blob: 4284cd31eba6dcb6f125f1fd574aa2213439b45f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/eventpoll.c ( Efficent event polling implementation )
3 * Copyright (C) 2001,...,2003 Davide Libenzi
4 *
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.
9 *
10 * Davide Libenzi <davidel@xmailserver.org>
11 *
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/kernel.h>
17#include <linux/sched.h>
18#include <linux/fs.h>
19#include <linux/file.h>
20#include <linux/signal.h>
21#include <linux/errno.h>
22#include <linux/mm.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 <asm/uaccess.h>
38#include <asm/system.h>
39#include <asm/io.h>
40#include <asm/mman.h>
41#include <asm/atomic.h>
42#include <asm/semaphore.h>
43
44
45/*
46 * LOCKING:
47 * There are three level of locking required by epoll :
48 *
49 * 1) epsem (semaphore)
50 * 2) ep->sem (rw_semaphore)
51 * 3) ep->lock (rw_lock)
52 *
53 * The acquire order is the one listed above, from 1 to 3.
54 * We need a spinlock (ep->lock) because we manipulate objects
55 * from inside the poll callback, that might be triggered from
56 * a wake_up() that in turn might be called from IRQ context.
57 * So we can't sleep inside the poll callback and hence we need
58 * a spinlock. During the event transfer loop (from kernel to
59 * user space) we could end up sleeping due a copy_to_user(), so
60 * we need a lock that will allow us to sleep. This lock is a
61 * read-write semaphore (ep->sem). It is acquired on read during
62 * the event transfer loop and in write during epoll_ctl(EPOLL_CTL_DEL)
63 * and during eventpoll_release_file(). Then we also need a global
64 * semaphore to serialize eventpoll_release_file() and ep_free().
65 * This semaphore is acquired by ep_free() during the epoll file
66 * cleanup path and it is also acquired by eventpoll_release_file()
67 * if a file has been pushed inside an epoll set and it is then
68 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
69 * It is possible to drop the "ep->sem" and to use the global
70 * semaphore "epsem" (together with "ep->lock") to have it working,
71 * but having "ep->sem" will make the interface more scalable.
72 * Events that require holding "epsem" are very rare, while for
73 * normal operations the epoll private "ep->sem" will guarantee
74 * a greater scalability.
75 */
76
77
78#define EVENTPOLLFS_MAGIC 0x03111965 /* My birthday should work for this :) */
79
80#define DEBUG_EPOLL 0
81
82#if DEBUG_EPOLL > 0
83#define DPRINTK(x) printk x
84#define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
85#else /* #if DEBUG_EPOLL > 0 */
86#define DPRINTK(x) (void) 0
87#define DNPRINTK(n, x) (void) 0
88#endif /* #if DEBUG_EPOLL > 0 */
89
90#define DEBUG_EPI 0
91
92#if DEBUG_EPI != 0
93#define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
94#else /* #if DEBUG_EPI != 0 */
95#define EPI_SLAB_DEBUG 0
96#endif /* #if DEBUG_EPI != 0 */
97
98/* Epoll private bits inside the event mask */
99#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
100
101/* Maximum number of poll wake up nests we are allowing */
102#define EP_MAX_POLLWAKE_NESTS 4
103
Davide Libenzie3306dd2005-09-27 21:45:33 -0700104/* Maximum msec timeout value storeable in a long int */
105#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
106
107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108struct epoll_filefd {
109 struct file *file;
110 int fd;
111};
112
113/*
114 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
115 * It is used to keep track on all tasks that are currently inside the wake_up() code
116 * to 1) short-circuit the one coming from the same task and same wait queue head
117 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
118 * 3) let go the ones coming from other tasks.
119 */
120struct wake_task_node {
121 struct list_head llink;
122 task_t *task;
123 wait_queue_head_t *wq;
124};
125
126/*
127 * This is used to implement the safe poll wake up avoiding to reenter
128 * the poll callback from inside wake_up().
129 */
130struct poll_safewake {
131 struct list_head wake_task_list;
132 spinlock_t lock;
133};
134
135/*
136 * This structure is stored inside the "private_data" member of the file
137 * structure and rapresent the main data sructure for the eventpoll
138 * interface.
139 */
140struct eventpoll {
141 /* Protect the this structure access */
142 rwlock_t lock;
143
144 /*
145 * This semaphore is used to ensure that files are not removed
146 * while epoll is using them. This is read-held during the event
147 * collection loop and it is write-held during the file cleanup
148 * path, the epoll file exit code and the ctl operations.
149 */
150 struct rw_semaphore sem;
151
152 /* Wait queue used by sys_epoll_wait() */
153 wait_queue_head_t wq;
154
155 /* Wait queue used by file->poll() */
156 wait_queue_head_t poll_wait;
157
158 /* List of ready file descriptors */
159 struct list_head rdllist;
160
161 /* RB-Tree root used to store monitored fd structs */
162 struct rb_root rbr;
163};
164
165/* Wait structure used by the poll hooks */
166struct eppoll_entry {
167 /* List header used to link this structure to the "struct epitem" */
168 struct list_head llink;
169
170 /* The "base" pointer is set to the container "struct epitem" */
171 void *base;
172
173 /*
174 * Wait queue item that will be linked to the target file wait
175 * queue head.
176 */
177 wait_queue_t wait;
178
179 /* The wait queue head that linked the "wait" wait queue item */
180 wait_queue_head_t *whead;
181};
182
183/*
184 * Each file descriptor added to the eventpoll interface will
185 * have an entry of this type linked to the hash.
186 */
187struct epitem {
188 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
189 struct rb_node rbn;
190
191 /* List header used to link this structure to the eventpoll ready list */
192 struct list_head rdllink;
193
194 /* The file descriptor information this item refers to */
195 struct epoll_filefd ffd;
196
197 /* Number of active wait queue attached to poll operations */
198 int nwait;
199
200 /* List containing poll wait queues */
201 struct list_head pwqlist;
202
203 /* The "container" of this item */
204 struct eventpoll *ep;
205
206 /* The structure that describe the interested events and the source fd */
207 struct epoll_event event;
208
209 /*
210 * Used to keep track of the usage count of the structure. This avoids
211 * that the structure will desappear from underneath our processing.
212 */
213 atomic_t usecnt;
214
215 /* List header used to link this item to the "struct file" items list */
216 struct list_head fllink;
217
218 /* List header used to link the item to the transfer list */
219 struct list_head txlink;
220
221 /*
222 * This is used during the collection/transfer of events to userspace
223 * to pin items empty events set.
224 */
225 unsigned int revents;
226};
227
228/* Wrapper struct used by poll queueing */
229struct ep_pqueue {
230 poll_table pt;
231 struct epitem *epi;
232};
233
234
235
236static void ep_poll_safewake_init(struct poll_safewake *psw);
237static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
Davide Libenzi53d2be72005-09-16 19:28:06 -0700238static int ep_getfd(int *efd, struct inode **einode, struct file **efile,
239 struct eventpoll *ep);
240static int ep_alloc(struct eventpoll **pep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241static void ep_free(struct eventpoll *ep);
242static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
243static void ep_use_epitem(struct epitem *epi);
244static void ep_release_epitem(struct epitem *epi);
245static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
246 poll_table *pt);
247static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi);
248static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
249 struct file *tfile, int fd);
250static int ep_modify(struct eventpoll *ep, struct epitem *epi,
251 struct epoll_event *event);
252static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
253static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
254static int ep_remove(struct eventpoll *ep, struct epitem *epi);
255static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
256static int ep_eventpoll_close(struct inode *inode, struct file *file);
257static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
258static int ep_collect_ready_items(struct eventpoll *ep,
259 struct list_head *txlist, int maxevents);
260static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
261 struct epoll_event __user *events);
262static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist);
263static int ep_events_transfer(struct eventpoll *ep,
264 struct epoll_event __user *events,
265 int maxevents);
266static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
267 int maxevents, long timeout);
268static int eventpollfs_delete_dentry(struct dentry *dentry);
269static struct inode *ep_eventpoll_inode(void);
270static struct super_block *eventpollfs_get_sb(struct file_system_type *fs_type,
271 int flags, const char *dev_name,
272 void *data);
273
274/*
275 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
276 */
Adrian Bunk75c96f82005-05-05 16:16:09 -0700277static struct semaphore epsem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278
279/* Safe wake up implementation */
280static struct poll_safewake psw;
281
282/* Slab cache used to allocate "struct epitem" */
283static kmem_cache_t *epi_cache;
284
285/* Slab cache used to allocate "struct eppoll_entry" */
286static kmem_cache_t *pwq_cache;
287
288/* Virtual fs used to allocate inodes for eventpoll files */
289static struct vfsmount *eventpoll_mnt;
290
291/* File callbacks that implement the eventpoll file behaviour */
292static struct file_operations eventpoll_fops = {
293 .release = ep_eventpoll_close,
294 .poll = ep_eventpoll_poll
295};
296
297/*
298 * This is used to register the virtual file system from where
299 * eventpoll inodes are allocated.
300 */
301static struct file_system_type eventpoll_fs_type = {
302 .name = "eventpollfs",
303 .get_sb = eventpollfs_get_sb,
304 .kill_sb = kill_anon_super,
305};
306
307/* Very basic directory entry operations for the eventpoll virtual file system */
308static struct dentry_operations eventpollfs_dentry_operations = {
309 .d_delete = eventpollfs_delete_dentry,
310};
311
312
313
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700314/* Fast test to see if the file is an evenpoll file */
315static inline int is_file_epoll(struct file *f)
316{
317 return f->f_op == &eventpoll_fops;
318}
319
320/* Setup the structure that is used as key for the rb-tree */
321static inline void ep_set_ffd(struct epoll_filefd *ffd,
322 struct file *file, int fd)
323{
324 ffd->file = file;
325 ffd->fd = fd;
326}
327
328/* Compare rb-tree keys */
329static inline int ep_cmp_ffd(struct epoll_filefd *p1,
330 struct epoll_filefd *p2)
331{
332 return (p1->file > p2->file ? +1:
333 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
334}
335
336/* Special initialization for the rb-tree node to detect linkage */
337static inline void ep_rb_initnode(struct rb_node *n)
338{
339 n->rb_parent = n;
340}
341
342/* Removes a node from the rb-tree and marks it for a fast is-linked check */
343static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
344{
345 rb_erase(n, r);
346 n->rb_parent = n;
347}
348
349/* Fast check to verify that the item is linked to the main rb-tree */
350static inline int ep_rb_linked(struct rb_node *n)
351{
352 return n->rb_parent != n;
353}
354
355/*
356 * Remove the item from the list and perform its initialization.
357 * This is useful for us because we can test if the item is linked
358 * using "ep_is_linked(p)".
359 */
360static inline void ep_list_del(struct list_head *p)
361{
362 list_del(p);
363 INIT_LIST_HEAD(p);
364}
365
366/* Tells us if the item is currently linked */
367static inline int ep_is_linked(struct list_head *p)
368{
369 return !list_empty(p);
370}
371
372/* Get the "struct epitem" from a wait queue pointer */
373static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
374{
375 return container_of(p, struct eppoll_entry, wait)->base;
376}
377
378/* Get the "struct epitem" from an epoll queue wrapper */
379static inline struct epitem * ep_item_from_epqueue(poll_table *p)
380{
381 return container_of(p, struct ep_pqueue, pt)->epi;
382}
383
384/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
385static inline int ep_op_hash_event(int op)
386{
387 return op != EPOLL_CTL_DEL;
388}
389
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390/* Initialize the poll safe wake up structure */
391static void ep_poll_safewake_init(struct poll_safewake *psw)
392{
393
394 INIT_LIST_HEAD(&psw->wake_task_list);
395 spin_lock_init(&psw->lock);
396}
397
398
399/*
400 * Perform a safe wake up of the poll wait list. The problem is that
401 * with the new callback'd wake up system, it is possible that the
402 * poll callback is reentered from inside the call to wake_up() done
403 * on the poll wait queue head. The rule is that we cannot reenter the
404 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
405 * and we cannot reenter the same wait queue head at all. This will
406 * enable to have a hierarchy of epoll file descriptor of no more than
407 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
408 * because this one gets called by the poll callback, that in turn is called
409 * from inside a wake_up(), that might be called from irq context.
410 */
411static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
412{
413 int wake_nests = 0;
414 unsigned long flags;
415 task_t *this_task = current;
416 struct list_head *lsthead = &psw->wake_task_list, *lnk;
417 struct wake_task_node *tncur;
418 struct wake_task_node tnode;
419
420 spin_lock_irqsave(&psw->lock, flags);
421
422 /* Try to see if the current task is already inside this wakeup call */
423 list_for_each(lnk, lsthead) {
424 tncur = list_entry(lnk, struct wake_task_node, llink);
425
426 if (tncur->wq == wq ||
427 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
428 /*
429 * Ops ... loop detected or maximum nest level reached.
430 * We abort this wake by breaking the cycle itself.
431 */
432 spin_unlock_irqrestore(&psw->lock, flags);
433 return;
434 }
435 }
436
437 /* Add the current task to the list */
438 tnode.task = this_task;
439 tnode.wq = wq;
440 list_add(&tnode.llink, lsthead);
441
442 spin_unlock_irqrestore(&psw->lock, flags);
443
444 /* Do really wake up now */
445 wake_up(wq);
446
447 /* Remove the current task from the list */
448 spin_lock_irqsave(&psw->lock, flags);
449 list_del(&tnode.llink);
450 spin_unlock_irqrestore(&psw->lock, flags);
451}
452
453
454/* Used to initialize the epoll bits inside the "struct file" */
455void eventpoll_init_file(struct file *file)
456{
457
458 INIT_LIST_HEAD(&file->f_ep_links);
459 spin_lock_init(&file->f_ep_lock);
460}
461
462
463/*
464 * This is called from eventpoll_release() to unlink files from the eventpoll
465 * interface. We need to have this facility to cleanup correctly files that are
466 * closed without being removed from the eventpoll interface.
467 */
468void eventpoll_release_file(struct file *file)
469{
470 struct list_head *lsthead = &file->f_ep_links;
471 struct eventpoll *ep;
472 struct epitem *epi;
473
474 /*
475 * We don't want to get "file->f_ep_lock" because it is not
476 * necessary. It is not necessary because we're in the "struct file"
477 * cleanup path, and this means that noone is using this file anymore.
478 * The only hit might come from ep_free() but by holding the semaphore
479 * will correctly serialize the operation. We do need to acquire
480 * "ep->sem" after "epsem" because ep_remove() requires it when called
481 * from anywhere but ep_free().
482 */
483 down(&epsem);
484
485 while (!list_empty(lsthead)) {
486 epi = list_entry(lsthead->next, struct epitem, fllink);
487
488 ep = epi->ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700489 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 down_write(&ep->sem);
491 ep_remove(ep, epi);
492 up_write(&ep->sem);
493 }
494
495 up(&epsem);
496}
497
498
499/*
500 * It opens an eventpoll file descriptor by suggesting a storage of "size"
501 * file descriptors. The size parameter is just an hint about how to size
502 * data structures. It won't prevent the user to store more than "size"
503 * file descriptors inside the epoll interface. It is the kernel part of
504 * the userspace epoll_create(2).
505 */
506asmlinkage long sys_epoll_create(int size)
507{
508 int error, fd;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700509 struct eventpoll *ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 struct inode *inode;
511 struct file *file;
512
513 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
514 current, size));
515
Davide Libenzi53d2be72005-09-16 19:28:06 -0700516 /*
517 * Sanity check on the size parameter, and create the internal data
518 * structure ( "struct eventpoll" ).
519 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 error = -EINVAL;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700521 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 goto eexit_1;
523
524 /*
525 * Creates all the items needed to setup an eventpoll file. That is,
526 * a file structure, and inode and a free file descriptor.
527 */
Davide Libenzi53d2be72005-09-16 19:28:06 -0700528 error = ep_getfd(&fd, &inode, &file, ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 if (error)
530 goto eexit_2;
531
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
533 current, size, fd));
534
535 return fd;
536
537eexit_2:
Davide Libenzi53d2be72005-09-16 19:28:06 -0700538 ep_free(ep);
539 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540eexit_1:
541 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
542 current, size, error));
543 return error;
544}
545
546
547/*
548 * The following function implements the controller interface for
549 * the eventpoll file that enables the insertion/removal/change of
550 * file descriptors inside the interest set. It represents
551 * the kernel part of the user space epoll_ctl(2).
552 */
553asmlinkage long
554sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
555{
556 int error;
557 struct file *file, *tfile;
558 struct eventpoll *ep;
559 struct epitem *epi;
560 struct epoll_event epds;
561
562 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
563 current, epfd, op, fd, event));
564
565 error = -EFAULT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700566 if (ep_op_hash_event(op) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 copy_from_user(&epds, event, sizeof(struct epoll_event)))
568 goto eexit_1;
569
570 /* Get the "struct file *" for the eventpoll file */
571 error = -EBADF;
572 file = fget(epfd);
573 if (!file)
574 goto eexit_1;
575
576 /* Get the "struct file *" for the target file */
577 tfile = fget(fd);
578 if (!tfile)
579 goto eexit_2;
580
581 /* The target file descriptor must support poll */
582 error = -EPERM;
583 if (!tfile->f_op || !tfile->f_op->poll)
584 goto eexit_3;
585
586 /*
587 * We have to check that the file structure underneath the file descriptor
588 * the user passed to us _is_ an eventpoll file. And also we do not permit
589 * adding an epoll file descriptor inside itself.
590 */
591 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700592 if (file == tfile || !is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 goto eexit_3;
594
595 /*
596 * At this point it is safe to assume that the "private_data" contains
597 * our own data structure.
598 */
599 ep = file->private_data;
600
601 down_write(&ep->sem);
602
603 /* Try to lookup the file inside our hash table */
604 epi = ep_find(ep, tfile, fd);
605
606 error = -EINVAL;
607 switch (op) {
608 case EPOLL_CTL_ADD:
609 if (!epi) {
610 epds.events |= POLLERR | POLLHUP;
611
612 error = ep_insert(ep, &epds, tfile, fd);
613 } else
614 error = -EEXIST;
615 break;
616 case EPOLL_CTL_DEL:
617 if (epi)
618 error = ep_remove(ep, epi);
619 else
620 error = -ENOENT;
621 break;
622 case EPOLL_CTL_MOD:
623 if (epi) {
624 epds.events |= POLLERR | POLLHUP;
625 error = ep_modify(ep, epi, &epds);
626 } else
627 error = -ENOENT;
628 break;
629 }
630
631 /*
632 * The function ep_find() increments the usage count of the structure
633 * so, if this is not NULL, we need to release it.
634 */
635 if (epi)
636 ep_release_epitem(epi);
637
638 up_write(&ep->sem);
639
640eexit_3:
641 fput(tfile);
642eexit_2:
643 fput(file);
644eexit_1:
645 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
646 current, epfd, op, fd, event, error));
647
648 return error;
649}
650
651#define MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
652
653/*
654 * Implement the event wait interface for the eventpoll file. It is the kernel
655 * part of the user space epoll_wait(2).
656 */
657asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
658 int maxevents, int timeout)
659{
660 int error;
661 struct file *file;
662 struct eventpoll *ep;
663
664 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
665 current, epfd, events, maxevents, timeout));
666
667 /* The maximum number of event must be greater than zero */
668 if (maxevents <= 0 || maxevents > MAX_EVENTS)
669 return -EINVAL;
670
671 /* Verify that the area passed by the user is writeable */
672 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
673 error = -EFAULT;
674 goto eexit_1;
675 }
676
677 /* Get the "struct file *" for the eventpoll file */
678 error = -EBADF;
679 file = fget(epfd);
680 if (!file)
681 goto eexit_1;
682
683 /*
684 * We have to check that the file structure underneath the fd
685 * the user passed to us _is_ an eventpoll file.
686 */
687 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700688 if (!is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 goto eexit_2;
690
691 /*
692 * At this point it is safe to assume that the "private_data" contains
693 * our own data structure.
694 */
695 ep = file->private_data;
696
697 /* Time to fish for events ... */
698 error = ep_poll(ep, events, maxevents, timeout);
699
700eexit_2:
701 fput(file);
702eexit_1:
703 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
704 current, epfd, events, maxevents, timeout, error));
705
706 return error;
707}
708
709
710/*
711 * Creates the file descriptor to be used by the epoll interface.
712 */
Davide Libenzi53d2be72005-09-16 19:28:06 -0700713static int ep_getfd(int *efd, struct inode **einode, struct file **efile,
714 struct eventpoll *ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715{
716 struct qstr this;
717 char name[32];
718 struct dentry *dentry;
719 struct inode *inode;
720 struct file *file;
721 int error, fd;
722
723 /* Get an ready to use file */
724 error = -ENFILE;
725 file = get_empty_filp();
726 if (!file)
727 goto eexit_1;
728
729 /* Allocates an inode from the eventpoll file system */
730 inode = ep_eventpoll_inode();
731 error = PTR_ERR(inode);
732 if (IS_ERR(inode))
733 goto eexit_2;
734
735 /* Allocates a free descriptor to plug the file onto */
736 error = get_unused_fd();
737 if (error < 0)
738 goto eexit_3;
739 fd = error;
740
741 /*
742 * Link the inode to a directory entry by creating a unique name
743 * using the inode number.
744 */
745 error = -ENOMEM;
746 sprintf(name, "[%lu]", inode->i_ino);
747 this.name = name;
748 this.len = strlen(name);
749 this.hash = inode->i_ino;
750 dentry = d_alloc(eventpoll_mnt->mnt_sb->s_root, &this);
751 if (!dentry)
752 goto eexit_4;
753 dentry->d_op = &eventpollfs_dentry_operations;
754 d_add(dentry, inode);
755 file->f_vfsmnt = mntget(eventpoll_mnt);
756 file->f_dentry = dentry;
757 file->f_mapping = inode->i_mapping;
758
759 file->f_pos = 0;
760 file->f_flags = O_RDONLY;
761 file->f_op = &eventpoll_fops;
762 file->f_mode = FMODE_READ;
763 file->f_version = 0;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700764 file->private_data = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765
766 /* Install the new setup file into the allocated fd. */
767 fd_install(fd, file);
768
769 *efd = fd;
770 *einode = inode;
771 *efile = file;
772 return 0;
773
774eexit_4:
775 put_unused_fd(fd);
776eexit_3:
777 iput(inode);
778eexit_2:
779 put_filp(file);
780eexit_1:
781 return error;
782}
783
784
Davide Libenzi53d2be72005-09-16 19:28:06 -0700785static int ep_alloc(struct eventpoll **pep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786{
Davide Libenzi53d2be72005-09-16 19:28:06 -0700787 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
Davide Libenzi53d2be72005-09-16 19:28:06 -0700789 if (!ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 return -ENOMEM;
791
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 rwlock_init(&ep->lock);
793 init_rwsem(&ep->sem);
794 init_waitqueue_head(&ep->wq);
795 init_waitqueue_head(&ep->poll_wait);
796 INIT_LIST_HEAD(&ep->rdllist);
797 ep->rbr = RB_ROOT;
798
Davide Libenzi53d2be72005-09-16 19:28:06 -0700799 *pep = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
Davide Libenzi53d2be72005-09-16 19:28:06 -0700801 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 current, ep));
803 return 0;
804}
805
806
807static void ep_free(struct eventpoll *ep)
808{
809 struct rb_node *rbp;
810 struct epitem *epi;
811
812 /* We need to release all tasks waiting for these file */
813 if (waitqueue_active(&ep->poll_wait))
814 ep_poll_safewake(&psw, &ep->poll_wait);
815
816 /*
817 * We need to lock this because we could be hit by
818 * eventpoll_release_file() while we're freeing the "struct eventpoll".
819 * We do not need to hold "ep->sem" here because the epoll file
820 * is on the way to be removed and no one has references to it
821 * anymore. The only hit might come from eventpoll_release_file() but
822 * holding "epsem" is sufficent here.
823 */
824 down(&epsem);
825
826 /*
827 * Walks through the whole tree by unregistering poll callbacks.
828 */
829 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
830 epi = rb_entry(rbp, struct epitem, rbn);
831
832 ep_unregister_pollwait(ep, epi);
833 }
834
835 /*
836 * Walks through the whole hash by freeing each "struct epitem". At this
837 * point we are sure no poll callbacks will be lingering around, and also by
838 * write-holding "sem" we can be sure that no file cleanup code will hit
839 * us during this operation. So we can avoid the lock on "ep->lock".
840 */
841 while ((rbp = rb_first(&ep->rbr)) != 0) {
842 epi = rb_entry(rbp, struct epitem, rbn);
843 ep_remove(ep, epi);
844 }
845
846 up(&epsem);
847}
848
849
850/*
851 * Search the file inside the eventpoll hash. It add usage count to
852 * the returned item, so the caller must call ep_release_epitem()
853 * after finished using the "struct epitem".
854 */
855static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
856{
857 int kcmp;
858 unsigned long flags;
859 struct rb_node *rbp;
860 struct epitem *epi, *epir = NULL;
861 struct epoll_filefd ffd;
862
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700863 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 read_lock_irqsave(&ep->lock, flags);
865 for (rbp = ep->rbr.rb_node; rbp; ) {
866 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700867 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 if (kcmp > 0)
869 rbp = rbp->rb_right;
870 else if (kcmp < 0)
871 rbp = rbp->rb_left;
872 else {
873 ep_use_epitem(epi);
874 epir = epi;
875 break;
876 }
877 }
878 read_unlock_irqrestore(&ep->lock, flags);
879
880 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
881 current, file, epir));
882
883 return epir;
884}
885
886
887/*
888 * Increment the usage count of the "struct epitem" making it sure
889 * that the user will have a valid pointer to reference.
890 */
891static void ep_use_epitem(struct epitem *epi)
892{
893
894 atomic_inc(&epi->usecnt);
895}
896
897
898/*
899 * Decrement ( release ) the usage count by signaling that the user
900 * has finished using the structure. It might lead to freeing the
901 * structure itself if the count goes to zero.
902 */
903static void ep_release_epitem(struct epitem *epi)
904{
905
906 if (atomic_dec_and_test(&epi->usecnt))
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700907 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908}
909
910
911/*
912 * This is the callback that is used to add our wait queue to the
913 * target file wakeup lists.
914 */
915static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
916 poll_table *pt)
917{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700918 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 struct eppoll_entry *pwq;
920
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700921 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, SLAB_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
923 pwq->whead = whead;
924 pwq->base = epi;
925 add_wait_queue(whead, &pwq->wait);
926 list_add_tail(&pwq->llink, &epi->pwqlist);
927 epi->nwait++;
928 } else {
929 /* We have to signal that an error occurred */
930 epi->nwait = -1;
931 }
932}
933
934
935static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
936{
937 int kcmp;
938 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
939 struct epitem *epic;
940
941 while (*p) {
942 parent = *p;
943 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700944 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 if (kcmp > 0)
946 p = &parent->rb_right;
947 else
948 p = &parent->rb_left;
949 }
950 rb_link_node(&epi->rbn, parent, p);
951 rb_insert_color(&epi->rbn, &ep->rbr);
952}
953
954
955static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
956 struct file *tfile, int fd)
957{
958 int error, revents, pwake = 0;
959 unsigned long flags;
960 struct epitem *epi;
961 struct ep_pqueue epq;
962
963 error = -ENOMEM;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700964 if (!(epi = kmem_cache_alloc(epi_cache, SLAB_KERNEL)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 goto eexit_1;
966
967 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700968 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 INIT_LIST_HEAD(&epi->rdllink);
970 INIT_LIST_HEAD(&epi->fllink);
971 INIT_LIST_HEAD(&epi->txlink);
972 INIT_LIST_HEAD(&epi->pwqlist);
973 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700974 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 epi->event = *event;
976 atomic_set(&epi->usecnt, 1);
977 epi->nwait = 0;
978
979 /* Initialize the poll table using the queue callback */
980 epq.epi = epi;
981 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
982
983 /*
984 * Attach the item to the poll hooks and get current event bits.
985 * We can safely use the file* here because its usage count has
986 * been increased by the caller of this function.
987 */
988 revents = tfile->f_op->poll(tfile, &epq.pt);
989
990 /*
991 * We have to check if something went wrong during the poll wait queue
992 * install process. Namely an allocation for a wait queue failed due
993 * high memory pressure.
994 */
995 if (epi->nwait < 0)
996 goto eexit_2;
997
998 /* Add the current item to the list of active epoll hook for this file */
999 spin_lock(&tfile->f_ep_lock);
1000 list_add_tail(&epi->fllink, &tfile->f_ep_links);
1001 spin_unlock(&tfile->f_ep_lock);
1002
1003 /* We have to drop the new item inside our item list to keep track of it */
1004 write_lock_irqsave(&ep->lock, flags);
1005
1006 /* Add the current item to the rb-tree */
1007 ep_rbtree_insert(ep, epi);
1008
1009 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001010 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 list_add_tail(&epi->rdllink, &ep->rdllist);
1012
1013 /* Notify waiting tasks that events are available */
1014 if (waitqueue_active(&ep->wq))
1015 wake_up(&ep->wq);
1016 if (waitqueue_active(&ep->poll_wait))
1017 pwake++;
1018 }
1019
1020 write_unlock_irqrestore(&ep->lock, flags);
1021
1022 /* We have to call this outside the lock */
1023 if (pwake)
1024 ep_poll_safewake(&psw, &ep->poll_wait);
1025
1026 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1027 current, ep, tfile, fd));
1028
1029 return 0;
1030
1031eexit_2:
1032 ep_unregister_pollwait(ep, epi);
1033
1034 /*
1035 * We need to do this because an event could have been arrived on some
1036 * allocated wait queue.
1037 */
1038 write_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001039 if (ep_is_linked(&epi->rdllink))
1040 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 write_unlock_irqrestore(&ep->lock, flags);
1042
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001043 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044eexit_1:
1045 return error;
1046}
1047
1048
1049/*
1050 * Modify the interest event mask by dropping an event if the new mask
1051 * has a match in the current file status.
1052 */
1053static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1054{
1055 int pwake = 0;
1056 unsigned int revents;
1057 unsigned long flags;
1058
1059 /*
1060 * Set the new event interest mask before calling f_op->poll(), otherwise
1061 * a potential race might occur. In fact if we do this operation inside
1062 * the lock, an event might happen between the f_op->poll() call and the
1063 * new event set registering.
1064 */
1065 epi->event.events = event->events;
1066
1067 /*
1068 * Get current event bits. We can safely use the file* here because
1069 * its usage count has been increased by the caller of this function.
1070 */
1071 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1072
1073 write_lock_irqsave(&ep->lock, flags);
1074
1075 /* Copy the data member from inside the lock */
1076 epi->event.data = event->data;
1077
1078 /*
1079 * If the item is not linked to the hash it means that it's on its
1080 * way toward the removal. Do nothing in this case.
1081 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001082 if (ep_rb_linked(&epi->rbn)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 /*
1084 * If the item is "hot" and it is not registered inside the ready
1085 * list, push it inside. If the item is not "hot" and it is currently
1086 * registered inside the ready list, unlink it.
1087 */
1088 if (revents & event->events) {
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001089 if (!ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 list_add_tail(&epi->rdllink, &ep->rdllist);
1091
1092 /* Notify waiting tasks that events are available */
1093 if (waitqueue_active(&ep->wq))
1094 wake_up(&ep->wq);
1095 if (waitqueue_active(&ep->poll_wait))
1096 pwake++;
1097 }
1098 }
1099 }
1100
1101 write_unlock_irqrestore(&ep->lock, flags);
1102
1103 /* We have to call this outside the lock */
1104 if (pwake)
1105 ep_poll_safewake(&psw, &ep->poll_wait);
1106
1107 return 0;
1108}
1109
1110
1111/*
1112 * This function unregister poll callbacks from the associated file descriptor.
1113 * Since this must be called without holding "ep->lock" the atomic exchange trick
1114 * will protect us from multiple unregister.
1115 */
1116static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1117{
1118 int nwait;
1119 struct list_head *lsthead = &epi->pwqlist;
1120 struct eppoll_entry *pwq;
1121
1122 /* This is called without locks, so we need the atomic exchange */
1123 nwait = xchg(&epi->nwait, 0);
1124
1125 if (nwait) {
1126 while (!list_empty(lsthead)) {
1127 pwq = list_entry(lsthead->next, struct eppoll_entry, llink);
1128
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001129 ep_list_del(&pwq->llink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 remove_wait_queue(pwq->whead, &pwq->wait);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001131 kmem_cache_free(pwq_cache, pwq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 }
1133 }
1134}
1135
1136
1137/*
1138 * Unlink the "struct epitem" from all places it might have been hooked up.
1139 * This function must be called with write IRQ lock on "ep->lock".
1140 */
1141static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1142{
1143 int error;
1144
1145 /*
1146 * It can happen that this one is called for an item already unlinked.
1147 * The check protect us from doing a double unlink ( crash ).
1148 */
1149 error = -ENOENT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001150 if (!ep_rb_linked(&epi->rbn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 goto eexit_1;
1152
1153 /*
1154 * Clear the event mask for the unlinked item. This will avoid item
1155 * notifications to be sent after the unlink operation from inside
1156 * the kernel->userspace event transfer loop.
1157 */
1158 epi->event.events = 0;
1159
1160 /*
1161 * At this point is safe to do the job, unlink the item from our rb-tree.
1162 * This operation togheter with the above check closes the door to
1163 * double unlinks.
1164 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001165 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166
1167 /*
1168 * If the item we are going to remove is inside the ready file descriptors
1169 * we want to remove it from this list to avoid stale events.
1170 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001171 if (ep_is_linked(&epi->rdllink))
1172 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173
1174 error = 0;
1175eexit_1:
1176
1177 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1178 current, ep, epi->file, error));
1179
1180 return error;
1181}
1182
1183
1184/*
1185 * Removes a "struct epitem" from the eventpoll hash and deallocates
1186 * all the associated resources.
1187 */
1188static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1189{
1190 int error;
1191 unsigned long flags;
1192 struct file *file = epi->ffd.file;
1193
1194 /*
1195 * Removes poll wait queue hooks. We _have_ to do this without holding
1196 * the "ep->lock" otherwise a deadlock might occur. This because of the
1197 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1198 * queue head lock when unregistering the wait queue. The wakeup callback
1199 * will run by holding the wait queue head lock and will call our callback
1200 * that will try to get "ep->lock".
1201 */
1202 ep_unregister_pollwait(ep, epi);
1203
1204 /* Remove the current item from the list of epoll hooks */
1205 spin_lock(&file->f_ep_lock);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001206 if (ep_is_linked(&epi->fllink))
1207 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 spin_unlock(&file->f_ep_lock);
1209
1210 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1211 write_lock_irqsave(&ep->lock, flags);
1212
1213 /* Really unlink the item from the hash */
1214 error = ep_unlink(ep, epi);
1215
1216 write_unlock_irqrestore(&ep->lock, flags);
1217
1218 if (error)
1219 goto eexit_1;
1220
1221 /* At this point it is safe to free the eventpoll item */
1222 ep_release_epitem(epi);
1223
1224 error = 0;
1225eexit_1:
1226 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1227 current, ep, file, error));
1228
1229 return error;
1230}
1231
1232
1233/*
1234 * This is the callback that is passed to the wait queue wakeup
1235 * machanism. It is called by the stored file descriptors when they
1236 * have events to report.
1237 */
1238static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1239{
1240 int pwake = 0;
1241 unsigned long flags;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001242 struct epitem *epi = ep_item_from_wait(wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 struct eventpoll *ep = epi->ep;
1244
1245 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1246 current, epi->file, epi, ep));
1247
1248 write_lock_irqsave(&ep->lock, flags);
1249
1250 /*
1251 * If the event mask does not contain any poll(2) event, we consider the
1252 * descriptor to be disabled. This condition is likely the effect of the
1253 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1254 * until the next EPOLL_CTL_MOD will be issued.
1255 */
1256 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1257 goto is_disabled;
1258
1259 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001260 if (ep_is_linked(&epi->rdllink))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 goto is_linked;
1262
1263 list_add_tail(&epi->rdllink, &ep->rdllist);
1264
1265is_linked:
1266 /*
1267 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1268 * wait list.
1269 */
1270 if (waitqueue_active(&ep->wq))
1271 wake_up(&ep->wq);
1272 if (waitqueue_active(&ep->poll_wait))
1273 pwake++;
1274
1275is_disabled:
1276 write_unlock_irqrestore(&ep->lock, flags);
1277
1278 /* We have to call this outside the lock */
1279 if (pwake)
1280 ep_poll_safewake(&psw, &ep->poll_wait);
1281
1282 return 1;
1283}
1284
1285
1286static int ep_eventpoll_close(struct inode *inode, struct file *file)
1287{
1288 struct eventpoll *ep = file->private_data;
1289
1290 if (ep) {
1291 ep_free(ep);
1292 kfree(ep);
1293 }
1294
1295 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1296 return 0;
1297}
1298
1299
1300static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1301{
1302 unsigned int pollflags = 0;
1303 unsigned long flags;
1304 struct eventpoll *ep = file->private_data;
1305
1306 /* Insert inside our poll wait queue */
1307 poll_wait(file, &ep->poll_wait, wait);
1308
1309 /* Check our condition */
1310 read_lock_irqsave(&ep->lock, flags);
1311 if (!list_empty(&ep->rdllist))
1312 pollflags = POLLIN | POLLRDNORM;
1313 read_unlock_irqrestore(&ep->lock, flags);
1314
1315 return pollflags;
1316}
1317
1318
1319/*
1320 * Since we have to release the lock during the __copy_to_user() operation and
1321 * during the f_op->poll() call, we try to collect the maximum number of items
1322 * by reducing the irqlock/irqunlock switching rate.
1323 */
1324static int ep_collect_ready_items(struct eventpoll *ep, struct list_head *txlist, int maxevents)
1325{
1326 int nepi;
1327 unsigned long flags;
1328 struct list_head *lsthead = &ep->rdllist, *lnk;
1329 struct epitem *epi;
1330
1331 write_lock_irqsave(&ep->lock, flags);
1332
1333 for (nepi = 0, lnk = lsthead->next; lnk != lsthead && nepi < maxevents;) {
1334 epi = list_entry(lnk, struct epitem, rdllink);
1335
1336 lnk = lnk->next;
1337
1338 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001339 if (!ep_is_linked(&epi->txlink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 /*
1341 * This is initialized in this way so that the default
1342 * behaviour of the reinjecting code will be to push back
1343 * the item inside the ready list.
1344 */
1345 epi->revents = epi->event.events;
1346
1347 /* Link the ready item into the transfer list */
1348 list_add(&epi->txlink, txlist);
1349 nepi++;
1350
1351 /*
1352 * Unlink the item from the ready list.
1353 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001354 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 }
1356 }
1357
1358 write_unlock_irqrestore(&ep->lock, flags);
1359
1360 return nepi;
1361}
1362
1363
1364/*
1365 * This function is called without holding the "ep->lock" since the call to
1366 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1367 * because of the way poll() is traditionally implemented in Linux.
1368 */
1369static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1370 struct epoll_event __user *events)
1371{
1372 int eventcnt = 0;
1373 unsigned int revents;
1374 struct list_head *lnk;
1375 struct epitem *epi;
1376
1377 /*
1378 * We can loop without lock because this is a task private list.
1379 * The test done during the collection loop will guarantee us that
1380 * another task will not try to collect this file. Also, items
1381 * cannot vanish during the loop because we are holding "sem".
1382 */
1383 list_for_each(lnk, txlist) {
1384 epi = list_entry(lnk, struct epitem, txlink);
1385
1386 /*
1387 * Get the ready file event set. We can safely use the file
1388 * because we are holding the "sem" in read and this will
1389 * guarantee that both the file and the item will not vanish.
1390 */
1391 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1392
1393 /*
1394 * Set the return event set for the current file descriptor.
1395 * Note that only the task task was successfully able to link
1396 * the item to its "txlist" will write this field.
1397 */
1398 epi->revents = revents & epi->event.events;
1399
1400 if (epi->revents) {
1401 if (__put_user(epi->revents,
1402 &events[eventcnt].events) ||
1403 __put_user(epi->event.data,
1404 &events[eventcnt].data))
1405 return -EFAULT;
1406 if (epi->event.events & EPOLLONESHOT)
1407 epi->event.events &= EP_PRIVATE_BITS;
1408 eventcnt++;
1409 }
1410 }
1411 return eventcnt;
1412}
1413
1414
1415/*
1416 * Walk through the transfer list we collected with ep_collect_ready_items()
1417 * and, if 1) the item is still "alive" 2) its event set is not empty 3) it's
1418 * not already linked, links it to the ready list. Same as above, we are holding
1419 * "sem" so items cannot vanish underneath our nose.
1420 */
1421static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist)
1422{
1423 int ricnt = 0, pwake = 0;
1424 unsigned long flags;
1425 struct epitem *epi;
1426
1427 write_lock_irqsave(&ep->lock, flags);
1428
1429 while (!list_empty(txlist)) {
1430 epi = list_entry(txlist->next, struct epitem, txlink);
1431
1432 /* Unlink the current item from the transfer list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001433 ep_list_del(&epi->txlink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
1435 /*
1436 * If the item is no more linked to the interest set, we don't
1437 * have to push it inside the ready list because the following
1438 * ep_release_epitem() is going to drop it. Also, if the current
1439 * item is set to have an Edge Triggered behaviour, we don't have
1440 * to push it back either.
1441 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001442 if (ep_rb_linked(&epi->rbn) && !(epi->event.events & EPOLLET) &&
1443 (epi->revents & epi->event.events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 list_add_tail(&epi->rdllink, &ep->rdllist);
1445 ricnt++;
1446 }
1447 }
1448
1449 if (ricnt) {
1450 /*
1451 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1452 * wait list.
1453 */
1454 if (waitqueue_active(&ep->wq))
1455 wake_up(&ep->wq);
1456 if (waitqueue_active(&ep->poll_wait))
1457 pwake++;
1458 }
1459
1460 write_unlock_irqrestore(&ep->lock, flags);
1461
1462 /* We have to call this outside the lock */
1463 if (pwake)
1464 ep_poll_safewake(&psw, &ep->poll_wait);
1465}
1466
1467
1468/*
1469 * Perform the transfer of events to user space.
1470 */
1471static int ep_events_transfer(struct eventpoll *ep,
1472 struct epoll_event __user *events, int maxevents)
1473{
1474 int eventcnt = 0;
1475 struct list_head txlist;
1476
1477 INIT_LIST_HEAD(&txlist);
1478
1479 /*
1480 * We need to lock this because we could be hit by
1481 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1482 */
1483 down_read(&ep->sem);
1484
1485 /* Collect/extract ready items */
1486 if (ep_collect_ready_items(ep, &txlist, maxevents) > 0) {
1487 /* Build result set in userspace */
1488 eventcnt = ep_send_events(ep, &txlist, events);
1489
1490 /* Reinject ready items into the ready list */
1491 ep_reinject_items(ep, &txlist);
1492 }
1493
1494 up_read(&ep->sem);
1495
1496 return eventcnt;
1497}
1498
1499
1500static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1501 int maxevents, long timeout)
1502{
1503 int res, eavail;
1504 unsigned long flags;
1505 long jtimeout;
1506 wait_queue_t wait;
1507
1508 /*
1509 * Calculate the timeout by checking for the "infinite" value ( -1 )
1510 * and the overflow condition. The passed timeout is in milliseconds,
1511 * that why (t * HZ) / 1000.
1512 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001513 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1514 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515
1516retry:
1517 write_lock_irqsave(&ep->lock, flags);
1518
1519 res = 0;
1520 if (list_empty(&ep->rdllist)) {
1521 /*
1522 * We don't have any available event to return to the caller.
1523 * We need to sleep here, and we will be wake up by
1524 * ep_poll_callback() when events will become available.
1525 */
1526 init_waitqueue_entry(&wait, current);
1527 add_wait_queue(&ep->wq, &wait);
1528
1529 for (;;) {
1530 /*
1531 * We don't want to sleep if the ep_poll_callback() sends us
1532 * a wakeup in between. That's why we set the task state
1533 * to TASK_INTERRUPTIBLE before doing the checks.
1534 */
1535 set_current_state(TASK_INTERRUPTIBLE);
1536 if (!list_empty(&ep->rdllist) || !jtimeout)
1537 break;
1538 if (signal_pending(current)) {
1539 res = -EINTR;
1540 break;
1541 }
1542
1543 write_unlock_irqrestore(&ep->lock, flags);
1544 jtimeout = schedule_timeout(jtimeout);
1545 write_lock_irqsave(&ep->lock, flags);
1546 }
1547 remove_wait_queue(&ep->wq, &wait);
1548
1549 set_current_state(TASK_RUNNING);
1550 }
1551
1552 /* Is it worth to try to dig for events ? */
1553 eavail = !list_empty(&ep->rdllist);
1554
1555 write_unlock_irqrestore(&ep->lock, flags);
1556
1557 /*
1558 * Try to transfer events to user space. In case we get 0 events and
1559 * there's still timeout left over, we go trying again in search of
1560 * more luck.
1561 */
1562 if (!res && eavail &&
1563 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1564 goto retry;
1565
1566 return res;
1567}
1568
1569
1570static int eventpollfs_delete_dentry(struct dentry *dentry)
1571{
1572
1573 return 1;
1574}
1575
1576
1577static struct inode *ep_eventpoll_inode(void)
1578{
1579 int error = -ENOMEM;
1580 struct inode *inode = new_inode(eventpoll_mnt->mnt_sb);
1581
1582 if (!inode)
1583 goto eexit_1;
1584
1585 inode->i_fop = &eventpoll_fops;
1586
1587 /*
1588 * Mark the inode dirty from the very beginning,
1589 * that way it will never be moved to the dirty
1590 * list because mark_inode_dirty() will think
1591 * that it already _is_ on the dirty list.
1592 */
1593 inode->i_state = I_DIRTY;
1594 inode->i_mode = S_IRUSR | S_IWUSR;
1595 inode->i_uid = current->fsuid;
1596 inode->i_gid = current->fsgid;
1597 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1598 inode->i_blksize = PAGE_SIZE;
1599 return inode;
1600
1601eexit_1:
1602 return ERR_PTR(error);
1603}
1604
1605
1606static struct super_block *
1607eventpollfs_get_sb(struct file_system_type *fs_type, int flags,
1608 const char *dev_name, void *data)
1609{
1610 return get_sb_pseudo(fs_type, "eventpoll:", NULL, EVENTPOLLFS_MAGIC);
1611}
1612
1613
1614static int __init eventpoll_init(void)
1615{
1616 int error;
1617
1618 init_MUTEX(&epsem);
1619
1620 /* Initialize the structure used to perform safe poll wait head wake ups */
1621 ep_poll_safewake_init(&psw);
1622
1623 /* Allocates slab cache used to allocate "struct epitem" items */
1624 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1625 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1626 NULL, NULL);
1627
1628 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1629 pwq_cache = kmem_cache_create("eventpoll_pwq",
1630 sizeof(struct eppoll_entry), 0,
1631 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1632
1633 /*
1634 * Register the virtual file system that will be the source of inodes
1635 * for the eventpoll files
1636 */
1637 error = register_filesystem(&eventpoll_fs_type);
1638 if (error)
1639 goto epanic;
1640
1641 /* Mount the above commented virtual file system */
1642 eventpoll_mnt = kern_mount(&eventpoll_fs_type);
1643 error = PTR_ERR(eventpoll_mnt);
1644 if (IS_ERR(eventpoll_mnt))
1645 goto epanic;
1646
1647 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: successfully initialized.\n",
1648 current));
1649 return 0;
1650
1651epanic:
1652 panic("eventpoll_init() failed\n");
1653}
1654
1655
1656static void __exit eventpoll_exit(void)
1657{
1658 /* Undo all operations done inside eventpoll_init() */
1659 unregister_filesystem(&eventpoll_fs_type);
1660 mntput(eventpoll_mnt);
1661 kmem_cache_destroy(pwq_cache);
1662 kmem_cache_destroy(epi_cache);
1663}
1664
1665module_init(eventpoll_init);
1666module_exit(eventpoll_exit);
1667
1668MODULE_LICENSE("GPL");