2 * fs/eventpoll.c ( Efficent event polling implementation )
3 * Copyright (C) 2001,...,2006 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/module.h>
15 #include <linux/init.h>
16 #include <linux/kernel.h>
17 #include <linux/sched.h>
19 #include <linux/file.h>
20 #include <linux/signal.h>
21 #include <linux/errno.h>
23 #include <linux/slab.h>
24 #include <linux/poll.h>
25 #include <linux/string.h>
26 #include <linux/list.h>
27 #include <linux/hash.h>
28 #include <linux/spinlock.h>
29 #include <linux/syscalls.h>
30 #include <linux/rwsem.h>
31 #include <linux/rbtree.h>
32 #include <linux/wait.h>
33 #include <linux/eventpoll.h>
34 #include <linux/mount.h>
35 #include <linux/bitops.h>
36 #include <linux/mutex.h>
37 #include <linux/anon_inodes.h>
38 #include <asm/uaccess.h>
39 #include <asm/system.h>
42 #include <asm/atomic.h>
43 #include <asm/semaphore.h>
48 * There are three level of locking required by epoll :
51 * 2) ep->sem (rw_semaphore)
52 * 3) ep->lock (rw_lock)
54 * The acquire order is the one listed above, from 1 to 3.
55 * We need a spinlock (ep->lock) because we manipulate objects
56 * from inside the poll callback, that might be triggered from
57 * a wake_up() that in turn might be called from IRQ context.
58 * So we can't sleep inside the poll callback and hence we need
59 * a spinlock. During the event transfer loop (from kernel to
60 * user space) we could end up sleeping due a copy_to_user(), so
61 * we need a lock that will allow us to sleep. This lock is a
62 * read-write semaphore (ep->sem). It is acquired on read during
63 * the event transfer loop and in write during epoll_ctl(EPOLL_CTL_DEL)
64 * and during eventpoll_release_file(). Then we also need a global
65 * semaphore to serialize eventpoll_release_file() and ep_free().
66 * This semaphore is acquired by ep_free() during the epoll file
67 * cleanup path and it is also acquired by eventpoll_release_file()
68 * if a file has been pushed inside an epoll set and it is then
69 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
70 * It is possible to drop the "ep->sem" and to use the global
71 * semaphore "epmutex" (together with "ep->lock") to have it working,
72 * but having "ep->sem" will make the interface more scalable.
73 * Events that require holding "epmutex" are very rare, while for
74 * normal operations the epoll private "ep->sem" will guarantee
75 * a greater scalability.
82 #define DPRINTK(x) printk x
83 #define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
84 #else /* #if DEBUG_EPOLL > 0 */
85 #define DPRINTK(x) (void) 0
86 #define DNPRINTK(n, x) (void) 0
87 #endif /* #if DEBUG_EPOLL > 0 */
92 #define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
93 #else /* #if DEBUG_EPI != 0 */
94 #define EPI_SLAB_DEBUG 0
95 #endif /* #if DEBUG_EPI != 0 */
97 /* Epoll private bits inside the event mask */
98 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
100 /* Maximum number of poll wake up nests we are allowing */
101 #define EP_MAX_POLLWAKE_NESTS 4
103 /* Maximum msec timeout value storeable in a long int */
104 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
106 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
109 struct epoll_filefd {
115 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
116 * It is used to keep track on all tasks that are currently inside the wake_up() code
117 * to 1) short-circuit the one coming from the same task and same wait queue head
118 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
119 * 3) let go the ones coming from other tasks.
121 struct wake_task_node {
122 struct list_head llink;
123 struct task_struct *task;
124 wait_queue_head_t *wq;
128 * This is used to implement the safe poll wake up avoiding to reenter
129 * the poll callback from inside wake_up().
131 struct poll_safewake {
132 struct list_head wake_task_list;
137 * This structure is stored inside the "private_data" member of the file
138 * structure and rapresent the main data sructure for the eventpoll
142 /* Protect the this structure access */
146 * This semaphore is used to ensure that files are not removed
147 * while epoll is using them. This is read-held during the event
148 * collection loop and it is write-held during the file cleanup
149 * path, the epoll file exit code and the ctl operations.
151 struct rw_semaphore sem;
153 /* Wait queue used by sys_epoll_wait() */
154 wait_queue_head_t wq;
156 /* Wait queue used by file->poll() */
157 wait_queue_head_t poll_wait;
159 /* List of ready file descriptors */
160 struct list_head rdllist;
162 /* RB-Tree root used to store monitored fd structs */
166 /* Wait structure used by the poll hooks */
167 struct eppoll_entry {
168 /* List header used to link this structure to the "struct epitem" */
169 struct list_head llink;
171 /* The "base" pointer is set to the container "struct epitem" */
175 * Wait queue item that will be linked to the target file wait
180 /* The wait queue head that linked the "wait" wait queue item */
181 wait_queue_head_t *whead;
185 * Each file descriptor added to the eventpoll interface will
186 * have an entry of this type linked to the "rbr" RB tree.
189 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
192 /* List header used to link this structure to the eventpoll ready list */
193 struct list_head rdllink;
195 /* The file descriptor information this item refers to */
196 struct epoll_filefd ffd;
198 /* Number of active wait queue attached to poll operations */
201 /* List containing poll wait queues */
202 struct list_head pwqlist;
204 /* The "container" of this item */
205 struct eventpoll *ep;
207 /* The structure that describe the interested events and the source fd */
208 struct epoll_event event;
211 * Used to keep track of the usage count of the structure. This avoids
212 * that the structure will desappear from underneath our processing.
216 /* List header used to link this item to the "struct file" items list */
217 struct list_head fllink;
220 /* Wrapper struct used by poll queueing */
228 static void ep_poll_safewake_init(struct poll_safewake *psw);
229 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
230 static int ep_alloc(struct eventpoll **pep);
231 static void ep_free(struct eventpoll *ep);
232 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
233 static void ep_use_epitem(struct epitem *epi);
234 static void ep_release_epitem(struct epitem *epi);
235 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
237 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi);
238 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
239 struct file *tfile, int fd);
240 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
241 struct epoll_event *event);
242 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
243 static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
244 static int ep_remove(struct eventpoll *ep, struct epitem *epi);
245 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
246 static int ep_eventpoll_close(struct inode *inode, struct file *file);
247 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
248 static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
249 struct epoll_event __user *events, int maxevents);
250 static int ep_events_transfer(struct eventpoll *ep,
251 struct epoll_event __user *events,
253 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
254 int maxevents, long timeout);
257 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
259 static struct mutex epmutex;
261 /* Safe wake up implementation */
262 static struct poll_safewake psw;
264 /* Slab cache used to allocate "struct epitem" */
265 static struct kmem_cache *epi_cache __read_mostly;
267 /* Slab cache used to allocate "struct eppoll_entry" */
268 static struct kmem_cache *pwq_cache __read_mostly;
270 /* File callbacks that implement the eventpoll file behaviour */
271 static const struct file_operations eventpoll_fops = {
272 .release = ep_eventpoll_close,
273 .poll = ep_eventpoll_poll
278 /* Fast test to see if the file is an evenpoll file */
279 static inline int is_file_epoll(struct file *f)
281 return f->f_op == &eventpoll_fops;
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 /* Special initialization for the rb-tree node to detect linkage */
301 static inline void ep_rb_initnode(struct rb_node *n)
306 /* Removes a node from the rb-tree and marks it for a fast is-linked check */
307 static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
313 /* Fast check to verify that the item is linked to the main rb-tree */
314 static inline int ep_rb_linked(struct rb_node *n)
316 return rb_parent(n) != n;
319 /* Tells us if the item is currently linked */
320 static inline int ep_is_linked(struct list_head *p)
322 return !list_empty(p);
325 /* Get the "struct epitem" from a wait queue pointer */
326 static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
328 return container_of(p, struct eppoll_entry, wait)->base;
331 /* Get the "struct epitem" from an epoll queue wrapper */
332 static inline struct epitem * ep_item_from_epqueue(poll_table *p)
334 return container_of(p, struct ep_pqueue, pt)->epi;
337 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
338 static inline int ep_op_has_event(int op)
340 return op != EPOLL_CTL_DEL;
343 /* Initialize the poll safe wake up structure */
344 static void ep_poll_safewake_init(struct poll_safewake *psw)
347 INIT_LIST_HEAD(&psw->wake_task_list);
348 spin_lock_init(&psw->lock);
353 * Perform a safe wake up of the poll wait list. The problem is that
354 * with the new callback'd wake up system, it is possible that the
355 * poll callback is reentered from inside the call to wake_up() done
356 * on the poll wait queue head. The rule is that we cannot reenter the
357 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
358 * and we cannot reenter the same wait queue head at all. This will
359 * enable to have a hierarchy of epoll file descriptor of no more than
360 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
361 * because this one gets called by the poll callback, that in turn is called
362 * from inside a wake_up(), that might be called from irq context.
364 static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
368 struct task_struct *this_task = current;
369 struct list_head *lsthead = &psw->wake_task_list, *lnk;
370 struct wake_task_node *tncur;
371 struct wake_task_node tnode;
373 spin_lock_irqsave(&psw->lock, flags);
375 /* Try to see if the current task is already inside this wakeup call */
376 list_for_each(lnk, lsthead) {
377 tncur = list_entry(lnk, struct wake_task_node, llink);
379 if (tncur->wq == wq ||
380 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
382 * Ops ... loop detected or maximum nest level reached.
383 * We abort this wake by breaking the cycle itself.
385 spin_unlock_irqrestore(&psw->lock, flags);
390 /* Add the current task to the list */
391 tnode.task = this_task;
393 list_add(&tnode.llink, lsthead);
395 spin_unlock_irqrestore(&psw->lock, flags);
397 /* Do really wake up now */
400 /* Remove the current task from the list */
401 spin_lock_irqsave(&psw->lock, flags);
402 list_del(&tnode.llink);
403 spin_unlock_irqrestore(&psw->lock, flags);
408 * This is called from eventpoll_release() to unlink files from the eventpoll
409 * interface. We need to have this facility to cleanup correctly files that are
410 * closed without being removed from the eventpoll interface.
412 void eventpoll_release_file(struct file *file)
414 struct list_head *lsthead = &file->f_ep_links;
415 struct eventpoll *ep;
419 * We don't want to get "file->f_ep_lock" because it is not
420 * necessary. It is not necessary because we're in the "struct file"
421 * cleanup path, and this means that noone is using this file anymore.
422 * The only hit might come from ep_free() but by holding the semaphore
423 * will correctly serialize the operation. We do need to acquire
424 * "ep->sem" after "epmutex" because ep_remove() requires it when called
425 * from anywhere but ep_free().
427 mutex_lock(&epmutex);
429 while (!list_empty(lsthead)) {
430 epi = list_first_entry(lsthead, struct epitem, fllink);
433 list_del_init(&epi->fllink);
434 down_write(&ep->sem);
439 mutex_unlock(&epmutex);
444 * It opens an eventpoll file descriptor by suggesting a storage of "size"
445 * file descriptors. The size parameter is just an hint about how to size
446 * data structures. It won't prevent the user to store more than "size"
447 * file descriptors inside the epoll interface. It is the kernel part of
448 * the userspace epoll_create(2).
450 asmlinkage long sys_epoll_create(int size)
453 struct eventpoll *ep;
457 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
461 * Sanity check on the size parameter, and create the internal data
462 * structure ( "struct eventpoll" ).
465 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
469 * Creates all the items needed to setup an eventpoll file. That is,
470 * a file structure, and inode and a free file descriptor.
472 error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]",
473 &eventpoll_fops, ep);
477 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
486 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
487 current, size, error));
493 * The following function implements the controller interface for
494 * the eventpoll file that enables the insertion/removal/change of
495 * file descriptors inside the interest set. It represents
496 * the kernel part of the user space epoll_ctl(2).
499 sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
502 struct file *file, *tfile;
503 struct eventpoll *ep;
505 struct epoll_event epds;
507 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
508 current, epfd, op, fd, event));
511 if (ep_op_has_event(op) &&
512 copy_from_user(&epds, event, sizeof(struct epoll_event)))
515 /* Get the "struct file *" for the eventpoll file */
521 /* Get the "struct file *" for the target file */
526 /* The target file descriptor must support poll */
528 if (!tfile->f_op || !tfile->f_op->poll)
532 * We have to check that the file structure underneath the file descriptor
533 * the user passed to us _is_ an eventpoll file. And also we do not permit
534 * adding an epoll file descriptor inside itself.
537 if (file == tfile || !is_file_epoll(file))
541 * At this point it is safe to assume that the "private_data" contains
542 * our own data structure.
544 ep = file->private_data;
546 down_write(&ep->sem);
548 /* Try to lookup the file inside our RB tree */
549 epi = ep_find(ep, tfile, fd);
555 epds.events |= POLLERR | POLLHUP;
557 error = ep_insert(ep, &epds, tfile, fd);
563 error = ep_remove(ep, epi);
569 epds.events |= POLLERR | POLLHUP;
570 error = ep_modify(ep, epi, &epds);
577 * The function ep_find() increments the usage count of the structure
578 * so, if this is not NULL, we need to release it.
581 ep_release_epitem(epi);
590 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
591 current, epfd, op, fd, event, error));
598 * Implement the event wait interface for the eventpoll file. It is the kernel
599 * part of the user space epoll_wait(2).
601 asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
602 int maxevents, int timeout)
606 struct eventpoll *ep;
608 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
609 current, epfd, events, maxevents, timeout));
611 /* The maximum number of event must be greater than zero */
612 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
615 /* Verify that the area passed by the user is writeable */
616 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
621 /* Get the "struct file *" for the eventpoll file */
628 * We have to check that the file structure underneath the fd
629 * the user passed to us _is_ an eventpoll file.
632 if (!is_file_epoll(file))
636 * At this point it is safe to assume that the "private_data" contains
637 * our own data structure.
639 ep = file->private_data;
641 /* Time to fish for events ... */
642 error = ep_poll(ep, events, maxevents, timeout);
647 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
648 current, epfd, events, maxevents, timeout, error));
654 #ifdef TIF_RESTORE_SIGMASK
657 * Implement the event wait interface for the eventpoll file. It is the kernel
658 * part of the user space epoll_pwait(2).
660 asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
661 int maxevents, int timeout, const sigset_t __user *sigmask,
665 sigset_t ksigmask, sigsaved;
668 * If the caller wants a certain signal mask to be set during the wait,
672 if (sigsetsize != sizeof(sigset_t))
674 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
676 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
677 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
680 error = sys_epoll_wait(epfd, events, maxevents, timeout);
683 * If we changed the signal mask, we need to restore the original one.
684 * In case we've got a signal while waiting, we do not restore the
685 * signal mask yet, and we allow do_signal() to deliver the signal on
686 * the way back to userspace, before the signal mask is restored.
689 if (error == -EINTR) {
690 memcpy(¤t->saved_sigmask, &sigsaved,
692 set_thread_flag(TIF_RESTORE_SIGMASK);
694 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
700 #endif /* #ifdef TIF_RESTORE_SIGMASK */
703 static int ep_alloc(struct eventpoll **pep)
705 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
710 rwlock_init(&ep->lock);
711 init_rwsem(&ep->sem);
712 init_waitqueue_head(&ep->wq);
713 init_waitqueue_head(&ep->poll_wait);
714 INIT_LIST_HEAD(&ep->rdllist);
719 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
725 static void ep_free(struct eventpoll *ep)
730 /* We need to release all tasks waiting for these file */
731 if (waitqueue_active(&ep->poll_wait))
732 ep_poll_safewake(&psw, &ep->poll_wait);
735 * We need to lock this because we could be hit by
736 * eventpoll_release_file() while we're freeing the "struct eventpoll".
737 * We do not need to hold "ep->sem" here because the epoll file
738 * is on the way to be removed and no one has references to it
739 * anymore. The only hit might come from eventpoll_release_file() but
740 * holding "epmutex" is sufficent here.
742 mutex_lock(&epmutex);
745 * Walks through the whole tree by unregistering poll callbacks.
747 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
748 epi = rb_entry(rbp, struct epitem, rbn);
750 ep_unregister_pollwait(ep, epi);
754 * Walks through the whole tree by freeing each "struct epitem". At this
755 * point we are sure no poll callbacks will be lingering around, and also by
756 * write-holding "sem" we can be sure that no file cleanup code will hit
757 * us during this operation. So we can avoid the lock on "ep->lock".
759 while ((rbp = rb_first(&ep->rbr)) != 0) {
760 epi = rb_entry(rbp, struct epitem, rbn);
764 mutex_unlock(&epmutex);
769 * Search the file inside the eventpoll tree. It add usage count to
770 * the returned item, so the caller must call ep_release_epitem()
771 * after finished using the "struct epitem".
773 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
778 struct epitem *epi, *epir = NULL;
779 struct epoll_filefd ffd;
781 ep_set_ffd(&ffd, file, fd);
782 read_lock_irqsave(&ep->lock, flags);
783 for (rbp = ep->rbr.rb_node; rbp; ) {
784 epi = rb_entry(rbp, struct epitem, rbn);
785 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
796 read_unlock_irqrestore(&ep->lock, flags);
798 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
799 current, file, epir));
806 * Increment the usage count of the "struct epitem" making it sure
807 * that the user will have a valid pointer to reference.
809 static void ep_use_epitem(struct epitem *epi)
812 atomic_inc(&epi->usecnt);
817 * Decrement ( release ) the usage count by signaling that the user
818 * has finished using the structure. It might lead to freeing the
819 * structure itself if the count goes to zero.
821 static void ep_release_epitem(struct epitem *epi)
824 if (atomic_dec_and_test(&epi->usecnt))
825 kmem_cache_free(epi_cache, epi);
830 * This is the callback that is used to add our wait queue to the
831 * target file wakeup lists.
833 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
836 struct epitem *epi = ep_item_from_epqueue(pt);
837 struct eppoll_entry *pwq;
839 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
840 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
843 add_wait_queue(whead, &pwq->wait);
844 list_add_tail(&pwq->llink, &epi->pwqlist);
847 /* We have to signal that an error occurred */
853 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
856 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
861 epic = rb_entry(parent, struct epitem, rbn);
862 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
864 p = &parent->rb_right;
866 p = &parent->rb_left;
868 rb_link_node(&epi->rbn, parent, p);
869 rb_insert_color(&epi->rbn, &ep->rbr);
873 static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
874 struct file *tfile, int fd)
876 int error, revents, pwake = 0;
879 struct ep_pqueue epq;
882 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
885 /* Item initialization follow here ... */
886 ep_rb_initnode(&epi->rbn);
887 INIT_LIST_HEAD(&epi->rdllink);
888 INIT_LIST_HEAD(&epi->fllink);
889 INIT_LIST_HEAD(&epi->pwqlist);
891 ep_set_ffd(&epi->ffd, tfile, fd);
893 atomic_set(&epi->usecnt, 1);
896 /* Initialize the poll table using the queue callback */
898 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
901 * Attach the item to the poll hooks and get current event bits.
902 * We can safely use the file* here because its usage count has
903 * been increased by the caller of this function.
905 revents = tfile->f_op->poll(tfile, &epq.pt);
908 * We have to check if something went wrong during the poll wait queue
909 * install process. Namely an allocation for a wait queue failed due
910 * high memory pressure.
915 /* Add the current item to the list of active epoll hook for this file */
916 spin_lock(&tfile->f_ep_lock);
917 list_add_tail(&epi->fllink, &tfile->f_ep_links);
918 spin_unlock(&tfile->f_ep_lock);
920 /* We have to drop the new item inside our item list to keep track of it */
921 write_lock_irqsave(&ep->lock, flags);
923 /* Add the current item to the rb-tree */
924 ep_rbtree_insert(ep, epi);
926 /* If the file is already "ready" we drop it inside the ready list */
927 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
928 list_add_tail(&epi->rdllink, &ep->rdllist);
930 /* Notify waiting tasks that events are available */
931 if (waitqueue_active(&ep->wq))
932 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
933 if (waitqueue_active(&ep->poll_wait))
937 write_unlock_irqrestore(&ep->lock, flags);
939 /* We have to call this outside the lock */
941 ep_poll_safewake(&psw, &ep->poll_wait);
943 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
944 current, ep, tfile, fd));
949 ep_unregister_pollwait(ep, epi);
952 * We need to do this because an event could have been arrived on some
953 * allocated wait queue.
955 write_lock_irqsave(&ep->lock, flags);
956 if (ep_is_linked(&epi->rdllink))
957 list_del_init(&epi->rdllink);
958 write_unlock_irqrestore(&ep->lock, flags);
960 kmem_cache_free(epi_cache, epi);
967 * Modify the interest event mask by dropping an event if the new mask
968 * has a match in the current file status.
970 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
973 unsigned int revents;
977 * Set the new event interest mask before calling f_op->poll(), otherwise
978 * a potential race might occur. In fact if we do this operation inside
979 * the lock, an event might happen between the f_op->poll() call and the
980 * new event set registering.
982 epi->event.events = event->events;
985 * Get current event bits. We can safely use the file* here because
986 * its usage count has been increased by the caller of this function.
988 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
990 write_lock_irqsave(&ep->lock, flags);
992 /* Copy the data member from inside the lock */
993 epi->event.data = event->data;
996 * If the item is not linked to the RB tree it means that it's on its
997 * way toward the removal. Do nothing in this case.
999 if (ep_rb_linked(&epi->rbn)) {
1001 * If the item is "hot" and it is not registered inside the ready
1002 * list, push it inside. If the item is not "hot" and it is currently
1003 * registered inside the ready list, unlink it.
1005 if (revents & event->events) {
1006 if (!ep_is_linked(&epi->rdllink)) {
1007 list_add_tail(&epi->rdllink, &ep->rdllist);
1009 /* Notify waiting tasks that events are available */
1010 if (waitqueue_active(&ep->wq))
1011 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1012 TASK_INTERRUPTIBLE);
1013 if (waitqueue_active(&ep->poll_wait))
1019 write_unlock_irqrestore(&ep->lock, flags);
1021 /* We have to call this outside the lock */
1023 ep_poll_safewake(&psw, &ep->poll_wait);
1030 * This function unregister poll callbacks from the associated file descriptor.
1031 * Since this must be called without holding "ep->lock" the atomic exchange trick
1032 * will protect us from multiple unregister.
1034 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1037 struct list_head *lsthead = &epi->pwqlist;
1038 struct eppoll_entry *pwq;
1040 /* This is called without locks, so we need the atomic exchange */
1041 nwait = xchg(&epi->nwait, 0);
1044 while (!list_empty(lsthead)) {
1045 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
1047 list_del_init(&pwq->llink);
1048 remove_wait_queue(pwq->whead, &pwq->wait);
1049 kmem_cache_free(pwq_cache, pwq);
1056 * Unlink the "struct epitem" from all places it might have been hooked up.
1057 * This function must be called with write IRQ lock on "ep->lock".
1059 static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1064 * It can happen that this one is called for an item already unlinked.
1065 * The check protect us from doing a double unlink ( crash ).
1068 if (!ep_rb_linked(&epi->rbn))
1072 * Clear the event mask for the unlinked item. This will avoid item
1073 * notifications to be sent after the unlink operation from inside
1074 * the kernel->userspace event transfer loop.
1076 epi->event.events = 0;
1079 * At this point is safe to do the job, unlink the item from our rb-tree.
1080 * This operation togheter with the above check closes the door to
1083 ep_rb_erase(&epi->rbn, &ep->rbr);
1086 * If the item we are going to remove is inside the ready file descriptors
1087 * we want to remove it from this list to avoid stale events.
1089 if (ep_is_linked(&epi->rdllink))
1090 list_del_init(&epi->rdllink);
1095 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1096 current, ep, epi->ffd.file, error));
1103 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
1104 * all the associated resources.
1106 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1109 unsigned long flags;
1110 struct file *file = epi->ffd.file;
1113 * Removes poll wait queue hooks. We _have_ to do this without holding
1114 * the "ep->lock" otherwise a deadlock might occur. This because of the
1115 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1116 * queue head lock when unregistering the wait queue. The wakeup callback
1117 * will run by holding the wait queue head lock and will call our callback
1118 * that will try to get "ep->lock".
1120 ep_unregister_pollwait(ep, epi);
1122 /* Remove the current item from the list of epoll hooks */
1123 spin_lock(&file->f_ep_lock);
1124 if (ep_is_linked(&epi->fllink))
1125 list_del_init(&epi->fllink);
1126 spin_unlock(&file->f_ep_lock);
1128 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1129 write_lock_irqsave(&ep->lock, flags);
1131 /* Really unlink the item from the RB tree */
1132 error = ep_unlink(ep, epi);
1134 write_unlock_irqrestore(&ep->lock, flags);
1139 /* At this point it is safe to free the eventpoll item */
1140 ep_release_epitem(epi);
1144 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1145 current, ep, file, error));
1152 * This is the callback that is passed to the wait queue wakeup
1153 * machanism. It is called by the stored file descriptors when they
1154 * have events to report.
1156 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1159 unsigned long flags;
1160 struct epitem *epi = ep_item_from_wait(wait);
1161 struct eventpoll *ep = epi->ep;
1163 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1164 current, epi->ffd.file, epi, ep));
1166 write_lock_irqsave(&ep->lock, flags);
1169 * If the event mask does not contain any poll(2) event, we consider the
1170 * descriptor to be disabled. This condition is likely the effect of the
1171 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1172 * until the next EPOLL_CTL_MOD will be issued.
1174 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1177 /* If this file is already in the ready list we exit soon */
1178 if (ep_is_linked(&epi->rdllink))
1181 list_add_tail(&epi->rdllink, &ep->rdllist);
1185 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1188 if (waitqueue_active(&ep->wq))
1189 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1190 TASK_INTERRUPTIBLE);
1191 if (waitqueue_active(&ep->poll_wait))
1195 write_unlock_irqrestore(&ep->lock, flags);
1197 /* We have to call this outside the lock */
1199 ep_poll_safewake(&psw, &ep->poll_wait);
1205 static int ep_eventpoll_close(struct inode *inode, struct file *file)
1207 struct eventpoll *ep = file->private_data;
1214 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1219 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1221 unsigned int pollflags = 0;
1222 unsigned long flags;
1223 struct eventpoll *ep = file->private_data;
1225 /* Insert inside our poll wait queue */
1226 poll_wait(file, &ep->poll_wait, wait);
1228 /* Check our condition */
1229 read_lock_irqsave(&ep->lock, flags);
1230 if (!list_empty(&ep->rdllist))
1231 pollflags = POLLIN | POLLRDNORM;
1232 read_unlock_irqrestore(&ep->lock, flags);
1239 * This function is called without holding the "ep->lock" since the call to
1240 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1241 * because of the way poll() is traditionally implemented in Linux.
1243 static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1244 struct epoll_event __user *events, int maxevents)
1246 int eventcnt, error = -EFAULT, pwake = 0;
1247 unsigned int revents;
1248 unsigned long flags;
1250 struct list_head injlist;
1252 INIT_LIST_HEAD(&injlist);
1255 * We can loop without lock because this is a task private list.
1256 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
1257 * Items cannot vanish during the loop because we are holding "sem" in
1260 for (eventcnt = 0; !list_empty(txlist) && eventcnt < maxevents;) {
1261 epi = list_first_entry(txlist, struct epitem, rdllink);
1262 prefetch(epi->rdllink.next);
1265 * Get the ready file event set. We can safely use the file
1266 * because we are holding the "sem" in read and this will
1267 * guarantee that both the file and the item will not vanish.
1269 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1270 revents &= epi->event.events;
1273 * Is the event mask intersect the caller-requested one,
1274 * deliver the event to userspace. Again, we are holding
1275 * "sem" in read, so no operations coming from userspace
1276 * can change the item.
1279 if (__put_user(revents,
1280 &events[eventcnt].events) ||
1281 __put_user(epi->event.data,
1282 &events[eventcnt].data))
1284 if (epi->event.events & EPOLLONESHOT)
1285 epi->event.events &= EP_PRIVATE_BITS;
1290 * This is tricky. We are holding the "sem" in read, and this
1291 * means that the operations that can change the "linked" status
1292 * of the epoll item (epi->rbn and epi->rdllink), cannot touch
1293 * them. Also, since we are "linked" from a epi->rdllink POV
1294 * (the item is linked to our transmission list we just
1295 * spliced), the ep_poll_callback() cannot touch us either,
1296 * because of the check present in there. Another parallel
1297 * epoll_wait() will not get the same result set, since we
1298 * spliced the ready list before. Note that list_del() still
1299 * shows the item as linked to the test in ep_poll_callback().
1301 list_del(&epi->rdllink);
1302 if (!(epi->event.events & EPOLLET) &&
1303 (revents & epi->event.events))
1304 list_add_tail(&epi->rdllink, &injlist);
1307 * Be sure the item is totally detached before re-init
1308 * the list_head. After INIT_LIST_HEAD() is committed,
1309 * the ep_poll_callback() can requeue the item again,
1310 * but we don't care since we are already past it.
1313 INIT_LIST_HEAD(&epi->rdllink);
1321 * If the re-injection list or the txlist are not empty, re-splice
1322 * them to the ready list and do proper wakeups.
1324 if (!list_empty(&injlist) || !list_empty(txlist)) {
1325 write_lock_irqsave(&ep->lock, flags);
1327 list_splice(txlist, &ep->rdllist);
1328 list_splice(&injlist, &ep->rdllist);
1330 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1333 if (waitqueue_active(&ep->wq))
1334 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1335 TASK_INTERRUPTIBLE);
1336 if (waitqueue_active(&ep->poll_wait))
1339 write_unlock_irqrestore(&ep->lock, flags);
1342 /* We have to call this outside the lock */
1344 ep_poll_safewake(&psw, &ep->poll_wait);
1346 return eventcnt == 0 ? error: eventcnt;
1351 * Perform the transfer of events to user space.
1353 static int ep_events_transfer(struct eventpoll *ep,
1354 struct epoll_event __user *events, int maxevents)
1357 unsigned long flags;
1358 struct list_head txlist;
1360 INIT_LIST_HEAD(&txlist);
1363 * We need to lock this because we could be hit by
1364 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1366 down_read(&ep->sem);
1369 * Steal the ready list, and re-init the original one to the
1372 write_lock_irqsave(&ep->lock, flags);
1373 list_splice(&ep->rdllist, &txlist);
1374 INIT_LIST_HEAD(&ep->rdllist);
1375 write_unlock_irqrestore(&ep->lock, flags);
1377 /* Build result set in userspace */
1378 eventcnt = ep_send_events(ep, &txlist, events, maxevents);
1386 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1387 int maxevents, long timeout)
1390 unsigned long flags;
1395 * Calculate the timeout by checking for the "infinite" value ( -1 )
1396 * and the overflow condition. The passed timeout is in milliseconds,
1397 * that why (t * HZ) / 1000.
1399 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1400 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1403 write_lock_irqsave(&ep->lock, flags);
1406 if (list_empty(&ep->rdllist)) {
1408 * We don't have any available event to return to the caller.
1409 * We need to sleep here, and we will be wake up by
1410 * ep_poll_callback() when events will become available.
1412 init_waitqueue_entry(&wait, current);
1413 __add_wait_queue(&ep->wq, &wait);
1417 * We don't want to sleep if the ep_poll_callback() sends us
1418 * a wakeup in between. That's why we set the task state
1419 * to TASK_INTERRUPTIBLE before doing the checks.
1421 set_current_state(TASK_INTERRUPTIBLE);
1422 if (!list_empty(&ep->rdllist) || !jtimeout)
1424 if (signal_pending(current)) {
1429 write_unlock_irqrestore(&ep->lock, flags);
1430 jtimeout = schedule_timeout(jtimeout);
1431 write_lock_irqsave(&ep->lock, flags);
1433 __remove_wait_queue(&ep->wq, &wait);
1435 set_current_state(TASK_RUNNING);
1438 /* Is it worth to try to dig for events ? */
1439 eavail = !list_empty(&ep->rdllist);
1441 write_unlock_irqrestore(&ep->lock, flags);
1444 * Try to transfer events to user space. In case we get 0 events and
1445 * there's still timeout left over, we go trying again in search of
1448 if (!res && eavail &&
1449 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1455 static int __init eventpoll_init(void)
1457 mutex_init(&epmutex);
1459 /* Initialize the structure used to perform safe poll wait head wake ups */
1460 ep_poll_safewake_init(&psw);
1462 /* Allocates slab cache used to allocate "struct epitem" items */
1463 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1464 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1467 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1468 pwq_cache = kmem_cache_create("eventpoll_pwq",
1469 sizeof(struct eppoll_entry), 0,
1470 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1476 static void __exit eventpoll_exit(void)
1478 /* Undo all operations done inside eventpoll_init() */
1479 kmem_cache_destroy(pwq_cache);
1480 kmem_cache_destroy(epi_cache);
1483 module_init(eventpoll_init);
1484 module_exit(eventpoll_exit);
1486 MODULE_LICENSE("GPL");