blob: 6ab1dd0ca904c4f851590cfed7dd8f6a3860e919 [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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104struct epoll_filefd {
105 struct file *file;
106 int fd;
107};
108
109/*
110 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
111 * It is used to keep track on all tasks that are currently inside the wake_up() code
112 * to 1) short-circuit the one coming from the same task and same wait queue head
113 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
114 * 3) let go the ones coming from other tasks.
115 */
116struct wake_task_node {
117 struct list_head llink;
118 task_t *task;
119 wait_queue_head_t *wq;
120};
121
122/*
123 * This is used to implement the safe poll wake up avoiding to reenter
124 * the poll callback from inside wake_up().
125 */
126struct poll_safewake {
127 struct list_head wake_task_list;
128 spinlock_t lock;
129};
130
131/*
132 * This structure is stored inside the "private_data" member of the file
133 * structure and rapresent the main data sructure for the eventpoll
134 * interface.
135 */
136struct eventpoll {
137 /* Protect the this structure access */
138 rwlock_t lock;
139
140 /*
141 * This semaphore is used to ensure that files are not removed
142 * while epoll is using them. This is read-held during the event
143 * collection loop and it is write-held during the file cleanup
144 * path, the epoll file exit code and the ctl operations.
145 */
146 struct rw_semaphore sem;
147
148 /* Wait queue used by sys_epoll_wait() */
149 wait_queue_head_t wq;
150
151 /* Wait queue used by file->poll() */
152 wait_queue_head_t poll_wait;
153
154 /* List of ready file descriptors */
155 struct list_head rdllist;
156
157 /* RB-Tree root used to store monitored fd structs */
158 struct rb_root rbr;
159};
160
161/* Wait structure used by the poll hooks */
162struct eppoll_entry {
163 /* List header used to link this structure to the "struct epitem" */
164 struct list_head llink;
165
166 /* The "base" pointer is set to the container "struct epitem" */
167 void *base;
168
169 /*
170 * Wait queue item that will be linked to the target file wait
171 * queue head.
172 */
173 wait_queue_t wait;
174
175 /* The wait queue head that linked the "wait" wait queue item */
176 wait_queue_head_t *whead;
177};
178
179/*
180 * Each file descriptor added to the eventpoll interface will
181 * have an entry of this type linked to the hash.
182 */
183struct epitem {
184 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
185 struct rb_node rbn;
186
187 /* List header used to link this structure to the eventpoll ready list */
188 struct list_head rdllink;
189
190 /* The file descriptor information this item refers to */
191 struct epoll_filefd ffd;
192
193 /* Number of active wait queue attached to poll operations */
194 int nwait;
195
196 /* List containing poll wait queues */
197 struct list_head pwqlist;
198
199 /* The "container" of this item */
200 struct eventpoll *ep;
201
202 /* The structure that describe the interested events and the source fd */
203 struct epoll_event event;
204
205 /*
206 * Used to keep track of the usage count of the structure. This avoids
207 * that the structure will desappear from underneath our processing.
208 */
209 atomic_t usecnt;
210
211 /* List header used to link this item to the "struct file" items list */
212 struct list_head fllink;
213
214 /* List header used to link the item to the transfer list */
215 struct list_head txlink;
216
217 /*
218 * This is used during the collection/transfer of events to userspace
219 * to pin items empty events set.
220 */
221 unsigned int revents;
222};
223
224/* Wrapper struct used by poll queueing */
225struct ep_pqueue {
226 poll_table pt;
227 struct epitem *epi;
228};
229
230
231
232static void ep_poll_safewake_init(struct poll_safewake *psw);
233static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
234static int ep_getfd(int *efd, struct inode **einode, struct file **efile);
235static int ep_file_init(struct file *file);
236static void ep_free(struct eventpoll *ep);
237static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
238static void ep_use_epitem(struct epitem *epi);
239static void ep_release_epitem(struct epitem *epi);
240static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
241 poll_table *pt);
242static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi);
243static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
244 struct file *tfile, int fd);
245static int ep_modify(struct eventpoll *ep, struct epitem *epi,
246 struct epoll_event *event);
247static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
248static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
249static int ep_remove(struct eventpoll *ep, struct epitem *epi);
250static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
251static int ep_eventpoll_close(struct inode *inode, struct file *file);
252static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
253static int ep_collect_ready_items(struct eventpoll *ep,
254 struct list_head *txlist, int maxevents);
255static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
256 struct epoll_event __user *events);
257static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist);
258static int ep_events_transfer(struct eventpoll *ep,
259 struct epoll_event __user *events,
260 int maxevents);
261static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
262 int maxevents, long timeout);
263static int eventpollfs_delete_dentry(struct dentry *dentry);
264static struct inode *ep_eventpoll_inode(void);
265static struct super_block *eventpollfs_get_sb(struct file_system_type *fs_type,
266 int flags, const char *dev_name,
267 void *data);
268
269/*
270 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
271 */
Adrian Bunk75c96f82005-05-05 16:16:09 -0700272static struct semaphore epsem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274/* Safe wake up implementation */
275static struct poll_safewake psw;
276
277/* Slab cache used to allocate "struct epitem" */
278static kmem_cache_t *epi_cache;
279
280/* Slab cache used to allocate "struct eppoll_entry" */
281static kmem_cache_t *pwq_cache;
282
283/* Virtual fs used to allocate inodes for eventpoll files */
284static struct vfsmount *eventpoll_mnt;
285
286/* File callbacks that implement the eventpoll file behaviour */
287static struct file_operations eventpoll_fops = {
288 .release = ep_eventpoll_close,
289 .poll = ep_eventpoll_poll
290};
291
292/*
293 * This is used to register the virtual file system from where
294 * eventpoll inodes are allocated.
295 */
296static struct file_system_type eventpoll_fs_type = {
297 .name = "eventpollfs",
298 .get_sb = eventpollfs_get_sb,
299 .kill_sb = kill_anon_super,
300};
301
302/* Very basic directory entry operations for the eventpoll virtual file system */
303static struct dentry_operations eventpollfs_dentry_operations = {
304 .d_delete = eventpollfs_delete_dentry,
305};
306
307
308
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700309/* Fast test to see if the file is an evenpoll file */
310static inline int is_file_epoll(struct file *f)
311{
312 return f->f_op == &eventpoll_fops;
313}
314
315/* Setup the structure that is used as key for the rb-tree */
316static inline void ep_set_ffd(struct epoll_filefd *ffd,
317 struct file *file, int fd)
318{
319 ffd->file = file;
320 ffd->fd = fd;
321}
322
323/* Compare rb-tree keys */
324static inline int ep_cmp_ffd(struct epoll_filefd *p1,
325 struct epoll_filefd *p2)
326{
327 return (p1->file > p2->file ? +1:
328 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
329}
330
331/* Special initialization for the rb-tree node to detect linkage */
332static inline void ep_rb_initnode(struct rb_node *n)
333{
334 n->rb_parent = n;
335}
336
337/* Removes a node from the rb-tree and marks it for a fast is-linked check */
338static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
339{
340 rb_erase(n, r);
341 n->rb_parent = n;
342}
343
344/* Fast check to verify that the item is linked to the main rb-tree */
345static inline int ep_rb_linked(struct rb_node *n)
346{
347 return n->rb_parent != n;
348}
349
350/*
351 * Remove the item from the list and perform its initialization.
352 * This is useful for us because we can test if the item is linked
353 * using "ep_is_linked(p)".
354 */
355static inline void ep_list_del(struct list_head *p)
356{
357 list_del(p);
358 INIT_LIST_HEAD(p);
359}
360
361/* Tells us if the item is currently linked */
362static inline int ep_is_linked(struct list_head *p)
363{
364 return !list_empty(p);
365}
366
367/* Get the "struct epitem" from a wait queue pointer */
368static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
369{
370 return container_of(p, struct eppoll_entry, wait)->base;
371}
372
373/* Get the "struct epitem" from an epoll queue wrapper */
374static inline struct epitem * ep_item_from_epqueue(poll_table *p)
375{
376 return container_of(p, struct ep_pqueue, pt)->epi;
377}
378
379/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
380static inline int ep_op_hash_event(int op)
381{
382 return op != EPOLL_CTL_DEL;
383}
384
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385/* Initialize the poll safe wake up structure */
386static void ep_poll_safewake_init(struct poll_safewake *psw)
387{
388
389 INIT_LIST_HEAD(&psw->wake_task_list);
390 spin_lock_init(&psw->lock);
391}
392
393
394/*
395 * Perform a safe wake up of the poll wait list. The problem is that
396 * with the new callback'd wake up system, it is possible that the
397 * poll callback is reentered from inside the call to wake_up() done
398 * on the poll wait queue head. The rule is that we cannot reenter the
399 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
400 * and we cannot reenter the same wait queue head at all. This will
401 * enable to have a hierarchy of epoll file descriptor of no more than
402 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
403 * because this one gets called by the poll callback, that in turn is called
404 * from inside a wake_up(), that might be called from irq context.
405 */
406static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
407{
408 int wake_nests = 0;
409 unsigned long flags;
410 task_t *this_task = current;
411 struct list_head *lsthead = &psw->wake_task_list, *lnk;
412 struct wake_task_node *tncur;
413 struct wake_task_node tnode;
414
415 spin_lock_irqsave(&psw->lock, flags);
416
417 /* Try to see if the current task is already inside this wakeup call */
418 list_for_each(lnk, lsthead) {
419 tncur = list_entry(lnk, struct wake_task_node, llink);
420
421 if (tncur->wq == wq ||
422 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
423 /*
424 * Ops ... loop detected or maximum nest level reached.
425 * We abort this wake by breaking the cycle itself.
426 */
427 spin_unlock_irqrestore(&psw->lock, flags);
428 return;
429 }
430 }
431
432 /* Add the current task to the list */
433 tnode.task = this_task;
434 tnode.wq = wq;
435 list_add(&tnode.llink, lsthead);
436
437 spin_unlock_irqrestore(&psw->lock, flags);
438
439 /* Do really wake up now */
440 wake_up(wq);
441
442 /* Remove the current task from the list */
443 spin_lock_irqsave(&psw->lock, flags);
444 list_del(&tnode.llink);
445 spin_unlock_irqrestore(&psw->lock, flags);
446}
447
448
449/* Used to initialize the epoll bits inside the "struct file" */
450void eventpoll_init_file(struct file *file)
451{
452
453 INIT_LIST_HEAD(&file->f_ep_links);
454 spin_lock_init(&file->f_ep_lock);
455}
456
457
458/*
459 * This is called from eventpoll_release() to unlink files from the eventpoll
460 * interface. We need to have this facility to cleanup correctly files that are
461 * closed without being removed from the eventpoll interface.
462 */
463void eventpoll_release_file(struct file *file)
464{
465 struct list_head *lsthead = &file->f_ep_links;
466 struct eventpoll *ep;
467 struct epitem *epi;
468
469 /*
470 * We don't want to get "file->f_ep_lock" because it is not
471 * necessary. It is not necessary because we're in the "struct file"
472 * cleanup path, and this means that noone is using this file anymore.
473 * The only hit might come from ep_free() but by holding the semaphore
474 * will correctly serialize the operation. We do need to acquire
475 * "ep->sem" after "epsem" because ep_remove() requires it when called
476 * from anywhere but ep_free().
477 */
478 down(&epsem);
479
480 while (!list_empty(lsthead)) {
481 epi = list_entry(lsthead->next, struct epitem, fllink);
482
483 ep = epi->ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700484 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 down_write(&ep->sem);
486 ep_remove(ep, epi);
487 up_write(&ep->sem);
488 }
489
490 up(&epsem);
491}
492
493
494/*
495 * It opens an eventpoll file descriptor by suggesting a storage of "size"
496 * file descriptors. The size parameter is just an hint about how to size
497 * data structures. It won't prevent the user to store more than "size"
498 * file descriptors inside the epoll interface. It is the kernel part of
499 * the userspace epoll_create(2).
500 */
501asmlinkage long sys_epoll_create(int size)
502{
503 int error, fd;
504 struct inode *inode;
505 struct file *file;
506
507 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
508 current, size));
509
510 /* Sanity check on the size parameter */
511 error = -EINVAL;
512 if (size <= 0)
513 goto eexit_1;
514
515 /*
516 * Creates all the items needed to setup an eventpoll file. That is,
517 * a file structure, and inode and a free file descriptor.
518 */
519 error = ep_getfd(&fd, &inode, &file);
520 if (error)
521 goto eexit_1;
522
523 /* Setup the file internal data structure ( "struct eventpoll" ) */
524 error = ep_file_init(file);
525 if (error)
526 goto eexit_2;
527
528
529 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
530 current, size, fd));
531
532 return fd;
533
534eexit_2:
535 sys_close(fd);
536eexit_1:
537 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
538 current, size, error));
539 return error;
540}
541
542
543/*
544 * The following function implements the controller interface for
545 * the eventpoll file that enables the insertion/removal/change of
546 * file descriptors inside the interest set. It represents
547 * the kernel part of the user space epoll_ctl(2).
548 */
549asmlinkage long
550sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
551{
552 int error;
553 struct file *file, *tfile;
554 struct eventpoll *ep;
555 struct epitem *epi;
556 struct epoll_event epds;
557
558 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
559 current, epfd, op, fd, event));
560
561 error = -EFAULT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700562 if (ep_op_hash_event(op) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 copy_from_user(&epds, event, sizeof(struct epoll_event)))
564 goto eexit_1;
565
566 /* Get the "struct file *" for the eventpoll file */
567 error = -EBADF;
568 file = fget(epfd);
569 if (!file)
570 goto eexit_1;
571
572 /* Get the "struct file *" for the target file */
573 tfile = fget(fd);
574 if (!tfile)
575 goto eexit_2;
576
577 /* The target file descriptor must support poll */
578 error = -EPERM;
579 if (!tfile->f_op || !tfile->f_op->poll)
580 goto eexit_3;
581
582 /*
583 * We have to check that the file structure underneath the file descriptor
584 * the user passed to us _is_ an eventpoll file. And also we do not permit
585 * adding an epoll file descriptor inside itself.
586 */
587 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700588 if (file == tfile || !is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 goto eexit_3;
590
591 /*
592 * At this point it is safe to assume that the "private_data" contains
593 * our own data structure.
594 */
595 ep = file->private_data;
596
597 down_write(&ep->sem);
598
599 /* Try to lookup the file inside our hash table */
600 epi = ep_find(ep, tfile, fd);
601
602 error = -EINVAL;
603 switch (op) {
604 case EPOLL_CTL_ADD:
605 if (!epi) {
606 epds.events |= POLLERR | POLLHUP;
607
608 error = ep_insert(ep, &epds, tfile, fd);
609 } else
610 error = -EEXIST;
611 break;
612 case EPOLL_CTL_DEL:
613 if (epi)
614 error = ep_remove(ep, epi);
615 else
616 error = -ENOENT;
617 break;
618 case EPOLL_CTL_MOD:
619 if (epi) {
620 epds.events |= POLLERR | POLLHUP;
621 error = ep_modify(ep, epi, &epds);
622 } else
623 error = -ENOENT;
624 break;
625 }
626
627 /*
628 * The function ep_find() increments the usage count of the structure
629 * so, if this is not NULL, we need to release it.
630 */
631 if (epi)
632 ep_release_epitem(epi);
633
634 up_write(&ep->sem);
635
636eexit_3:
637 fput(tfile);
638eexit_2:
639 fput(file);
640eexit_1:
641 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
642 current, epfd, op, fd, event, error));
643
644 return error;
645}
646
647#define MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
648
649/*
650 * Implement the event wait interface for the eventpoll file. It is the kernel
651 * part of the user space epoll_wait(2).
652 */
653asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
654 int maxevents, int timeout)
655{
656 int error;
657 struct file *file;
658 struct eventpoll *ep;
659
660 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
661 current, epfd, events, maxevents, timeout));
662
663 /* The maximum number of event must be greater than zero */
664 if (maxevents <= 0 || maxevents > MAX_EVENTS)
665 return -EINVAL;
666
667 /* Verify that the area passed by the user is writeable */
668 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
669 error = -EFAULT;
670 goto eexit_1;
671 }
672
673 /* Get the "struct file *" for the eventpoll file */
674 error = -EBADF;
675 file = fget(epfd);
676 if (!file)
677 goto eexit_1;
678
679 /*
680 * We have to check that the file structure underneath the fd
681 * the user passed to us _is_ an eventpoll file.
682 */
683 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700684 if (!is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 goto eexit_2;
686
687 /*
688 * At this point it is safe to assume that the "private_data" contains
689 * our own data structure.
690 */
691 ep = file->private_data;
692
693 /* Time to fish for events ... */
694 error = ep_poll(ep, events, maxevents, timeout);
695
696eexit_2:
697 fput(file);
698eexit_1:
699 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
700 current, epfd, events, maxevents, timeout, error));
701
702 return error;
703}
704
705
706/*
707 * Creates the file descriptor to be used by the epoll interface.
708 */
709static int ep_getfd(int *efd, struct inode **einode, struct file **efile)
710{
711 struct qstr this;
712 char name[32];
713 struct dentry *dentry;
714 struct inode *inode;
715 struct file *file;
716 int error, fd;
717
718 /* Get an ready to use file */
719 error = -ENFILE;
720 file = get_empty_filp();
721 if (!file)
722 goto eexit_1;
723
724 /* Allocates an inode from the eventpoll file system */
725 inode = ep_eventpoll_inode();
726 error = PTR_ERR(inode);
727 if (IS_ERR(inode))
728 goto eexit_2;
729
730 /* Allocates a free descriptor to plug the file onto */
731 error = get_unused_fd();
732 if (error < 0)
733 goto eexit_3;
734 fd = error;
735
736 /*
737 * Link the inode to a directory entry by creating a unique name
738 * using the inode number.
739 */
740 error = -ENOMEM;
741 sprintf(name, "[%lu]", inode->i_ino);
742 this.name = name;
743 this.len = strlen(name);
744 this.hash = inode->i_ino;
745 dentry = d_alloc(eventpoll_mnt->mnt_sb->s_root, &this);
746 if (!dentry)
747 goto eexit_4;
748 dentry->d_op = &eventpollfs_dentry_operations;
749 d_add(dentry, inode);
750 file->f_vfsmnt = mntget(eventpoll_mnt);
751 file->f_dentry = dentry;
752 file->f_mapping = inode->i_mapping;
753
754 file->f_pos = 0;
755 file->f_flags = O_RDONLY;
756 file->f_op = &eventpoll_fops;
757 file->f_mode = FMODE_READ;
758 file->f_version = 0;
759 file->private_data = NULL;
760
761 /* Install the new setup file into the allocated fd. */
762 fd_install(fd, file);
763
764 *efd = fd;
765 *einode = inode;
766 *efile = file;
767 return 0;
768
769eexit_4:
770 put_unused_fd(fd);
771eexit_3:
772 iput(inode);
773eexit_2:
774 put_filp(file);
775eexit_1:
776 return error;
777}
778
779
780static int ep_file_init(struct file *file)
781{
782 struct eventpoll *ep;
783
784 if (!(ep = kmalloc(sizeof(struct eventpoll), GFP_KERNEL)))
785 return -ENOMEM;
786
787 memset(ep, 0, sizeof(*ep));
788 rwlock_init(&ep->lock);
789 init_rwsem(&ep->sem);
790 init_waitqueue_head(&ep->wq);
791 init_waitqueue_head(&ep->poll_wait);
792 INIT_LIST_HEAD(&ep->rdllist);
793 ep->rbr = RB_ROOT;
794
795 file->private_data = ep;
796
797 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_file_init() ep=%p\n",
798 current, ep));
799 return 0;
800}
801
802
803static void ep_free(struct eventpoll *ep)
804{
805 struct rb_node *rbp;
806 struct epitem *epi;
807
808 /* We need to release all tasks waiting for these file */
809 if (waitqueue_active(&ep->poll_wait))
810 ep_poll_safewake(&psw, &ep->poll_wait);
811
812 /*
813 * We need to lock this because we could be hit by
814 * eventpoll_release_file() while we're freeing the "struct eventpoll".
815 * We do not need to hold "ep->sem" here because the epoll file
816 * is on the way to be removed and no one has references to it
817 * anymore. The only hit might come from eventpoll_release_file() but
818 * holding "epsem" is sufficent here.
819 */
820 down(&epsem);
821
822 /*
823 * Walks through the whole tree by unregistering poll callbacks.
824 */
825 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
826 epi = rb_entry(rbp, struct epitem, rbn);
827
828 ep_unregister_pollwait(ep, epi);
829 }
830
831 /*
832 * Walks through the whole hash by freeing each "struct epitem". At this
833 * point we are sure no poll callbacks will be lingering around, and also by
834 * write-holding "sem" we can be sure that no file cleanup code will hit
835 * us during this operation. So we can avoid the lock on "ep->lock".
836 */
837 while ((rbp = rb_first(&ep->rbr)) != 0) {
838 epi = rb_entry(rbp, struct epitem, rbn);
839 ep_remove(ep, epi);
840 }
841
842 up(&epsem);
843}
844
845
846/*
847 * Search the file inside the eventpoll hash. It add usage count to
848 * the returned item, so the caller must call ep_release_epitem()
849 * after finished using the "struct epitem".
850 */
851static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
852{
853 int kcmp;
854 unsigned long flags;
855 struct rb_node *rbp;
856 struct epitem *epi, *epir = NULL;
857 struct epoll_filefd ffd;
858
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700859 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 read_lock_irqsave(&ep->lock, flags);
861 for (rbp = ep->rbr.rb_node; rbp; ) {
862 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700863 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 if (kcmp > 0)
865 rbp = rbp->rb_right;
866 else if (kcmp < 0)
867 rbp = rbp->rb_left;
868 else {
869 ep_use_epitem(epi);
870 epir = epi;
871 break;
872 }
873 }
874 read_unlock_irqrestore(&ep->lock, flags);
875
876 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
877 current, file, epir));
878
879 return epir;
880}
881
882
883/*
884 * Increment the usage count of the "struct epitem" making it sure
885 * that the user will have a valid pointer to reference.
886 */
887static void ep_use_epitem(struct epitem *epi)
888{
889
890 atomic_inc(&epi->usecnt);
891}
892
893
894/*
895 * Decrement ( release ) the usage count by signaling that the user
896 * has finished using the structure. It might lead to freeing the
897 * structure itself if the count goes to zero.
898 */
899static void ep_release_epitem(struct epitem *epi)
900{
901
902 if (atomic_dec_and_test(&epi->usecnt))
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700903 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904}
905
906
907/*
908 * This is the callback that is used to add our wait queue to the
909 * target file wakeup lists.
910 */
911static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
912 poll_table *pt)
913{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700914 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 struct eppoll_entry *pwq;
916
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700917 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, SLAB_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
919 pwq->whead = whead;
920 pwq->base = epi;
921 add_wait_queue(whead, &pwq->wait);
922 list_add_tail(&pwq->llink, &epi->pwqlist);
923 epi->nwait++;
924 } else {
925 /* We have to signal that an error occurred */
926 epi->nwait = -1;
927 }
928}
929
930
931static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
932{
933 int kcmp;
934 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
935 struct epitem *epic;
936
937 while (*p) {
938 parent = *p;
939 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700940 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 if (kcmp > 0)
942 p = &parent->rb_right;
943 else
944 p = &parent->rb_left;
945 }
946 rb_link_node(&epi->rbn, parent, p);
947 rb_insert_color(&epi->rbn, &ep->rbr);
948}
949
950
951static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
952 struct file *tfile, int fd)
953{
954 int error, revents, pwake = 0;
955 unsigned long flags;
956 struct epitem *epi;
957 struct ep_pqueue epq;
958
959 error = -ENOMEM;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700960 if (!(epi = kmem_cache_alloc(epi_cache, SLAB_KERNEL)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 goto eexit_1;
962
963 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700964 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 INIT_LIST_HEAD(&epi->rdllink);
966 INIT_LIST_HEAD(&epi->fllink);
967 INIT_LIST_HEAD(&epi->txlink);
968 INIT_LIST_HEAD(&epi->pwqlist);
969 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700970 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 epi->event = *event;
972 atomic_set(&epi->usecnt, 1);
973 epi->nwait = 0;
974
975 /* Initialize the poll table using the queue callback */
976 epq.epi = epi;
977 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
978
979 /*
980 * Attach the item to the poll hooks and get current event bits.
981 * We can safely use the file* here because its usage count has
982 * been increased by the caller of this function.
983 */
984 revents = tfile->f_op->poll(tfile, &epq.pt);
985
986 /*
987 * We have to check if something went wrong during the poll wait queue
988 * install process. Namely an allocation for a wait queue failed due
989 * high memory pressure.
990 */
991 if (epi->nwait < 0)
992 goto eexit_2;
993
994 /* Add the current item to the list of active epoll hook for this file */
995 spin_lock(&tfile->f_ep_lock);
996 list_add_tail(&epi->fllink, &tfile->f_ep_links);
997 spin_unlock(&tfile->f_ep_lock);
998
999 /* We have to drop the new item inside our item list to keep track of it */
1000 write_lock_irqsave(&ep->lock, flags);
1001
1002 /* Add the current item to the rb-tree */
1003 ep_rbtree_insert(ep, epi);
1004
1005 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001006 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 list_add_tail(&epi->rdllink, &ep->rdllist);
1008
1009 /* Notify waiting tasks that events are available */
1010 if (waitqueue_active(&ep->wq))
1011 wake_up(&ep->wq);
1012 if (waitqueue_active(&ep->poll_wait))
1013 pwake++;
1014 }
1015
1016 write_unlock_irqrestore(&ep->lock, flags);
1017
1018 /* We have to call this outside the lock */
1019 if (pwake)
1020 ep_poll_safewake(&psw, &ep->poll_wait);
1021
1022 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
1023 current, ep, tfile, fd));
1024
1025 return 0;
1026
1027eexit_2:
1028 ep_unregister_pollwait(ep, epi);
1029
1030 /*
1031 * We need to do this because an event could have been arrived on some
1032 * allocated wait queue.
1033 */
1034 write_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001035 if (ep_is_linked(&epi->rdllink))
1036 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 write_unlock_irqrestore(&ep->lock, flags);
1038
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001039 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040eexit_1:
1041 return error;
1042}
1043
1044
1045/*
1046 * Modify the interest event mask by dropping an event if the new mask
1047 * has a match in the current file status.
1048 */
1049static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1050{
1051 int pwake = 0;
1052 unsigned int revents;
1053 unsigned long flags;
1054
1055 /*
1056 * Set the new event interest mask before calling f_op->poll(), otherwise
1057 * a potential race might occur. In fact if we do this operation inside
1058 * the lock, an event might happen between the f_op->poll() call and the
1059 * new event set registering.
1060 */
1061 epi->event.events = event->events;
1062
1063 /*
1064 * Get current event bits. We can safely use the file* here because
1065 * its usage count has been increased by the caller of this function.
1066 */
1067 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1068
1069 write_lock_irqsave(&ep->lock, flags);
1070
1071 /* Copy the data member from inside the lock */
1072 epi->event.data = event->data;
1073
1074 /*
1075 * If the item is not linked to the hash it means that it's on its
1076 * way toward the removal. Do nothing in this case.
1077 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001078 if (ep_rb_linked(&epi->rbn)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 /*
1080 * If the item is "hot" and it is not registered inside the ready
1081 * list, push it inside. If the item is not "hot" and it is currently
1082 * registered inside the ready list, unlink it.
1083 */
1084 if (revents & event->events) {
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001085 if (!ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 list_add_tail(&epi->rdllink, &ep->rdllist);
1087
1088 /* Notify waiting tasks that events are available */
1089 if (waitqueue_active(&ep->wq))
1090 wake_up(&ep->wq);
1091 if (waitqueue_active(&ep->poll_wait))
1092 pwake++;
1093 }
1094 }
1095 }
1096
1097 write_unlock_irqrestore(&ep->lock, flags);
1098
1099 /* We have to call this outside the lock */
1100 if (pwake)
1101 ep_poll_safewake(&psw, &ep->poll_wait);
1102
1103 return 0;
1104}
1105
1106
1107/*
1108 * This function unregister poll callbacks from the associated file descriptor.
1109 * Since this must be called without holding "ep->lock" the atomic exchange trick
1110 * will protect us from multiple unregister.
1111 */
1112static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1113{
1114 int nwait;
1115 struct list_head *lsthead = &epi->pwqlist;
1116 struct eppoll_entry *pwq;
1117
1118 /* This is called without locks, so we need the atomic exchange */
1119 nwait = xchg(&epi->nwait, 0);
1120
1121 if (nwait) {
1122 while (!list_empty(lsthead)) {
1123 pwq = list_entry(lsthead->next, struct eppoll_entry, llink);
1124
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001125 ep_list_del(&pwq->llink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 remove_wait_queue(pwq->whead, &pwq->wait);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001127 kmem_cache_free(pwq_cache, pwq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 }
1129 }
1130}
1131
1132
1133/*
1134 * Unlink the "struct epitem" from all places it might have been hooked up.
1135 * This function must be called with write IRQ lock on "ep->lock".
1136 */
1137static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1138{
1139 int error;
1140
1141 /*
1142 * It can happen that this one is called for an item already unlinked.
1143 * The check protect us from doing a double unlink ( crash ).
1144 */
1145 error = -ENOENT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001146 if (!ep_rb_linked(&epi->rbn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 goto eexit_1;
1148
1149 /*
1150 * Clear the event mask for the unlinked item. This will avoid item
1151 * notifications to be sent after the unlink operation from inside
1152 * the kernel->userspace event transfer loop.
1153 */
1154 epi->event.events = 0;
1155
1156 /*
1157 * At this point is safe to do the job, unlink the item from our rb-tree.
1158 * This operation togheter with the above check closes the door to
1159 * double unlinks.
1160 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001161 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162
1163 /*
1164 * If the item we are going to remove is inside the ready file descriptors
1165 * we want to remove it from this list to avoid stale events.
1166 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001167 if (ep_is_linked(&epi->rdllink))
1168 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
1170 error = 0;
1171eexit_1:
1172
1173 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
1174 current, ep, epi->file, error));
1175
1176 return error;
1177}
1178
1179
1180/*
1181 * Removes a "struct epitem" from the eventpoll hash and deallocates
1182 * all the associated resources.
1183 */
1184static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1185{
1186 int error;
1187 unsigned long flags;
1188 struct file *file = epi->ffd.file;
1189
1190 /*
1191 * Removes poll wait queue hooks. We _have_ to do this without holding
1192 * the "ep->lock" otherwise a deadlock might occur. This because of the
1193 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1194 * queue head lock when unregistering the wait queue. The wakeup callback
1195 * will run by holding the wait queue head lock and will call our callback
1196 * that will try to get "ep->lock".
1197 */
1198 ep_unregister_pollwait(ep, epi);
1199
1200 /* Remove the current item from the list of epoll hooks */
1201 spin_lock(&file->f_ep_lock);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001202 if (ep_is_linked(&epi->fllink))
1203 ep_list_del(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 spin_unlock(&file->f_ep_lock);
1205
1206 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1207 write_lock_irqsave(&ep->lock, flags);
1208
1209 /* Really unlink the item from the hash */
1210 error = ep_unlink(ep, epi);
1211
1212 write_unlock_irqrestore(&ep->lock, flags);
1213
1214 if (error)
1215 goto eexit_1;
1216
1217 /* At this point it is safe to free the eventpoll item */
1218 ep_release_epitem(epi);
1219
1220 error = 0;
1221eexit_1:
1222 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1223 current, ep, file, error));
1224
1225 return error;
1226}
1227
1228
1229/*
1230 * This is the callback that is passed to the wait queue wakeup
1231 * machanism. It is called by the stored file descriptors when they
1232 * have events to report.
1233 */
1234static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1235{
1236 int pwake = 0;
1237 unsigned long flags;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001238 struct epitem *epi = ep_item_from_wait(wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 struct eventpoll *ep = epi->ep;
1240
1241 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
1242 current, epi->file, epi, ep));
1243
1244 write_lock_irqsave(&ep->lock, flags);
1245
1246 /*
1247 * If the event mask does not contain any poll(2) event, we consider the
1248 * descriptor to be disabled. This condition is likely the effect of the
1249 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1250 * until the next EPOLL_CTL_MOD will be issued.
1251 */
1252 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1253 goto is_disabled;
1254
1255 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001256 if (ep_is_linked(&epi->rdllink))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 goto is_linked;
1258
1259 list_add_tail(&epi->rdllink, &ep->rdllist);
1260
1261is_linked:
1262 /*
1263 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1264 * wait list.
1265 */
1266 if (waitqueue_active(&ep->wq))
1267 wake_up(&ep->wq);
1268 if (waitqueue_active(&ep->poll_wait))
1269 pwake++;
1270
1271is_disabled:
1272 write_unlock_irqrestore(&ep->lock, flags);
1273
1274 /* We have to call this outside the lock */
1275 if (pwake)
1276 ep_poll_safewake(&psw, &ep->poll_wait);
1277
1278 return 1;
1279}
1280
1281
1282static int ep_eventpoll_close(struct inode *inode, struct file *file)
1283{
1284 struct eventpoll *ep = file->private_data;
1285
1286 if (ep) {
1287 ep_free(ep);
1288 kfree(ep);
1289 }
1290
1291 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1292 return 0;
1293}
1294
1295
1296static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1297{
1298 unsigned int pollflags = 0;
1299 unsigned long flags;
1300 struct eventpoll *ep = file->private_data;
1301
1302 /* Insert inside our poll wait queue */
1303 poll_wait(file, &ep->poll_wait, wait);
1304
1305 /* Check our condition */
1306 read_lock_irqsave(&ep->lock, flags);
1307 if (!list_empty(&ep->rdllist))
1308 pollflags = POLLIN | POLLRDNORM;
1309 read_unlock_irqrestore(&ep->lock, flags);
1310
1311 return pollflags;
1312}
1313
1314
1315/*
1316 * Since we have to release the lock during the __copy_to_user() operation and
1317 * during the f_op->poll() call, we try to collect the maximum number of items
1318 * by reducing the irqlock/irqunlock switching rate.
1319 */
1320static int ep_collect_ready_items(struct eventpoll *ep, struct list_head *txlist, int maxevents)
1321{
1322 int nepi;
1323 unsigned long flags;
1324 struct list_head *lsthead = &ep->rdllist, *lnk;
1325 struct epitem *epi;
1326
1327 write_lock_irqsave(&ep->lock, flags);
1328
1329 for (nepi = 0, lnk = lsthead->next; lnk != lsthead && nepi < maxevents;) {
1330 epi = list_entry(lnk, struct epitem, rdllink);
1331
1332 lnk = lnk->next;
1333
1334 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001335 if (!ep_is_linked(&epi->txlink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 /*
1337 * This is initialized in this way so that the default
1338 * behaviour of the reinjecting code will be to push back
1339 * the item inside the ready list.
1340 */
1341 epi->revents = epi->event.events;
1342
1343 /* Link the ready item into the transfer list */
1344 list_add(&epi->txlink, txlist);
1345 nepi++;
1346
1347 /*
1348 * Unlink the item from the ready list.
1349 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001350 ep_list_del(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 }
1352 }
1353
1354 write_unlock_irqrestore(&ep->lock, flags);
1355
1356 return nepi;
1357}
1358
1359
1360/*
1361 * This function is called without holding the "ep->lock" since the call to
1362 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1363 * because of the way poll() is traditionally implemented in Linux.
1364 */
1365static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
1366 struct epoll_event __user *events)
1367{
1368 int eventcnt = 0;
1369 unsigned int revents;
1370 struct list_head *lnk;
1371 struct epitem *epi;
1372
1373 /*
1374 * We can loop without lock because this is a task private list.
1375 * The test done during the collection loop will guarantee us that
1376 * another task will not try to collect this file. Also, items
1377 * cannot vanish during the loop because we are holding "sem".
1378 */
1379 list_for_each(lnk, txlist) {
1380 epi = list_entry(lnk, struct epitem, txlink);
1381
1382 /*
1383 * Get the ready file event set. We can safely use the file
1384 * because we are holding the "sem" in read and this will
1385 * guarantee that both the file and the item will not vanish.
1386 */
1387 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1388
1389 /*
1390 * Set the return event set for the current file descriptor.
1391 * Note that only the task task was successfully able to link
1392 * the item to its "txlist" will write this field.
1393 */
1394 epi->revents = revents & epi->event.events;
1395
1396 if (epi->revents) {
1397 if (__put_user(epi->revents,
1398 &events[eventcnt].events) ||
1399 __put_user(epi->event.data,
1400 &events[eventcnt].data))
1401 return -EFAULT;
1402 if (epi->event.events & EPOLLONESHOT)
1403 epi->event.events &= EP_PRIVATE_BITS;
1404 eventcnt++;
1405 }
1406 }
1407 return eventcnt;
1408}
1409
1410
1411/*
1412 * Walk through the transfer list we collected with ep_collect_ready_items()
1413 * and, if 1) the item is still "alive" 2) its event set is not empty 3) it's
1414 * not already linked, links it to the ready list. Same as above, we are holding
1415 * "sem" so items cannot vanish underneath our nose.
1416 */
1417static void ep_reinject_items(struct eventpoll *ep, struct list_head *txlist)
1418{
1419 int ricnt = 0, pwake = 0;
1420 unsigned long flags;
1421 struct epitem *epi;
1422
1423 write_lock_irqsave(&ep->lock, flags);
1424
1425 while (!list_empty(txlist)) {
1426 epi = list_entry(txlist->next, struct epitem, txlink);
1427
1428 /* Unlink the current item from the transfer list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001429 ep_list_del(&epi->txlink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430
1431 /*
1432 * If the item is no more linked to the interest set, we don't
1433 * have to push it inside the ready list because the following
1434 * ep_release_epitem() is going to drop it. Also, if the current
1435 * item is set to have an Edge Triggered behaviour, we don't have
1436 * to push it back either.
1437 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001438 if (ep_rb_linked(&epi->rbn) && !(epi->event.events & EPOLLET) &&
1439 (epi->revents & epi->event.events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 list_add_tail(&epi->rdllink, &ep->rdllist);
1441 ricnt++;
1442 }
1443 }
1444
1445 if (ricnt) {
1446 /*
1447 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1448 * wait list.
1449 */
1450 if (waitqueue_active(&ep->wq))
1451 wake_up(&ep->wq);
1452 if (waitqueue_active(&ep->poll_wait))
1453 pwake++;
1454 }
1455
1456 write_unlock_irqrestore(&ep->lock, flags);
1457
1458 /* We have to call this outside the lock */
1459 if (pwake)
1460 ep_poll_safewake(&psw, &ep->poll_wait);
1461}
1462
1463
1464/*
1465 * Perform the transfer of events to user space.
1466 */
1467static int ep_events_transfer(struct eventpoll *ep,
1468 struct epoll_event __user *events, int maxevents)
1469{
1470 int eventcnt = 0;
1471 struct list_head txlist;
1472
1473 INIT_LIST_HEAD(&txlist);
1474
1475 /*
1476 * We need to lock this because we could be hit by
1477 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1478 */
1479 down_read(&ep->sem);
1480
1481 /* Collect/extract ready items */
1482 if (ep_collect_ready_items(ep, &txlist, maxevents) > 0) {
1483 /* Build result set in userspace */
1484 eventcnt = ep_send_events(ep, &txlist, events);
1485
1486 /* Reinject ready items into the ready list */
1487 ep_reinject_items(ep, &txlist);
1488 }
1489
1490 up_read(&ep->sem);
1491
1492 return eventcnt;
1493}
1494
1495
1496static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1497 int maxevents, long timeout)
1498{
1499 int res, eavail;
1500 unsigned long flags;
1501 long jtimeout;
1502 wait_queue_t wait;
1503
1504 /*
1505 * Calculate the timeout by checking for the "infinite" value ( -1 )
1506 * and the overflow condition. The passed timeout is in milliseconds,
1507 * that why (t * HZ) / 1000.
1508 */
1509 jtimeout = timeout == -1 || timeout > (MAX_SCHEDULE_TIMEOUT - 1000) / HZ ?
1510 MAX_SCHEDULE_TIMEOUT: (timeout * HZ + 999) / 1000;
1511
1512retry:
1513 write_lock_irqsave(&ep->lock, flags);
1514
1515 res = 0;
1516 if (list_empty(&ep->rdllist)) {
1517 /*
1518 * We don't have any available event to return to the caller.
1519 * We need to sleep here, and we will be wake up by
1520 * ep_poll_callback() when events will become available.
1521 */
1522 init_waitqueue_entry(&wait, current);
1523 add_wait_queue(&ep->wq, &wait);
1524
1525 for (;;) {
1526 /*
1527 * We don't want to sleep if the ep_poll_callback() sends us
1528 * a wakeup in between. That's why we set the task state
1529 * to TASK_INTERRUPTIBLE before doing the checks.
1530 */
1531 set_current_state(TASK_INTERRUPTIBLE);
1532 if (!list_empty(&ep->rdllist) || !jtimeout)
1533 break;
1534 if (signal_pending(current)) {
1535 res = -EINTR;
1536 break;
1537 }
1538
1539 write_unlock_irqrestore(&ep->lock, flags);
1540 jtimeout = schedule_timeout(jtimeout);
1541 write_lock_irqsave(&ep->lock, flags);
1542 }
1543 remove_wait_queue(&ep->wq, &wait);
1544
1545 set_current_state(TASK_RUNNING);
1546 }
1547
1548 /* Is it worth to try to dig for events ? */
1549 eavail = !list_empty(&ep->rdllist);
1550
1551 write_unlock_irqrestore(&ep->lock, flags);
1552
1553 /*
1554 * Try to transfer events to user space. In case we get 0 events and
1555 * there's still timeout left over, we go trying again in search of
1556 * more luck.
1557 */
1558 if (!res && eavail &&
1559 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1560 goto retry;
1561
1562 return res;
1563}
1564
1565
1566static int eventpollfs_delete_dentry(struct dentry *dentry)
1567{
1568
1569 return 1;
1570}
1571
1572
1573static struct inode *ep_eventpoll_inode(void)
1574{
1575 int error = -ENOMEM;
1576 struct inode *inode = new_inode(eventpoll_mnt->mnt_sb);
1577
1578 if (!inode)
1579 goto eexit_1;
1580
1581 inode->i_fop = &eventpoll_fops;
1582
1583 /*
1584 * Mark the inode dirty from the very beginning,
1585 * that way it will never be moved to the dirty
1586 * list because mark_inode_dirty() will think
1587 * that it already _is_ on the dirty list.
1588 */
1589 inode->i_state = I_DIRTY;
1590 inode->i_mode = S_IRUSR | S_IWUSR;
1591 inode->i_uid = current->fsuid;
1592 inode->i_gid = current->fsgid;
1593 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1594 inode->i_blksize = PAGE_SIZE;
1595 return inode;
1596
1597eexit_1:
1598 return ERR_PTR(error);
1599}
1600
1601
1602static struct super_block *
1603eventpollfs_get_sb(struct file_system_type *fs_type, int flags,
1604 const char *dev_name, void *data)
1605{
1606 return get_sb_pseudo(fs_type, "eventpoll:", NULL, EVENTPOLLFS_MAGIC);
1607}
1608
1609
1610static int __init eventpoll_init(void)
1611{
1612 int error;
1613
1614 init_MUTEX(&epsem);
1615
1616 /* Initialize the structure used to perform safe poll wait head wake ups */
1617 ep_poll_safewake_init(&psw);
1618
1619 /* Allocates slab cache used to allocate "struct epitem" items */
1620 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1621 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1622 NULL, NULL);
1623
1624 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1625 pwq_cache = kmem_cache_create("eventpoll_pwq",
1626 sizeof(struct eppoll_entry), 0,
1627 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1628
1629 /*
1630 * Register the virtual file system that will be the source of inodes
1631 * for the eventpoll files
1632 */
1633 error = register_filesystem(&eventpoll_fs_type);
1634 if (error)
1635 goto epanic;
1636
1637 /* Mount the above commented virtual file system */
1638 eventpoll_mnt = kern_mount(&eventpoll_fs_type);
1639 error = PTR_ERR(eventpoll_mnt);
1640 if (IS_ERR(eventpoll_mnt))
1641 goto epanic;
1642
1643 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: successfully initialized.\n",
1644 current));
1645 return 0;
1646
1647epanic:
1648 panic("eventpoll_init() failed\n");
1649}
1650
1651
1652static void __exit eventpoll_exit(void)
1653{
1654 /* Undo all operations done inside eventpoll_init() */
1655 unregister_filesystem(&eventpoll_fs_type);
1656 mntput(eventpoll_mnt);
1657 kmem_cache_destroy(pwq_cache);
1658 kmem_cache_destroy(epi_cache);
1659}
1660
1661module_init(eventpoll_init);
1662module_exit(eventpoll_exit);
1663
1664MODULE_LICENSE("GPL");