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/smp_lock.h>
26 #include <linux/string.h>
27 #include <linux/list.h>
28 #include <linux/hash.h>
29 #include <linux/spinlock.h>
30 #include <linux/syscalls.h>
31 #include <linux/rwsem.h>
32 #include <linux/rbtree.h>
33 #include <linux/wait.h>
34 #include <linux/eventpoll.h>
35 #include <linux/mount.h>
36 #include <linux/bitops.h>
37 #include <linux/mutex.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.
79 #define EVENTPOLLFS_MAGIC 0x03111965 /* My birthday should work for this :) */
84 #define DPRINTK(x) printk x
85 #define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
86 #else /* #if DEBUG_EPOLL > 0 */
87 #define DPRINTK(x) (void) 0
88 #define DNPRINTK(n, x) (void) 0
89 #endif /* #if DEBUG_EPOLL > 0 */
94 #define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
95 #else /* #if DEBUG_EPI != 0 */
96 #define EPI_SLAB_DEBUG 0
97 #endif /* #if DEBUG_EPI != 0 */
99 /* Epoll private bits inside the event mask */
100 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
102 /* Maximum number of poll wake up nests we are allowing */
103 #define EP_MAX_POLLWAKE_NESTS 4
105 /* Maximum msec timeout value storeable in a long int */
106 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
108 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
111 struct epoll_filefd
{
117 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
118 * It is used to keep track on all tasks that are currently inside the wake_up() code
119 * to 1) short-circuit the one coming from the same task and same wait queue head
120 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
121 * 3) let go the ones coming from other tasks.
123 struct wake_task_node
{
124 struct list_head llink
;
125 struct task_struct
*task
;
126 wait_queue_head_t
*wq
;
130 * This is used to implement the safe poll wake up avoiding to reenter
131 * the poll callback from inside wake_up().
133 struct poll_safewake
{
134 struct list_head wake_task_list
;
139 * This structure is stored inside the "private_data" member of the file
140 * structure and rapresent the main data sructure for the eventpoll
144 /* Protect the this structure access */
148 * This semaphore is used to ensure that files are not removed
149 * while epoll is using them. This is read-held during the event
150 * collection loop and it is write-held during the file cleanup
151 * path, the epoll file exit code and the ctl operations.
153 struct rw_semaphore sem
;
155 /* Wait queue used by sys_epoll_wait() */
156 wait_queue_head_t wq
;
158 /* Wait queue used by file->poll() */
159 wait_queue_head_t poll_wait
;
161 /* List of ready file descriptors */
162 struct list_head rdllist
;
164 /* RB-Tree root used to store monitored fd structs */
168 /* Wait structure used by the poll hooks */
169 struct eppoll_entry
{
170 /* List header used to link this structure to the "struct epitem" */
171 struct list_head llink
;
173 /* The "base" pointer is set to the container "struct epitem" */
177 * Wait queue item that will be linked to the target file wait
182 /* The wait queue head that linked the "wait" wait queue item */
183 wait_queue_head_t
*whead
;
187 * Each file descriptor added to the eventpoll interface will
188 * have an entry of this type linked to the "rbr" RB tree.
191 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
194 /* List header used to link this structure to the eventpoll ready list */
195 struct list_head rdllink
;
197 /* The file descriptor information this item refers to */
198 struct epoll_filefd ffd
;
200 /* Number of active wait queue attached to poll operations */
203 /* List containing poll wait queues */
204 struct list_head pwqlist
;
206 /* The "container" of this item */
207 struct eventpoll
*ep
;
209 /* The structure that describe the interested events and the source fd */
210 struct epoll_event event
;
213 * Used to keep track of the usage count of the structure. This avoids
214 * that the structure will desappear from underneath our processing.
218 /* List header used to link this item to the "struct file" items list */
219 struct list_head fllink
;
222 /* Wrapper struct used by poll queueing */
230 static void ep_poll_safewake_init(struct poll_safewake
*psw
);
231 static void ep_poll_safewake(struct poll_safewake
*psw
, wait_queue_head_t
*wq
);
232 static int ep_getfd(int *efd
, struct inode
**einode
, struct file
**efile
,
233 struct eventpoll
*ep
);
234 static int ep_alloc(struct eventpoll
**pep
);
235 static void ep_free(struct eventpoll
*ep
);
236 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
);
237 static void ep_use_epitem(struct epitem
*epi
);
238 static void ep_release_epitem(struct epitem
*epi
);
239 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
241 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
);
242 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
243 struct file
*tfile
, int fd
);
244 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
,
245 struct epoll_event
*event
);
246 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
);
247 static int ep_unlink(struct eventpoll
*ep
, struct epitem
*epi
);
248 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
);
249 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
);
250 static int ep_eventpoll_close(struct inode
*inode
, struct file
*file
);
251 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
);
252 static int ep_send_events(struct eventpoll
*ep
, struct list_head
*txlist
,
253 struct epoll_event __user
*events
, int maxevents
);
254 static int ep_events_transfer(struct eventpoll
*ep
,
255 struct epoll_event __user
*events
,
257 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
258 int maxevents
, long timeout
);
259 static int eventpollfs_delete_dentry(struct dentry
*dentry
);
260 static struct inode
*ep_eventpoll_inode(void);
261 static int eventpollfs_get_sb(struct file_system_type
*fs_type
,
262 int flags
, const char *dev_name
,
263 void *data
, struct vfsmount
*mnt
);
266 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
268 static struct mutex epmutex
;
270 /* Safe wake up implementation */
271 static struct poll_safewake psw
;
273 /* Slab cache used to allocate "struct epitem" */
274 static struct kmem_cache
*epi_cache __read_mostly
;
276 /* Slab cache used to allocate "struct eppoll_entry" */
277 static struct kmem_cache
*pwq_cache __read_mostly
;
279 /* Virtual fs used to allocate inodes for eventpoll files */
280 static struct vfsmount
*eventpoll_mnt __read_mostly
;
282 /* File callbacks that implement the eventpoll file behaviour */
283 static const struct file_operations eventpoll_fops
= {
284 .release
= ep_eventpoll_close
,
285 .poll
= ep_eventpoll_poll
289 * This is used to register the virtual file system from where
290 * eventpoll inodes are allocated.
292 static struct file_system_type eventpoll_fs_type
= {
293 .name
= "eventpollfs",
294 .get_sb
= eventpollfs_get_sb
,
295 .kill_sb
= kill_anon_super
,
298 /* Very basic directory entry operations for the eventpoll virtual file system */
299 static struct dentry_operations eventpollfs_dentry_operations
= {
300 .d_delete
= eventpollfs_delete_dentry
,
305 /* Fast test to see if the file is an evenpoll file */
306 static inline int is_file_epoll(struct file
*f
)
308 return f
->f_op
== &eventpoll_fops
;
311 /* Setup the structure that is used as key for the rb-tree */
312 static inline void ep_set_ffd(struct epoll_filefd
*ffd
,
313 struct file
*file
, int fd
)
319 /* Compare rb-tree keys */
320 static inline int ep_cmp_ffd(struct epoll_filefd
*p1
,
321 struct epoll_filefd
*p2
)
323 return (p1
->file
> p2
->file
? +1:
324 (p1
->file
< p2
->file
? -1 : p1
->fd
- p2
->fd
));
327 /* Special initialization for the rb-tree node to detect linkage */
328 static inline void ep_rb_initnode(struct rb_node
*n
)
333 /* Removes a node from the rb-tree and marks it for a fast is-linked check */
334 static inline void ep_rb_erase(struct rb_node
*n
, struct rb_root
*r
)
340 /* Fast check to verify that the item is linked to the main rb-tree */
341 static inline int ep_rb_linked(struct rb_node
*n
)
343 return rb_parent(n
) != n
;
346 /* Tells us if the item is currently linked */
347 static inline int ep_is_linked(struct list_head
*p
)
349 return !list_empty(p
);
352 /* Get the "struct epitem" from a wait queue pointer */
353 static inline struct epitem
* ep_item_from_wait(wait_queue_t
*p
)
355 return container_of(p
, struct eppoll_entry
, wait
)->base
;
358 /* Get the "struct epitem" from an epoll queue wrapper */
359 static inline struct epitem
* ep_item_from_epqueue(poll_table
*p
)
361 return container_of(p
, struct ep_pqueue
, pt
)->epi
;
364 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
365 static inline int ep_op_has_event(int op
)
367 return op
!= EPOLL_CTL_DEL
;
370 /* Initialize the poll safe wake up structure */
371 static void ep_poll_safewake_init(struct poll_safewake
*psw
)
374 INIT_LIST_HEAD(&psw
->wake_task_list
);
375 spin_lock_init(&psw
->lock
);
380 * Perform a safe wake up of the poll wait list. The problem is that
381 * with the new callback'd wake up system, it is possible that the
382 * poll callback is reentered from inside the call to wake_up() done
383 * on the poll wait queue head. The rule is that we cannot reenter the
384 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
385 * and we cannot reenter the same wait queue head at all. This will
386 * enable to have a hierarchy of epoll file descriptor of no more than
387 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
388 * because this one gets called by the poll callback, that in turn is called
389 * from inside a wake_up(), that might be called from irq context.
391 static void ep_poll_safewake(struct poll_safewake
*psw
, wait_queue_head_t
*wq
)
395 struct task_struct
*this_task
= current
;
396 struct list_head
*lsthead
= &psw
->wake_task_list
, *lnk
;
397 struct wake_task_node
*tncur
;
398 struct wake_task_node tnode
;
400 spin_lock_irqsave(&psw
->lock
, flags
);
402 /* Try to see if the current task is already inside this wakeup call */
403 list_for_each(lnk
, lsthead
) {
404 tncur
= list_entry(lnk
, struct wake_task_node
, llink
);
406 if (tncur
->wq
== wq
||
407 (tncur
->task
== this_task
&& ++wake_nests
> EP_MAX_POLLWAKE_NESTS
)) {
409 * Ops ... loop detected or maximum nest level reached.
410 * We abort this wake by breaking the cycle itself.
412 spin_unlock_irqrestore(&psw
->lock
, flags
);
417 /* Add the current task to the list */
418 tnode
.task
= this_task
;
420 list_add(&tnode
.llink
, lsthead
);
422 spin_unlock_irqrestore(&psw
->lock
, flags
);
424 /* Do really wake up now */
427 /* Remove the current task from the list */
428 spin_lock_irqsave(&psw
->lock
, flags
);
429 list_del(&tnode
.llink
);
430 spin_unlock_irqrestore(&psw
->lock
, flags
);
435 * This is called from eventpoll_release() to unlink files from the eventpoll
436 * interface. We need to have this facility to cleanup correctly files that are
437 * closed without being removed from the eventpoll interface.
439 void eventpoll_release_file(struct file
*file
)
441 struct list_head
*lsthead
= &file
->f_ep_links
;
442 struct eventpoll
*ep
;
446 * We don't want to get "file->f_ep_lock" because it is not
447 * necessary. It is not necessary because we're in the "struct file"
448 * cleanup path, and this means that noone is using this file anymore.
449 * The only hit might come from ep_free() but by holding the semaphore
450 * will correctly serialize the operation. We do need to acquire
451 * "ep->sem" after "epmutex" because ep_remove() requires it when called
452 * from anywhere but ep_free().
454 mutex_lock(&epmutex
);
456 while (!list_empty(lsthead
)) {
457 epi
= list_entry(lsthead
->next
, struct epitem
, fllink
);
460 list_del_init(&epi
->fllink
);
461 down_write(&ep
->sem
);
466 mutex_unlock(&epmutex
);
471 * It opens an eventpoll file descriptor by suggesting a storage of "size"
472 * file descriptors. The size parameter is just an hint about how to size
473 * data structures. It won't prevent the user to store more than "size"
474 * file descriptors inside the epoll interface. It is the kernel part of
475 * the userspace epoll_create(2).
477 asmlinkage
long sys_epoll_create(int size
)
480 struct eventpoll
*ep
;
484 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_create(%d)\n",
488 * Sanity check on the size parameter, and create the internal data
489 * structure ( "struct eventpoll" ).
492 if (size
<= 0 || (error
= ep_alloc(&ep
)) != 0)
496 * Creates all the items needed to setup an eventpoll file. That is,
497 * a file structure, and inode and a free file descriptor.
499 error
= ep_getfd(&fd
, &inode
, &file
, ep
);
503 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_create(%d) = %d\n",
512 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_create(%d) = %d\n",
513 current
, size
, error
));
519 * The following function implements the controller interface for
520 * the eventpoll file that enables the insertion/removal/change of
521 * file descriptors inside the interest set. It represents
522 * the kernel part of the user space epoll_ctl(2).
525 sys_epoll_ctl(int epfd
, int op
, int fd
, struct epoll_event __user
*event
)
528 struct file
*file
, *tfile
;
529 struct eventpoll
*ep
;
531 struct epoll_event epds
;
533 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
534 current
, epfd
, op
, fd
, event
));
537 if (ep_op_has_event(op
) &&
538 copy_from_user(&epds
, event
, sizeof(struct epoll_event
)))
541 /* Get the "struct file *" for the eventpoll file */
547 /* Get the "struct file *" for the target file */
552 /* The target file descriptor must support poll */
554 if (!tfile
->f_op
|| !tfile
->f_op
->poll
)
558 * We have to check that the file structure underneath the file descriptor
559 * the user passed to us _is_ an eventpoll file. And also we do not permit
560 * adding an epoll file descriptor inside itself.
563 if (file
== tfile
|| !is_file_epoll(file
))
567 * At this point it is safe to assume that the "private_data" contains
568 * our own data structure.
570 ep
= file
->private_data
;
572 down_write(&ep
->sem
);
574 /* Try to lookup the file inside our RB tree */
575 epi
= ep_find(ep
, tfile
, fd
);
581 epds
.events
|= POLLERR
| POLLHUP
;
583 error
= ep_insert(ep
, &epds
, tfile
, fd
);
589 error
= ep_remove(ep
, epi
);
595 epds
.events
|= POLLERR
| POLLHUP
;
596 error
= ep_modify(ep
, epi
, &epds
);
603 * The function ep_find() increments the usage count of the structure
604 * so, if this is not NULL, we need to release it.
607 ep_release_epitem(epi
);
616 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
617 current
, epfd
, op
, fd
, event
, error
));
624 * Implement the event wait interface for the eventpoll file. It is the kernel
625 * part of the user space epoll_wait(2).
627 asmlinkage
long sys_epoll_wait(int epfd
, struct epoll_event __user
*events
,
628 int maxevents
, int timeout
)
632 struct eventpoll
*ep
;
634 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
635 current
, epfd
, events
, maxevents
, timeout
));
637 /* The maximum number of event must be greater than zero */
638 if (maxevents
<= 0 || maxevents
> EP_MAX_EVENTS
)
641 /* Verify that the area passed by the user is writeable */
642 if (!access_ok(VERIFY_WRITE
, events
, maxevents
* sizeof(struct epoll_event
))) {
647 /* Get the "struct file *" for the eventpoll file */
654 * We have to check that the file structure underneath the fd
655 * the user passed to us _is_ an eventpoll file.
658 if (!is_file_epoll(file
))
662 * At this point it is safe to assume that the "private_data" contains
663 * our own data structure.
665 ep
= file
->private_data
;
667 /* Time to fish for events ... */
668 error
= ep_poll(ep
, events
, maxevents
, timeout
);
673 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
674 current
, epfd
, events
, maxevents
, timeout
, error
));
680 #ifdef TIF_RESTORE_SIGMASK
683 * Implement the event wait interface for the eventpoll file. It is the kernel
684 * part of the user space epoll_pwait(2).
686 asmlinkage
long sys_epoll_pwait(int epfd
, struct epoll_event __user
*events
,
687 int maxevents
, int timeout
, const sigset_t __user
*sigmask
,
691 sigset_t ksigmask
, sigsaved
;
694 * If the caller wants a certain signal mask to be set during the wait,
698 if (sigsetsize
!= sizeof(sigset_t
))
700 if (copy_from_user(&ksigmask
, sigmask
, sizeof(ksigmask
)))
702 sigdelsetmask(&ksigmask
, sigmask(SIGKILL
) | sigmask(SIGSTOP
));
703 sigprocmask(SIG_SETMASK
, &ksigmask
, &sigsaved
);
706 error
= sys_epoll_wait(epfd
, events
, maxevents
, timeout
);
709 * If we changed the signal mask, we need to restore the original one.
710 * In case we've got a signal while waiting, we do not restore the
711 * signal mask yet, and we allow do_signal() to deliver the signal on
712 * the way back to userspace, before the signal mask is restored.
715 if (error
== -EINTR
) {
716 memcpy(¤t
->saved_sigmask
, &sigsaved
,
718 set_thread_flag(TIF_RESTORE_SIGMASK
);
720 sigprocmask(SIG_SETMASK
, &sigsaved
, NULL
);
726 #endif /* #ifdef TIF_RESTORE_SIGMASK */
730 * Creates the file descriptor to be used by the epoll interface.
732 static int ep_getfd(int *efd
, struct inode
**einode
, struct file
**efile
,
733 struct eventpoll
*ep
)
737 struct dentry
*dentry
;
742 /* Get an ready to use file */
744 file
= get_empty_filp();
748 /* Allocates an inode from the eventpoll file system */
749 inode
= ep_eventpoll_inode();
751 error
= PTR_ERR(inode
);
755 /* Allocates a free descriptor to plug the file onto */
756 error
= get_unused_fd();
762 * Link the inode to a directory entry by creating a unique name
763 * using the inode number.
766 sprintf(name
, "[%lu]", inode
->i_ino
);
768 this.len
= strlen(name
);
769 this.hash
= inode
->i_ino
;
770 dentry
= d_alloc(eventpoll_mnt
->mnt_sb
->s_root
, &this);
773 dentry
->d_op
= &eventpollfs_dentry_operations
;
774 d_add(dentry
, inode
);
775 file
->f_path
.mnt
= mntget(eventpoll_mnt
);
776 file
->f_path
.dentry
= dentry
;
777 file
->f_mapping
= inode
->i_mapping
;
780 file
->f_flags
= O_RDONLY
;
781 file
->f_op
= &eventpoll_fops
;
782 file
->f_mode
= FMODE_READ
;
784 file
->private_data
= ep
;
786 /* Install the new setup file into the allocated fd. */
787 fd_install(fd
, file
);
805 static int ep_alloc(struct eventpoll
**pep
)
807 struct eventpoll
*ep
= kzalloc(sizeof(*ep
), GFP_KERNEL
);
812 rwlock_init(&ep
->lock
);
813 init_rwsem(&ep
->sem
);
814 init_waitqueue_head(&ep
->wq
);
815 init_waitqueue_head(&ep
->poll_wait
);
816 INIT_LIST_HEAD(&ep
->rdllist
);
821 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: ep_alloc() ep=%p\n",
827 static void ep_free(struct eventpoll
*ep
)
832 /* We need to release all tasks waiting for these file */
833 if (waitqueue_active(&ep
->poll_wait
))
834 ep_poll_safewake(&psw
, &ep
->poll_wait
);
837 * We need to lock this because we could be hit by
838 * eventpoll_release_file() while we're freeing the "struct eventpoll".
839 * We do not need to hold "ep->sem" here because the epoll file
840 * is on the way to be removed and no one has references to it
841 * anymore. The only hit might come from eventpoll_release_file() but
842 * holding "epmutex" is sufficent here.
844 mutex_lock(&epmutex
);
847 * Walks through the whole tree by unregistering poll callbacks.
849 for (rbp
= rb_first(&ep
->rbr
); rbp
; rbp
= rb_next(rbp
)) {
850 epi
= rb_entry(rbp
, struct epitem
, rbn
);
852 ep_unregister_pollwait(ep
, epi
);
856 * Walks through the whole tree by freeing each "struct epitem". At this
857 * point we are sure no poll callbacks will be lingering around, and also by
858 * write-holding "sem" we can be sure that no file cleanup code will hit
859 * us during this operation. So we can avoid the lock on "ep->lock".
861 while ((rbp
= rb_first(&ep
->rbr
)) != 0) {
862 epi
= rb_entry(rbp
, struct epitem
, rbn
);
866 mutex_unlock(&epmutex
);
871 * Search the file inside the eventpoll tree. It add usage count to
872 * the returned item, so the caller must call ep_release_epitem()
873 * after finished using the "struct epitem".
875 static struct epitem
*ep_find(struct eventpoll
*ep
, struct file
*file
, int fd
)
880 struct epitem
*epi
, *epir
= NULL
;
881 struct epoll_filefd ffd
;
883 ep_set_ffd(&ffd
, file
, fd
);
884 read_lock_irqsave(&ep
->lock
, flags
);
885 for (rbp
= ep
->rbr
.rb_node
; rbp
; ) {
886 epi
= rb_entry(rbp
, struct epitem
, rbn
);
887 kcmp
= ep_cmp_ffd(&ffd
, &epi
->ffd
);
898 read_unlock_irqrestore(&ep
->lock
, flags
);
900 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: ep_find(%p) -> %p\n",
901 current
, file
, epir
));
908 * Increment the usage count of the "struct epitem" making it sure
909 * that the user will have a valid pointer to reference.
911 static void ep_use_epitem(struct epitem
*epi
)
914 atomic_inc(&epi
->usecnt
);
919 * Decrement ( release ) the usage count by signaling that the user
920 * has finished using the structure. It might lead to freeing the
921 * structure itself if the count goes to zero.
923 static void ep_release_epitem(struct epitem
*epi
)
926 if (atomic_dec_and_test(&epi
->usecnt
))
927 kmem_cache_free(epi_cache
, epi
);
932 * This is the callback that is used to add our wait queue to the
933 * target file wakeup lists.
935 static void ep_ptable_queue_proc(struct file
*file
, wait_queue_head_t
*whead
,
938 struct epitem
*epi
= ep_item_from_epqueue(pt
);
939 struct eppoll_entry
*pwq
;
941 if (epi
->nwait
>= 0 && (pwq
= kmem_cache_alloc(pwq_cache
, GFP_KERNEL
))) {
942 init_waitqueue_func_entry(&pwq
->wait
, ep_poll_callback
);
945 add_wait_queue(whead
, &pwq
->wait
);
946 list_add_tail(&pwq
->llink
, &epi
->pwqlist
);
949 /* We have to signal that an error occurred */
955 static void ep_rbtree_insert(struct eventpoll
*ep
, struct epitem
*epi
)
958 struct rb_node
**p
= &ep
->rbr
.rb_node
, *parent
= NULL
;
963 epic
= rb_entry(parent
, struct epitem
, rbn
);
964 kcmp
= ep_cmp_ffd(&epi
->ffd
, &epic
->ffd
);
966 p
= &parent
->rb_right
;
968 p
= &parent
->rb_left
;
970 rb_link_node(&epi
->rbn
, parent
, p
);
971 rb_insert_color(&epi
->rbn
, &ep
->rbr
);
975 static int ep_insert(struct eventpoll
*ep
, struct epoll_event
*event
,
976 struct file
*tfile
, int fd
)
978 int error
, revents
, pwake
= 0;
981 struct ep_pqueue epq
;
984 if (!(epi
= kmem_cache_alloc(epi_cache
, GFP_KERNEL
)))
987 /* Item initialization follow here ... */
988 ep_rb_initnode(&epi
->rbn
);
989 INIT_LIST_HEAD(&epi
->rdllink
);
990 INIT_LIST_HEAD(&epi
->fllink
);
991 INIT_LIST_HEAD(&epi
->pwqlist
);
993 ep_set_ffd(&epi
->ffd
, tfile
, fd
);
995 atomic_set(&epi
->usecnt
, 1);
998 /* Initialize the poll table using the queue callback */
1000 init_poll_funcptr(&epq
.pt
, ep_ptable_queue_proc
);
1003 * Attach the item to the poll hooks and get current event bits.
1004 * We can safely use the file* here because its usage count has
1005 * been increased by the caller of this function.
1007 revents
= tfile
->f_op
->poll(tfile
, &epq
.pt
);
1010 * We have to check if something went wrong during the poll wait queue
1011 * install process. Namely an allocation for a wait queue failed due
1012 * high memory pressure.
1017 /* Add the current item to the list of active epoll hook for this file */
1018 spin_lock(&tfile
->f_ep_lock
);
1019 list_add_tail(&epi
->fllink
, &tfile
->f_ep_links
);
1020 spin_unlock(&tfile
->f_ep_lock
);
1022 /* We have to drop the new item inside our item list to keep track of it */
1023 write_lock_irqsave(&ep
->lock
, flags
);
1025 /* Add the current item to the rb-tree */
1026 ep_rbtree_insert(ep
, epi
);
1028 /* If the file is already "ready" we drop it inside the ready list */
1029 if ((revents
& event
->events
) && !ep_is_linked(&epi
->rdllink
)) {
1030 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1032 /* Notify waiting tasks that events are available */
1033 if (waitqueue_active(&ep
->wq
))
1034 __wake_up_locked(&ep
->wq
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
);
1035 if (waitqueue_active(&ep
->poll_wait
))
1039 write_unlock_irqrestore(&ep
->lock
, flags
);
1041 /* We have to call this outside the lock */
1043 ep_poll_safewake(&psw
, &ep
->poll_wait
);
1045 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1046 current
, ep
, tfile
, fd
));
1051 ep_unregister_pollwait(ep
, epi
);
1054 * We need to do this because an event could have been arrived on some
1055 * allocated wait queue.
1057 write_lock_irqsave(&ep
->lock
, flags
);
1058 if (ep_is_linked(&epi
->rdllink
))
1059 list_del_init(&epi
->rdllink
);
1060 write_unlock_irqrestore(&ep
->lock
, flags
);
1062 kmem_cache_free(epi_cache
, epi
);
1069 * Modify the interest event mask by dropping an event if the new mask
1070 * has a match in the current file status.
1072 static int ep_modify(struct eventpoll
*ep
, struct epitem
*epi
, struct epoll_event
*event
)
1075 unsigned int revents
;
1076 unsigned long flags
;
1079 * Set the new event interest mask before calling f_op->poll(), otherwise
1080 * a potential race might occur. In fact if we do this operation inside
1081 * the lock, an event might happen between the f_op->poll() call and the
1082 * new event set registering.
1084 epi
->event
.events
= event
->events
;
1087 * Get current event bits. We can safely use the file* here because
1088 * its usage count has been increased by the caller of this function.
1090 revents
= epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
);
1092 write_lock_irqsave(&ep
->lock
, flags
);
1094 /* Copy the data member from inside the lock */
1095 epi
->event
.data
= event
->data
;
1098 * If the item is not linked to the RB tree it means that it's on its
1099 * way toward the removal. Do nothing in this case.
1101 if (ep_rb_linked(&epi
->rbn
)) {
1103 * If the item is "hot" and it is not registered inside the ready
1104 * list, push it inside. If the item is not "hot" and it is currently
1105 * registered inside the ready list, unlink it.
1107 if (revents
& event
->events
) {
1108 if (!ep_is_linked(&epi
->rdllink
)) {
1109 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1111 /* Notify waiting tasks that events are available */
1112 if (waitqueue_active(&ep
->wq
))
1113 __wake_up_locked(&ep
->wq
, TASK_UNINTERRUPTIBLE
|
1114 TASK_INTERRUPTIBLE
);
1115 if (waitqueue_active(&ep
->poll_wait
))
1121 write_unlock_irqrestore(&ep
->lock
, flags
);
1123 /* We have to call this outside the lock */
1125 ep_poll_safewake(&psw
, &ep
->poll_wait
);
1132 * This function unregister poll callbacks from the associated file descriptor.
1133 * Since this must be called without holding "ep->lock" the atomic exchange trick
1134 * will protect us from multiple unregister.
1136 static void ep_unregister_pollwait(struct eventpoll
*ep
, struct epitem
*epi
)
1139 struct list_head
*lsthead
= &epi
->pwqlist
;
1140 struct eppoll_entry
*pwq
;
1142 /* This is called without locks, so we need the atomic exchange */
1143 nwait
= xchg(&epi
->nwait
, 0);
1146 while (!list_empty(lsthead
)) {
1147 pwq
= list_entry(lsthead
->next
, struct eppoll_entry
, llink
);
1149 list_del_init(&pwq
->llink
);
1150 remove_wait_queue(pwq
->whead
, &pwq
->wait
);
1151 kmem_cache_free(pwq_cache
, pwq
);
1158 * Unlink the "struct epitem" from all places it might have been hooked up.
1159 * This function must be called with write IRQ lock on "ep->lock".
1161 static int ep_unlink(struct eventpoll
*ep
, struct epitem
*epi
)
1166 * It can happen that this one is called for an item already unlinked.
1167 * The check protect us from doing a double unlink ( crash ).
1170 if (!ep_rb_linked(&epi
->rbn
))
1174 * Clear the event mask for the unlinked item. This will avoid item
1175 * notifications to be sent after the unlink operation from inside
1176 * the kernel->userspace event transfer loop.
1178 epi
->event
.events
= 0;
1181 * At this point is safe to do the job, unlink the item from our rb-tree.
1182 * This operation togheter with the above check closes the door to
1185 ep_rb_erase(&epi
->rbn
, &ep
->rbr
);
1188 * If the item we are going to remove is inside the ready file descriptors
1189 * we want to remove it from this list to avoid stale events.
1191 if (ep_is_linked(&epi
->rdllink
))
1192 list_del_init(&epi
->rdllink
);
1197 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1198 current
, ep
, epi
->ffd
.file
, error
));
1205 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
1206 * all the associated resources.
1208 static int ep_remove(struct eventpoll
*ep
, struct epitem
*epi
)
1211 unsigned long flags
;
1212 struct file
*file
= epi
->ffd
.file
;
1215 * Removes poll wait queue hooks. We _have_ to do this without holding
1216 * the "ep->lock" otherwise a deadlock might occur. This because of the
1217 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1218 * queue head lock when unregistering the wait queue. The wakeup callback
1219 * will run by holding the wait queue head lock and will call our callback
1220 * that will try to get "ep->lock".
1222 ep_unregister_pollwait(ep
, epi
);
1224 /* Remove the current item from the list of epoll hooks */
1225 spin_lock(&file
->f_ep_lock
);
1226 if (ep_is_linked(&epi
->fllink
))
1227 list_del_init(&epi
->fllink
);
1228 spin_unlock(&file
->f_ep_lock
);
1230 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1231 write_lock_irqsave(&ep
->lock
, flags
);
1233 /* Really unlink the item from the RB tree */
1234 error
= ep_unlink(ep
, epi
);
1236 write_unlock_irqrestore(&ep
->lock
, flags
);
1241 /* At this point it is safe to free the eventpoll item */
1242 ep_release_epitem(epi
);
1246 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1247 current
, ep
, file
, error
));
1254 * This is the callback that is passed to the wait queue wakeup
1255 * machanism. It is called by the stored file descriptors when they
1256 * have events to report.
1258 static int ep_poll_callback(wait_queue_t
*wait
, unsigned mode
, int sync
, void *key
)
1261 unsigned long flags
;
1262 struct epitem
*epi
= ep_item_from_wait(wait
);
1263 struct eventpoll
*ep
= epi
->ep
;
1265 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1266 current
, epi
->ffd
.file
, epi
, ep
));
1268 write_lock_irqsave(&ep
->lock
, flags
);
1271 * If the event mask does not contain any poll(2) event, we consider the
1272 * descriptor to be disabled. This condition is likely the effect of the
1273 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1274 * until the next EPOLL_CTL_MOD will be issued.
1276 if (!(epi
->event
.events
& ~EP_PRIVATE_BITS
))
1279 /* If this file is already in the ready list we exit soon */
1280 if (ep_is_linked(&epi
->rdllink
))
1283 list_add_tail(&epi
->rdllink
, &ep
->rdllist
);
1287 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1290 if (waitqueue_active(&ep
->wq
))
1291 __wake_up_locked(&ep
->wq
, TASK_UNINTERRUPTIBLE
|
1292 TASK_INTERRUPTIBLE
);
1293 if (waitqueue_active(&ep
->poll_wait
))
1297 write_unlock_irqrestore(&ep
->lock
, flags
);
1299 /* We have to call this outside the lock */
1301 ep_poll_safewake(&psw
, &ep
->poll_wait
);
1307 static int ep_eventpoll_close(struct inode
*inode
, struct file
*file
)
1309 struct eventpoll
*ep
= file
->private_data
;
1316 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: close() ep=%p\n", current
, ep
));
1321 static unsigned int ep_eventpoll_poll(struct file
*file
, poll_table
*wait
)
1323 unsigned int pollflags
= 0;
1324 unsigned long flags
;
1325 struct eventpoll
*ep
= file
->private_data
;
1327 /* Insert inside our poll wait queue */
1328 poll_wait(file
, &ep
->poll_wait
, wait
);
1330 /* Check our condition */
1331 read_lock_irqsave(&ep
->lock
, flags
);
1332 if (!list_empty(&ep
->rdllist
))
1333 pollflags
= POLLIN
| POLLRDNORM
;
1334 read_unlock_irqrestore(&ep
->lock
, flags
);
1341 * This function is called without holding the "ep->lock" since the call to
1342 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1343 * because of the way poll() is traditionally implemented in Linux.
1345 static int ep_send_events(struct eventpoll
*ep
, struct list_head
*txlist
,
1346 struct epoll_event __user
*events
, int maxevents
)
1348 int eventcnt
, error
= -EFAULT
, pwake
= 0;
1349 unsigned int revents
;
1350 unsigned long flags
;
1352 struct list_head injlist
;
1354 INIT_LIST_HEAD(&injlist
);
1357 * We can loop without lock because this is a task private list.
1358 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
1359 * Items cannot vanish during the loop because we are holding "sem" in
1362 for (eventcnt
= 0; !list_empty(txlist
) && eventcnt
< maxevents
;) {
1363 epi
= list_entry(txlist
->next
, struct epitem
, rdllink
);
1364 prefetch(epi
->rdllink
.next
);
1367 * Get the ready file event set. We can safely use the file
1368 * because we are holding the "sem" in read and this will
1369 * guarantee that both the file and the item will not vanish.
1371 revents
= epi
->ffd
.file
->f_op
->poll(epi
->ffd
.file
, NULL
);
1372 revents
&= epi
->event
.events
;
1375 * Is the event mask intersect the caller-requested one,
1376 * deliver the event to userspace. Again, we are holding
1377 * "sem" in read, so no operations coming from userspace
1378 * can change the item.
1381 if (__put_user(revents
,
1382 &events
[eventcnt
].events
) ||
1383 __put_user(epi
->event
.data
,
1384 &events
[eventcnt
].data
))
1386 if (epi
->event
.events
& EPOLLONESHOT
)
1387 epi
->event
.events
&= EP_PRIVATE_BITS
;
1392 * This is tricky. We are holding the "sem" in read, and this
1393 * means that the operations that can change the "linked" status
1394 * of the epoll item (epi->rbn and epi->rdllink), cannot touch
1395 * them. Also, since we are "linked" from a epi->rdllink POV
1396 * (the item is linked to our transmission list we just
1397 * spliced), the ep_poll_callback() cannot touch us either,
1398 * because of the check present in there. Another parallel
1399 * epoll_wait() will not get the same result set, since we
1400 * spliced the ready list before. Note that list_del() still
1401 * shows the item as linked to the test in ep_poll_callback().
1403 list_del(&epi
->rdllink
);
1404 if (!(epi
->event
.events
& EPOLLET
) &&
1405 (revents
& epi
->event
.events
))
1406 list_add_tail(&epi
->rdllink
, &injlist
);
1409 * Be sure the item is totally detached before re-init
1410 * the list_head. After INIT_LIST_HEAD() is committed,
1411 * the ep_poll_callback() can requeue the item again,
1412 * but we don't care since we are already past it.
1415 INIT_LIST_HEAD(&epi
->rdllink
);
1423 * If the re-injection list or the txlist are not empty, re-splice
1424 * them to the ready list and do proper wakeups.
1426 if (!list_empty(&injlist
) || !list_empty(txlist
)) {
1427 write_lock_irqsave(&ep
->lock
, flags
);
1429 list_splice(txlist
, &ep
->rdllist
);
1430 list_splice(&injlist
, &ep
->rdllist
);
1432 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1435 if (waitqueue_active(&ep
->wq
))
1436 __wake_up_locked(&ep
->wq
, TASK_UNINTERRUPTIBLE
|
1437 TASK_INTERRUPTIBLE
);
1438 if (waitqueue_active(&ep
->poll_wait
))
1441 write_unlock_irqrestore(&ep
->lock
, flags
);
1444 /* We have to call this outside the lock */
1446 ep_poll_safewake(&psw
, &ep
->poll_wait
);
1448 return eventcnt
== 0 ? error
: eventcnt
;
1453 * Perform the transfer of events to user space.
1455 static int ep_events_transfer(struct eventpoll
*ep
,
1456 struct epoll_event __user
*events
, int maxevents
)
1459 unsigned long flags
;
1460 struct list_head txlist
;
1462 INIT_LIST_HEAD(&txlist
);
1465 * We need to lock this because we could be hit by
1466 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1468 down_read(&ep
->sem
);
1471 * Steal the ready list, and re-init the original one to the
1474 write_lock_irqsave(&ep
->lock
, flags
);
1475 list_splice(&ep
->rdllist
, &txlist
);
1476 INIT_LIST_HEAD(&ep
->rdllist
);
1477 write_unlock_irqrestore(&ep
->lock
, flags
);
1479 /* Build result set in userspace */
1480 eventcnt
= ep_send_events(ep
, &txlist
, events
, maxevents
);
1488 static int ep_poll(struct eventpoll
*ep
, struct epoll_event __user
*events
,
1489 int maxevents
, long timeout
)
1492 unsigned long flags
;
1497 * Calculate the timeout by checking for the "infinite" value ( -1 )
1498 * and the overflow condition. The passed timeout is in milliseconds,
1499 * that why (t * HZ) / 1000.
1501 jtimeout
= (timeout
< 0 || timeout
>= EP_MAX_MSTIMEO
) ?
1502 MAX_SCHEDULE_TIMEOUT
: (timeout
* HZ
+ 999) / 1000;
1505 write_lock_irqsave(&ep
->lock
, flags
);
1508 if (list_empty(&ep
->rdllist
)) {
1510 * We don't have any available event to return to the caller.
1511 * We need to sleep here, and we will be wake up by
1512 * ep_poll_callback() when events will become available.
1514 init_waitqueue_entry(&wait
, current
);
1515 __add_wait_queue(&ep
->wq
, &wait
);
1519 * We don't want to sleep if the ep_poll_callback() sends us
1520 * a wakeup in between. That's why we set the task state
1521 * to TASK_INTERRUPTIBLE before doing the checks.
1523 set_current_state(TASK_INTERRUPTIBLE
);
1524 if (!list_empty(&ep
->rdllist
) || !jtimeout
)
1526 if (signal_pending(current
)) {
1531 write_unlock_irqrestore(&ep
->lock
, flags
);
1532 jtimeout
= schedule_timeout(jtimeout
);
1533 write_lock_irqsave(&ep
->lock
, flags
);
1535 __remove_wait_queue(&ep
->wq
, &wait
);
1537 set_current_state(TASK_RUNNING
);
1540 /* Is it worth to try to dig for events ? */
1541 eavail
= !list_empty(&ep
->rdllist
);
1543 write_unlock_irqrestore(&ep
->lock
, flags
);
1546 * Try to transfer events to user space. In case we get 0 events and
1547 * there's still timeout left over, we go trying again in search of
1550 if (!res
&& eavail
&&
1551 !(res
= ep_events_transfer(ep
, events
, maxevents
)) && jtimeout
)
1557 static int eventpollfs_delete_dentry(struct dentry
*dentry
)
1563 static struct inode
*ep_eventpoll_inode(void)
1565 int error
= -ENOMEM
;
1566 struct inode
*inode
= new_inode(eventpoll_mnt
->mnt_sb
);
1571 inode
->i_fop
= &eventpoll_fops
;
1574 * Mark the inode dirty from the very beginning,
1575 * that way it will never be moved to the dirty
1576 * list because mark_inode_dirty() will think
1577 * that it already _is_ on the dirty list.
1579 inode
->i_state
= I_DIRTY
;
1580 inode
->i_mode
= S_IRUSR
| S_IWUSR
;
1581 inode
->i_uid
= current
->fsuid
;
1582 inode
->i_gid
= current
->fsgid
;
1583 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
1587 return ERR_PTR(error
);
1591 eventpollfs_get_sb(struct file_system_type
*fs_type
, int flags
,
1592 const char *dev_name
, void *data
, struct vfsmount
*mnt
)
1594 return get_sb_pseudo(fs_type
, "eventpoll:", NULL
, EVENTPOLLFS_MAGIC
,
1599 static int __init
eventpoll_init(void)
1603 mutex_init(&epmutex
);
1605 /* Initialize the structure used to perform safe poll wait head wake ups */
1606 ep_poll_safewake_init(&psw
);
1608 /* Allocates slab cache used to allocate "struct epitem" items */
1609 epi_cache
= kmem_cache_create("eventpoll_epi", sizeof(struct epitem
),
1610 0, SLAB_HWCACHE_ALIGN
|EPI_SLAB_DEBUG
|SLAB_PANIC
,
1613 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1614 pwq_cache
= kmem_cache_create("eventpoll_pwq",
1615 sizeof(struct eppoll_entry
), 0,
1616 EPI_SLAB_DEBUG
|SLAB_PANIC
, NULL
, NULL
);
1619 * Register the virtual file system that will be the source of inodes
1620 * for the eventpoll files
1622 error
= register_filesystem(&eventpoll_fs_type
);
1626 /* Mount the above commented virtual file system */
1627 eventpoll_mnt
= kern_mount(&eventpoll_fs_type
);
1628 error
= PTR_ERR(eventpoll_mnt
);
1629 if (IS_ERR(eventpoll_mnt
))
1632 DNPRINTK(3, (KERN_INFO
"[%p] eventpoll: successfully initialized.\n",
1637 panic("eventpoll_init() failed\n");
1641 static void __exit
eventpoll_exit(void)
1643 /* Undo all operations done inside eventpoll_init() */
1644 unregister_filesystem(&eventpoll_fs_type
);
1645 mntput(eventpoll_mnt
);
1646 kmem_cache_destroy(pwq_cache
);
1647 kmem_cache_destroy(epi_cache
);
1650 module_init(eventpoll_init
);
1651 module_exit(eventpoll_exit
);
1653 MODULE_LICENSE("GPL");