x86: fix ldt limit for 64 bit
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / eventpoll.c
blob990c01d2d66bcce64a1c2a993cd6cc01ea5951d5
1 /*
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>
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 #define DEBUG_EPOLL 0
76 #if DEBUG_EPOLL > 0
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 */
84 #define DEBUG_EPI 0
86 #if DEBUG_EPI != 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 {
106 struct file *file;
107 int fd;
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;
129 spinlock_t lock;
133 * Each file descriptor added to the eventpoll interface will
134 * have an entry of this type linked to the "rbr" RB tree.
136 struct epitem {
137 /* RB tree node used to link this structure to the eventpoll RB tree */
138 struct rb_node rbn;
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.
147 struct epitem *next;
149 /* The file descriptor information this item refers to */
150 struct epoll_filefd ffd;
152 /* Number of active wait queue attached to poll operations */
153 int nwait;
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
171 * interface.
173 struct eventpoll {
174 /* Protect the this structure access */
175 spinlock_t lock;
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.
183 struct mutex mtx;
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 */
195 struct rb_root rbr;
198 * This is a single linked list that chains all the "struct epitem" that
199 * happened while transfering ready events to userspace w/out
200 * holding ->lock.
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" */
211 void *base;
214 * Wait queue item that will be linked to the target file wait
215 * queue head.
217 wait_queue_t 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 */
224 struct ep_pqueue {
225 poll_table pt;
226 struct epitem *epi;
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)
248 ffd->file = file;
249 ffd->fd = 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 /* Tells us if the item is currently linked */
261 static inline int ep_is_linked(struct list_head *p)
263 return !list_empty(p);
266 /* Get the "struct epitem" from a wait queue pointer */
267 static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
269 return container_of(p, struct eppoll_entry, wait)->base;
272 /* Get the "struct epitem" from an epoll queue wrapper */
273 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
275 return container_of(p, struct ep_pqueue, pt)->epi;
278 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
279 static inline int ep_op_has_event(int op)
281 return op != EPOLL_CTL_DEL;
284 /* Initialize the poll safe wake up structure */
285 static void ep_poll_safewake_init(struct poll_safewake *psw)
288 INIT_LIST_HEAD(&psw->wake_task_list);
289 spin_lock_init(&psw->lock);
293 * Perform a safe wake up of the poll wait list. The problem is that
294 * with the new callback'd wake up system, it is possible that the
295 * poll callback is reentered from inside the call to wake_up() done
296 * on the poll wait queue head. The rule is that we cannot reenter the
297 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
298 * and we cannot reenter the same wait queue head at all. This will
299 * enable to have a hierarchy of epoll file descriptor of no more than
300 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
301 * because this one gets called by the poll callback, that in turn is called
302 * from inside a wake_up(), that might be called from irq context.
304 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
306 int wake_nests = 0;
307 unsigned long flags;
308 struct task_struct *this_task = current;
309 struct list_head *lsthead = &psw->wake_task_list;
310 struct wake_task_node *tncur;
311 struct wake_task_node tnode;
313 spin_lock_irqsave(&psw->lock, flags);
315 /* Try to see if the current task is already inside this wakeup call */
316 list_for_each_entry(tncur, lsthead, llink) {
318 if (tncur->wq == wq ||
319 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
321 * Ops ... loop detected or maximum nest level reached.
322 * We abort this wake by breaking the cycle itself.
324 spin_unlock_irqrestore(&psw->lock, flags);
325 return;
329 /* Add the current task to the list */
330 tnode.task = this_task;
331 tnode.wq = wq;
332 list_add(&tnode.llink, lsthead);
334 spin_unlock_irqrestore(&psw->lock, flags);
336 /* Do really wake up now */
337 wake_up_nested(wq, 1 + wake_nests);
339 /* Remove the current task from the list */
340 spin_lock_irqsave(&psw->lock, flags);
341 list_del(&tnode.llink);
342 spin_unlock_irqrestore(&psw->lock, flags);
346 * This function unregister poll callbacks from the associated file descriptor.
347 * Since this must be called without holding "ep->lock" the atomic exchange trick
348 * will protect us from multiple unregister.
350 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
352 int nwait;
353 struct list_head *lsthead = &epi->pwqlist;
354 struct eppoll_entry *pwq;
356 /* This is called without locks, so we need the atomic exchange */
357 nwait = xchg(&epi->nwait, 0);
359 if (nwait) {
360 while (!list_empty(lsthead)) {
361 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
363 list_del_init(&pwq->llink);
364 remove_wait_queue(pwq->whead, &pwq->wait);
365 kmem_cache_free(pwq_cache, pwq);
371 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
372 * all the associated resources. Must be called with "mtx" held.
374 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
376 unsigned long flags;
377 struct file *file = epi->ffd.file;
380 * Removes poll wait queue hooks. We _have_ to do this without holding
381 * the "ep->lock" otherwise a deadlock might occur. This because of the
382 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
383 * queue head lock when unregistering the wait queue. The wakeup callback
384 * will run by holding the wait queue head lock and will call our callback
385 * that will try to get "ep->lock".
387 ep_unregister_pollwait(ep, epi);
389 /* Remove the current item from the list of epoll hooks */
390 spin_lock(&file->f_ep_lock);
391 if (ep_is_linked(&epi->fllink))
392 list_del_init(&epi->fllink);
393 spin_unlock(&file->f_ep_lock);
395 rb_erase(&epi->rbn, &ep->rbr);
397 spin_lock_irqsave(&ep->lock, flags);
398 if (ep_is_linked(&epi->rdllink))
399 list_del_init(&epi->rdllink);
400 spin_unlock_irqrestore(&ep->lock, flags);
402 /* At this point it is safe to free the eventpoll item */
403 kmem_cache_free(epi_cache, epi);
405 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
406 current, ep, file));
408 return 0;
411 static void ep_free(struct eventpoll *ep)
413 struct rb_node *rbp;
414 struct epitem *epi;
416 /* We need to release all tasks waiting for these file */
417 if (waitqueue_active(&ep->poll_wait))
418 ep_poll_safewake(&psw, &ep->poll_wait);
421 * We need to lock this because we could be hit by
422 * eventpoll_release_file() while we're freeing the "struct eventpoll".
423 * We do not need to hold "ep->mtx" here because the epoll file
424 * is on the way to be removed and no one has references to it
425 * anymore. The only hit might come from eventpoll_release_file() but
426 * holding "epmutex" is sufficent here.
428 mutex_lock(&epmutex);
431 * Walks through the whole tree by unregistering poll callbacks.
433 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
434 epi = rb_entry(rbp, struct epitem, rbn);
436 ep_unregister_pollwait(ep, epi);
440 * Walks through the whole tree by freeing each "struct epitem". At this
441 * point we are sure no poll callbacks will be lingering around, and also by
442 * holding "epmutex" we can be sure that no file cleanup code will hit
443 * us during this operation. So we can avoid the lock on "ep->lock".
445 while ((rbp = rb_first(&ep->rbr)) != NULL) {
446 epi = rb_entry(rbp, struct epitem, rbn);
447 ep_remove(ep, epi);
450 mutex_unlock(&epmutex);
451 mutex_destroy(&ep->mtx);
452 kfree(ep);
455 static int ep_eventpoll_release(struct inode *inode, struct file *file)
457 struct eventpoll *ep = file->private_data;
459 if (ep)
460 ep_free(ep);
462 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
463 return 0;
466 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
468 unsigned int pollflags = 0;
469 unsigned long flags;
470 struct eventpoll *ep = file->private_data;
472 /* Insert inside our poll wait queue */
473 poll_wait(file, &ep->poll_wait, wait);
475 /* Check our condition */
476 spin_lock_irqsave(&ep->lock, flags);
477 if (!list_empty(&ep->rdllist))
478 pollflags = POLLIN | POLLRDNORM;
479 spin_unlock_irqrestore(&ep->lock, flags);
481 return pollflags;
484 /* File callbacks that implement the eventpoll file behaviour */
485 static const struct file_operations eventpoll_fops = {
486 .release = ep_eventpoll_release,
487 .poll = ep_eventpoll_poll
490 /* Fast test to see if the file is an evenpoll file */
491 static inline int is_file_epoll(struct file *f)
493 return f->f_op == &eventpoll_fops;
497 * This is called from eventpoll_release() to unlink files from the eventpoll
498 * interface. We need to have this facility to cleanup correctly files that are
499 * closed without being removed from the eventpoll interface.
501 void eventpoll_release_file(struct file *file)
503 struct list_head *lsthead = &file->f_ep_links;
504 struct eventpoll *ep;
505 struct epitem *epi;
508 * We don't want to get "file->f_ep_lock" because it is not
509 * necessary. It is not necessary because we're in the "struct file"
510 * cleanup path, and this means that noone is using this file anymore.
511 * So, for example, epoll_ctl() cannot hit here sicne if we reach this
512 * point, the file counter already went to zero and fget() would fail.
513 * The only hit might come from ep_free() but by holding the mutex
514 * will correctly serialize the operation. We do need to acquire
515 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
516 * from anywhere but ep_free().
518 mutex_lock(&epmutex);
520 while (!list_empty(lsthead)) {
521 epi = list_first_entry(lsthead, struct epitem, fllink);
523 ep = epi->ep;
524 list_del_init(&epi->fllink);
525 mutex_lock(&ep->mtx);
526 ep_remove(ep, epi);
527 mutex_unlock(&ep->mtx);
530 mutex_unlock(&epmutex);
533 static int ep_alloc(struct eventpoll **pep)
535 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
537 if (!ep)
538 return -ENOMEM;
540 spin_lock_init(&ep->lock);
541 mutex_init(&ep->mtx);
542 init_waitqueue_head(&ep->wq);
543 init_waitqueue_head(&ep->poll_wait);
544 INIT_LIST_HEAD(&ep->rdllist);
545 ep->rbr = RB_ROOT;
546 ep->ovflist = EP_UNACTIVE_PTR;
548 *pep = ep;
550 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
551 current, ep));
552 return 0;
556 * Search the file inside the eventpoll tree. The RB tree operations
557 * are protected by the "mtx" mutex, and ep_find() must be called with
558 * "mtx" held.
560 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
562 int kcmp;
563 struct rb_node *rbp;
564 struct epitem *epi, *epir = NULL;
565 struct epoll_filefd ffd;
567 ep_set_ffd(&ffd, file, fd);
568 for (rbp = ep->rbr.rb_node; rbp; ) {
569 epi = rb_entry(rbp, struct epitem, rbn);
570 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
571 if (kcmp > 0)
572 rbp = rbp->rb_right;
573 else if (kcmp < 0)
574 rbp = rbp->rb_left;
575 else {
576 epir = epi;
577 break;
581 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
582 current, file, epir));
584 return epir;
588 * This is the callback that is passed to the wait queue wakeup
589 * machanism. It is called by the stored file descriptors when they
590 * have events to report.
592 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
594 int pwake = 0;
595 unsigned long flags;
596 struct epitem *epi = ep_item_from_wait(wait);
597 struct eventpoll *ep = epi->ep;
599 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
600 current, epi->ffd.file, epi, ep));
602 spin_lock_irqsave(&ep->lock, flags);
605 * If the event mask does not contain any poll(2) event, we consider the
606 * descriptor to be disabled. This condition is likely the effect of the
607 * EPOLLONESHOT bit that disables the descriptor when an event is received,
608 * until the next EPOLL_CTL_MOD will be issued.
610 if (!(epi->event.events & ~EP_PRIVATE_BITS))
611 goto out_unlock;
614 * If we are trasfering events to userspace, we can hold no locks
615 * (because we're accessing user memory, and because of linux f_op->poll()
616 * semantics). All the events that happens during that period of time are
617 * chained in ep->ovflist and requeued later on.
619 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
620 if (epi->next == EP_UNACTIVE_PTR) {
621 epi->next = ep->ovflist;
622 ep->ovflist = epi;
624 goto out_unlock;
627 /* If this file is already in the ready list we exit soon */
628 if (ep_is_linked(&epi->rdllink))
629 goto is_linked;
631 list_add_tail(&epi->rdllink, &ep->rdllist);
633 is_linked:
635 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
636 * wait list.
638 if (waitqueue_active(&ep->wq))
639 wake_up_locked(&ep->wq);
640 if (waitqueue_active(&ep->poll_wait))
641 pwake++;
643 out_unlock:
644 spin_unlock_irqrestore(&ep->lock, flags);
646 /* We have to call this outside the lock */
647 if (pwake)
648 ep_poll_safewake(&psw, &ep->poll_wait);
650 return 1;
654 * This is the callback that is used to add our wait queue to the
655 * target file wakeup lists.
657 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
658 poll_table *pt)
660 struct epitem *epi = ep_item_from_epqueue(pt);
661 struct eppoll_entry *pwq;
663 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
664 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
665 pwq->whead = whead;
666 pwq->base = epi;
667 add_wait_queue(whead, &pwq->wait);
668 list_add_tail(&pwq->llink, &epi->pwqlist);
669 epi->nwait++;
670 } else {
671 /* We have to signal that an error occurred */
672 epi->nwait = -1;
676 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
678 int kcmp;
679 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
680 struct epitem *epic;
682 while (*p) {
683 parent = *p;
684 epic = rb_entry(parent, struct epitem, rbn);
685 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
686 if (kcmp > 0)
687 p = &parent->rb_right;
688 else
689 p = &parent->rb_left;
691 rb_link_node(&epi->rbn, parent, p);
692 rb_insert_color(&epi->rbn, &ep->rbr);
696 * Must be called with "mtx" held.
698 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
699 struct file *tfile, int fd)
701 int error, revents, pwake = 0;
702 unsigned long flags;
703 struct epitem *epi;
704 struct ep_pqueue epq;
706 error = -ENOMEM;
707 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
708 goto error_return;
710 /* Item initialization follow here ... */
711 INIT_LIST_HEAD(&epi->rdllink);
712 INIT_LIST_HEAD(&epi->fllink);
713 INIT_LIST_HEAD(&epi->pwqlist);
714 epi->ep = ep;
715 ep_set_ffd(&epi->ffd, tfile, fd);
716 epi->event = *event;
717 epi->nwait = 0;
718 epi->next = EP_UNACTIVE_PTR;
720 /* Initialize the poll table using the queue callback */
721 epq.epi = epi;
722 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
725 * Attach the item to the poll hooks and get current event bits.
726 * We can safely use the file* here because its usage count has
727 * been increased by the caller of this function. Note that after
728 * this operation completes, the poll callback can start hitting
729 * the new item.
731 revents = tfile->f_op->poll(tfile, &epq.pt);
734 * We have to check if something went wrong during the poll wait queue
735 * install process. Namely an allocation for a wait queue failed due
736 * high memory pressure.
738 if (epi->nwait < 0)
739 goto error_unregister;
741 /* Add the current item to the list of active epoll hook for this file */
742 spin_lock(&tfile->f_ep_lock);
743 list_add_tail(&epi->fllink, &tfile->f_ep_links);
744 spin_unlock(&tfile->f_ep_lock);
747 * Add the current item to the RB tree. All RB tree operations are
748 * protected by "mtx", and ep_insert() is called with "mtx" held.
750 ep_rbtree_insert(ep, epi);
752 /* We have to drop the new item inside our item list to keep track of it */
753 spin_lock_irqsave(&ep->lock, flags);
755 /* If the file is already "ready" we drop it inside the ready list */
756 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
757 list_add_tail(&epi->rdllink, &ep->rdllist);
759 /* Notify waiting tasks that events are available */
760 if (waitqueue_active(&ep->wq))
761 wake_up_locked(&ep->wq);
762 if (waitqueue_active(&ep->poll_wait))
763 pwake++;
766 spin_unlock_irqrestore(&ep->lock, flags);
768 /* We have to call this outside the lock */
769 if (pwake)
770 ep_poll_safewake(&psw, &ep->poll_wait);
772 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
773 current, ep, tfile, fd));
775 return 0;
777 error_unregister:
778 ep_unregister_pollwait(ep, epi);
781 * We need to do this because an event could have been arrived on some
782 * allocated wait queue. Note that we don't care about the ep->ovflist
783 * list, since that is used/cleaned only inside a section bound by "mtx".
784 * And ep_insert() is called with "mtx" held.
786 spin_lock_irqsave(&ep->lock, flags);
787 if (ep_is_linked(&epi->rdllink))
788 list_del_init(&epi->rdllink);
789 spin_unlock_irqrestore(&ep->lock, flags);
791 kmem_cache_free(epi_cache, epi);
792 error_return:
793 return error;
797 * Modify the interest event mask by dropping an event if the new mask
798 * has a match in the current file status. Must be called with "mtx" held.
800 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
802 int pwake = 0;
803 unsigned int revents;
804 unsigned long flags;
807 * Set the new event interest mask before calling f_op->poll(), otherwise
808 * a potential race might occur. In fact if we do this operation inside
809 * the lock, an event might happen between the f_op->poll() call and the
810 * new event set registering.
812 epi->event.events = event->events;
815 * Get current event bits. We can safely use the file* here because
816 * its usage count has been increased by the caller of this function.
818 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
820 spin_lock_irqsave(&ep->lock, flags);
822 /* Copy the data member from inside the lock */
823 epi->event.data = event->data;
826 * If the item is "hot" and it is not registered inside the ready
827 * list, push it inside.
829 if (revents & event->events) {
830 if (!ep_is_linked(&epi->rdllink)) {
831 list_add_tail(&epi->rdllink, &ep->rdllist);
833 /* Notify waiting tasks that events are available */
834 if (waitqueue_active(&ep->wq))
835 wake_up_locked(&ep->wq);
836 if (waitqueue_active(&ep->poll_wait))
837 pwake++;
840 spin_unlock_irqrestore(&ep->lock, flags);
842 /* We have to call this outside the lock */
843 if (pwake)
844 ep_poll_safewake(&psw, &ep->poll_wait);
846 return 0;
849 static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
850 int maxevents)
852 int eventcnt, error = -EFAULT, pwake = 0;
853 unsigned int revents;
854 unsigned long flags;
855 struct epitem *epi, *nepi;
856 struct list_head txlist;
858 INIT_LIST_HEAD(&txlist);
861 * We need to lock this because we could be hit by
862 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
864 mutex_lock(&ep->mtx);
867 * Steal the ready list, and re-init the original one to the
868 * empty list. Also, set ep->ovflist to NULL so that events
869 * happening while looping w/out locks, are not lost. We cannot
870 * have the poll callback to queue directly on ep->rdllist,
871 * because we are doing it in the loop below, in a lockless way.
873 spin_lock_irqsave(&ep->lock, flags);
874 list_splice(&ep->rdllist, &txlist);
875 INIT_LIST_HEAD(&ep->rdllist);
876 ep->ovflist = NULL;
877 spin_unlock_irqrestore(&ep->lock, flags);
880 * We can loop without lock because this is a task private list.
881 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
882 * Items cannot vanish during the loop because we are holding "mtx".
884 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
885 epi = list_first_entry(&txlist, struct epitem, rdllink);
887 list_del_init(&epi->rdllink);
890 * Get the ready file event set. We can safely use the file
891 * because we are holding the "mtx" and this will guarantee
892 * that both the file and the item will not vanish.
894 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
895 revents &= epi->event.events;
898 * Is the event mask intersect the caller-requested one,
899 * deliver the event to userspace. Again, we are holding
900 * "mtx", so no operations coming from userspace can change
901 * the item.
903 if (revents) {
904 if (__put_user(revents,
905 &events[eventcnt].events) ||
906 __put_user(epi->event.data,
907 &events[eventcnt].data))
908 goto errxit;
909 if (epi->event.events & EPOLLONESHOT)
910 epi->event.events &= EP_PRIVATE_BITS;
911 eventcnt++;
914 * At this point, noone can insert into ep->rdllist besides
915 * us. The epoll_ctl() callers are locked out by us holding
916 * "mtx" and the poll callback will queue them in ep->ovflist.
918 if (!(epi->event.events & EPOLLET) &&
919 (revents & epi->event.events))
920 list_add_tail(&epi->rdllink, &ep->rdllist);
922 error = 0;
924 errxit:
926 spin_lock_irqsave(&ep->lock, flags);
928 * During the time we spent in the loop above, some other events
929 * might have been queued by the poll callback. We re-insert them
930 * here (in case they are not already queued, or they're one-shot).
932 for (nepi = ep->ovflist; (epi = nepi) != NULL;
933 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
934 if (!ep_is_linked(&epi->rdllink) &&
935 (epi->event.events & ~EP_PRIVATE_BITS))
936 list_add_tail(&epi->rdllink, &ep->rdllist);
939 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
940 * releasing the lock, events will be queued in the normal way inside
941 * ep->rdllist.
943 ep->ovflist = EP_UNACTIVE_PTR;
946 * In case of error in the event-send loop, or in case the number of
947 * ready events exceeds the userspace limit, we need to splice the
948 * "txlist" back inside ep->rdllist.
950 list_splice(&txlist, &ep->rdllist);
952 if (!list_empty(&ep->rdllist)) {
954 * Wake up (if active) both the eventpoll wait list and the ->poll()
955 * wait list (delayed after we release the lock).
957 if (waitqueue_active(&ep->wq))
958 wake_up_locked(&ep->wq);
959 if (waitqueue_active(&ep->poll_wait))
960 pwake++;
962 spin_unlock_irqrestore(&ep->lock, flags);
964 mutex_unlock(&ep->mtx);
966 /* We have to call this outside the lock */
967 if (pwake)
968 ep_poll_safewake(&psw, &ep->poll_wait);
970 return eventcnt == 0 ? error: eventcnt;
973 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
974 int maxevents, long timeout)
976 int res, eavail;
977 unsigned long flags;
978 long jtimeout;
979 wait_queue_t wait;
982 * Calculate the timeout by checking for the "infinite" value ( -1 )
983 * and the overflow condition. The passed timeout is in milliseconds,
984 * that why (t * HZ) / 1000.
986 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
987 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
989 retry:
990 spin_lock_irqsave(&ep->lock, flags);
992 res = 0;
993 if (list_empty(&ep->rdllist)) {
995 * We don't have any available event to return to the caller.
996 * We need to sleep here, and we will be wake up by
997 * ep_poll_callback() when events will become available.
999 init_waitqueue_entry(&wait, current);
1000 wait.flags |= WQ_FLAG_EXCLUSIVE;
1001 __add_wait_queue(&ep->wq, &wait);
1003 for (;;) {
1005 * We don't want to sleep if the ep_poll_callback() sends us
1006 * a wakeup in between. That's why we set the task state
1007 * to TASK_INTERRUPTIBLE before doing the checks.
1009 set_current_state(TASK_INTERRUPTIBLE);
1010 if (!list_empty(&ep->rdllist) || !jtimeout)
1011 break;
1012 if (signal_pending(current)) {
1013 res = -EINTR;
1014 break;
1017 spin_unlock_irqrestore(&ep->lock, flags);
1018 jtimeout = schedule_timeout(jtimeout);
1019 spin_lock_irqsave(&ep->lock, flags);
1021 __remove_wait_queue(&ep->wq, &wait);
1023 set_current_state(TASK_RUNNING);
1026 /* Is it worth to try to dig for events ? */
1027 eavail = !list_empty(&ep->rdllist);
1029 spin_unlock_irqrestore(&ep->lock, flags);
1032 * Try to transfer events to user space. In case we get 0 events and
1033 * there's still timeout left over, we go trying again in search of
1034 * more luck.
1036 if (!res && eavail &&
1037 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1038 goto retry;
1040 return res;
1044 * It opens an eventpoll file descriptor. The "size" parameter is there
1045 * for historical reasons, when epoll was using an hash instead of an
1046 * RB tree. With the current implementation, the "size" parameter is ignored
1047 * (besides sanity checks).
1049 asmlinkage long sys_epoll_create(int size)
1051 int error, fd = -1;
1052 struct eventpoll *ep;
1054 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
1055 current, size));
1058 * Sanity check on the size parameter, and create the internal data
1059 * structure ( "struct eventpoll" ).
1061 error = -EINVAL;
1062 if (size <= 0 || (error = ep_alloc(&ep)) < 0) {
1063 fd = error;
1064 goto error_return;
1068 * Creates all the items needed to setup an eventpoll file. That is,
1069 * a file structure and a free file descriptor.
1071 fd = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep);
1072 if (fd < 0)
1073 ep_free(ep);
1075 error_return:
1076 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1077 current, size, fd));
1079 return fd;
1083 * The following function implements the controller interface for
1084 * the eventpoll file that enables the insertion/removal/change of
1085 * file descriptors inside the interest set.
1087 asmlinkage long sys_epoll_ctl(int epfd, int op, int fd,
1088 struct epoll_event __user *event)
1090 int error;
1091 struct file *file, *tfile;
1092 struct eventpoll *ep;
1093 struct epitem *epi;
1094 struct epoll_event epds;
1096 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1097 current, epfd, op, fd, event));
1099 error = -EFAULT;
1100 if (ep_op_has_event(op) &&
1101 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1102 goto error_return;
1104 /* Get the "struct file *" for the eventpoll file */
1105 error = -EBADF;
1106 file = fget(epfd);
1107 if (!file)
1108 goto error_return;
1110 /* Get the "struct file *" for the target file */
1111 tfile = fget(fd);
1112 if (!tfile)
1113 goto error_fput;
1115 /* The target file descriptor must support poll */
1116 error = -EPERM;
1117 if (!tfile->f_op || !tfile->f_op->poll)
1118 goto error_tgt_fput;
1121 * We have to check that the file structure underneath the file descriptor
1122 * the user passed to us _is_ an eventpoll file. And also we do not permit
1123 * adding an epoll file descriptor inside itself.
1125 error = -EINVAL;
1126 if (file == tfile || !is_file_epoll(file))
1127 goto error_tgt_fput;
1130 * At this point it is safe to assume that the "private_data" contains
1131 * our own data structure.
1133 ep = file->private_data;
1135 mutex_lock(&ep->mtx);
1138 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1139 * above, we can be sure to be able to use the item looked up by
1140 * ep_find() till we release the mutex.
1142 epi = ep_find(ep, tfile, fd);
1144 error = -EINVAL;
1145 switch (op) {
1146 case EPOLL_CTL_ADD:
1147 if (!epi) {
1148 epds.events |= POLLERR | POLLHUP;
1150 error = ep_insert(ep, &epds, tfile, fd);
1151 } else
1152 error = -EEXIST;
1153 break;
1154 case EPOLL_CTL_DEL:
1155 if (epi)
1156 error = ep_remove(ep, epi);
1157 else
1158 error = -ENOENT;
1159 break;
1160 case EPOLL_CTL_MOD:
1161 if (epi) {
1162 epds.events |= POLLERR | POLLHUP;
1163 error = ep_modify(ep, epi, &epds);
1164 } else
1165 error = -ENOENT;
1166 break;
1168 mutex_unlock(&ep->mtx);
1170 error_tgt_fput:
1171 fput(tfile);
1172 error_fput:
1173 fput(file);
1174 error_return:
1175 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1176 current, epfd, op, fd, event, error));
1178 return error;
1182 * Implement the event wait interface for the eventpoll file. It is the kernel
1183 * part of the user space epoll_wait(2).
1185 asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
1186 int maxevents, int timeout)
1188 int error;
1189 struct file *file;
1190 struct eventpoll *ep;
1192 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1193 current, epfd, events, maxevents, timeout));
1195 /* The maximum number of event must be greater than zero */
1196 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1197 return -EINVAL;
1199 /* Verify that the area passed by the user is writeable */
1200 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1201 error = -EFAULT;
1202 goto error_return;
1205 /* Get the "struct file *" for the eventpoll file */
1206 error = -EBADF;
1207 file = fget(epfd);
1208 if (!file)
1209 goto error_return;
1212 * We have to check that the file structure underneath the fd
1213 * the user passed to us _is_ an eventpoll file.
1215 error = -EINVAL;
1216 if (!is_file_epoll(file))
1217 goto error_fput;
1220 * At this point it is safe to assume that the "private_data" contains
1221 * our own data structure.
1223 ep = file->private_data;
1225 /* Time to fish for events ... */
1226 error = ep_poll(ep, events, maxevents, timeout);
1228 error_fput:
1229 fput(file);
1230 error_return:
1231 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1232 current, epfd, events, maxevents, timeout, error));
1234 return error;
1237 #ifdef HAVE_SET_RESTORE_SIGMASK
1240 * Implement the event wait interface for the eventpoll file. It is the kernel
1241 * part of the user space epoll_pwait(2).
1243 asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
1244 int maxevents, int timeout, const sigset_t __user *sigmask,
1245 size_t sigsetsize)
1247 int error;
1248 sigset_t ksigmask, sigsaved;
1251 * If the caller wants a certain signal mask to be set during the wait,
1252 * we apply it here.
1254 if (sigmask) {
1255 if (sigsetsize != sizeof(sigset_t))
1256 return -EINVAL;
1257 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1258 return -EFAULT;
1259 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1260 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1263 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1266 * If we changed the signal mask, we need to restore the original one.
1267 * In case we've got a signal while waiting, we do not restore the
1268 * signal mask yet, and we allow do_signal() to deliver the signal on
1269 * the way back to userspace, before the signal mask is restored.
1271 if (sigmask) {
1272 if (error == -EINTR) {
1273 memcpy(&current->saved_sigmask, &sigsaved,
1274 sizeof(sigsaved));
1275 set_restore_sigmask();
1276 } else
1277 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1280 return error;
1283 #endif /* HAVE_SET_RESTORE_SIGMASK */
1285 static int __init eventpoll_init(void)
1287 mutex_init(&epmutex);
1289 /* Initialize the structure used to perform safe poll wait head wake ups */
1290 ep_poll_safewake_init(&psw);
1292 /* Allocates slab cache used to allocate "struct epitem" items */
1293 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1294 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1295 NULL);
1297 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1298 pwq_cache = kmem_cache_create("eventpoll_pwq",
1299 sizeof(struct eppoll_entry), 0,
1300 EPI_SLAB_DEBUG|SLAB_PANIC, NULL);
1302 return 0;
1304 fs_initcall(eventpoll_init);