ready list we exit soon */
828 if (!ep_is_linked(&epi->rdllink))
829 list_add_tail(&epi->rdllink, &ep->rdllist);
832 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
833 * wait list.
835 if (waitqueue_active(&ep->wq))
836 wake_up_locked(&ep->wq);
837 if (waitqueue_active(&ep->poll_wait))
838 pwake++;
840 out_unlock:
841 spin_unlock_irqrestore(&ep->lock, flags);
843 /* We have to call this outside the lock */
844 if (pwake)
845 ep_poll_safewake(&ep->poll_wait);
847 return 1;
851 * This is the callback that is used to add our wait queue to the
852 * target file wakeup lists.
854 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
855 poll_table *pt)
857 struct epitem *epi = ep_item_from_epqueue(pt);
858 struct eppoll_entry *pwq;
860 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
861 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
862 pwq->whead = whead;
863 pwq->base = epi;
864 add_wait_queue(whead, &pwq->wait);
865 list_add_tail(&pwq->llink, &epi->pwqlist);
866 epi->nwait++;
867 } else {
868 /* We have to signal that an error occurred */
869 epi->nwait = -1;
873 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
875 int kcmp;
876 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
877 struct epitem *epic;
879 while (*p) {
880 parent = *p;
881 epic = rb_entry(parent, struct epitem, rbn);
882 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
883 if (kcmp > 0)
884 p = &parent->rb_right;
885 else
886 p = &parent->rb_left;
888 rb_link_node(&epi->rbn, parent, p);
889 rb_insert_color(&epi->rbn, &ep->rbr);
893 * Must be called with "mtx" held.
895 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
896 struct file *tfile, int fd)
898 int error, revents, pwake = 0;
899 unsigned long flags;
900 struct epitem *epi;
901 struct ep_pqueue epq;
903 if (unlikely(atomic_read(&ep->user->epoll_watches) >=
904 max_user_watches))
905 return -ENOSPC;
906 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
907 return -ENOMEM;
909 /* Item initialization follow here ... */
910 INIT_LIST_HEAD(&epi->rdllink);
911 INIT_LIST_HEAD(&epi->fllink);
912 INIT_LIST_HEAD(&epi->pwqlist);
913 epi->ep = ep;
914 ep_set_ffd(&epi->ffd, tfile, fd);
915 epi->event = *event;
916 epi->nwait = 0;
917 epi->next = EP_UNACTIVE_PTR;
919 /* Initialize the poll table using the queue callback */
920 epq.epi = epi;
921 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
924 * Attach the item to the poll hooks and get current event bits.
925 * We can safely use the file* here because its usage count has
926 * been increased by the caller of this function. Note that after
927 * this operation completes, the poll callback can start hitting
928 * the new item.
930 revents = tfile->f_op->poll(tfile, &epq.pt);
933 * We have to check if something went wrong during the poll wait queue
934 * install process. Namely an allocation for a wait queue failed due
935 * high memory pressure.
937 error = -ENOMEM;
938 if (epi->nwait < 0)
939 goto error_unregister;
941 /* Add the current item to the list of active epoll hook for this file */
942 spin_lock(&tfile->f_lock);
943 list_add_tail(&epi->fllink, &tfile->f_ep_links);
944 spin_unlock(&tfile->f_lock);
947 * Add the current item to the RB tree. All RB tree operations are
948 * protected by "mtx", and ep_insert() is called with "mtx" held.
950 ep_rbtree_insert(ep, epi);
952 /* We have to drop the new item inside our item list to keep track of it */
953 spin_lock_irqsave(&ep->lock, flags);
955 /* If the file is already "ready" we drop it inside the ready list */
956 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
957 list_add_tail(&epi->rdllink, &ep->rdllist);
959 /* Notify waiting tasks that events are available */
960 if (waitqueue_active(&ep->wq))
961 wake_up_locked(&ep->wq);
962 if (waitqueue_active(&ep->poll_wait))
963 pwake++;
966 spin_unlock_irqrestore(&ep->lock, flags);
968 atomic_inc(&ep->user->epoll_watches);
970 /* We have to call this outside the lock */
971 if (pwake)
972 ep_poll_safewake(&ep->poll_wait);
974 return 0;
976 error_unregister:
977 ep_unregister_pollwait(ep, epi);
980 * We need to do this because an event could have been arrived on some
981 * allocated wait queue. Note that we don't care about the ep->ovflist
982 * list, since that is used/cleaned only inside a section bound by "mtx".
983 * And ep_insert() is called with "mtx" held.
985 spin_lock_irqsave(&ep->lock, flags);
986 if (ep_is_linked(&epi->rdllink))
987 list_del_init(&epi->rdllink);
988 spin_unlock_irqrestore(&ep->lock, flags);
990 kmem_cache_free(epi_cache, epi);
992 return error;
996 * Modify the interest event mask by dropping an event if the new mask
997 * has a match in the current file status. Must be called with "mtx" held.
999 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1001 int pwake = 0;
1002 unsigned int revents;
1005 * Set the new event interest mask before calling f_op->poll();
1006 * otherwise we might miss an event that happens between the
1007 * f_op->poll() call and the new event set registering.
1009 epi->event.events = event->events;
1010 epi->event.data = event->data; /* protected by mtx */
1013 * Get current event bits. We can safely use the file* here because
1014 * its usage count has been increased by the caller of this function.
1016 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1019 * If the item is "hot" and it is not registered inside the ready
1020 * list, push it inside.
1022 if (revents & event->events) {
1023 spin_lock_irq(&ep->lock);
1024 if (!ep_is_linked(&epi->rdllink)) {
1025 list_add_tail(&epi->rdllink, &ep->rdllist);
1027 /* Notify waiting tasks that events are available */
1028 if (waitqueue_active(&ep->wq))
1029 wake_up_locked(&ep->wq);
1030 if (waitqueue_active(&ep->poll_wait))
1031 pwake++;
1033 spin_unlock_irq(&ep->lock);
1036 /* We have to call this outside the lock */
1037 if (pwake)
1038 ep_poll_safewake(&ep->poll_wait);
1040 return 0;
1043 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1044 void *priv)
1046 struct ep_send_events_data *esed = priv;
1047 int eventcnt;
1048 unsigned int revents;
1049 struct epitem *epi;
1050 struct epoll_event __user *uevent;
1053 * We can loop without lock because we are passed a task private list.
1054 * Items cannot vanish during the loop because ep_scan_ready_list() is
1055 * holding "mtx" during this call.
1057 for (eventcnt = 0, uevent = esed->events;
1058 !list_empty(head) && eventcnt < esed->maxevents;) {
1059 epi = list_first_entry(head, struct epitem, rdllink);
1061 list_del_init(&epi->rdllink);
1063 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1064 epi->event.events;
1067 * If the event mask intersect the caller-requested one,
1068 * deliver the event to userspace. Again, ep_scan_ready_list()
1069 * is holding "mtx", so no operations coming from userspace
1070 * can change the item.
1072 if (revents) {
1073 if (__put_user(revents, &uevent->events) ||
1074 __put_user(epi->event.data, &uevent->data)) {
1075 list_add(&epi->rdllink, head);
1076 return eventcnt ? eventcnt : -EFAULT;
1078 eventcnt++;
1079 uevent++;
1080 if (epi->event.events & EPOLLONESHOT)
1081 epi->event.events &= EP_PRIVATE_BITS;
1082 else if (!(epi->event.events & EPOLLET)) {
1084 * If this file has been added with Level
1085 * Trigger mode, we need to insert back inside
1086 * the ready list, so that the next call to
1087 * epoll_wait() will check again the events
1088 * availability. At this point, noone can insert
1089 * into ep->rdllist besides us. The epoll_ctl()
1090 * callers are locked out by
1091 * ep_scan_ready_list() holding "mtx" and the
1092 * poll callback will queue them in ep->ovflist.
1094 list_add_tail(&epi->rdllink, &ep->rdllist);
1099 return eventcnt;
1102 static int ep_send_events(struct eventpoll *ep,
1103 struct epoll_event __user *events, int maxevents)
1105 struct ep_send_events_data esed;
1107 esed.maxevents = maxevents;
1108 esed.events = events;
1110 return ep_scan_ready_list(ep, ep_send_events_proc, &esed);
1113 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1114 int maxevents, long timeout)
1116 int res, eavail;
1117 unsigned long flags;
1118 long jtimeout;
1119 wait_queue_t wait;
1122 * Calculate the timeout by checking for the "infinite" value (-1)
1123 * and the overflow condition. The passed timeout is in milliseconds,
1124 * that why (t * HZ) / 1000.
1126 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1127 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1129 retry:
1130 spin_lock_irqsave(&ep->lock, flags);
1132 res = 0;
1133 if (list_empty(&ep->rdllist)) {
1135 * We don't have any available event to return to the caller.
1136 * We need to sleep here, and we will be wake up by
1137 * ep_poll_callback() when events will become available.
1139 init_waitqueue_entry(&wait, current);
1140 wait.flags |= WQ_FLAG_EXCLUSIVE;
1141 __add_wait_queue(&ep->wq, &wait);
1143 for (;;) {
1145 * We don't want to sleep if the ep_poll_callback() sends us
1146 * a wakeup in between. That's why we set the task state
1147 * to TASK_INTERRUPTIBLE before doing the checks.
1149 set_current_state(TASK_INTERRUPTIBLE);
1150 if (!list_empty(&ep->rdllist) || !jtimeout)
1151 break;
1152 if (signal_pending(current)) {
1153 res = -EINTR;
1154 break;
1157 spin_unlock_irqrestore(&ep->lock, flags);
1158 jtimeout = schedule_timeout(jtimeout);
1159 spin_lock_irqsave(&ep->lock, flags);
1161 __remove_wait_queue(&ep->wq, &wait);
1163 set_current_state(TASK_RUNNING);
1165 /* Is it worth to try to dig for events ? */
1166 eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
1168 spin_unlock_irqrestore(&ep->lock, flags);
1171 * Try to transfer events to user space. In case we get 0 events and
1172 * there's still timeout left over, we go trying again in search of
1173 * more luck.
1175 if (!res && eavail &&
1176 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1177 goto retry;
1179 return res;
1183 * Open an eventpoll file descriptor.
1185 SYSCALL_DEFINE1(epoll_create1, int, flags)
1187 int error;
1188 struct eventpoll *ep = NULL;
1190 /* Check the EPOLL_* constant for consistency. */
1191 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1193 if (flags & ~EPOLL_CLOEXEC)
1194 return -EINVAL;
1196 * Create the internal data structure ("struct eventpoll").
1198 error = ep_alloc(&ep);
1199 if (error < 0)
1200 return error;
1202 * Creates all the items needed to setup an eventpoll file. That is,
1203 * a file structure and a free file descriptor.
1205 error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1206 flags & O_CLOEXEC);
1207 if (error < 0)
1208 ep_free(ep);
1210 return error;
1213 SYSCALL_DEFINE1(epoll_create, int, size)
1215 if (size <= 0)
1216 return -EINVAL;
1218 return sys_epoll_create1(0);
1222 * The following function implements the controller interface for
1223 * the eventpoll file that enables the insertion/removal/change of
1224 * file descriptors inside the interest set.
1226 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1227 struct epoll_event __user *, event)
1229 int error;
1230 struct file *file, *tfile;
1231 struct eventpoll *ep;
1232 struct epitem *epi;
1233 struct epoll_event epds;
1235 error = -EFAULT;
1236 if (ep_op_has_event(op) &&
1237 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1238 goto error_return;
1240 /* Get the "struct file *" for the eventpoll file */
1241 error = -EBADF;
1242 file = fget(epfd);
1243 if (!file)
1244 goto error_return;
1246 /* Get the "struct file *" for the target file */
1247 tfile = fget(fd);
1248 if (!tfile)
1249 goto error_fput;
1251 /* The target file descriptor must support poll */
1252 error = -EPERM;
1253 if (!tfile->f_op || !tfile->f_op->poll)
1254 goto error_tgt_fput;
1257 * We have to check that the file structure underneath the file descriptor
1258 * the user passed to us _is_ an eventpoll file. And also we do not permit
1259 * adding an epoll file descriptor inside itself.
1261 error = -EINVAL;
1262 if (file == tfile || !is_file_epoll(file))
1263 goto error_tgt_fput;
1266 * At this point it is safe to assume that the "private_data" contains
1267 * our own data structure.
1269 ep = file->private_data;
1271 mutex_lock(&ep->mtx);
1274 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1275 * above, we can be sure to be able to use the item looked up by
1276 * ep_find() till we release the mutex.
1278 epi = ep_find(ep, tfile, fd);
1280 error = -EINVAL;
1281 switch (op) {
1282 case EPOLL_CTL_ADD:
1283 if (!epi) {
1284 epds.events |= POLLERR | POLLHUP;
1285 error = ep_insert(ep, &epds, tfile, fd);
1286 } else
1287 error = -EEXIST;
1288 break;
1289 case EPOLL_CTL_DEL:
1290 if (epi)
1291 error = ep_remove(ep, epi);
1292 else
1293 error = -ENOENT;
1294 break;
1295 case EPOLL_CTL_MOD:
1296 if (epi) {
1297 epds.events |= POLLERR | POLLHUP;
1298 error = ep_modify(ep, epi, &epds);
1299 } else
1300 error = -ENOENT;
1301 break;
1303 mutex_unlock(&ep->mtx);
1305 error_tgt_fput:
1306 fput(tfile);
1307 error_fput:
1308 fput(file);
1309 error_return:
1311 return error;
1315 * Implement the event wait interface for the eventpoll file. It is the kernel
1316 * part of the user space epoll_wait(2).
1318 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1319 int, maxevents, int, timeout)
1321 int error;
1322 struct file *file;
1323 struct eventpoll *ep;
1325 /* The maximum number of event must be greater than zero */
1326 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1327 return -EINVAL;
1329 /* Verify that the area passed by the user is writeable */
1330 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1331 error = -EFAULT;
1332 goto error_return;
1335 /* Get the "struct file *" for the eventpoll file */
1336 error = -EBADF;
1337 file = fget(epfd);
1338 if (!file)
1339 goto error_return;
1342 * We have to check that the file structure underneath the fd
1343 * the user passed to us _is_ an eventpoll file.
1345 error = -EINVAL;
1346 if (!is_file_epoll(file))
1347 goto error_fput;
1350 * At this point it is safe to assume that the "private_data" contains
1351 * our own data structure.
1353 ep = file->private_data;
1355 /* Time to fish for events ... */
1356 error = ep_poll(ep, events, maxevents, timeout);
1358 error_fput:
1359 fput(file);
1360 error_return:
1362 return error;
1365 #ifdef HAVE_SET_RESTORE_SIGMASK
1368 * Implement the event wait interface for the eventpoll file. It is the kernel
1369 * part of the user space epoll_pwait(2).
1371 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1372 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1373 size_t, sigsetsize)
1375 int error;
1376 sigset_t ksigmask, sigsaved;
1379 * If the caller wants a certain signal mask to be set during the wait,
1380 * we apply it here.
1382 if (sigmask) {
1383 if (sigsetsize != sizeof(sigset_t))
1384 return -EINVAL;
1385 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1386 return -EFAULT;
1387 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1388 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1391 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1394 * If we changed the signal mask, we need to restore the original one.
1395 * In case we've got a signal while waiting, we do not restore the
1396 * signal mask yet, and we allow do_signal() to deliver the signal on
1397 * the way back to userspace, before the signal mask is restored.
1399 if (sigmask) {
1400 if (error == -EINTR) {
1401 memcpy(&current->saved_sigmask, &sigsaved,
1402 sizeof(sigsaved));
1403 set_restore_sigmask();
1404 } else
1405 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1408 return error;
1411 #endif /* HAVE_SET_RESTORE_SIGMASK */
1413 static int __init eventpoll_init(void)
1415 struct sysinfo si;
1417 si_meminfo(&si);
1419 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1421 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1422 EP_ITEM_COST;
1424 /* Initialize the structure used to perform safe poll wait head wake ups */
1425 ep_nested_calls_init(&poll_safewake_ncalls);
1427 /* Initialize the structure used to perform file's f_op->poll() calls */
1428 ep_nested_calls_init(&poll_readywalk_ncalls);
1430 /* Allocates slab cache used to allocate "struct epitem" items */
1431 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1432 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1434 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1435 pwq_cache = kmem_cache_create("eventpoll_pwq",
1436 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1438 return 0;
1440 fs_initcall(eventpoll_init);