blob: 2831c8f9f3e356ff10b7f76f95a2c0c1addcdd08 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/eventpoll.c ( Efficent event polling implementation )
Davide Libenzi3419b232006-06-25 05:48:14 -07003 * Copyright (C) 2001,...,2006 Davide Libenzi
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/string.h>
26#include <linux/list.h>
27#include <linux/hash.h>
28#include <linux/spinlock.h>
29#include <linux/syscalls.h>
30#include <linux/rwsem.h>
31#include <linux/rbtree.h>
32#include <linux/wait.h>
33#include <linux/eventpoll.h>
34#include <linux/mount.h>
35#include <linux/bitops.h>
Arjan van de Ven144efe32006-03-23 03:00:32 -080036#include <linux/mutex.h>
Davide Libenzida66f7c2007-05-10 22:23:21 -070037#include <linux/anon_inodes.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/uaccess.h>
39#include <asm/system.h>
40#include <asm/io.h>
41#include <asm/mman.h>
42#include <asm/atomic.h>
43#include <asm/semaphore.h>
44
45
46/*
47 * LOCKING:
48 * There are three level of locking required by epoll :
49 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080050 * 1) epmutex (mutex)
Linus Torvalds1da177e2005-04-16 15:20:36 -070051 * 2) ep->sem (rw_semaphore)
52 * 3) ep->lock (rw_lock)
53 *
54 * The acquire order is the one listed above, from 1 to 3.
55 * We need a spinlock (ep->lock) because we manipulate objects
56 * from inside the poll callback, that might be triggered from
57 * a wake_up() that in turn might be called from IRQ context.
58 * So we can't sleep inside the poll callback and hence we need
59 * a spinlock. During the event transfer loop (from kernel to
60 * user space) we could end up sleeping due a copy_to_user(), so
61 * we need a lock that will allow us to sleep. This lock is a
62 * read-write semaphore (ep->sem). It is acquired on read during
63 * the event transfer loop and in write during epoll_ctl(EPOLL_CTL_DEL)
64 * and during eventpoll_release_file(). Then we also need a global
65 * semaphore to serialize eventpoll_release_file() and ep_free().
66 * This semaphore is acquired by ep_free() during the epoll file
67 * cleanup path and it is also acquired by eventpoll_release_file()
68 * if a file has been pushed inside an epoll set and it is then
69 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
70 * It is possible to drop the "ep->sem" and to use the global
Arjan van de Ven144efe32006-03-23 03:00:32 -080071 * semaphore "epmutex" (together with "ep->lock") to have it working,
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * but having "ep->sem" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080073 * Events that require holding "epmutex" are very rare, while for
Linus Torvalds1da177e2005-04-16 15:20:36 -070074 * normal operations the epoll private "ep->sem" will guarantee
75 * a greater scalability.
76 */
77
78
Linus Torvalds1da177e2005-04-16 15:20:36 -070079#define DEBUG_EPOLL 0
80
81#if DEBUG_EPOLL > 0
82#define DPRINTK(x) printk x
83#define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
84#else /* #if DEBUG_EPOLL > 0 */
85#define DPRINTK(x) (void) 0
86#define DNPRINTK(n, x) (void) 0
87#endif /* #if DEBUG_EPOLL > 0 */
88
89#define DEBUG_EPI 0
90
91#if DEBUG_EPI != 0
92#define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
93#else /* #if DEBUG_EPI != 0 */
94#define EPI_SLAB_DEBUG 0
95#endif /* #if DEBUG_EPI != 0 */
96
97/* Epoll private bits inside the event mask */
98#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
99
100/* Maximum number of poll wake up nests we are allowing */
101#define EP_MAX_POLLWAKE_NESTS 4
102
Davide Libenzie3306dd2005-09-27 21:45:33 -0700103/* Maximum msec timeout value storeable in a long int */
104#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
105
Davide Libenzib6119672006-10-11 01:21:44 -0700106#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
107
Davide Libenzie3306dd2005-09-27 21:45:33 -0700108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109struct epoll_filefd {
110 struct file *file;
111 int fd;
112};
113
114/*
115 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
116 * It is used to keep track on all tasks that are currently inside the wake_up() code
117 * to 1) short-circuit the one coming from the same task and same wait queue head
118 * ( loop ) 2) allow a maximum number of epoll descriptors inclusion nesting
119 * 3) let go the ones coming from other tasks.
120 */
121struct wake_task_node {
122 struct list_head llink;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700123 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 wait_queue_head_t *wq;
125};
126
127/*
128 * This is used to implement the safe poll wake up avoiding to reenter
129 * the poll callback from inside wake_up().
130 */
131struct poll_safewake {
132 struct list_head wake_task_list;
133 spinlock_t lock;
134};
135
136/*
137 * This structure is stored inside the "private_data" member of the file
138 * structure and rapresent the main data sructure for the eventpoll
139 * interface.
140 */
141struct eventpoll {
142 /* Protect the this structure access */
143 rwlock_t lock;
144
145 /*
146 * This semaphore is used to ensure that files are not removed
147 * while epoll is using them. This is read-held during the event
148 * collection loop and it is write-held during the file cleanup
149 * path, the epoll file exit code and the ctl operations.
150 */
151 struct rw_semaphore sem;
152
153 /* Wait queue used by sys_epoll_wait() */
154 wait_queue_head_t wq;
155
156 /* Wait queue used by file->poll() */
157 wait_queue_head_t poll_wait;
158
159 /* List of ready file descriptors */
160 struct list_head rdllist;
161
162 /* RB-Tree root used to store monitored fd structs */
163 struct rb_root rbr;
164};
165
166/* Wait structure used by the poll hooks */
167struct eppoll_entry {
168 /* List header used to link this structure to the "struct epitem" */
169 struct list_head llink;
170
171 /* The "base" pointer is set to the container "struct epitem" */
172 void *base;
173
174 /*
175 * Wait queue item that will be linked to the target file wait
176 * queue head.
177 */
178 wait_queue_t wait;
179
180 /* The wait queue head that linked the "wait" wait queue item */
181 wait_queue_head_t *whead;
182};
183
184/*
185 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700186 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187 */
188struct epitem {
189 /* RB-Tree node used to link this structure to the eventpoll rb-tree */
190 struct rb_node rbn;
191
192 /* List header used to link this structure to the eventpoll ready list */
193 struct list_head rdllink;
194
195 /* The file descriptor information this item refers to */
196 struct epoll_filefd ffd;
197
198 /* Number of active wait queue attached to poll operations */
199 int nwait;
200
201 /* List containing poll wait queues */
202 struct list_head pwqlist;
203
204 /* The "container" of this item */
205 struct eventpoll *ep;
206
207 /* The structure that describe the interested events and the source fd */
208 struct epoll_event event;
209
210 /*
211 * Used to keep track of the usage count of the structure. This avoids
212 * that the structure will desappear from underneath our processing.
213 */
214 atomic_t usecnt;
215
216 /* List header used to link this item to the "struct file" items list */
217 struct list_head fllink;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218};
219
220/* Wrapper struct used by poll queueing */
221struct ep_pqueue {
222 poll_table pt;
223 struct epitem *epi;
224};
225
226
227
228static void ep_poll_safewake_init(struct poll_safewake *psw);
229static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq);
Davide Libenzi53d2be72005-09-16 19:28:06 -0700230static int ep_alloc(struct eventpoll **pep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231static void ep_free(struct eventpoll *ep);
232static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd);
233static void ep_use_epitem(struct epitem *epi);
234static void ep_release_epitem(struct epitem *epi);
235static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
236 poll_table *pt);
237static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi);
238static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
239 struct file *tfile, int fd);
240static int ep_modify(struct eventpoll *ep, struct epitem *epi,
241 struct epoll_event *event);
242static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi);
243static int ep_unlink(struct eventpoll *ep, struct epitem *epi);
244static int ep_remove(struct eventpoll *ep, struct epitem *epi);
245static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key);
246static int ep_eventpoll_close(struct inode *inode, struct file *file);
247static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
Davide Libenzi6192bd52007-05-08 00:25:41 -0700249 struct epoll_event __user *events, int maxevents);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250static int ep_events_transfer(struct eventpoll *ep,
251 struct epoll_event __user *events,
252 int maxevents);
253static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
254 int maxevents, long timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255
256/*
257 * This semaphore is used to serialize ep_free() and eventpoll_release_file().
258 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800259static struct mutex epmutex;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
261/* Safe wake up implementation */
262static struct poll_safewake psw;
263
264/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800265static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
267/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800268static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270/* File callbacks that implement the eventpoll file behaviour */
Arjan van de Ven4b6f5d22006-03-28 01:56:42 -0800271static const struct file_operations eventpoll_fops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 .release = ep_eventpoll_close,
273 .poll = ep_eventpoll_poll
274};
275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276
277
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700278/* Fast test to see if the file is an evenpoll file */
279static inline int is_file_epoll(struct file *f)
280{
281 return f->f_op == &eventpoll_fops;
282}
283
284/* Setup the structure that is used as key for the rb-tree */
285static inline void ep_set_ffd(struct epoll_filefd *ffd,
286 struct file *file, int fd)
287{
288 ffd->file = file;
289 ffd->fd = fd;
290}
291
292/* Compare rb-tree keys */
293static inline int ep_cmp_ffd(struct epoll_filefd *p1,
294 struct epoll_filefd *p2)
295{
296 return (p1->file > p2->file ? +1:
297 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
298}
299
300/* Special initialization for the rb-tree node to detect linkage */
301static inline void ep_rb_initnode(struct rb_node *n)
302{
David Woodhousec5698822006-04-21 13:17:24 +0100303 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700304}
305
306/* Removes a node from the rb-tree and marks it for a fast is-linked check */
307static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
308{
309 rb_erase(n, r);
David Woodhousec5698822006-04-21 13:17:24 +0100310 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700311}
312
313/* Fast check to verify that the item is linked to the main rb-tree */
314static inline int ep_rb_linked(struct rb_node *n)
315{
David Woodhousec5698822006-04-21 13:17:24 +0100316 return rb_parent(n) != n;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700317}
318
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700319/* Tells us if the item is currently linked */
320static inline int ep_is_linked(struct list_head *p)
321{
322 return !list_empty(p);
323}
324
325/* Get the "struct epitem" from a wait queue pointer */
326static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
327{
328 return container_of(p, struct eppoll_entry, wait)->base;
329}
330
331/* Get the "struct epitem" from an epoll queue wrapper */
332static inline struct epitem * ep_item_from_epqueue(poll_table *p)
333{
334 return container_of(p, struct ep_pqueue, pt)->epi;
335}
336
337/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700338static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700339{
340 return op != EPOLL_CTL_DEL;
341}
342
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343/* Initialize the poll safe wake up structure */
344static void ep_poll_safewake_init(struct poll_safewake *psw)
345{
346
347 INIT_LIST_HEAD(&psw->wake_task_list);
348 spin_lock_init(&psw->lock);
349}
350
351
352/*
353 * Perform a safe wake up of the poll wait list. The problem is that
354 * with the new callback'd wake up system, it is possible that the
355 * poll callback is reentered from inside the call to wake_up() done
356 * on the poll wait queue head. The rule is that we cannot reenter the
357 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
358 * and we cannot reenter the same wait queue head at all. This will
359 * enable to have a hierarchy of epoll file descriptor of no more than
360 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
361 * because this one gets called by the poll callback, that in turn is called
362 * from inside a wake_up(), that might be called from irq context.
363 */
364static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
365{
366 int wake_nests = 0;
367 unsigned long flags;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700368 struct task_struct *this_task = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 struct list_head *lsthead = &psw->wake_task_list, *lnk;
370 struct wake_task_node *tncur;
371 struct wake_task_node tnode;
372
373 spin_lock_irqsave(&psw->lock, flags);
374
375 /* Try to see if the current task is already inside this wakeup call */
376 list_for_each(lnk, lsthead) {
377 tncur = list_entry(lnk, struct wake_task_node, llink);
378
379 if (tncur->wq == wq ||
380 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
381 /*
382 * Ops ... loop detected or maximum nest level reached.
383 * We abort this wake by breaking the cycle itself.
384 */
385 spin_unlock_irqrestore(&psw->lock, flags);
386 return;
387 }
388 }
389
390 /* Add the current task to the list */
391 tnode.task = this_task;
392 tnode.wq = wq;
393 list_add(&tnode.llink, lsthead);
394
395 spin_unlock_irqrestore(&psw->lock, flags);
396
397 /* Do really wake up now */
398 wake_up(wq);
399
400 /* Remove the current task from the list */
401 spin_lock_irqsave(&psw->lock, flags);
402 list_del(&tnode.llink);
403 spin_unlock_irqrestore(&psw->lock, flags);
404}
405
406
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407/*
408 * This is called from eventpoll_release() to unlink files from the eventpoll
409 * interface. We need to have this facility to cleanup correctly files that are
410 * closed without being removed from the eventpoll interface.
411 */
412void eventpoll_release_file(struct file *file)
413{
414 struct list_head *lsthead = &file->f_ep_links;
415 struct eventpoll *ep;
416 struct epitem *epi;
417
418 /*
419 * We don't want to get "file->f_ep_lock" because it is not
420 * necessary. It is not necessary because we're in the "struct file"
421 * cleanup path, and this means that noone is using this file anymore.
422 * The only hit might come from ep_free() but by holding the semaphore
423 * will correctly serialize the operation. We do need to acquire
Arjan van de Ven144efe32006-03-23 03:00:32 -0800424 * "ep->sem" after "epmutex" because ep_remove() requires it when called
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 * from anywhere but ep_free().
426 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800427 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
429 while (!list_empty(lsthead)) {
Pavel Emelianovb5e61812007-05-08 00:30:19 -0700430 epi = list_first_entry(lsthead, struct epitem, fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
432 ep = epi->ep;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700433 list_del_init(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 down_write(&ep->sem);
435 ep_remove(ep, epi);
436 up_write(&ep->sem);
437 }
438
Arjan van de Ven144efe32006-03-23 03:00:32 -0800439 mutex_unlock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440}
441
442
443/*
444 * It opens an eventpoll file descriptor by suggesting a storage of "size"
445 * file descriptors. The size parameter is just an hint about how to size
446 * data structures. It won't prevent the user to store more than "size"
447 * file descriptors inside the epoll interface. It is the kernel part of
448 * the userspace epoll_create(2).
449 */
450asmlinkage long sys_epoll_create(int size)
451{
Davide Libenzib6119672006-10-11 01:21:44 -0700452 int error, fd = -1;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700453 struct eventpoll *ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 struct inode *inode;
455 struct file *file;
456
457 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
458 current, size));
459
Davide Libenzi53d2be72005-09-16 19:28:06 -0700460 /*
461 * Sanity check on the size parameter, and create the internal data
462 * structure ( "struct eventpoll" ).
463 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 error = -EINVAL;
Davide Libenzi53d2be72005-09-16 19:28:06 -0700465 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 goto eexit_1;
467
468 /*
469 * Creates all the items needed to setup an eventpoll file. That is,
470 * a file structure, and inode and a free file descriptor.
471 */
Davide Libenzida66f7c2007-05-10 22:23:21 -0700472 error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]",
473 &eventpoll_fops, ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 if (error)
475 goto eexit_2;
476
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
478 current, size, fd));
479
480 return fd;
481
482eexit_2:
Davide Libenzi53d2be72005-09-16 19:28:06 -0700483 ep_free(ep);
484 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485eexit_1:
486 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
487 current, size, error));
488 return error;
489}
490
491
492/*
493 * The following function implements the controller interface for
494 * the eventpoll file that enables the insertion/removal/change of
495 * file descriptors inside the interest set. It represents
496 * the kernel part of the user space epoll_ctl(2).
497 */
498asmlinkage long
499sys_epoll_ctl(int epfd, int op, int fd, struct epoll_event __user *event)
500{
501 int error;
502 struct file *file, *tfile;
503 struct eventpoll *ep;
504 struct epitem *epi;
505 struct epoll_event epds;
506
507 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
508 current, epfd, op, fd, event));
509
510 error = -EFAULT;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700511 if (ep_op_has_event(op) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 copy_from_user(&epds, event, sizeof(struct epoll_event)))
513 goto eexit_1;
514
515 /* Get the "struct file *" for the eventpoll file */
516 error = -EBADF;
517 file = fget(epfd);
518 if (!file)
519 goto eexit_1;
520
521 /* Get the "struct file *" for the target file */
522 tfile = fget(fd);
523 if (!tfile)
524 goto eexit_2;
525
526 /* The target file descriptor must support poll */
527 error = -EPERM;
528 if (!tfile->f_op || !tfile->f_op->poll)
529 goto eexit_3;
530
531 /*
532 * We have to check that the file structure underneath the file descriptor
533 * the user passed to us _is_ an eventpoll file. And also we do not permit
534 * adding an epoll file descriptor inside itself.
535 */
536 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700537 if (file == tfile || !is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 goto eexit_3;
539
540 /*
541 * At this point it is safe to assume that the "private_data" contains
542 * our own data structure.
543 */
544 ep = file->private_data;
545
546 down_write(&ep->sem);
547
Davide Libenzi6192bd52007-05-08 00:25:41 -0700548 /* Try to lookup the file inside our RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 epi = ep_find(ep, tfile, fd);
550
551 error = -EINVAL;
552 switch (op) {
553 case EPOLL_CTL_ADD:
554 if (!epi) {
Davide Libenzi23951402006-04-10 22:54:12 -0700555 epds.events |= POLLERR | POLLHUP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
557 error = ep_insert(ep, &epds, tfile, fd);
558 } else
559 error = -EEXIST;
560 break;
561 case EPOLL_CTL_DEL:
562 if (epi)
563 error = ep_remove(ep, epi);
564 else
565 error = -ENOENT;
566 break;
567 case EPOLL_CTL_MOD:
568 if (epi) {
Davide Libenzi23951402006-04-10 22:54:12 -0700569 epds.events |= POLLERR | POLLHUP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 error = ep_modify(ep, epi, &epds);
571 } else
572 error = -ENOENT;
573 break;
574 }
575
576 /*
577 * The function ep_find() increments the usage count of the structure
578 * so, if this is not NULL, we need to release it.
579 */
580 if (epi)
581 ep_release_epitem(epi);
582
583 up_write(&ep->sem);
584
585eexit_3:
586 fput(tfile);
587eexit_2:
588 fput(file);
589eexit_1:
590 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
591 current, epfd, op, fd, event, error));
592
593 return error;
594}
595
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596
597/*
598 * Implement the event wait interface for the eventpoll file. It is the kernel
599 * part of the user space epoll_wait(2).
600 */
601asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
602 int maxevents, int timeout)
603{
604 int error;
605 struct file *file;
606 struct eventpoll *ep;
607
608 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
609 current, epfd, events, maxevents, timeout));
610
611 /* The maximum number of event must be greater than zero */
Davide Libenzib6119672006-10-11 01:21:44 -0700612 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 return -EINVAL;
614
615 /* Verify that the area passed by the user is writeable */
616 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
617 error = -EFAULT;
618 goto eexit_1;
619 }
620
621 /* Get the "struct file *" for the eventpoll file */
622 error = -EBADF;
623 file = fget(epfd);
624 if (!file)
625 goto eexit_1;
626
627 /*
628 * We have to check that the file structure underneath the fd
629 * the user passed to us _is_ an eventpoll file.
630 */
631 error = -EINVAL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700632 if (!is_file_epoll(file))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 goto eexit_2;
634
635 /*
636 * At this point it is safe to assume that the "private_data" contains
637 * our own data structure.
638 */
639 ep = file->private_data;
640
641 /* Time to fish for events ... */
642 error = ep_poll(ep, events, maxevents, timeout);
643
644eexit_2:
645 fput(file);
646eexit_1:
647 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
648 current, epfd, events, maxevents, timeout, error));
649
650 return error;
651}
652
653
Davide Libenzib6119672006-10-11 01:21:44 -0700654#ifdef TIF_RESTORE_SIGMASK
655
656/*
657 * Implement the event wait interface for the eventpoll file. It is the kernel
658 * part of the user space epoll_pwait(2).
659 */
660asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
661 int maxevents, int timeout, const sigset_t __user *sigmask,
662 size_t sigsetsize)
663{
664 int error;
665 sigset_t ksigmask, sigsaved;
666
667 /*
668 * If the caller wants a certain signal mask to be set during the wait,
669 * we apply it here.
670 */
671 if (sigmask) {
672 if (sigsetsize != sizeof(sigset_t))
673 return -EINVAL;
674 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
675 return -EFAULT;
676 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
677 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
678 }
679
680 error = sys_epoll_wait(epfd, events, maxevents, timeout);
681
682 /*
683 * If we changed the signal mask, we need to restore the original one.
684 * In case we've got a signal while waiting, we do not restore the
685 * signal mask yet, and we allow do_signal() to deliver the signal on
686 * the way back to userspace, before the signal mask is restored.
687 */
688 if (sigmask) {
689 if (error == -EINTR) {
690 memcpy(&current->saved_sigmask, &sigsaved,
691 sizeof(sigsaved));
692 set_thread_flag(TIF_RESTORE_SIGMASK);
693 } else
694 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
695 }
696
697 return error;
698}
699
700#endif /* #ifdef TIF_RESTORE_SIGMASK */
701
702
Davide Libenzi53d2be72005-09-16 19:28:06 -0700703static int ep_alloc(struct eventpoll **pep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704{
Davide Libenzi53d2be72005-09-16 19:28:06 -0700705 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
Davide Libenzi53d2be72005-09-16 19:28:06 -0700707 if (!ep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 return -ENOMEM;
709
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 rwlock_init(&ep->lock);
711 init_rwsem(&ep->sem);
712 init_waitqueue_head(&ep->wq);
713 init_waitqueue_head(&ep->poll_wait);
714 INIT_LIST_HEAD(&ep->rdllist);
715 ep->rbr = RB_ROOT;
716
Davide Libenzi53d2be72005-09-16 19:28:06 -0700717 *pep = ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
Davide Libenzi53d2be72005-09-16 19:28:06 -0700719 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 current, ep));
721 return 0;
722}
723
724
725static void ep_free(struct eventpoll *ep)
726{
727 struct rb_node *rbp;
728 struct epitem *epi;
729
730 /* We need to release all tasks waiting for these file */
731 if (waitqueue_active(&ep->poll_wait))
732 ep_poll_safewake(&psw, &ep->poll_wait);
733
734 /*
735 * We need to lock this because we could be hit by
736 * eventpoll_release_file() while we're freeing the "struct eventpoll".
737 * We do not need to hold "ep->sem" here because the epoll file
738 * is on the way to be removed and no one has references to it
739 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800740 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800742 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743
744 /*
745 * Walks through the whole tree by unregistering poll callbacks.
746 */
747 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
748 epi = rb_entry(rbp, struct epitem, rbn);
749
750 ep_unregister_pollwait(ep, epi);
751 }
752
753 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700754 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 * point we are sure no poll callbacks will be lingering around, and also by
756 * write-holding "sem" we can be sure that no file cleanup code will hit
757 * us during this operation. So we can avoid the lock on "ep->lock".
758 */
759 while ((rbp = rb_first(&ep->rbr)) != 0) {
760 epi = rb_entry(rbp, struct epitem, rbn);
761 ep_remove(ep, epi);
762 }
763
Arjan van de Ven144efe32006-03-23 03:00:32 -0800764 mutex_unlock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765}
766
767
768/*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700769 * Search the file inside the eventpoll tree. It add usage count to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 * the returned item, so the caller must call ep_release_epitem()
771 * after finished using the "struct epitem".
772 */
773static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
774{
775 int kcmp;
776 unsigned long flags;
777 struct rb_node *rbp;
778 struct epitem *epi, *epir = NULL;
779 struct epoll_filefd ffd;
780
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700781 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 read_lock_irqsave(&ep->lock, flags);
783 for (rbp = ep->rbr.rb_node; rbp; ) {
784 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700785 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 if (kcmp > 0)
787 rbp = rbp->rb_right;
788 else if (kcmp < 0)
789 rbp = rbp->rb_left;
790 else {
791 ep_use_epitem(epi);
792 epir = epi;
793 break;
794 }
795 }
796 read_unlock_irqrestore(&ep->lock, flags);
797
798 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
799 current, file, epir));
800
801 return epir;
802}
803
804
805/*
806 * Increment the usage count of the "struct epitem" making it sure
807 * that the user will have a valid pointer to reference.
808 */
809static void ep_use_epitem(struct epitem *epi)
810{
811
812 atomic_inc(&epi->usecnt);
813}
814
815
816/*
817 * Decrement ( release ) the usage count by signaling that the user
818 * has finished using the structure. It might lead to freeing the
819 * structure itself if the count goes to zero.
820 */
821static void ep_release_epitem(struct epitem *epi)
822{
823
824 if (atomic_dec_and_test(&epi->usecnt))
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700825 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826}
827
828
829/*
830 * This is the callback that is used to add our wait queue to the
831 * target file wakeup lists.
832 */
833static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
834 poll_table *pt)
835{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700836 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 struct eppoll_entry *pwq;
838
Christoph Lametere94b1762006-12-06 20:33:17 -0800839 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
841 pwq->whead = whead;
842 pwq->base = epi;
843 add_wait_queue(whead, &pwq->wait);
844 list_add_tail(&pwq->llink, &epi->pwqlist);
845 epi->nwait++;
846 } else {
847 /* We have to signal that an error occurred */
848 epi->nwait = -1;
849 }
850}
851
852
853static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
854{
855 int kcmp;
856 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
857 struct epitem *epic;
858
859 while (*p) {
860 parent = *p;
861 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700862 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 if (kcmp > 0)
864 p = &parent->rb_right;
865 else
866 p = &parent->rb_left;
867 }
868 rb_link_node(&epi->rbn, parent, p);
869 rb_insert_color(&epi->rbn, &ep->rbr);
870}
871
872
873static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
874 struct file *tfile, int fd)
875{
876 int error, revents, pwake = 0;
877 unsigned long flags;
878 struct epitem *epi;
879 struct ep_pqueue epq;
880
881 error = -ENOMEM;
Christoph Lametere94b1762006-12-06 20:33:17 -0800882 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 goto eexit_1;
884
885 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700886 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887 INIT_LIST_HEAD(&epi->rdllink);
888 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 INIT_LIST_HEAD(&epi->pwqlist);
890 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700891 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 epi->event = *event;
893 atomic_set(&epi->usecnt, 1);
894 epi->nwait = 0;
895
896 /* Initialize the poll table using the queue callback */
897 epq.epi = epi;
898 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
899
900 /*
901 * Attach the item to the poll hooks and get current event bits.
902 * We can safely use the file* here because its usage count has
903 * been increased by the caller of this function.
904 */
905 revents = tfile->f_op->poll(tfile, &epq.pt);
906
907 /*
908 * We have to check if something went wrong during the poll wait queue
909 * install process. Namely an allocation for a wait queue failed due
910 * high memory pressure.
911 */
912 if (epi->nwait < 0)
913 goto eexit_2;
914
915 /* Add the current item to the list of active epoll hook for this file */
916 spin_lock(&tfile->f_ep_lock);
917 list_add_tail(&epi->fllink, &tfile->f_ep_links);
918 spin_unlock(&tfile->f_ep_lock);
919
920 /* We have to drop the new item inside our item list to keep track of it */
921 write_lock_irqsave(&ep->lock, flags);
922
923 /* Add the current item to the rb-tree */
924 ep_rbtree_insert(ep, epi);
925
926 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700927 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 list_add_tail(&epi->rdllink, &ep->rdllist);
929
930 /* Notify waiting tasks that events are available */
931 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -0700932 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 if (waitqueue_active(&ep->poll_wait))
934 pwake++;
935 }
936
937 write_unlock_irqrestore(&ep->lock, flags);
938
939 /* We have to call this outside the lock */
940 if (pwake)
941 ep_poll_safewake(&psw, &ep->poll_wait);
942
943 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
944 current, ep, tfile, fd));
945
946 return 0;
947
948eexit_2:
949 ep_unregister_pollwait(ep, epi);
950
951 /*
952 * We need to do this because an event could have been arrived on some
953 * allocated wait queue.
954 */
955 write_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700956 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700957 list_del_init(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 write_unlock_irqrestore(&ep->lock, flags);
959
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700960 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961eexit_1:
962 return error;
963}
964
965
966/*
967 * Modify the interest event mask by dropping an event if the new mask
968 * has a match in the current file status.
969 */
970static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
971{
972 int pwake = 0;
973 unsigned int revents;
974 unsigned long flags;
975
976 /*
977 * Set the new event interest mask before calling f_op->poll(), otherwise
978 * a potential race might occur. In fact if we do this operation inside
979 * the lock, an event might happen between the f_op->poll() call and the
980 * new event set registering.
981 */
982 epi->event.events = event->events;
983
984 /*
985 * Get current event bits. We can safely use the file* here because
986 * its usage count has been increased by the caller of this function.
987 */
988 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
989
990 write_lock_irqsave(&ep->lock, flags);
991
992 /* Copy the data member from inside the lock */
993 epi->event.data = event->data;
994
995 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700996 * If the item is not linked to the RB tree it means that it's on its
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 * way toward the removal. Do nothing in this case.
998 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700999 if (ep_rb_linked(&epi->rbn)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 /*
1001 * If the item is "hot" and it is not registered inside the ready
1002 * list, push it inside. If the item is not "hot" and it is currently
1003 * registered inside the ready list, unlink it.
1004 */
1005 if (revents & event->events) {
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001006 if (!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))
Davide Libenzi3419b232006-06-25 05:48:14 -07001011 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1012 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 if (waitqueue_active(&ep->poll_wait))
1014 pwake++;
1015 }
1016 }
1017 }
1018
1019 write_unlock_irqrestore(&ep->lock, flags);
1020
1021 /* We have to call this outside the lock */
1022 if (pwake)
1023 ep_poll_safewake(&psw, &ep->poll_wait);
1024
1025 return 0;
1026}
1027
1028
1029/*
1030 * This function unregister poll callbacks from the associated file descriptor.
1031 * Since this must be called without holding "ep->lock" the atomic exchange trick
1032 * will protect us from multiple unregister.
1033 */
1034static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
1035{
1036 int nwait;
1037 struct list_head *lsthead = &epi->pwqlist;
1038 struct eppoll_entry *pwq;
1039
1040 /* This is called without locks, so we need the atomic exchange */
1041 nwait = xchg(&epi->nwait, 0);
1042
1043 if (nwait) {
1044 while (!list_empty(lsthead)) {
Pavel Emelianovb5e61812007-05-08 00:30:19 -07001045 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046
Davide Libenzi6192bd52007-05-08 00:25:41 -07001047 list_del_init(&pwq->llink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 remove_wait_queue(pwq->whead, &pwq->wait);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001049 kmem_cache_free(pwq_cache, pwq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 }
1051 }
1052}
1053
1054
1055/*
1056 * Unlink the "struct epitem" from all places it might have been hooked up.
1057 * This function must be called with write IRQ lock on "ep->lock".
1058 */
1059static int ep_unlink(struct eventpoll *ep, struct epitem *epi)
1060{
1061 int error;
1062
1063 /*
1064 * It can happen that this one is called for an item already unlinked.
1065 * The check protect us from doing a double unlink ( crash ).
1066 */
1067 error = -ENOENT;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001068 if (!ep_rb_linked(&epi->rbn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 goto eexit_1;
1070
1071 /*
1072 * Clear the event mask for the unlinked item. This will avoid item
1073 * notifications to be sent after the unlink operation from inside
1074 * the kernel->userspace event transfer loop.
1075 */
1076 epi->event.events = 0;
1077
1078 /*
1079 * At this point is safe to do the job, unlink the item from our rb-tree.
1080 * This operation togheter with the above check closes the door to
1081 * double unlinks.
1082 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001083 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
1085 /*
1086 * If the item we are going to remove is inside the ready file descriptors
1087 * we want to remove it from this list to avoid stale events.
1088 */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001089 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001090 list_del_init(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091
1092 error = 0;
1093eexit_1:
1094
1095 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_unlink(%p, %p) = %d\n",
Masoud Asgharifard Sharbiani45f17e02006-08-27 01:23:48 -07001096 current, ep, epi->ffd.file, error));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097
1098 return error;
1099}
1100
1101
1102/*
Davide Libenzi6192bd52007-05-08 00:25:41 -07001103 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 * all the associated resources.
1105 */
1106static int ep_remove(struct eventpoll *ep, struct epitem *epi)
1107{
1108 int error;
1109 unsigned long flags;
1110 struct file *file = epi->ffd.file;
1111
1112 /*
1113 * Removes poll wait queue hooks. We _have_ to do this without holding
1114 * the "ep->lock" otherwise a deadlock might occur. This because of the
1115 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
1116 * queue head lock when unregistering the wait queue. The wakeup callback
1117 * will run by holding the wait queue head lock and will call our callback
1118 * that will try to get "ep->lock".
1119 */
1120 ep_unregister_pollwait(ep, epi);
1121
1122 /* Remove the current item from the list of epoll hooks */
1123 spin_lock(&file->f_ep_lock);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001124 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001125 list_del_init(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 spin_unlock(&file->f_ep_lock);
1127
1128 /* We need to acquire the write IRQ lock before calling ep_unlink() */
1129 write_lock_irqsave(&ep->lock, flags);
1130
Davide Libenzi6192bd52007-05-08 00:25:41 -07001131 /* Really unlink the item from the RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 error = ep_unlink(ep, epi);
1133
1134 write_unlock_irqrestore(&ep->lock, flags);
1135
1136 if (error)
1137 goto eexit_1;
1138
1139 /* At this point it is safe to free the eventpoll item */
1140 ep_release_epitem(epi);
1141
1142 error = 0;
1143eexit_1:
1144 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p) = %d\n",
1145 current, ep, file, error));
1146
1147 return error;
1148}
1149
1150
1151/*
1152 * This is the callback that is passed to the wait queue wakeup
1153 * machanism. It is called by the stored file descriptors when they
1154 * have events to report.
1155 */
1156static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1157{
1158 int pwake = 0;
1159 unsigned long flags;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001160 struct epitem *epi = ep_item_from_wait(wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 struct eventpoll *ep = epi->ep;
1162
1163 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
Masoud Asgharifard Sharbiani45f17e02006-08-27 01:23:48 -07001164 current, epi->ffd.file, epi, ep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
1166 write_lock_irqsave(&ep->lock, flags);
1167
1168 /*
1169 * If the event mask does not contain any poll(2) event, we consider the
1170 * descriptor to be disabled. This condition is likely the effect of the
1171 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1172 * until the next EPOLL_CTL_MOD will be issued.
1173 */
1174 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1175 goto is_disabled;
1176
1177 /* If this file is already in the ready list we exit soon */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001178 if (ep_is_linked(&epi->rdllink))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 goto is_linked;
1180
1181 list_add_tail(&epi->rdllink, &ep->rdllist);
1182
1183is_linked:
1184 /*
1185 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1186 * wait list.
1187 */
1188 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001189 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1190 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 if (waitqueue_active(&ep->poll_wait))
1192 pwake++;
1193
1194is_disabled:
1195 write_unlock_irqrestore(&ep->lock, flags);
1196
1197 /* We have to call this outside the lock */
1198 if (pwake)
1199 ep_poll_safewake(&psw, &ep->poll_wait);
1200
1201 return 1;
1202}
1203
1204
1205static int ep_eventpoll_close(struct inode *inode, struct file *file)
1206{
1207 struct eventpoll *ep = file->private_data;
1208
1209 if (ep) {
1210 ep_free(ep);
1211 kfree(ep);
1212 }
1213
1214 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
1215 return 0;
1216}
1217
1218
1219static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
1220{
1221 unsigned int pollflags = 0;
1222 unsigned long flags;
1223 struct eventpoll *ep = file->private_data;
1224
1225 /* Insert inside our poll wait queue */
1226 poll_wait(file, &ep->poll_wait, wait);
1227
1228 /* Check our condition */
1229 read_lock_irqsave(&ep->lock, flags);
1230 if (!list_empty(&ep->rdllist))
1231 pollflags = POLLIN | POLLRDNORM;
1232 read_unlock_irqrestore(&ep->lock, flags);
1233
1234 return pollflags;
1235}
1236
1237
1238/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 * This function is called without holding the "ep->lock" since the call to
1240 * __copy_to_user() might sleep, and also f_op->poll() might reenable the IRQ
1241 * because of the way poll() is traditionally implemented in Linux.
1242 */
1243static int ep_send_events(struct eventpoll *ep, struct list_head *txlist,
Davide Libenzi6192bd52007-05-08 00:25:41 -07001244 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245{
Davide Libenzi6192bd52007-05-08 00:25:41 -07001246 int eventcnt, error = -EFAULT, pwake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 unsigned int revents;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001248 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 struct epitem *epi;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001250 struct list_head injlist;
1251
1252 INIT_LIST_HEAD(&injlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
1254 /*
1255 * We can loop without lock because this is a task private list.
Davide Libenzi6192bd52007-05-08 00:25:41 -07001256 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
1257 * Items cannot vanish during the loop because we are holding "sem" in
1258 * read.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 */
Davide Libenzi6192bd52007-05-08 00:25:41 -07001260 for (eventcnt = 0; !list_empty(txlist) && eventcnt < maxevents;) {
Pavel Emelianovb5e61812007-05-08 00:30:19 -07001261 epi = list_first_entry(txlist, struct epitem, rdllink);
Davide Libenzi6192bd52007-05-08 00:25:41 -07001262 prefetch(epi->rdllink.next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263
1264 /*
1265 * Get the ready file event set. We can safely use the file
1266 * because we are holding the "sem" in read and this will
1267 * guarantee that both the file and the item will not vanish.
1268 */
1269 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
Davide Libenzi6192bd52007-05-08 00:25:41 -07001270 revents &= epi->event.events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271
1272 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -07001273 * Is the event mask intersect the caller-requested one,
1274 * deliver the event to userspace. Again, we are holding
1275 * "sem" in read, so no operations coming from userspace
1276 * can change the item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 */
Davide Libenzi6192bd52007-05-08 00:25:41 -07001278 if (revents) {
1279 if (__put_user(revents,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 &events[eventcnt].events) ||
1281 __put_user(epi->event.data,
1282 &events[eventcnt].data))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001283 goto errxit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 if (epi->event.events & EPOLLONESHOT)
1285 epi->event.events &= EP_PRIVATE_BITS;
1286 eventcnt++;
1287 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288
1289 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -07001290 * This is tricky. We are holding the "sem" in read, and this
1291 * means that the operations that can change the "linked" status
1292 * of the epoll item (epi->rbn and epi->rdllink), cannot touch
1293 * them. Also, since we are "linked" from a epi->rdllink POV
1294 * (the item is linked to our transmission list we just
1295 * spliced), the ep_poll_callback() cannot touch us either,
1296 * because of the check present in there. Another parallel
1297 * epoll_wait() will not get the same result set, since we
1298 * spliced the ready list before. Note that list_del() still
1299 * shows the item as linked to the test in ep_poll_callback().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 */
Davide Libenzi6192bd52007-05-08 00:25:41 -07001301 list_del(&epi->rdllink);
1302 if (!(epi->event.events & EPOLLET) &&
1303 (revents & epi->event.events))
1304 list_add_tail(&epi->rdllink, &injlist);
1305 else {
1306 /*
1307 * Be sure the item is totally detached before re-init
1308 * the list_head. After INIT_LIST_HEAD() is committed,
1309 * the ep_poll_callback() can requeue the item again,
1310 * but we don't care since we are already past it.
1311 */
1312 smp_mb();
1313 INIT_LIST_HEAD(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 }
1315 }
Davide Libenzi6192bd52007-05-08 00:25:41 -07001316 error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317
Davide Libenzi6192bd52007-05-08 00:25:41 -07001318 errxit:
1319
1320 /*
1321 * If the re-injection list or the txlist are not empty, re-splice
1322 * them to the ready list and do proper wakeups.
1323 */
1324 if (!list_empty(&injlist) || !list_empty(txlist)) {
1325 write_lock_irqsave(&ep->lock, flags);
1326
1327 list_splice(txlist, &ep->rdllist);
1328 list_splice(&injlist, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 /*
1330 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1331 * wait list.
1332 */
1333 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -07001334 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
1335 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 if (waitqueue_active(&ep->poll_wait))
1337 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338
Davide Libenzi6192bd52007-05-08 00:25:41 -07001339 write_unlock_irqrestore(&ep->lock, flags);
1340 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341
1342 /* We have to call this outside the lock */
1343 if (pwake)
1344 ep_poll_safewake(&psw, &ep->poll_wait);
Davide Libenzi6192bd52007-05-08 00:25:41 -07001345
1346 return eventcnt == 0 ? error: eventcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347}
1348
1349
1350/*
1351 * Perform the transfer of events to user space.
1352 */
1353static int ep_events_transfer(struct eventpoll *ep,
1354 struct epoll_event __user *events, int maxevents)
1355{
Davide Libenzi6192bd52007-05-08 00:25:41 -07001356 int eventcnt;
1357 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 struct list_head txlist;
1359
1360 INIT_LIST_HEAD(&txlist);
1361
1362 /*
1363 * We need to lock this because we could be hit by
1364 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
1365 */
1366 down_read(&ep->sem);
1367
Davide Libenzi6192bd52007-05-08 00:25:41 -07001368 /*
1369 * Steal the ready list, and re-init the original one to the
1370 * empty list.
1371 */
1372 write_lock_irqsave(&ep->lock, flags);
1373 list_splice(&ep->rdllist, &txlist);
1374 INIT_LIST_HEAD(&ep->rdllist);
1375 write_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
Davide Libenzi6192bd52007-05-08 00:25:41 -07001377 /* Build result set in userspace */
1378 eventcnt = ep_send_events(ep, &txlist, events, maxevents);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379
1380 up_read(&ep->sem);
1381
1382 return eventcnt;
1383}
1384
1385
1386static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1387 int maxevents, long timeout)
1388{
1389 int res, eavail;
1390 unsigned long flags;
1391 long jtimeout;
1392 wait_queue_t wait;
1393
1394 /*
1395 * Calculate the timeout by checking for the "infinite" value ( -1 )
1396 * and the overflow condition. The passed timeout is in milliseconds,
1397 * that why (t * HZ) / 1000.
1398 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001399 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1400 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401
1402retry:
1403 write_lock_irqsave(&ep->lock, flags);
1404
1405 res = 0;
1406 if (list_empty(&ep->rdllist)) {
1407 /*
1408 * We don't have any available event to return to the caller.
1409 * We need to sleep here, and we will be wake up by
1410 * ep_poll_callback() when events will become available.
1411 */
1412 init_waitqueue_entry(&wait, current);
Davide Libenzi3419b232006-06-25 05:48:14 -07001413 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414
1415 for (;;) {
1416 /*
1417 * We don't want to sleep if the ep_poll_callback() sends us
1418 * a wakeup in between. That's why we set the task state
1419 * to TASK_INTERRUPTIBLE before doing the checks.
1420 */
1421 set_current_state(TASK_INTERRUPTIBLE);
1422 if (!list_empty(&ep->rdllist) || !jtimeout)
1423 break;
1424 if (signal_pending(current)) {
1425 res = -EINTR;
1426 break;
1427 }
1428
1429 write_unlock_irqrestore(&ep->lock, flags);
1430 jtimeout = schedule_timeout(jtimeout);
1431 write_lock_irqsave(&ep->lock, flags);
1432 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001433 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
1435 set_current_state(TASK_RUNNING);
1436 }
1437
1438 /* Is it worth to try to dig for events ? */
1439 eavail = !list_empty(&ep->rdllist);
1440
1441 write_unlock_irqrestore(&ep->lock, flags);
1442
1443 /*
1444 * Try to transfer events to user space. In case we get 0 events and
1445 * there's still timeout left over, we go trying again in search of
1446 * more luck.
1447 */
1448 if (!res && eavail &&
1449 !(res = ep_events_transfer(ep, events, maxevents)) && jtimeout)
1450 goto retry;
1451
1452 return res;
1453}
1454
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455static int __init eventpoll_init(void)
1456{
Arjan van de Ven144efe32006-03-23 03:00:32 -08001457 mutex_init(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458
1459 /* Initialize the structure used to perform safe poll wait head wake ups */
1460 ep_poll_safewake_init(&psw);
1461
1462 /* Allocates slab cache used to allocate "struct epitem" items */
1463 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1464 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1465 NULL, NULL);
1466
1467 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1468 pwq_cache = kmem_cache_create("eventpoll_pwq",
1469 sizeof(struct eppoll_entry), 0,
1470 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473}
1474
1475
1476static void __exit eventpoll_exit(void)
1477{
1478 /* Undo all operations done inside eventpoll_init() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 kmem_cache_destroy(pwq_cache);
1480 kmem_cache_destroy(epi_cache);
1481}
1482
1483module_init(eventpoll_init);
1484module_exit(eventpoll_exit);
1485
1486MODULE_LICENSE("GPL");