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 struct epoll_filefd {
111 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
112 * It is used to keep track on all tasks that are currently inside the wake_up() code
113 * to 1) short-circuit the one coming from the same task and same wait queue head
114 * (loop) 2) allow a maximum number of epoll descriptors inclusion nesting
115 * 3) let go the ones coming from other tasks.
117 struct wake_task_node {
118 struct list_head llink;
119 struct task_struct *task;
120 wait_queue_head_t *wq;
124 * This is used to implement the safe poll wake up avoiding to reenter
125 * the poll callback from inside wake_up().
127 struct poll_safewake {
128 struct list_head wake_task_list;
133 * Each file descriptor added to the eventpoll interface will
134 * have an entry of this type linked to the "rbr" RB tree.
137 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
140 /* List header used to link this structure to the eventpoll ready list */
141 struct list_head rdllink;
144 * Works together "struct eventpoll"->ovflist in keeping the
145 * single linked chain of items.
149 /* The file descriptor information this item refers to */
150 struct epoll_filefd ffd;
152 /* Number of active wait queue attached to poll operations */
155 /* List containing poll wait queues */
156 struct list_head pwqlist;
158 /* The "container" of this item */
159 struct eventpoll *ep;
161 /* List header used to link this item to the "struct file" items list */
162 struct list_head fllink;
164 /* The structure that describe the interested events and the source fd */
165 struct epoll_event event;
169 * This structure is stored inside the "private_data" member of the file
170 * structure and rapresent the main data sructure for the eventpoll
174 /* Protect the this structure access */
178 * This mutex is used to ensure that files are not removed
179 * while epoll is using them. This is held during the event
180 * collection loop, the file cleanup path, the epoll file exit
181 * code and the ctl operations.
185 /* Wait queue used by sys_epoll_wait() */
186 wait_queue_head_t wq;
188 /* Wait queue used by file->poll() */
189 wait_queue_head_t poll_wait;
191 /* List of ready file descriptors */
192 struct list_head rdllist;
194 /* RB-Tree root used to store monitored fd structs */
198 * This is a single linked list that chains all the "struct epitem" that
199 * happened while transfering ready events to userspace w/out
202 struct epitem *ovflist;
205 /* Wait structure used by the poll hooks */
206 struct eppoll_entry {
207 /* List header used to link this structure to the "struct epitem" */
208 struct list_head llink;
210 /* The "base" pointer is set to the container "struct epitem" */
214 * Wait queue item that will be linked to the target file wait
219 /* The wait queue head that linked the "wait" wait queue item */
220 wait_queue_head_t *whead;
223 /* Wrapper struct used by poll queueing */
230 * This mutex is used to serialize ep_free() and eventpoll_release_file().
232 static struct mutex epmutex;
234 /* Safe wake up implementation */
235 static struct poll_safewake psw;
237 /* Slab cache used to allocate "struct epitem" */
238 static struct kmem_cache *epi_cache __read_mostly;
240 /* Slab cache used to allocate "struct eppoll_entry" */
241 static struct kmem_cache *pwq_cache __read_mostly;
244 /* Setup the structure that is used as key for the rb-tree */
245 static inline void ep_set_ffd(struct epoll_filefd *ffd,
246 struct file *file, int fd)
252 /* Compare rb-tree keys */
253 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
254 struct epoll_filefd *p2)
256 return (p1->file > p2->file ? +1:
257 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
260 /* Special initialization for the rb-tree node to detect linkage */
261 static inline void ep_rb_initnode(struct rb_node *n)
266 /* Removes a node from the rb-tree and marks it for a fast is-linked check */
267 static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
273 /* Fast check to verify that the item is linked to the main rb-tree */
274 static inline int ep_rb_linked(struct rb_node *n)
276 return rb_parent(n) != n;
279 /* Tells us if the item is currently linked */
280 static inline int ep_is_linked(struct list_head *p)
282 return !list_empty(p);
285 /* Get the "struct epitem" from a wait queue pointer */
286 static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
288 return container_of(p, struct eppoll_entry, wait)->base;
291 /* Get the "struct epitem" from an epoll queue wrapper */
292 static inline struct epitem * ep_item_from_epqueue(poll_table *p)
294 return container_of(p, struct ep_pqueue, pt)->epi;
297 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
298 static inline int ep_op_has_event(int op)
300 return op != EPOLL_CTL_DEL;
303 /* Initialize the poll safe wake up structure */
304 static void ep_poll_safewake_init(struct poll_safewake *psw)
307 INIT_LIST_HEAD(&psw->wake_task_list);
308 spin_lock_init(&psw->lock);
312 * Perform a safe wake up of the poll wait list. The problem is that
313 * with the new callback'd wake up system, it is possible that the
314 * poll callback is reentered from inside the call to wake_up() done
315 * on the poll wait queue head. The rule is that we cannot reenter the
316 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
317 * and we cannot reenter the same wait queue head at all. This will
318 * enable to have a hierarchy of epoll file descriptor of no more than
319 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
320 * because this one gets called by the poll callback, that in turn is called
321 * from inside a wake_up(), that might be called from irq context.
323 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
327 struct task_struct *this_task = current;
328 struct list_head *lsthead = &psw->wake_task_list, *lnk;
329 struct wake_task_node *tncur;
330 struct wake_task_node tnode;
332 spin_lock_irqsave(&psw->lock, flags);
334 /* Try to see if the current task is already inside this wakeup call */
335 list_for_each(lnk, lsthead) {
336 tncur = list_entry(lnk, struct wake_task_node, llink);
338 if (tncur->wq == wq ||
339 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
341 * Ops ... loop detected or maximum nest level reached.
342 * We abort this wake by breaking the cycle itself.
344 spin_unlock_irqrestore(&psw->lock, flags);
349 /* Add the current task to the list */
350 tnode.task = this_task;
352 list_add(&tnode.llink, lsthead);
354 spin_unlock_irqrestore(&psw->lock, flags);
356 /* Do really wake up now */
359 /* Remove the current task from the list */
360 spin_lock_irqsave(&psw->lock, flags);
361 list_del(&tnode.llink);
362 spin_unlock_irqrestore(&psw->lock, flags);
366 * This function unregister poll callbacks from the associated file descriptor.
367 * Since this must be called without holding "ep->lock" the atomic exchange trick
368 * will protect us from multiple unregister.
370 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
373 struct list_head *lsthead = &epi->pwqlist;
374 struct eppoll_entry *pwq;
376 /* This is called without locks, so we need the atomic exchange */
377 nwait = xchg(&epi->nwait, 0);
380 while (!list_empty(lsthead)) {
381 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
383 list_del_init(&pwq->llink);
384 remove_wait_queue(pwq->whead, &pwq->wait);
385 kmem_cache_free(pwq_cache, pwq);
391 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
392 * all the associated resources. Must be called with "mtx" held.
394 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
397 struct file *file = epi->ffd.file;
400 * Removes poll wait queue hooks. We _have_ to do this without holding
401 * the "ep->lock" otherwise a deadlock might occur. This because of the
402 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
403 * queue head lock when unregistering the wait queue. The wakeup callback
404 * will run by holding the wait queue head lock and will call our callback
405 * that will try to get "ep->lock".
407 ep_unregister_pollwait(ep, epi);
409 /* Remove the current item from the list of epoll hooks */
410 spin_lock(&file->f_ep_lock);
411 if (ep_is_linked(&epi->fllink))
412 list_del_init(&epi->fllink);
413 spin_unlock(&file->f_ep_lock);
415 if (ep_rb_linked(&epi->rbn))
416 ep_rb_erase(&epi->rbn, &ep->rbr);
418 spin_lock_irqsave(&ep->lock, flags);
419 if (ep_is_linked(&epi->rdllink))
420 list_del_init(&epi->rdllink);
421 spin_unlock_irqrestore(&ep->lock, flags);
423 /* At this point it is safe to free the eventpoll item */
424 kmem_cache_free(epi_cache, epi);
426 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
432 static void ep_free(struct eventpoll *ep)
437 /* We need to release all tasks waiting for these file */
438 if (waitqueue_active(&ep->poll_wait))
439 ep_poll_safewake(&psw, &ep->poll_wait);
442 * We need to lock this because we could be hit by
443 * eventpoll_release_file() while we're freeing the "struct eventpoll".
444 * We do not need to hold "ep->mtx" here because the epoll file
445 * is on the way to be removed and no one has references to it
446 * anymore. The only hit might come from eventpoll_release_file() but
447 * holding "epmutex" is sufficent here.
449 mutex_lock(&epmutex);
452 * Walks through the whole tree by unregistering poll callbacks.
454 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
455 epi = rb_entry(rbp, struct epitem, rbn);
457 ep_unregister_pollwait(ep, epi);
461 * Walks through the whole tree by freeing each "struct epitem". At this
462 * point we are sure no poll callbacks will be lingering around, and also by
463 * holding "epmutex" we can be sure that no file cleanup code will hit
464 * us during this operation. So we can avoid the lock on "ep->lock".
466 while ((rbp = rb_first(&ep->rbr)) != 0) {
467 epi = rb_entry(rbp, struct epitem, rbn);
471 mutex_unlock(&epmutex);
473 mutex_destroy(&ep->mtx);
476 static int ep_eventpoll_release(struct inode *inode, struct file *file)
478 struct eventpoll *ep = file->private_data;
485 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
489 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
491 unsigned int pollflags = 0;
493 struct eventpoll *ep = file->private_data;
495 /* Insert inside our poll wait queue */
496 poll_wait(file, &ep->poll_wait, wait);
498 /* Check our condition */
499 spin_lock_irqsave(&ep->lock, flags);
500 if (!list_empty(&ep->rdllist))
501 pollflags = POLLIN | POLLRDNORM;
502 spin_unlock_irqrestore(&ep->lock, flags);
507 /* File callbacks that implement the eventpoll file behaviour */
508 static const struct file_operations eventpoll_fops = {
509 .release = ep_eventpoll_release,
510 .poll = ep_eventpoll_poll
513 /* Fast test to see if the file is an evenpoll file */
514 static inline int is_file_epoll(struct file *f)
516 return f->f_op == &eventpoll_fops;
520 * This is called from eventpoll_release() to unlink files from the eventpoll
521 * interface. We need to have this facility to cleanup correctly files that are
522 * closed without being removed from the eventpoll interface.
524 void eventpoll_release_file(struct file *file)
526 struct list_head *lsthead = &file->f_ep_links;
527 struct eventpoll *ep;
531 * We don't want to get "file->f_ep_lock" because it is not
532 * necessary. It is not necessary because we're in the "struct file"
533 * cleanup path, and this means that noone is using this file anymore.
534 * The only hit might come from ep_free() but by holding the mutex
535 * will correctly serialize the operation. We do need to acquire
536 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
537 * from anywhere but ep_free().
539 mutex_lock(&epmutex);
541 while (!list_empty(lsthead)) {
542 epi = list_first_entry(lsthead, struct epitem, fllink);
545 list_del_init(&epi->fllink);
546 mutex_lock(&ep->mtx);
548 mutex_unlock(&ep->mtx);
551 mutex_unlock(&epmutex);
554 static int ep_alloc(struct eventpoll **pep)
556 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
561 spin_lock_init(&ep->lock);
562 mutex_init(&ep->mtx);
563 init_waitqueue_head(&ep->wq);
564 init_waitqueue_head(&ep->poll_wait);
565 INIT_LIST_HEAD(&ep->rdllist);
567 ep->ovflist = EP_UNACTIVE_PTR;
571 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
577 * Search the file inside the eventpoll tree. The RB tree operations
578 * are protected by the "mtx" mutex, and ep_find() must be called with
581 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
585 struct epitem *epi, *epir = NULL;
586 struct epoll_filefd ffd;
588 ep_set_ffd(&ffd, file, fd);
589 for (rbp = ep->rbr.rb_node; rbp; ) {
590 epi = rb_entry(rbp, struct epitem, rbn);
591 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
602 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
603 current, file, epir));
609 * This is the callback that is passed to the wait queue wakeup
610 * machanism. It is called by the stored file descriptors when they
611 * have events to report.
613 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
617 struct epitem *epi = ep_item_from_wait(wait);
618 struct eventpoll *ep = epi->ep;
620 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
621 current, epi->ffd.file, epi, ep));
623 spin_lock_irqsave(&ep->lock, flags);
626 * If the event mask does not contain any poll(2) event, we consider the
627 * descriptor to be disabled. This condition is likely the effect of the
628 * EPOLLONESHOT bit that disables the descriptor when an event is received,
629 * until the next EPOLL_CTL_MOD will be issued.
631 if (!(epi->event.events & ~EP_PRIVATE_BITS))
635 * If we are trasfering events to userspace, we can hold no locks
636 * (because we're accessing user memory, and because of linux f_op->poll()
637 * semantics). All the events that happens during that period of time are
638 * chained in ep->ovflist and requeued later on.
640 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
641 if (epi->next == EP_UNACTIVE_PTR) {
642 epi->next = ep->ovflist;
648 /* If this file is already in the ready list we exit soon */
649 if (ep_is_linked(&epi->rdllink))
652 list_add_tail(&epi->rdllink, &ep->rdllist);
656 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
659 if (waitqueue_active(&ep->wq))
660 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
662 if (waitqueue_active(&ep->poll_wait))
666 spin_unlock_irqrestore(&ep->lock, flags);
668 /* We have to call this outside the lock */
670 ep_poll_safewake(&psw, &ep->poll_wait);
676 * This is the callback that is used to add our wait queue to the
677 * target file wakeup lists.
679 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
682 struct epitem *epi = ep_item_from_epqueue(pt);
683 struct eppoll_entry *pwq;
685 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
686 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
689 add_wait_queue(whead, &pwq->wait);
690 list_add_tail(&pwq->llink, &epi->pwqlist);
693 /* We have to signal that an error occurred */
698 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
701 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
706 epic = rb_entry(parent, struct epitem, rbn);
707 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
709 p = &parent->rb_right;
711 p = &parent->rb_left;
713 rb_link_node(&epi->rbn, parent, p);
714 rb_insert_color(&epi->rbn, &ep->rbr);
718 * Must be called with "mtx" held.
720 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
721 struct file *tfile, int fd)
723 int error, revents, pwake = 0;
726 struct ep_pqueue epq;
729 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
732 /* Item initialization follow here ... */
733 ep_rb_initnode(&epi->rbn);
734 INIT_LIST_HEAD(&epi->rdllink);
735 INIT_LIST_HEAD(&epi->fllink);
736 INIT_LIST_HEAD(&epi->pwqlist);
738 ep_set_ffd(&epi->ffd, tfile, fd);
741 epi->next = EP_UNACTIVE_PTR;
743 /* Initialize the poll table using the queue callback */
745 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
748 * Attach the item to the poll hooks and get current event bits.
749 * We can safely use the file* here because its usage count has
750 * been increased by the caller of this function. Note that after
751 * this operation completes, the poll callback can start hitting
754 revents = tfile->f_op->poll(tfile, &epq.pt);
757 * We have to check if something went wrong during the poll wait queue
758 * install process. Namely an allocation for a wait queue failed due
759 * high memory pressure.
762 goto error_unregister;
764 /* Add the current item to the list of active epoll hook for this file */
765 spin_lock(&tfile->f_ep_lock);
766 list_add_tail(&epi->fllink, &tfile->f_ep_links);
767 spin_unlock(&tfile->f_ep_lock);
770 * Add the current item to the RB tree. All RB tree operations are
771 * protected by "mtx", and ep_insert() is called with "mtx" held.
773 ep_rbtree_insert(ep, epi);
775 /* We have to drop the new item inside our item list to keep track of it */
776 spin_lock_irqsave(&ep->lock, flags);
778 /* If the file is already "ready" we drop it inside the ready list */
779 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
780 list_add_tail(&epi->rdllink, &ep->rdllist);
782 /* Notify waiting tasks that events are available */
783 if (waitqueue_active(&ep->wq))
784 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
785 if (waitqueue_active(&ep->poll_wait))
789 spin_unlock_irqrestore(&ep->lock, flags);
791 /* We have to call this outside the lock */
793 ep_poll_safewake(&psw, &ep->poll_wait);
795 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
796 current, ep, tfile, fd));
801 ep_unregister_pollwait(ep, epi);
804 * We need to do this because an event could have been arrived on some
805 * allocated wait queue.
807 spin_lock_irqsave(&ep->lock, flags);
808 if (ep_is_linked(&epi->rdllink))
809 list_del_init(&epi->rdllink);
810 spin_unlock_irqrestore(&ep->lock, flags);
812 kmem_cache_free(epi_cache, epi);
818 * Modify the interest event mask by dropping an event if the new mask
819 * has a match in the current file status. Must be called with "mtx" held.
821 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
824 unsigned int revents;
828 * Set the new event interest mask before calling f_op->poll(), otherwise
829 * a potential race might occur. In fact if we do this operation inside
830 * the lock, an event might happen between the f_op->poll() call and the
831 * new event set registering.
833 epi->event.events = event->events;
836 * Get current event bits. We can safely use the file* here because
837 * its usage count has been increased by the caller of this function.
839 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
841 spin_lock_irqsave(&ep->lock, flags);
843 /* Copy the data member from inside the lock */
844 epi->event.data = event->data;
847 * If the item is "hot" and it is not registered inside the ready
848 * list, push it inside. If the item is not "hot" and it is currently
849 * registered inside the ready list, unlink it.
851 if (revents & event->events) {
852 if (!ep_is_linked(&epi->rdllink)) {
853 list_add_tail(&epi->rdllink, &ep->rdllist);
855 /* Notify waiting tasks that events are available */
856 if (waitqueue_active(&ep->wq))
857 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
859 if (waitqueue_active(&ep->poll_wait))
863 spin_unlock_irqrestore(&ep->lock, flags);
865 /* We have to call this outside the lock */
867 ep_poll_safewake(&psw, &ep->poll_wait);
872 static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
875 int eventcnt, error = -EFAULT, pwake = 0;
876 unsigned int revents;
878 struct epitem *epi, *nepi;
879 struct list_head txlist;
881 INIT_LIST_HEAD(&txlist);
884 * We need to lock this because we could be hit by
885 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
887 mutex_lock(&ep->mtx);
890 * Steal the ready list, and re-init the original one to the
891 * empty list. Also, set ep->ovflist to NULL so that events
892 * happening while looping w/out locks, are not lost. We cannot
893 * have the poll callback to queue directly on ep->rdllist,
894 * because we are doing it in the loop below, in a lockless way.
896 spin_lock_irqsave(&ep->lock, flags);
897 list_splice(&ep->rdllist, &txlist);
898 INIT_LIST_HEAD(&ep->rdllist);
900 spin_unlock_irqrestore(&ep->lock, flags);
903 * We can loop without lock because this is a task private list.
904 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
905 * Items cannot vanish during the loop because we are holding "mtx".
907 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
908 epi = list_first_entry(&txlist, struct epitem, rdllink);
910 list_del_init(&epi->rdllink);
913 * Get the ready file event set. We can safely use the file
914 * because we are holding the "mtx" and this will guarantee
915 * that both the file and the item will not vanish.
917 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
918 revents &= epi->event.events;
921 * Is the event mask intersect the caller-requested one,
922 * deliver the event to userspace. Again, we are holding
923 * "mtx", so no operations coming from userspace can change
927 if (__put_user(revents,
928 &events[eventcnt].events) ||
929 __put_user(epi->event.data,
930 &events[eventcnt].data))
932 if (epi->event.events & EPOLLONESHOT)
933 epi->event.events &= EP_PRIVATE_BITS;
937 * At this point, noone can insert into ep->rdllist besides
938 * us. The epoll_ctl() callers are locked out by us holding
939 * "mtx" and the poll callback will queue them in ep->ovflist.
941 if (!(epi->event.events & EPOLLET) &&
942 (revents & epi->event.events))
943 list_add_tail(&epi->rdllink, &ep->rdllist);
949 spin_lock_irqsave(&ep->lock, flags);
951 * During the time we spent in the loop above, some other events
952 * might have been queued by the poll callback. We re-insert them
953 * here (in case they are not already queued, or they're one-shot).
955 for (nepi = ep->ovflist; (epi = nepi) != NULL;
956 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
957 if (!ep_is_linked(&epi->rdllink) &&
958 (epi->event.events & ~EP_PRIVATE_BITS))
959 list_add_tail(&epi->rdllink, &ep->rdllist);
962 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
963 * releasing the lock, events will be queued in the normal way inside
966 ep->ovflist = EP_UNACTIVE_PTR;
969 * In case of error in the event-send loop, we might still have items
970 * inside the "txlist". We need to splice them back inside ep->rdllist.
972 list_splice(&txlist, &ep->rdllist);
974 if (!list_empty(&ep->rdllist)) {
976 * Wake up (if active) both the eventpoll wait list and the ->poll()
979 if (waitqueue_active(&ep->wq))
980 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
982 if (waitqueue_active(&ep->poll_wait))
985 spin_unlock_irqrestore(&ep->lock, flags);
987 mutex_unlock(&ep->mtx);
989 /* We have to call this outside the lock */
991 ep_poll_safewake(&psw, &ep->poll_wait);
993 return eventcnt == 0 ? error: eventcnt;
996 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
997 int maxevents, long timeout)
1000 unsigned long flags;
1005 * Calculate the timeout by checking for the "infinite" value ( -1 )
1006 * and the overflow condition. The passed timeout is in milliseconds,
1007 * that why (t * HZ) / 1000.
1009 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1010 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1013 spin_lock_irqsave(&ep->lock, flags);
1016 if (list_empty(&ep->rdllist)) {
1018 * We don't have any available event to return to the caller.
1019 * We need to sleep here, and we will be wake up by
1020 * ep_poll_callback() when events will become available.
1022 init_waitqueue_entry(&wait, current);
1023 wait.flags |= WQ_FLAG_EXCLUSIVE;
1024 __add_wait_queue(&ep->wq, &wait);
1028 * We don't want to sleep if the ep_poll_callback() sends us
1029 * a wakeup in between. That's why we set the task state
1030 * to TASK_INTERRUPTIBLE before doing the checks.
1032 set_current_state(TASK_INTERRUPTIBLE);
1033 if (!list_empty(&ep->rdllist) || !jtimeout)
1035 if (signal_pending(current)) {
1040 spin_unlock_irqrestore(&ep->lock, flags);
1041 jtimeout = schedule_timeout(jtimeout);
1042 spin_lock_irqsave(&ep->lock, flags);
1044 __remove_wait_queue(&ep->wq, &wait);
1046 set_current_state(TASK_RUNNING);
1049 /* Is it worth to try to dig for events ? */
1050 eavail = !list_empty(&ep->rdllist);
1052 spin_unlock_irqrestore(&ep->lock, flags);
1055 * Try to transfer events to user space. In case we get 0 events and
1056 * there's still timeout left over, we go trying again in search of
1059 if (!res && eavail &&
1060 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1067 * It opens an eventpoll file descriptor by suggesting a storage of "size"
1068 * file descriptors. The size parameter is just an hint about how to size
1069 * data structures. It won't prevent the user to store more than "size"
1070 * file descriptors inside the epoll interface. It is the kernel part of
1071 * the userspace epoll_create(2).
1073 asmlinkage long sys_epoll_create(int size)
1076 struct eventpoll *ep;
1077 struct inode *inode;
1080 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
1084 * Sanity check on the size parameter, and create the internal data
1085 * structure ( "struct eventpoll" ).
1088 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
1092 * Creates all the items needed to setup an eventpoll file. That is,
1093 * a file structure, and inode and a free file descriptor.
1095 error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]",
1096 &eventpoll_fops, ep);
1100 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1101 current, size, fd));
1109 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1110 current, size, error));
1115 * The following function implements the controller interface for
1116 * the eventpoll file that enables the insertion/removal/change of
1117 * file descriptors inside the interest set. It represents
1118 * the kernel part of the user space epoll_ctl(2).
1120 asmlinkage long sys_epoll_ctl(int epfd, int op, int fd,
1121 struct epoll_event __user *event)
1124 struct file *file, *tfile;
1125 struct eventpoll *ep;
1127 struct epoll_event epds;
1129 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1130 current, epfd, op, fd, event));
1133 if (ep_op_has_event(op) &&
1134 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1137 /* Get the "struct file *" for the eventpoll file */
1143 /* Get the "struct file *" for the target file */
1148 /* The target file descriptor must support poll */
1150 if (!tfile->f_op || !tfile->f_op->poll)
1151 goto error_tgt_fput;
1154 * We have to check that the file structure underneath the file descriptor
1155 * the user passed to us _is_ an eventpoll file. And also we do not permit
1156 * adding an epoll file descriptor inside itself.
1159 if (file == tfile || !is_file_epoll(file))
1160 goto error_tgt_fput;
1163 * At this point it is safe to assume that the "private_data" contains
1164 * our own data structure.
1166 ep = file->private_data;
1168 mutex_lock(&ep->mtx);
1170 /* Try to lookup the file inside our RB tree */
1171 epi = ep_find(ep, tfile, fd);
1177 epds.events |= POLLERR | POLLHUP;
1179 error = ep_insert(ep, &epds, tfile, fd);
1185 error = ep_remove(ep, epi);
1191 epds.events |= POLLERR | POLLHUP;
1192 error = ep_modify(ep, epi, &epds);
1197 mutex_unlock(&ep->mtx);
1204 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1205 current, epfd, op, fd, event, error));
1211 * Implement the event wait interface for the eventpoll file. It is the kernel
1212 * part of the user space epoll_wait(2).
1214 asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
1215 int maxevents, int timeout)
1219 struct eventpoll *ep;
1221 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1222 current, epfd, events, maxevents, timeout));
1224 /* The maximum number of event must be greater than zero */
1225 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1228 /* Verify that the area passed by the user is writeable */
1229 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1234 /* Get the "struct file *" for the eventpoll file */
1241 * We have to check that the file structure underneath the fd
1242 * the user passed to us _is_ an eventpoll file.
1245 if (!is_file_epoll(file))
1249 * At this point it is safe to assume that the "private_data" contains
1250 * our own data structure.
1252 ep = file->private_data;
1254 /* Time to fish for events ... */
1255 error = ep_poll(ep, events, maxevents, timeout);
1260 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1261 current, epfd, events, maxevents, timeout, error));
1266 #ifdef TIF_RESTORE_SIGMASK
1269 * Implement the event wait interface for the eventpoll file. It is the kernel
1270 * part of the user space epoll_pwait(2).
1272 asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
1273 int maxevents, int timeout, const sigset_t __user *sigmask,
1277 sigset_t ksigmask, sigsaved;
1280 * If the caller wants a certain signal mask to be set during the wait,
1284 if (sigsetsize != sizeof(sigset_t))
1286 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1288 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1289 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1292 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1295 * If we changed the signal mask, we need to restore the original one.
1296 * In case we've got a signal while waiting, we do not restore the
1297 * signal mask yet, and we allow do_signal() to deliver the signal on
1298 * the way back to userspace, before the signal mask is restored.
1301 if (error == -EINTR) {
1302 memcpy(¤t->saved_sigmask, &sigsaved,
1304 set_thread_flag(TIF_RESTORE_SIGMASK);
1306 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1312 #endif /* #ifdef TIF_RESTORE_SIGMASK */
1314 static int __init eventpoll_init(void)
1316 mutex_init(&epmutex);
1318 /* Initialize the structure used to perform safe poll wait head wake ups */
1319 ep_poll_safewake_init(&psw);
1321 /* Allocates slab cache used to allocate "struct epitem" items */
1322 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1323 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1326 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1327 pwq_cache = kmem_cache_create("eventpoll_pwq",
1328 sizeof(struct eppoll_entry), 0,
1329 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1333 fs_initcall(eventpoll_init);