blob: 011b9b8c90c6976bb15e5bf50ffaf865fd31892b [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Davide Libenzic7ea7632007-05-15 01:40:47 -07002 * fs/eventpoll.c (Efficent event polling implementation)
3 * Copyright (C) 2001,...,2007 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
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/init.h>
15#include <linux/kernel.h>
16#include <linux/sched.h>
17#include <linux/fs.h>
18#include <linux/file.h>
19#include <linux/signal.h>
20#include <linux/errno.h>
21#include <linux/mm.h>
22#include <linux/slab.h>
23#include <linux/poll.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/string.h>
25#include <linux/list.h>
26#include <linux/hash.h>
27#include <linux/spinlock.h>
28#include <linux/syscalls.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/rbtree.h>
30#include <linux/wait.h>
31#include <linux/eventpoll.h>
32#include <linux/mount.h>
33#include <linux/bitops.h>
Arjan van de Ven144efe32006-03-23 03:00:32 -080034#include <linux/mutex.h>
Davide Libenzida66f7c2007-05-10 22:23:21 -070035#include <linux/anon_inodes.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <asm/uaccess.h>
37#include <asm/system.h>
38#include <asm/io.h>
39#include <asm/mman.h>
40#include <asm/atomic.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
Linus Torvalds1da177e2005-04-16 15:20:36 -070042/*
43 * LOCKING:
44 * There are three level of locking required by epoll :
45 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080046 * 1) epmutex (mutex)
Davide Libenzic7ea7632007-05-15 01:40:47 -070047 * 2) ep->mtx (mutex)
48 * 3) ep->lock (spinlock)
Linus Torvalds1da177e2005-04-16 15:20:36 -070049 *
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a spinlock (ep->lock) because we manipulate objects
52 * from inside the poll callback, that might be triggered from
53 * a wake_up() that in turn might be called from IRQ context.
54 * So we can't sleep inside the poll callback and hence we need
55 * a spinlock. During the event transfer loop (from kernel to
56 * user space) we could end up sleeping due a copy_to_user(), so
57 * we need a lock that will allow us to sleep. This lock is a
Davide Libenzid47de162007-05-15 01:40:41 -070058 * mutex (ep->mtx). It is acquired during the event transfer loop,
59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60 * Then we also need a global mutex to serialize eventpoll_release_file()
61 * and ep_free().
62 * This mutex is acquired by ep_free() during the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
Davide Libenzid47de162007-05-15 01:40:41 -070066 * It is possible to drop the "ep->mtx" and to use the global
67 * mutex "epmutex" (together with "ep->lock") to have it working,
68 * but having "ep->mtx" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080069 * Events that require holding "epmutex" are very rare, while for
Davide Libenzid47de162007-05-15 01:40:41 -070070 * normal operations the epoll private "ep->mtx" will guarantee
71 * a better scalability.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 */
73
Linus Torvalds1da177e2005-04-16 15:20:36 -070074#define DEBUG_EPOLL 0
75
76#if DEBUG_EPOLL > 0
77#define DPRINTK(x) printk x
78#define DNPRINTK(n, x) do { if ((n) <= DEBUG_EPOLL) printk x; } while (0)
79#else /* #if DEBUG_EPOLL > 0 */
80#define DPRINTK(x) (void) 0
81#define DNPRINTK(n, x) (void) 0
82#endif /* #if DEBUG_EPOLL > 0 */
83
84#define DEBUG_EPI 0
85
86#if DEBUG_EPI != 0
87#define EPI_SLAB_DEBUG (SLAB_DEBUG_FREE | SLAB_RED_ZONE /* | SLAB_POISON */)
88#else /* #if DEBUG_EPI != 0 */
89#define EPI_SLAB_DEBUG 0
90#endif /* #if DEBUG_EPI != 0 */
91
92/* Epoll private bits inside the event mask */
93#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
94
95/* Maximum number of poll wake up nests we are allowing */
96#define EP_MAX_POLLWAKE_NESTS 4
97
Davide Libenzie3306dd2005-09-27 21:45:33 -070098/* Maximum msec timeout value storeable in a long int */
99#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
100
Davide Libenzib6119672006-10-11 01:21:44 -0700101#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
102
Davide Libenzid47de162007-05-15 01:40:41 -0700103#define EP_UNACTIVE_PTR ((void *) -1L)
104
Davide Libenzi7ef99642008-12-01 13:13:55 -0800105#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
106
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107struct epoll_filefd {
108 struct file *file;
109 int fd;
110};
111
112/*
113 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
114 * It is used to keep track on all tasks that are currently inside the wake_up() code
115 * to 1) short-circuit the one coming from the same task and same wait queue head
Davide Libenzid47de162007-05-15 01:40:41 -0700116 * (loop) 2) allow a maximum number of epoll descriptors inclusion nesting
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 * 3) let go the ones coming from other tasks.
118 */
119struct wake_task_node {
120 struct list_head llink;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700121 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 wait_queue_head_t *wq;
123};
124
125/*
126 * This is used to implement the safe poll wake up avoiding to reenter
127 * the poll callback from inside wake_up().
128 */
129struct poll_safewake {
130 struct list_head wake_task_list;
131 spinlock_t lock;
132};
133
134/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700136 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 */
138struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700139 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140 struct rb_node rbn;
141
142 /* List header used to link this structure to the eventpoll ready list */
143 struct list_head rdllink;
144
Davide Libenzic7ea7632007-05-15 01:40:47 -0700145 /*
146 * Works together "struct eventpoll"->ovflist in keeping the
147 * single linked chain of items.
148 */
149 struct epitem *next;
150
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 /* The file descriptor information this item refers to */
152 struct epoll_filefd ffd;
153
154 /* Number of active wait queue attached to poll operations */
155 int nwait;
156
157 /* List containing poll wait queues */
158 struct list_head pwqlist;
159
160 /* The "container" of this item */
161 struct eventpoll *ep;
162
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 /* List header used to link this item to the "struct file" items list */
164 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700165
Davide Libenzic7ea7632007-05-15 01:40:47 -0700166 /* The structure that describe the interested events and the source fd */
167 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700168};
169
170/*
171 * This structure is stored inside the "private_data" member of the file
172 * structure and rapresent the main data sructure for the eventpoll
173 * interface.
174 */
175struct eventpoll {
176 /* Protect the this structure access */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700177 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700178
179 /*
180 * This mutex is used to ensure that files are not removed
181 * while epoll is using them. This is held during the event
182 * collection loop, the file cleanup path, the epoll file exit
183 * code and the ctl operations.
184 */
185 struct mutex mtx;
186
187 /* Wait queue used by sys_epoll_wait() */
188 wait_queue_head_t wq;
189
190 /* Wait queue used by file->poll() */
191 wait_queue_head_t poll_wait;
192
193 /* List of ready file descriptors */
194 struct list_head rdllist;
195
Davide Libenzi67647d02007-05-15 01:40:52 -0700196 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700197 struct rb_root rbr;
198
199 /*
200 * This is a single linked list that chains all the "struct epitem" that
201 * happened while transfering ready events to userspace w/out
202 * holding ->lock.
203 */
204 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800205
206 /* The user that created the eventpoll descriptor */
207 struct user_struct *user;
Davide Libenzid47de162007-05-15 01:40:41 -0700208};
209
210/* Wait structure used by the poll hooks */
211struct eppoll_entry {
212 /* List header used to link this structure to the "struct epitem" */
213 struct list_head llink;
214
215 /* The "base" pointer is set to the container "struct epitem" */
216 void *base;
217
218 /*
219 * Wait queue item that will be linked to the target file wait
220 * queue head.
221 */
222 wait_queue_t wait;
223
224 /* The wait queue head that linked the "wait" wait queue item */
225 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226};
227
228/* Wrapper struct used by poll queueing */
229struct ep_pqueue {
230 poll_table pt;
231 struct epitem *epi;
232};
233
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800235 * Configuration options available inside /proc/sys/fs/epoll/
236 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800237/* Maximum number of epoll watched descriptors, per user */
238static int max_user_watches __read_mostly;
239
240/*
Davide Libenzid47de162007-05-15 01:40:41 -0700241 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800243static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
245/* Safe wake up implementation */
246static struct poll_safewake psw;
247
248/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800249static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250
251/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800252static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
Davide Libenzi7ef99642008-12-01 13:13:55 -0800254#ifdef CONFIG_SYSCTL
255
256#include <linux/sysctl.h>
257
258static int zero;
259
260ctl_table epoll_table[] = {
261 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800262 .procname = "max_user_watches",
263 .data = &max_user_watches,
264 .maxlen = sizeof(int),
265 .mode = 0644,
266 .proc_handler = &proc_dointvec_minmax,
267 .extra1 = &zero,
268 },
269 { .ctl_name = 0 }
270};
271#endif /* CONFIG_SYSCTL */
272
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700273
Davide Libenzi67647d02007-05-15 01:40:52 -0700274/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700275static inline void ep_set_ffd(struct epoll_filefd *ffd,
276 struct file *file, int fd)
277{
278 ffd->file = file;
279 ffd->fd = fd;
280}
281
Davide Libenzi67647d02007-05-15 01:40:52 -0700282/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700283static inline int ep_cmp_ffd(struct epoll_filefd *p1,
284 struct epoll_filefd *p2)
285{
286 return (p1->file > p2->file ? +1:
287 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
288}
289
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700290/* Tells us if the item is currently linked */
291static inline int ep_is_linked(struct list_head *p)
292{
293 return !list_empty(p);
294}
295
296/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700297static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700298{
299 return container_of(p, struct eppoll_entry, wait)->base;
300}
301
302/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700303static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700304{
305 return container_of(p, struct ep_pqueue, pt)->epi;
306}
307
308/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700309static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700310{
311 return op != EPOLL_CTL_DEL;
312}
313
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314/* Initialize the poll safe wake up structure */
315static void ep_poll_safewake_init(struct poll_safewake *psw)
316{
317
318 INIT_LIST_HEAD(&psw->wake_task_list);
319 spin_lock_init(&psw->lock);
320}
321
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322/*
323 * Perform a safe wake up of the poll wait list. The problem is that
324 * with the new callback'd wake up system, it is possible that the
325 * poll callback is reentered from inside the call to wake_up() done
326 * on the poll wait queue head. The rule is that we cannot reenter the
327 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
328 * and we cannot reenter the same wait queue head at all. This will
329 * enable to have a hierarchy of epoll file descriptor of no more than
330 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
331 * because this one gets called by the poll callback, that in turn is called
332 * from inside a wake_up(), that might be called from irq context.
333 */
334static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
335{
336 int wake_nests = 0;
337 unsigned long flags;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700338 struct task_struct *this_task = current;
Matthias Kaehlckeb70c3942007-10-18 23:39:56 -0700339 struct list_head *lsthead = &psw->wake_task_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 struct wake_task_node *tncur;
341 struct wake_task_node tnode;
342
343 spin_lock_irqsave(&psw->lock, flags);
344
345 /* Try to see if the current task is already inside this wakeup call */
Matthias Kaehlckeb70c3942007-10-18 23:39:56 -0700346 list_for_each_entry(tncur, lsthead, llink) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
348 if (tncur->wq == wq ||
349 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
350 /*
351 * Ops ... loop detected or maximum nest level reached.
352 * We abort this wake by breaking the cycle itself.
353 */
354 spin_unlock_irqrestore(&psw->lock, flags);
355 return;
356 }
357 }
358
359 /* Add the current task to the list */
360 tnode.task = this_task;
361 tnode.wq = wq;
362 list_add(&tnode.llink, lsthead);
363
364 spin_unlock_irqrestore(&psw->lock, flags);
365
366 /* Do really wake up now */
Peter Zijlstra0ccf8312008-02-04 22:27:20 -0800367 wake_up_nested(wq, 1 + wake_nests);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368
369 /* Remove the current task from the list */
370 spin_lock_irqsave(&psw->lock, flags);
371 list_del(&tnode.llink);
372 spin_unlock_irqrestore(&psw->lock, flags);
373}
374
Davide Libenzi7699acd2007-05-10 22:23:23 -0700375/*
376 * This function unregister poll callbacks from the associated file descriptor.
377 * Since this must be called without holding "ep->lock" the atomic exchange trick
378 * will protect us from multiple unregister.
379 */
380static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
381{
382 int nwait;
383 struct list_head *lsthead = &epi->pwqlist;
384 struct eppoll_entry *pwq;
385
386 /* This is called without locks, so we need the atomic exchange */
387 nwait = xchg(&epi->nwait, 0);
388
389 if (nwait) {
390 while (!list_empty(lsthead)) {
391 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
392
393 list_del_init(&pwq->llink);
394 remove_wait_queue(pwq->whead, &pwq->wait);
395 kmem_cache_free(pwq_cache, pwq);
396 }
397 }
398}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700401 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700402 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700403 */
404static int ep_remove(struct eventpoll *ep, struct epitem *epi)
405{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700406 unsigned long flags;
407 struct file *file = epi->ffd.file;
408
409 /*
410 * Removes poll wait queue hooks. We _have_ to do this without holding
411 * the "ep->lock" otherwise a deadlock might occur. This because of the
412 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
413 * queue head lock when unregistering the wait queue. The wakeup callback
414 * will run by holding the wait queue head lock and will call our callback
415 * that will try to get "ep->lock".
416 */
417 ep_unregister_pollwait(ep, epi);
418
419 /* Remove the current item from the list of epoll hooks */
420 spin_lock(&file->f_ep_lock);
421 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700422 list_del_init(&epi->fllink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700423 spin_unlock(&file->f_ep_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424
Davide Libenzicdac75e2008-04-29 00:58:34 -0700425 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426
Davide Libenzic7ea7632007-05-15 01:40:47 -0700427 spin_lock_irqsave(&ep->lock, flags);
428 if (ep_is_linked(&epi->rdllink))
429 list_del_init(&epi->rdllink);
430 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
Davide Libenzi7699acd2007-05-10 22:23:23 -0700432 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700433 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
Davide Libenzi7ef99642008-12-01 13:13:55 -0800435 atomic_dec(&ep->user->epoll_watches);
436
Davide Libenzic7ea7632007-05-15 01:40:47 -0700437 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
438 current, ep, file));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
Davide Libenzic7ea7632007-05-15 01:40:47 -0700440 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441}
442
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443static void ep_free(struct eventpoll *ep)
444{
445 struct rb_node *rbp;
446 struct epitem *epi;
447
448 /* We need to release all tasks waiting for these file */
449 if (waitqueue_active(&ep->poll_wait))
450 ep_poll_safewake(&psw, &ep->poll_wait);
451
452 /*
453 * We need to lock this because we could be hit by
454 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700455 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 * is on the way to be removed and no one has references to it
457 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800458 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800460 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
462 /*
463 * Walks through the whole tree by unregistering poll callbacks.
464 */
465 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
466 epi = rb_entry(rbp, struct epitem, rbn);
467
468 ep_unregister_pollwait(ep, epi);
469 }
470
471 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700472 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700474 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 * us during this operation. So we can avoid the lock on "ep->lock".
476 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700477 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 epi = rb_entry(rbp, struct epitem, rbn);
479 ep_remove(ep, epi);
480 }
481
Arjan van de Ven144efe32006-03-23 03:00:32 -0800482 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700483 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800484 free_uid(ep->user);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700485 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486}
487
Davide Libenzi7699acd2007-05-10 22:23:23 -0700488static int ep_eventpoll_release(struct inode *inode, struct file *file)
489{
490 struct eventpoll *ep = file->private_data;
491
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700492 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700493 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700494
495 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
496 return 0;
497}
498
499static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
500{
501 unsigned int pollflags = 0;
502 unsigned long flags;
503 struct eventpoll *ep = file->private_data;
504
505 /* Insert inside our poll wait queue */
506 poll_wait(file, &ep->poll_wait, wait);
507
508 /* Check our condition */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700509 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700510 if (!list_empty(&ep->rdllist))
511 pollflags = POLLIN | POLLRDNORM;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700512 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700513
514 return pollflags;
515}
516
517/* File callbacks that implement the eventpoll file behaviour */
518static const struct file_operations eventpoll_fops = {
519 .release = ep_eventpoll_release,
520 .poll = ep_eventpoll_poll
521};
522
523/* Fast test to see if the file is an evenpoll file */
524static inline int is_file_epoll(struct file *f)
525{
526 return f->f_op == &eventpoll_fops;
527}
528
529/*
530 * This is called from eventpoll_release() to unlink files from the eventpoll
531 * interface. We need to have this facility to cleanup correctly files that are
532 * closed without being removed from the eventpoll interface.
533 */
534void eventpoll_release_file(struct file *file)
535{
536 struct list_head *lsthead = &file->f_ep_links;
537 struct eventpoll *ep;
538 struct epitem *epi;
539
540 /*
541 * We don't want to get "file->f_ep_lock" because it is not
542 * necessary. It is not necessary because we're in the "struct file"
543 * cleanup path, and this means that noone is using this file anymore.
Davide Libenzi67647d02007-05-15 01:40:52 -0700544 * So, for example, epoll_ctl() cannot hit here sicne if we reach this
545 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700546 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700547 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700548 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700549 * from anywhere but ep_free().
550 */
551 mutex_lock(&epmutex);
552
553 while (!list_empty(lsthead)) {
554 epi = list_first_entry(lsthead, struct epitem, fllink);
555
556 ep = epi->ep;
557 list_del_init(&epi->fllink);
Davide Libenzid47de162007-05-15 01:40:41 -0700558 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700559 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700560 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700561 }
562
563 mutex_unlock(&epmutex);
564}
565
566static int ep_alloc(struct eventpoll **pep)
567{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800568 int error;
569 struct user_struct *user;
570 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700571
Davide Libenzi7ef99642008-12-01 13:13:55 -0800572 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800573 error = -ENOMEM;
574 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
575 if (unlikely(!ep))
576 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700577
Davide Libenzic7ea7632007-05-15 01:40:47 -0700578 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700579 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700580 init_waitqueue_head(&ep->wq);
581 init_waitqueue_head(&ep->poll_wait);
582 INIT_LIST_HEAD(&ep->rdllist);
583 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700584 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800585 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700586
587 *pep = ep;
588
589 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
590 current, ep));
591 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800592
593free_uid:
594 free_uid(user);
595 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700596}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597
598/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700599 * Search the file inside the eventpoll tree. The RB tree operations
600 * are protected by the "mtx" mutex, and ep_find() must be called with
601 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 */
603static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
604{
605 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 struct rb_node *rbp;
607 struct epitem *epi, *epir = NULL;
608 struct epoll_filefd ffd;
609
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700610 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 for (rbp = ep->rbr.rb_node; rbp; ) {
612 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700613 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 if (kcmp > 0)
615 rbp = rbp->rb_right;
616 else if (kcmp < 0)
617 rbp = rbp->rb_left;
618 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 epir = epi;
620 break;
621 }
622 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623
624 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
625 current, file, epir));
626
627 return epir;
628}
629
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700631 * This is the callback that is passed to the wait queue wakeup
632 * machanism. It is called by the stored file descriptors when they
633 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700635static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700637 int pwake = 0;
638 unsigned long flags;
639 struct epitem *epi = ep_item_from_wait(wait);
640 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641
Davide Libenzi7699acd2007-05-10 22:23:23 -0700642 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
643 current, epi->ffd.file, epi, ep));
644
Davide Libenzic7ea7632007-05-15 01:40:47 -0700645 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700646
647 /*
648 * If the event mask does not contain any poll(2) event, we consider the
649 * descriptor to be disabled. This condition is likely the effect of the
650 * EPOLLONESHOT bit that disables the descriptor when an event is received,
651 * until the next EPOLL_CTL_MOD will be issued.
652 */
653 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700654 goto out_unlock;
655
656 /*
657 * If we are trasfering events to userspace, we can hold no locks
658 * (because we're accessing user memory, and because of linux f_op->poll()
659 * semantics). All the events that happens during that period of time are
660 * chained in ep->ovflist and requeued later on.
661 */
662 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
663 if (epi->next == EP_UNACTIVE_PTR) {
664 epi->next = ep->ovflist;
665 ep->ovflist = epi;
666 }
667 goto out_unlock;
668 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700669
670 /* If this file is already in the ready list we exit soon */
671 if (ep_is_linked(&epi->rdllink))
672 goto is_linked;
673
674 list_add_tail(&epi->rdllink, &ep->rdllist);
675
676is_linked:
677 /*
678 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
679 * wait list.
680 */
681 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400682 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700683 if (waitqueue_active(&ep->poll_wait))
684 pwake++;
685
Davide Libenzid47de162007-05-15 01:40:41 -0700686out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700687 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700688
689 /* We have to call this outside the lock */
690 if (pwake)
691 ep_poll_safewake(&psw, &ep->poll_wait);
692
693 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694}
695
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696/*
697 * This is the callback that is used to add our wait queue to the
698 * target file wakeup lists.
699 */
700static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
701 poll_table *pt)
702{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700703 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 struct eppoll_entry *pwq;
705
Christoph Lametere94b1762006-12-06 20:33:17 -0800706 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
708 pwq->whead = whead;
709 pwq->base = epi;
710 add_wait_queue(whead, &pwq->wait);
711 list_add_tail(&pwq->llink, &epi->pwqlist);
712 epi->nwait++;
713 } else {
714 /* We have to signal that an error occurred */
715 epi->nwait = -1;
716 }
717}
718
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
720{
721 int kcmp;
722 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
723 struct epitem *epic;
724
725 while (*p) {
726 parent = *p;
727 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700728 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 if (kcmp > 0)
730 p = &parent->rb_right;
731 else
732 p = &parent->rb_left;
733 }
734 rb_link_node(&epi->rbn, parent, p);
735 rb_insert_color(&epi->rbn, &ep->rbr);
736}
737
Davide Libenzic7ea7632007-05-15 01:40:47 -0700738/*
739 * Must be called with "mtx" held.
740 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
742 struct file *tfile, int fd)
743{
744 int error, revents, pwake = 0;
745 unsigned long flags;
746 struct epitem *epi;
747 struct ep_pqueue epq;
748
Davide Libenzi7ef99642008-12-01 13:13:55 -0800749 if (unlikely(atomic_read(&ep->user->epoll_watches) >=
750 max_user_watches))
751 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -0800752 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -0800753 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754
755 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 INIT_LIST_HEAD(&epi->rdllink);
757 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 INIT_LIST_HEAD(&epi->pwqlist);
759 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700760 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -0700763 epi->next = EP_UNACTIVE_PTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764
765 /* Initialize the poll table using the queue callback */
766 epq.epi = epi;
767 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
768
769 /*
770 * Attach the item to the poll hooks and get current event bits.
771 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -0700772 * been increased by the caller of this function. Note that after
773 * this operation completes, the poll callback can start hitting
774 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 */
776 revents = tfile->f_op->poll(tfile, &epq.pt);
777
778 /*
779 * We have to check if something went wrong during the poll wait queue
780 * install process. Namely an allocation for a wait queue failed due
781 * high memory pressure.
782 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800783 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700785 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
787 /* Add the current item to the list of active epoll hook for this file */
788 spin_lock(&tfile->f_ep_lock);
789 list_add_tail(&epi->fllink, &tfile->f_ep_links);
790 spin_unlock(&tfile->f_ep_lock);
791
Davide Libenzic7ea7632007-05-15 01:40:47 -0700792 /*
793 * Add the current item to the RB tree. All RB tree operations are
794 * protected by "mtx", and ep_insert() is called with "mtx" held.
795 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 ep_rbtree_insert(ep, epi);
797
Davide Libenzic7ea7632007-05-15 01:40:47 -0700798 /* We have to drop the new item inside our item list to keep track of it */
799 spin_lock_irqsave(&ep->lock, flags);
800
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700802 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 list_add_tail(&epi->rdllink, &ep->rdllist);
804
805 /* Notify waiting tasks that events are available */
806 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400807 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 if (waitqueue_active(&ep->poll_wait))
809 pwake++;
810 }
811
Davide Libenzic7ea7632007-05-15 01:40:47 -0700812 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
Davide Libenzi7ef99642008-12-01 13:13:55 -0800814 atomic_inc(&ep->user->epoll_watches);
815
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 /* We have to call this outside the lock */
817 if (pwake)
818 ep_poll_safewake(&psw, &ep->poll_wait);
819
820 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
821 current, ep, tfile, fd));
822
823 return 0;
824
Davide Libenzi7699acd2007-05-10 22:23:23 -0700825error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 ep_unregister_pollwait(ep, epi);
827
828 /*
829 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -0700830 * allocated wait queue. Note that we don't care about the ep->ovflist
831 * list, since that is used/cleaned only inside a section bound by "mtx".
832 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700834 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700835 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700836 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700837 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700839 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800840
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 return error;
842}
843
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844/*
845 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -0700846 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 */
848static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
849{
850 int pwake = 0;
851 unsigned int revents;
852 unsigned long flags;
853
854 /*
855 * Set the new event interest mask before calling f_op->poll(), otherwise
856 * a potential race might occur. In fact if we do this operation inside
857 * the lock, an event might happen between the f_op->poll() call and the
858 * new event set registering.
859 */
860 epi->event.events = event->events;
861
862 /*
863 * Get current event bits. We can safely use the file* here because
864 * its usage count has been increased by the caller of this function.
865 */
866 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
867
Davide Libenzic7ea7632007-05-15 01:40:47 -0700868 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869
870 /* Copy the data member from inside the lock */
871 epi->event.data = event->data;
872
873 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700874 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -0700875 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700877 if (revents & event->events) {
878 if (!ep_is_linked(&epi->rdllink)) {
879 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880
Davide Libenzic7ea7632007-05-15 01:40:47 -0700881 /* Notify waiting tasks that events are available */
882 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400883 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700884 if (waitqueue_active(&ep->poll_wait))
885 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 }
887 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700888 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
890 /* We have to call this outside the lock */
891 if (pwake)
892 ep_poll_safewake(&psw, &ep->poll_wait);
893
894 return 0;
895}
896
Davide Libenzid47de162007-05-15 01:40:41 -0700897static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
898 int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899{
Davide Libenzi6192bd52007-05-08 00:25:41 -0700900 int eventcnt, error = -EFAULT, pwake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 unsigned int revents;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700902 unsigned long flags;
Davide Libenzid47de162007-05-15 01:40:41 -0700903 struct epitem *epi, *nepi;
904 struct list_head txlist;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700905
Davide Libenzid47de162007-05-15 01:40:41 -0700906 INIT_LIST_HEAD(&txlist);
907
908 /*
909 * We need to lock this because we could be hit by
910 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
911 */
912 mutex_lock(&ep->mtx);
913
914 /*
915 * Steal the ready list, and re-init the original one to the
916 * empty list. Also, set ep->ovflist to NULL so that events
917 * happening while looping w/out locks, are not lost. We cannot
918 * have the poll callback to queue directly on ep->rdllist,
919 * because we are doing it in the loop below, in a lockless way.
920 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700921 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700922 list_splice(&ep->rdllist, &txlist);
923 INIT_LIST_HEAD(&ep->rdllist);
924 ep->ovflist = NULL;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700925 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926
927 /*
928 * We can loop without lock because this is a task private list.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700929 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
Davide Libenzid47de162007-05-15 01:40:41 -0700930 * Items cannot vanish during the loop because we are holding "mtx".
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 */
Davide Libenzid47de162007-05-15 01:40:41 -0700932 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
933 epi = list_first_entry(&txlist, struct epitem, rdllink);
934
935 list_del_init(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
937 /*
938 * Get the ready file event set. We can safely use the file
Davide Libenzid47de162007-05-15 01:40:41 -0700939 * because we are holding the "mtx" and this will guarantee
940 * that both the file and the item will not vanish.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 */
942 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700943 revents &= epi->event.events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
945 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700946 * Is the event mask intersect the caller-requested one,
947 * deliver the event to userspace. Again, we are holding
Davide Libenzid47de162007-05-15 01:40:41 -0700948 * "mtx", so no operations coming from userspace can change
949 * the item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700951 if (revents) {
952 if (__put_user(revents,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 &events[eventcnt].events) ||
954 __put_user(epi->event.data,
955 &events[eventcnt].data))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700956 goto errxit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 if (epi->event.events & EPOLLONESHOT)
958 epi->event.events &= EP_PRIVATE_BITS;
959 eventcnt++;
960 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700962 * At this point, noone can insert into ep->rdllist besides
963 * us. The epoll_ctl() callers are locked out by us holding
964 * "mtx" and the poll callback will queue them in ep->ovflist.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700966 if (!(epi->event.events & EPOLLET) &&
Davide Libenzid47de162007-05-15 01:40:41 -0700967 (revents & epi->event.events))
968 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 }
Davide Libenzi6192bd52007-05-08 00:25:41 -0700970 error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Davide Libenzid47de162007-05-15 01:40:41 -0700972errxit:
973
Davide Libenzic7ea7632007-05-15 01:40:47 -0700974 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700975 /*
976 * During the time we spent in the loop above, some other events
977 * might have been queued by the poll callback. We re-insert them
Davide Libenzif337b9c2008-10-15 22:01:56 -0700978 * inside the main ready-list here.
Davide Libenzid47de162007-05-15 01:40:41 -0700979 */
980 for (nepi = ep->ovflist; (epi = nepi) != NULL;
Davide Libenzi9ce209d2008-10-17 16:17:40 -0700981 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
982 /*
983 * If the above loop quit with errors, the epoll item might still
984 * be linked to "txlist", and the list_splice() done below will
985 * take care of those cases.
986 */
987 if (!ep_is_linked(&epi->rdllink))
988 list_add_tail(&epi->rdllink, &ep->rdllist);
989 }
Davide Libenzid47de162007-05-15 01:40:41 -0700990 /*
991 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
992 * releasing the lock, events will be queued in the normal way inside
993 * ep->rdllist.
994 */
995 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700996
997 /*
Davide Libenzi67647d02007-05-15 01:40:52 -0700998 * In case of error in the event-send loop, or in case the number of
999 * ready events exceeds the userspace limit, we need to splice the
1000 * "txlist" back inside ep->rdllist.
Davide Libenzi6192bd52007-05-08 00:25:41 -07001001 */
Davide Libenzid47de162007-05-15 01:40:41 -07001002 list_splice(&txlist, &ep->rdllist);
Davide Libenzi6192bd52007-05-08 00:25:41 -07001003
Davide Libenzid47de162007-05-15 01:40:41 -07001004 if (!list_empty(&ep->rdllist)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 /*
Davide Libenzid47de162007-05-15 01:40:41 -07001006 * Wake up (if active) both the eventpoll wait list and the ->poll()
Davide Libenzi67647d02007-05-15 01:40:52 -07001007 * wait list (delayed after we release the lock).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 */
1009 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -04001010 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 if (waitqueue_active(&ep->poll_wait))
1012 pwake++;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001013 }
Davide Libenzic7ea7632007-05-15 01:40:47 -07001014 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -07001015
1016 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
1018 /* We have to call this outside the lock */
1019 if (pwake)
1020 ep_poll_safewake(&psw, &ep->poll_wait);
Davide Libenzi6192bd52007-05-08 00:25:41 -07001021
1022 return eventcnt == 0 ? error: eventcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023}
1024
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1026 int maxevents, long timeout)
1027{
1028 int res, eavail;
1029 unsigned long flags;
1030 long jtimeout;
1031 wait_queue_t wait;
1032
1033 /*
1034 * Calculate the timeout by checking for the "infinite" value ( -1 )
1035 * and the overflow condition. The passed timeout is in milliseconds,
1036 * that why (t * HZ) / 1000.
1037 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001038 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1039 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040
1041retry:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001042 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043
1044 res = 0;
1045 if (list_empty(&ep->rdllist)) {
1046 /*
1047 * We don't have any available event to return to the caller.
1048 * We need to sleep here, and we will be wake up by
1049 * ep_poll_callback() when events will become available.
1050 */
1051 init_waitqueue_entry(&wait, current);
Davide Libenzid47de162007-05-15 01:40:41 -07001052 wait.flags |= WQ_FLAG_EXCLUSIVE;
Davide Libenzi3419b232006-06-25 05:48:14 -07001053 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054
1055 for (;;) {
1056 /*
1057 * We don't want to sleep if the ep_poll_callback() sends us
1058 * a wakeup in between. That's why we set the task state
1059 * to TASK_INTERRUPTIBLE before doing the checks.
1060 */
1061 set_current_state(TASK_INTERRUPTIBLE);
1062 if (!list_empty(&ep->rdllist) || !jtimeout)
1063 break;
1064 if (signal_pending(current)) {
1065 res = -EINTR;
1066 break;
1067 }
1068
Davide Libenzic7ea7632007-05-15 01:40:47 -07001069 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070 jtimeout = schedule_timeout(jtimeout);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001071 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001073 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074
1075 set_current_state(TASK_RUNNING);
1076 }
1077
1078 /* Is it worth to try to dig for events ? */
1079 eavail = !list_empty(&ep->rdllist);
1080
Davide Libenzic7ea7632007-05-15 01:40:47 -07001081 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082
1083 /*
1084 * Try to transfer events to user space. In case we get 0 events and
1085 * there's still timeout left over, we go trying again in search of
1086 * more luck.
1087 */
1088 if (!res && eavail &&
Davide Libenzid47de162007-05-15 01:40:41 -07001089 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 goto retry;
1091
1092 return res;
1093}
1094
Davide Libenzi7699acd2007-05-10 22:23:23 -07001095/*
Andrew Morton523723b2008-08-12 15:09:01 -07001096 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001097 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001098SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001099{
1100 int error, fd = -1;
1101 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001102
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001103 /* Check the EPOLL_* constant for consistency. */
1104 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1105
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001106 if (flags & ~EPOLL_CLOEXEC)
1107 return -EINVAL;
1108
Davide Libenzi7699acd2007-05-10 22:23:23 -07001109 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001110 current, flags));
Davide Libenzi7699acd2007-05-10 22:23:23 -07001111
1112 /*
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001113 * Create the internal data structure ( "struct eventpoll" ).
Davide Libenzi7699acd2007-05-10 22:23:23 -07001114 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001115 error = ep_alloc(&ep);
1116 if (error < 0) {
Al Viro2030a422008-02-23 06:46:49 -05001117 fd = error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001118 goto error_return;
Al Viro2030a422008-02-23 06:46:49 -05001119 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001120
1121 /*
1122 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001123 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001124 */
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001125 fd = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1126 flags & O_CLOEXEC);
Al Viro2030a422008-02-23 06:46:49 -05001127 if (fd < 0)
1128 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001129
Al Viro2030a422008-02-23 06:46:49 -05001130error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001131 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001132 current, flags, fd));
Davide Libenzi7699acd2007-05-10 22:23:23 -07001133
1134 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001135}
1136
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001137SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001138{
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001139 if (size < 0)
1140 return -EINVAL;
1141
1142 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001143}
1144
Davide Libenzi7699acd2007-05-10 22:23:23 -07001145/*
1146 * The following function implements the controller interface for
1147 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001148 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001149 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001150SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1151 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001152{
1153 int error;
1154 struct file *file, *tfile;
1155 struct eventpoll *ep;
1156 struct epitem *epi;
1157 struct epoll_event epds;
1158
1159 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1160 current, epfd, op, fd, event));
1161
1162 error = -EFAULT;
1163 if (ep_op_has_event(op) &&
1164 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1165 goto error_return;
1166
1167 /* Get the "struct file *" for the eventpoll file */
1168 error = -EBADF;
1169 file = fget(epfd);
1170 if (!file)
1171 goto error_return;
1172
1173 /* Get the "struct file *" for the target file */
1174 tfile = fget(fd);
1175 if (!tfile)
1176 goto error_fput;
1177
1178 /* The target file descriptor must support poll */
1179 error = -EPERM;
1180 if (!tfile->f_op || !tfile->f_op->poll)
1181 goto error_tgt_fput;
1182
1183 /*
1184 * We have to check that the file structure underneath the file descriptor
1185 * the user passed to us _is_ an eventpoll file. And also we do not permit
1186 * adding an epoll file descriptor inside itself.
1187 */
1188 error = -EINVAL;
1189 if (file == tfile || !is_file_epoll(file))
1190 goto error_tgt_fput;
1191
1192 /*
1193 * At this point it is safe to assume that the "private_data" contains
1194 * our own data structure.
1195 */
1196 ep = file->private_data;
1197
Davide Libenzid47de162007-05-15 01:40:41 -07001198 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001199
Davide Libenzi67647d02007-05-15 01:40:52 -07001200 /*
1201 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1202 * above, we can be sure to be able to use the item looked up by
1203 * ep_find() till we release the mutex.
1204 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001205 epi = ep_find(ep, tfile, fd);
1206
1207 error = -EINVAL;
1208 switch (op) {
1209 case EPOLL_CTL_ADD:
1210 if (!epi) {
1211 epds.events |= POLLERR | POLLHUP;
1212
1213 error = ep_insert(ep, &epds, tfile, fd);
1214 } else
1215 error = -EEXIST;
1216 break;
1217 case EPOLL_CTL_DEL:
1218 if (epi)
1219 error = ep_remove(ep, epi);
1220 else
1221 error = -ENOENT;
1222 break;
1223 case EPOLL_CTL_MOD:
1224 if (epi) {
1225 epds.events |= POLLERR | POLLHUP;
1226 error = ep_modify(ep, epi, &epds);
1227 } else
1228 error = -ENOENT;
1229 break;
1230 }
Davide Libenzid47de162007-05-15 01:40:41 -07001231 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001232
1233error_tgt_fput:
1234 fput(tfile);
1235error_fput:
1236 fput(file);
1237error_return:
1238 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1239 current, epfd, op, fd, event, error));
1240
1241 return error;
1242}
1243
1244/*
1245 * Implement the event wait interface for the eventpoll file. It is the kernel
1246 * part of the user space epoll_wait(2).
1247 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001248SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1249 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001250{
1251 int error;
1252 struct file *file;
1253 struct eventpoll *ep;
1254
1255 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1256 current, epfd, events, maxevents, timeout));
1257
1258 /* The maximum number of event must be greater than zero */
1259 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1260 return -EINVAL;
1261
1262 /* Verify that the area passed by the user is writeable */
1263 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1264 error = -EFAULT;
1265 goto error_return;
1266 }
1267
1268 /* Get the "struct file *" for the eventpoll file */
1269 error = -EBADF;
1270 file = fget(epfd);
1271 if (!file)
1272 goto error_return;
1273
1274 /*
1275 * We have to check that the file structure underneath the fd
1276 * the user passed to us _is_ an eventpoll file.
1277 */
1278 error = -EINVAL;
1279 if (!is_file_epoll(file))
1280 goto error_fput;
1281
1282 /*
1283 * At this point it is safe to assume that the "private_data" contains
1284 * our own data structure.
1285 */
1286 ep = file->private_data;
1287
1288 /* Time to fish for events ... */
1289 error = ep_poll(ep, events, maxevents, timeout);
1290
1291error_fput:
1292 fput(file);
1293error_return:
1294 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1295 current, epfd, events, maxevents, timeout, error));
1296
1297 return error;
1298}
1299
Roland McGrathf3de2722008-04-30 00:53:09 -07001300#ifdef HAVE_SET_RESTORE_SIGMASK
Davide Libenzi7699acd2007-05-10 22:23:23 -07001301
1302/*
1303 * Implement the event wait interface for the eventpoll file. It is the kernel
1304 * part of the user space epoll_pwait(2).
1305 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001306SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1307 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1308 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001309{
1310 int error;
1311 sigset_t ksigmask, sigsaved;
1312
1313 /*
1314 * If the caller wants a certain signal mask to be set during the wait,
1315 * we apply it here.
1316 */
1317 if (sigmask) {
1318 if (sigsetsize != sizeof(sigset_t))
1319 return -EINVAL;
1320 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1321 return -EFAULT;
1322 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1323 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1324 }
1325
1326 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1327
1328 /*
1329 * If we changed the signal mask, we need to restore the original one.
1330 * In case we've got a signal while waiting, we do not restore the
1331 * signal mask yet, and we allow do_signal() to deliver the signal on
1332 * the way back to userspace, before the signal mask is restored.
1333 */
1334 if (sigmask) {
1335 if (error == -EINTR) {
1336 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001337 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001338 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001339 } else
1340 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1341 }
1342
1343 return error;
1344}
1345
Roland McGrathf3de2722008-04-30 00:53:09 -07001346#endif /* HAVE_SET_RESTORE_SIGMASK */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001347
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348static int __init eventpoll_init(void)
1349{
Davide Libenzi7ef99642008-12-01 13:13:55 -08001350 struct sysinfo si;
1351
1352 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08001353 /*
1354 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1355 */
1356 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08001357 EP_ITEM_COST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358
1359 /* Initialize the structure used to perform safe poll wait head wake ups */
1360 ep_poll_safewake_init(&psw);
1361
1362 /* Allocates slab cache used to allocate "struct epitem" items */
1363 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1364 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001365 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
1367 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1368 pwq_cache = kmem_cache_create("eventpoll_pwq",
1369 sizeof(struct eppoll_entry), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09001370 EPI_SLAB_DEBUG|SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373}
Davide Libenzicea69242007-05-10 22:23:22 -07001374fs_initcall(eventpoll_init);