2 * fs/eventpoll.c (Efficent event polling implementation)
3 * Copyright (C) 2001,...,2007 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.
77 #define DPRINTK(x) printk x
78 #define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
79 #else /* #if DEBUG_EPOLL > 0 */
80 #define DPRINTK(x) (void) 0
81 #define DNPRINTK(n, x) (void) 0
82 #endif /* #if DEBUG_EPOLL > 0 */
87 #define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
88 #else /* #if DEBUG_EPI != 0 */
89 #define EPI_SLAB_DEBUG 0
90 #endif /* #if DEBUG_EPI != 0 */
92 /* Epoll private bits inside the event mask */
93 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
95 /* Maximum number of poll wake up nests we are allowing */
96 #define EP_MAX_POLLWAKE_NESTS 4
98 /* Maximum msec timeout value storeable in a long int */
99 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
101 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
103 #define EP_UNACTIVE_PTR ((void *) -1L)
105 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
107 struct epoll_filefd {
113 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
114 * It is used to keep track on all tasks that are currently inside the wake_up() code
115 * to 1) short-circuit the one coming from the same task and same wait queue head
116 * (loop) 2) allow a maximum number of epoll descriptors inclusion nesting
117 * 3) let go the ones coming from other tasks.
119 struct wake_task_node {
120 struct list_head llink;
121 struct task_struct *task;
122 wait_queue_head_t *wq;
126 * This is used to implement the safe poll wake up avoiding to reenter
127 * the poll callback from inside wake_up().
129 struct poll_safewake {
130 struct list_head wake_task_list;
135 * Each file descriptor added to the eventpoll interface will
136 * have an entry of this type linked to the "rbr" RB tree.
139 /* RB tree node used to link this structure to the eventpoll RB tree */
142 /* List header used to link this structure to the eventpoll ready list */
143 struct list_head rdllink;
146 * Works together "struct eventpoll"->ovflist in keeping the
147 * single linked chain of items.
151 /* The file descriptor information this item refers to */
152 struct epoll_filefd ffd;
154 /* Number of active wait queue attached to poll operations */
157 /* List containing poll wait queues */
158 struct list_head pwqlist;
160 /* The "container" of this item */
161 struct eventpoll *ep;
163 /* List header used to link this item to the "struct file" items list */
164 struct list_head fllink;
166 /* The structure that describe the interested events and the source fd */
167 struct epoll_event event;
171 * This structure is stored inside the "private_data" member of the file
172 * structure and rapresent the main data sructure for the eventpoll
176 /* Protect the this structure access */
180 * This mutex is used to ensure that files are not removed
181 * while epoll is using them. This is held during the event
182 * collection loop, the file cleanup path, the epoll file exit
183 * code and the ctl operations.
187 /* Wait queue used by sys_epoll_wait() */
188 wait_queue_head_t wq;
190 /* Wait queue used by file->poll() */
191 wait_queue_head_t poll_wait;
193 /* List of ready file descriptors */
194 struct list_head rdllist;
196 /* RB tree root used to store monitored fd structs */
200 * This is a single linked list that chains all the "struct epitem" that
201 * happened while transfering ready events to userspace w/out
204 struct epitem *ovflist;
206 /* The user that created the eventpoll descriptor */
207 struct user_struct *user;
210 /* Wait structure used by the poll hooks */
211 struct eppoll_entry {
212 /* List header used to link this structure to the "struct epitem" */
213 struct list_head llink;
215 /* The "base" pointer is set to the container "struct epitem" */
219 * Wait queue item that will be linked to the target file wait
224 /* The wait queue head that linked the "wait" wait queue item */
225 wait_queue_head_t *whead;
228 /* Wrapper struct used by poll queueing */
235 * Configuration options available inside /proc/sys/fs/epoll/
237 /* Maximum number of epoll devices, per user */
238 static int max_user_instances __read_mostly;
239 /* Maximum number of epoll watched descriptors, per user */
240 static int max_user_watches __read_mostly;
243 * This mutex is used to serialize ep_free() and eventpoll_release_file().
245 static DEFINE_MUTEX(epmutex);
247 /* Safe wake up implementation */
248 static struct poll_safewake psw;
250 /* Slab cache used to allocate "struct epitem" */
251 static struct kmem_cache *epi_cache __read_mostly;
253 /* Slab cache used to allocate "struct eppoll_entry" */
254 static struct kmem_cache *pwq_cache __read_mostly;
258 #include <linux/sysctl.h>
262 ctl_table epoll_table[] = {
264 .procname = "max_user_instances",
265 .data = &max_user_instances,
266 .maxlen = sizeof(int),
268 .proc_handler = &proc_dointvec_minmax,
272 .procname = "max_user_watches",
273 .data = &max_user_watches,
274 .maxlen = sizeof(int),
276 .proc_handler = &proc_dointvec_minmax,
281 #endif /* CONFIG_SYSCTL */
284 /* Setup the structure that is used as key for the RB tree */
285 static inline void ep_set_ffd(struct epoll_filefd *ffd,
286 struct file *file, int fd)
292 /* Compare RB tree keys */
293 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
294 struct epoll_filefd *p2)
296 return (p1->file > p2->file ? +1:
297 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
300 /* Tells us if the item is currently linked */
301 static inline int ep_is_linked(struct list_head *p)
303 return !list_empty(p);
306 /* Get the "struct epitem" from a wait queue pointer */
307 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
309 return container_of(p, struct eppoll_entry, wait)->base;
312 /* Get the "struct epitem" from an epoll queue wrapper */
313 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
315 return container_of(p, struct ep_pqueue, pt)->epi;
318 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
319 static inline int ep_op_has_event(int op)
321 return op != EPOLL_CTL_DEL;
324 /* Initialize the poll safe wake up structure */
325 static void ep_poll_safewake_init(struct poll_safewake *psw)
328 INIT_LIST_HEAD(&psw->wake_task_list);
329 spin_lock_init(&psw->lock);
333 * Perform a safe wake up of the poll wait list. The problem is that
334 * with the new callback'd wake up system, it is possible that the
335 * poll callback is reentered from inside the call to wake_up() done
336 * on the poll wait queue head. The rule is that we cannot reenter the
337 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
338 * and we cannot reenter the same wait queue head at all. This will
339 * enable to have a hierarchy of epoll file descriptor of no more than
340 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
341 * because this one gets called by the poll callback, that in turn is called
342 * from inside a wake_up(), that might be called from irq context.
344 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
348 struct task_struct *this_task = current;
349 struct list_head *lsthead = &psw->wake_task_list;
350 struct wake_task_node *tncur;
351 struct wake_task_node tnode;
353 spin_lock_irqsave(&psw->lock, flags);
355 /* Try to see if the current task is already inside this wakeup call */
356 list_for_each_entry(tncur, lsthead, llink) {
358 if (tncur->wq == wq ||
359 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
361 * Ops ... loop detected or maximum nest level reached.
362 * We abort this wake by breaking the cycle itself.
364 spin_unlock_irqrestore(&psw->lock, flags);
369 /* Add the current task to the list */
370 tnode.task = this_task;
372 list_add(&tnode.llink, lsthead);
374 spin_unlock_irqrestore(&psw->lock, flags);
376 /* Do really wake up now */
377 wake_up_nested(wq, 1 + wake_nests);
379 /* Remove the current task from the list */
380 spin_lock_irqsave(&psw->lock, flags);
381 list_del(&tnode.llink);
382 spin_unlock_irqrestore(&psw->lock, flags);
386 * This function unregister poll callbacks from the associated file descriptor.
387 * Since this must be called without holding "ep->lock" the atomic exchange trick
388 * will protect us from multiple unregister.
390 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
393 struct list_head *lsthead = &epi->pwqlist;
394 struct eppoll_entry *pwq;
396 /* This is called without locks, so we need the atomic exchange */
397 nwait = xchg(&epi->nwait, 0);
400 while (!list_empty(lsthead)) {
401 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
403 list_del_init(&pwq->llink);
404 remove_wait_queue(pwq->whead, &pwq->wait);
405 kmem_cache_free(pwq_cache, pwq);
411 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
412 * all the associated resources. Must be called with "mtx" held.
414 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
417 struct file *file = epi->ffd.file;
420 * Removes poll wait queue hooks. We _have_ to do this without holding
421 * the "ep->lock" otherwise a deadlock might occur. This because of the
422 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
423 * queue head lock when unregistering the wait queue. The wakeup callback
424 * will run by holding the wait queue head lock and will call our callback
425 * that will try to get "ep->lock".
427 ep_unregister_pollwait(ep, epi);
429 /* Remove the current item from the list of epoll hooks */
430 spin_lock(&file->f_ep_lock);
431 if (ep_is_linked(&epi->fllink))
432 list_del_init(&epi->fllink);
433 spin_unlock(&file->f_ep_lock);
435 rb_erase(&epi->rbn, &ep->rbr);
437 spin_lock_irqsave(&ep->lock, flags);
438 if (ep_is_linked(&epi->rdllink))
439 list_del_init(&epi->rdllink);
440 spin_unlock_irqrestore(&ep->lock, flags);
442 /* At this point it is safe to free the eventpoll item */
443 kmem_cache_free(epi_cache, epi);
445 atomic_dec(&ep->user->epoll_watches);
447 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
453 static void ep_free(struct eventpoll *ep)
458 /* We need to release all tasks waiting for these file */
459 if (waitqueue_active(&ep->poll_wait))
460 ep_poll_safewake(&psw, &ep->poll_wait);
463 * We need to lock this because we could be hit by
464 * eventpoll_release_file() while we're freeing the "struct eventpoll".
465 * We do not need to hold "ep->mtx" here because the epoll file
466 * is on the way to be removed and no one has references to it
467 * anymore. The only hit might come from eventpoll_release_file() but
468 * holding "epmutex" is sufficent here.
470 mutex_lock(&epmutex);
473 * Walks through the whole tree by unregistering poll callbacks.
475 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
476 epi = rb_entry(rbp, struct epitem, rbn);
478 ep_unregister_pollwait(ep, epi);
482 * Walks through the whole tree by freeing each "struct epitem". At this
483 * point we are sure no poll callbacks will be lingering around, and also by
484 * holding "epmutex" we can be sure that no file cleanup code will hit
485 * us during this operation. So we can avoid the lock on "ep->lock".
487 while ((rbp = rb_first(&ep->rbr)) != NULL) {
488 epi = rb_entry(rbp, struct epitem, rbn);
492 mutex_unlock(&epmutex);
493 mutex_destroy(&ep->mtx);
494 atomic_dec(&ep->user->epoll_devs);
499 static int ep_eventpoll_release(struct inode *inode, struct file *file)
501 struct eventpoll *ep = file->private_data;
506 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
510 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
512 unsigned int pollflags = 0;
514 struct eventpoll *ep = file->private_data;
516 /* Insert inside our poll wait queue */
517 poll_wait(file, &ep->poll_wait, wait);
519 /* Check our condition */
520 spin_lock_irqsave(&ep->lock, flags);
521 if (!list_empty(&ep->rdllist))
522 pollflags = POLLIN | POLLRDNORM;
523 spin_unlock_irqrestore(&ep->lock, flags);
528 /* File callbacks that implement the eventpoll file behaviour */
529 static const struct file_operations eventpoll_fops = {
530 .release = ep_eventpoll_release,
531 .poll = ep_eventpoll_poll
534 /* Fast test to see if the file is an evenpoll file */
535 static inline int is_file_epoll(struct file *f)
537 return f->f_op == &eventpoll_fops;
541 * This is called from eventpoll_release() to unlink files from the eventpoll
542 * interface. We need to have this facility to cleanup correctly files that are
543 * closed without being removed from the eventpoll interface.
545 void eventpoll_release_file(struct file *file)
547 struct list_head *lsthead = &file->f_ep_links;
548 struct eventpoll *ep;
552 * We don't want to get "file->f_ep_lock" because it is not
553 * necessary. It is not necessary because we're in the "struct file"
554 * cleanup path, and this means that noone is using this file anymore.
555 * So, for example, epoll_ctl() cannot hit here sicne if we reach this
556 * point, the file counter already went to zero and fget() would fail.
557 * The only hit might come from ep_free() but by holding the mutex
558 * will correctly serialize the operation. We do need to acquire
559 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
560 * from anywhere but ep_free().
562 mutex_lock(&epmutex);
564 while (!list_empty(lsthead)) {
565 epi = list_first_entry(lsthead, struct epitem, fllink);
568 list_del_init(&epi->fllink);
569 mutex_lock(&ep->mtx);
571 mutex_unlock(&ep->mtx);
574 mutex_unlock(&epmutex);
577 static int ep_alloc(struct eventpoll **pep)
580 struct user_struct *user;
581 struct eventpoll *ep;
583 user = get_current_user();
585 if (unlikely(atomic_read(&user->epoll_devs) >=
589 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
593 spin_lock_init(&ep->lock);
594 mutex_init(&ep->mtx);
595 init_waitqueue_head(&ep->wq);
596 init_waitqueue_head(&ep->poll_wait);
597 INIT_LIST_HEAD(&ep->rdllist);
599 ep->ovflist = EP_UNACTIVE_PTR;
604 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
614 * Search the file inside the eventpoll tree. The RB tree operations
615 * are protected by the "mtx" mutex, and ep_find() must be called with
618 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
622 struct epitem *epi, *epir = NULL;
623 struct epoll_filefd ffd;
625 ep_set_ffd(&ffd, file, fd);
626 for (rbp = ep->rbr.rb_node; rbp; ) {
627 epi = rb_entry(rbp, struct epitem, rbn);
628 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
639 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
640 current, file, epir));
646 * This is the callback that is passed to the wait queue wakeup
647 * machanism. It is called by the stored file descriptors when they
648 * have events to report.
650 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
654 struct epitem *epi = ep_item_from_wait(wait);
655 struct eventpoll *ep = epi->ep;
657 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
658 current, epi->ffd.file, epi, ep));
660 spin_lock_irqsave(&ep->lock, flags);
663 * If the event mask does not contain any poll(2) event, we consider the
664 * descriptor to be disabled. This condition is likely the effect of the
665 * EPOLLONESHOT bit that disables the descriptor when an event is received,
666 * until the next EPOLL_CTL_MOD will be issued.
668 if (!(epi->event.events & ~EP_PRIVATE_BITS))
672 * If we are trasfering events to userspace, we can hold no locks
673 * (because we're accessing user memory, and because of linux f_op->poll()
674 * semantics). All the events that happens during that period of time are
675 * chained in ep->ovflist and requeued later on.
677 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
678 if (epi->next == EP_UNACTIVE_PTR) {
679 epi->next = ep->ovflist;
685 /* If this file is already in the ready list we exit soon */
686 if (ep_is_linked(&epi->rdllink))
689 list_add_tail(&epi->rdllink, &ep->rdllist);
693 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
696 if (waitqueue_active(&ep->wq))
697 wake_up_locked(&ep->wq);
698 if (waitqueue_active(&ep->poll_wait))
702 spin_unlock_irqrestore(&ep->lock, flags);
704 /* We have to call this outside the lock */
706 ep_poll_safewake(&psw, &ep->poll_wait);
712 * This is the callback that is used to add our wait queue to the
713 * target file wakeup lists.
715 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
718 struct epitem *epi = ep_item_from_epqueue(pt);
719 struct eppoll_entry *pwq;
721 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
722 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
725 add_wait_queue(whead, &pwq->wait);
726 list_add_tail(&pwq->llink, &epi->pwqlist);
729 /* We have to signal that an error occurred */
734 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
737 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
742 epic = rb_entry(parent, struct epitem, rbn);
743 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
745 p = &parent->rb_right;
747 p = &parent->rb_left;
749 rb_link_node(&epi->rbn, parent, p);
750 rb_insert_color(&epi->rbn, &ep->rbr);
754 * Must be called with "mtx" held.
756 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
757 struct file *tfile, int fd)
759 int error, revents, pwake = 0;
762 struct ep_pqueue epq;
764 if (unlikely(atomic_read(&ep->user->epoll_watches) >=
767 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
770 /* Item initialization follow here ... */
771 INIT_LIST_HEAD(&epi->rdllink);
772 INIT_LIST_HEAD(&epi->fllink);
773 INIT_LIST_HEAD(&epi->pwqlist);
775 ep_set_ffd(&epi->ffd, tfile, fd);
778 epi->next = EP_UNACTIVE_PTR;
780 /* Initialize the poll table using the queue callback */
782 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
785 * Attach the item to the poll hooks and get current event bits.
786 * We can safely use the file* here because its usage count has
787 * been increased by the caller of this function. Note that after
788 * this operation completes, the poll callback can start hitting
791 revents = tfile->f_op->poll(tfile, &epq.pt);
794 * We have to check if something went wrong during the poll wait queue
795 * install process. Namely an allocation for a wait queue failed due
796 * high memory pressure.
800 goto error_unregister;
802 /* Add the current item to the list of active epoll hook for this file */
803 spin_lock(&tfile->f_ep_lock);
804 list_add_tail(&epi->fllink, &tfile->f_ep_links);
805 spin_unlock(&tfile->f_ep_lock);
808 * Add the current item to the RB tree. All RB tree operations are
809 * protected by "mtx", and ep_insert() is called with "mtx" held.
811 ep_rbtree_insert(ep, epi);
813 /* We have to drop the new item inside our item list to keep track of it */
814 spin_lock_irqsave(&ep->lock, flags);
816 /* If the file is already "ready" we drop it inside the ready list */
817 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
818 list_add_tail(&epi->rdllink, &ep->rdllist);
820 /* Notify waiting tasks that events are available */
821 if (waitqueue_active(&ep->wq))
822 wake_up_locked(&ep->wq);
823 if (waitqueue_active(&ep->poll_wait))
827 spin_unlock_irqrestore(&ep->lock, flags);
829 atomic_inc(&ep->user->epoll_watches);
831 /* We have to call this outside the lock */
833 ep_poll_safewake(&psw, &ep->poll_wait);
835 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
836 current, ep, tfile, fd));
841 ep_unregister_pollwait(ep, epi);
844 * We need to do this because an event could have been arrived on some
845 * allocated wait queue. Note that we don't care about the ep->ovflist
846 * list, since that is used/cleaned only inside a section bound by "mtx".
847 * And ep_insert() is called with "mtx" held.
849 spin_lock_irqsave(&ep->lock, flags);
850 if (ep_is_linked(&epi->rdllink))
851 list_del_init(&epi->rdllink);
852 spin_unlock_irqrestore(&ep->lock, flags);
854 kmem_cache_free(epi_cache, epi);
860 * Modify the interest event mask by dropping an event if the new mask
861 * has a match in the current file status. Must be called with "mtx" held.
863 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
866 unsigned int revents;
870 * Set the new event interest mask before calling f_op->poll(), otherwise
871 * a potential race might occur. In fact if we do this operation inside
872 * the lock, an event might happen between the f_op->poll() call and the
873 * new event set registering.
875 epi->event.events = event->events;
878 * Get current event bits. We can safely use the file* here because
879 * its usage count has been increased by the caller of this function.
881 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
883 spin_lock_irqsave(&ep->lock, flags);
885 /* Copy the data member from inside the lock */
886 epi->event.data = event->data;
889 * If the item is "hot" and it is not registered inside the ready
890 * list, push it inside.
892 if (revents & event->events) {
893 if (!ep_is_linked(&epi->rdllink)) {
894 list_add_tail(&epi->rdllink, &ep->rdllist);
896 /* Notify waiting tasks that events are available */
897 if (waitqueue_active(&ep->wq))
898 wake_up_locked(&ep->wq);
899 if (waitqueue_active(&ep->poll_wait))
903 spin_unlock_irqrestore(&ep->lock, flags);
905 /* We have to call this outside the lock */
907 ep_poll_safewake(&psw, &ep->poll_wait);
912 static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
915 int eventcnt, error = -EFAULT, pwake = 0;
916 unsigned int revents;
918 struct epitem *epi, *nepi;
919 struct list_head txlist;
921 INIT_LIST_HEAD(&txlist);
924 * We need to lock this because we could be hit by
925 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
927 mutex_lock(&ep->mtx);
930 * Steal the ready list, and re-init the original one to the
931 * empty list. Also, set ep->ovflist to NULL so that events
932 * happening while looping w/out locks, are not lost. We cannot
933 * have the poll callback to queue directly on ep->rdllist,
934 * because we are doing it in the loop below, in a lockless way.
936 spin_lock_irqsave(&ep->lock, flags);
937 list_splice(&ep->rdllist, &txlist);
938 INIT_LIST_HEAD(&ep->rdllist);
940 spin_unlock_irqrestore(&ep->lock, flags);
943 * We can loop without lock because this is a task private list.
944 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
945 * Items cannot vanish during the loop because we are holding "mtx".
947 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
948 epi = list_first_entry(&txlist, struct epitem, rdllink);
950 list_del_init(&epi->rdllink);
953 * Get the ready file event set. We can safely use the file
954 * because we are holding the "mtx" and this will guarantee
955 * that both the file and the item will not vanish.
957 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
958 revents &= epi->event.events;
961 * Is the event mask intersect the caller-requested one,
962 * deliver the event to userspace. Again, we are holding
963 * "mtx", so no operations coming from userspace can change
967 if (__put_user(revents,
968 &events[eventcnt].events) ||
969 __put_user(epi->event.data,
970 &events[eventcnt].data))
972 if (epi->event.events & EPOLLONESHOT)
973 epi->event.events &= EP_PRIVATE_BITS;
977 * At this point, noone can insert into ep->rdllist besides
978 * us. The epoll_ctl() callers are locked out by us holding
979 * "mtx" and the poll callback will queue them in ep->ovflist.
981 if (!(epi->event.events & EPOLLET) &&
982 (revents & epi->event.events))
983 list_add_tail(&epi->rdllink, &ep->rdllist);
989 spin_lock_irqsave(&ep->lock, flags);
991 * During the time we spent in the loop above, some other events
992 * might have been queued by the poll callback. We re-insert them
993 * inside the main ready-list here.
995 for (nepi = ep->ovflist; (epi = nepi) != NULL;
996 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
998 * If the above loop quit with errors, the epoll item might still
999 * be linked to "txlist", and the list_splice() done below will
1000 * take care of those cases.
1002 if (!ep_is_linked(&epi->rdllink))
1003 list_add_tail(&epi->rdllink, &ep->rdllist);
1006 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
1007 * releasing the lock, events will be queued in the normal way inside
1010 ep->ovflist = EP_UNACTIVE_PTR;
1013 * In case of error in the event-send loop, or in case the number of
1014 * ready events exceeds the userspace limit, we need to splice the
1015 * "txlist" back inside ep->rdllist.
1017 list_splice(&txlist, &ep->rdllist);
1019 if (!list_empty(&ep->rdllist)) {
1021 * Wake up (if active) both the eventpoll wait list and the ->poll()
1022 * wait list (delayed after we release the lock).
1024 if (waitqueue_active(&ep->wq))
1025 wake_up_locked(&ep->wq);
1026 if (waitqueue_active(&ep->poll_wait))
1029 spin_unlock_irqrestore(&ep->lock, flags);
1031 mutex_unlock(&ep->mtx);
1033 /* We have to call this outside the lock */
1035 ep_poll_safewake(&psw, &ep->poll_wait);
1037 return eventcnt == 0 ? error: eventcnt;
1040 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1041 int maxevents, long timeout)
1044 unsigned long flags;
1049 * Calculate the timeout by checking for the "infinite" value ( -1 )
1050 * and the overflow condition. The passed timeout is in milliseconds,
1051 * that why (t * HZ) / 1000.
1053 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1054 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1057 spin_lock_irqsave(&ep->lock, flags);
1060 if (list_empty(&ep->rdllist)) {
1062 * We don't have any available event to return to the caller.
1063 * We need to sleep here, and we will be wake up by
1064 * ep_poll_callback() when events will become available.
1066 init_waitqueue_entry(&wait, current);
1067 wait.flags |= WQ_FLAG_EXCLUSIVE;
1068 __add_wait_queue(&ep->wq, &wait);
1072 * We don't want to sleep if the ep_poll_callback() sends us
1073 * a wakeup in between. That's why we set the task state
1074 * to TASK_INTERRUPTIBLE before doing the checks.
1076 set_current_state(TASK_INTERRUPTIBLE);
1077 if (!list_empty(&ep->rdllist) || !jtimeout)
1079 if (signal_pending(current)) {
1084 spin_unlock_irqrestore(&ep->lock, flags);
1085 jtimeout = schedule_timeout(jtimeout);
1086 spin_lock_irqsave(&ep->lock, flags);
1088 __remove_wait_queue(&ep->wq, &wait);
1090 set_current_state(TASK_RUNNING);
1093 /* Is it worth to try to dig for events ? */
1094 eavail = !list_empty(&ep->rdllist);
1096 spin_unlock_irqrestore(&ep->lock, flags);
1099 * Try to transfer events to user space. In case we get 0 events and
1100 * there's still timeout left over, we go trying again in search of
1103 if (!res && eavail &&
1104 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1111 * Open an eventpoll file descriptor.
1113 asmlinkage long sys_epoll_create1(int flags)
1116 struct eventpoll *ep;
1118 /* Check the EPOLL_* constant for consistency. */
1119 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1121 if (flags & ~EPOLL_CLOEXEC)
1124 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
1128 * Create the internal data structure ( "struct eventpoll" ).
1130 error = ep_alloc(&ep);
1137 * Creates all the items needed to setup an eventpoll file. That is,
1138 * a file structure and a free file descriptor.
1140 fd = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1144 atomic_inc(&ep->user->epoll_devs);
1147 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1148 current, flags, fd));
1153 asmlinkage long sys_epoll_create(int size)
1158 return sys_epoll_create1(0);
1162 * The following function implements the controller interface for
1163 * the eventpoll file that enables the insertion/removal/change of
1164 * file descriptors inside the interest set.
1166 asmlinkage long sys_epoll_ctl(int epfd, int op, int fd,
1167 struct epoll_event __user *event)
1170 struct file *file, *tfile;
1171 struct eventpoll *ep;
1173 struct epoll_event epds;
1175 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1176 current, epfd, op, fd, event));
1179 if (ep_op_has_event(op) &&
1180 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1183 /* Get the "struct file *" for the eventpoll file */
1189 /* Get the "struct file *" for the target file */
1194 /* The target file descriptor must support poll */
1196 if (!tfile->f_op || !tfile->f_op->poll)
1197 goto error_tgt_fput;
1200 * We have to check that the file structure underneath the file descriptor
1201 * the user passed to us _is_ an eventpoll file. And also we do not permit
1202 * adding an epoll file descriptor inside itself.
1205 if (file == tfile || !is_file_epoll(file))
1206 goto error_tgt_fput;
1209 * At this point it is safe to assume that the "private_data" contains
1210 * our own data structure.
1212 ep = file->private_data;
1214 mutex_lock(&ep->mtx);
1217 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1218 * above, we can be sure to be able to use the item looked up by
1219 * ep_find() till we release the mutex.
1221 epi = ep_find(ep, tfile, fd);
1227 epds.events |= POLLERR | POLLHUP;
1229 error = ep_insert(ep, &epds, tfile, fd);
1235 error = ep_remove(ep, epi);
1241 epds.events |= POLLERR | POLLHUP;
1242 error = ep_modify(ep, epi, &epds);
1247 mutex_unlock(&ep->mtx);
1254 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1255 current, epfd, op, fd, event, error));
1261 * Implement the event wait interface for the eventpoll file. It is the kernel
1262 * part of the user space epoll_wait(2).
1264 asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
1265 int maxevents, int timeout)
1269 struct eventpoll *ep;
1271 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1272 current, epfd, events, maxevents, timeout));
1274 /* The maximum number of event must be greater than zero */
1275 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1278 /* Verify that the area passed by the user is writeable */
1279 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1284 /* Get the "struct file *" for the eventpoll file */
1291 * We have to check that the file structure underneath the fd
1292 * the user passed to us _is_ an eventpoll file.
1295 if (!is_file_epoll(file))
1299 * At this point it is safe to assume that the "private_data" contains
1300 * our own data structure.
1302 ep = file->private_data;
1304 /* Time to fish for events ... */
1305 error = ep_poll(ep, events, maxevents, timeout);
1310 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1311 current, epfd, events, maxevents, timeout, error));
1316 #ifdef HAVE_SET_RESTORE_SIGMASK
1319 * Implement the event wait interface for the eventpoll file. It is the kernel
1320 * part of the user space epoll_pwait(2).
1322 asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
1323 int maxevents, int timeout, const sigset_t __user *sigmask,
1327 sigset_t ksigmask, sigsaved;
1330 * If the caller wants a certain signal mask to be set during the wait,
1334 if (sigsetsize != sizeof(sigset_t))
1336 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1338 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1339 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1342 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1345 * If we changed the signal mask, we need to restore the original one.
1346 * In case we've got a signal while waiting, we do not restore the
1347 * signal mask yet, and we allow do_signal() to deliver the signal on
1348 * the way back to userspace, before the signal mask is restored.
1351 if (error == -EINTR) {
1352 memcpy(¤t->saved_sigmask, &sigsaved,
1354 set_restore_sigmask();
1356 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1362 #endif /* HAVE_SET_RESTORE_SIGMASK */
1364 static int __init eventpoll_init(void)
1369 max_user_instances = 128;
1370 max_user_watches = (((si.totalram - si.totalhigh) / 32) << PAGE_SHIFT) /
1373 /* Initialize the structure used to perform safe poll wait head wake ups */
1374 ep_poll_safewake_init(&psw);
1376 /* Allocates slab cache used to allocate "struct epitem" items */
1377 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1378 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1381 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1382 pwq_cache = kmem_cache_create("eventpoll_pwq",
1383 sizeof(struct eppoll_entry), 0,
1384 EPI_SLAB_DEBUG|SLAB_PANIC, NULL);
1388 fs_initcall(eventpoll_init);