blob: 0c87474f7917776d92fb1427c93f118535ef624f [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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105struct epoll_filefd {
106 struct file *file;
107 int fd;
108};
109
110/*
111 * Node that is linked into the "wake_task_list" member of the "struct poll_safewake".
112 * It is used to keep track on all tasks that are currently inside the wake_up() code
113 * to 1) short-circuit the one coming from the same task and same wait queue head
Davide Libenzid47de162007-05-15 01:40:41 -0700114 * (loop) 2) allow a maximum number of epoll descriptors inclusion nesting
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 * 3) let go the ones coming from other tasks.
116 */
117struct wake_task_node {
118 struct list_head llink;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700119 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 wait_queue_head_t *wq;
121};
122
123/*
124 * This is used to implement the safe poll wake up avoiding to reenter
125 * the poll callback from inside wake_up().
126 */
127struct poll_safewake {
128 struct list_head wake_task_list;
129 spinlock_t lock;
130};
131
132/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700134 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 */
136struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700137 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 struct rb_node rbn;
139
140 /* List header used to link this structure to the eventpoll ready list */
141 struct list_head rdllink;
142
Davide Libenzic7ea7632007-05-15 01:40:47 -0700143 /*
144 * Works together "struct eventpoll"->ovflist in keeping the
145 * single linked chain of items.
146 */
147 struct epitem *next;
148
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 /* The file descriptor information this item refers to */
150 struct epoll_filefd ffd;
151
152 /* Number of active wait queue attached to poll operations */
153 int nwait;
154
155 /* List containing poll wait queues */
156 struct list_head pwqlist;
157
158 /* The "container" of this item */
159 struct eventpoll *ep;
160
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 /* List header used to link this item to the "struct file" items list */
162 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700163
Davide Libenzic7ea7632007-05-15 01:40:47 -0700164 /* The structure that describe the interested events and the source fd */
165 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700166};
167
168/*
169 * This structure is stored inside the "private_data" member of the file
170 * structure and rapresent the main data sructure for the eventpoll
171 * interface.
172 */
173struct eventpoll {
174 /* Protect the this structure access */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700175 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700176
177 /*
178 * This mutex is used to ensure that files are not removed
179 * while epoll is using them. This is held during the event
180 * collection loop, the file cleanup path, the epoll file exit
181 * code and the ctl operations.
182 */
183 struct mutex mtx;
184
185 /* Wait queue used by sys_epoll_wait() */
186 wait_queue_head_t wq;
187
188 /* Wait queue used by file->poll() */
189 wait_queue_head_t poll_wait;
190
191 /* List of ready file descriptors */
192 struct list_head rdllist;
193
Davide Libenzi67647d02007-05-15 01:40:52 -0700194 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700195 struct rb_root rbr;
196
197 /*
198 * This is a single linked list that chains all the "struct epitem" that
199 * happened while transfering ready events to userspace w/out
200 * holding ->lock.
201 */
202 struct epitem *ovflist;
203};
204
205/* Wait structure used by the poll hooks */
206struct eppoll_entry {
207 /* List header used to link this structure to the "struct epitem" */
208 struct list_head llink;
209
210 /* The "base" pointer is set to the container "struct epitem" */
211 void *base;
212
213 /*
214 * Wait queue item that will be linked to the target file wait
215 * queue head.
216 */
217 wait_queue_t wait;
218
219 /* The wait queue head that linked the "wait" wait queue item */
220 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221};
222
223/* Wrapper struct used by poll queueing */
224struct ep_pqueue {
225 poll_table pt;
226 struct epitem *epi;
227};
228
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229/*
Davide Libenzid47de162007-05-15 01:40:41 -0700230 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800232static struct mutex epmutex;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233
234/* Safe wake up implementation */
235static struct poll_safewake psw;
236
237/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800238static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
240/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800241static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700243
Davide Libenzi67647d02007-05-15 01:40:52 -0700244/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700245static inline void ep_set_ffd(struct epoll_filefd *ffd,
246 struct file *file, int fd)
247{
248 ffd->file = file;
249 ffd->fd = fd;
250}
251
Davide Libenzi67647d02007-05-15 01:40:52 -0700252/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700253static inline int ep_cmp_ffd(struct epoll_filefd *p1,
254 struct epoll_filefd *p2)
255{
256 return (p1->file > p2->file ? +1:
257 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
258}
259
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700260/* Tells us if the item is currently linked */
261static inline int ep_is_linked(struct list_head *p)
262{
263 return !list_empty(p);
264}
265
266/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700267static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700268{
269 return container_of(p, struct eppoll_entry, wait)->base;
270}
271
272/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700273static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700274{
275 return container_of(p, struct ep_pqueue, pt)->epi;
276}
277
278/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700279static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700280{
281 return op != EPOLL_CTL_DEL;
282}
283
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284/* Initialize the poll safe wake up structure */
285static void ep_poll_safewake_init(struct poll_safewake *psw)
286{
287
288 INIT_LIST_HEAD(&psw->wake_task_list);
289 spin_lock_init(&psw->lock);
290}
291
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292/*
293 * Perform a safe wake up of the poll wait list. The problem is that
294 * with the new callback'd wake up system, it is possible that the
295 * poll callback is reentered from inside the call to wake_up() done
296 * on the poll wait queue head. The rule is that we cannot reenter the
297 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
298 * and we cannot reenter the same wait queue head at all. This will
299 * enable to have a hierarchy of epoll file descriptor of no more than
300 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
301 * because this one gets called by the poll callback, that in turn is called
302 * from inside a wake_up(), that might be called from irq context.
303 */
304static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
305{
306 int wake_nests = 0;
307 unsigned long flags;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700308 struct task_struct *this_task = current;
Matthias Kaehlckeb70c3942007-10-18 23:39:56 -0700309 struct list_head *lsthead = &psw->wake_task_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310 struct wake_task_node *tncur;
311 struct wake_task_node tnode;
312
313 spin_lock_irqsave(&psw->lock, flags);
314
315 /* Try to see if the current task is already inside this wakeup call */
Matthias Kaehlckeb70c3942007-10-18 23:39:56 -0700316 list_for_each_entry(tncur, lsthead, llink) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317
318 if (tncur->wq == wq ||
319 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
320 /*
321 * Ops ... loop detected or maximum nest level reached.
322 * We abort this wake by breaking the cycle itself.
323 */
324 spin_unlock_irqrestore(&psw->lock, flags);
325 return;
326 }
327 }
328
329 /* Add the current task to the list */
330 tnode.task = this_task;
331 tnode.wq = wq;
332 list_add(&tnode.llink, lsthead);
333
334 spin_unlock_irqrestore(&psw->lock, flags);
335
336 /* Do really wake up now */
Peter Zijlstra0ccf8312008-02-04 22:27:20 -0800337 wake_up_nested(wq, 1 + wake_nests);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338
339 /* Remove the current task from the list */
340 spin_lock_irqsave(&psw->lock, flags);
341 list_del(&tnode.llink);
342 spin_unlock_irqrestore(&psw->lock, flags);
343}
344
Davide Libenzi7699acd2007-05-10 22:23:23 -0700345/*
346 * This function unregister poll callbacks from the associated file descriptor.
347 * Since this must be called without holding "ep->lock" the atomic exchange trick
348 * will protect us from multiple unregister.
349 */
350static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
351{
352 int nwait;
353 struct list_head *lsthead = &epi->pwqlist;
354 struct eppoll_entry *pwq;
355
356 /* This is called without locks, so we need the atomic exchange */
357 nwait = xchg(&epi->nwait, 0);
358
359 if (nwait) {
360 while (!list_empty(lsthead)) {
361 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
362
363 list_del_init(&pwq->llink);
364 remove_wait_queue(pwq->whead, &pwq->wait);
365 kmem_cache_free(pwq_cache, pwq);
366 }
367 }
368}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700371 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700372 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700373 */
374static int ep_remove(struct eventpoll *ep, struct epitem *epi)
375{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700376 unsigned long flags;
377 struct file *file = epi->ffd.file;
378
379 /*
380 * Removes poll wait queue hooks. We _have_ to do this without holding
381 * the "ep->lock" otherwise a deadlock might occur. This because of the
382 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
383 * queue head lock when unregistering the wait queue. The wakeup callback
384 * will run by holding the wait queue head lock and will call our callback
385 * that will try to get "ep->lock".
386 */
387 ep_unregister_pollwait(ep, epi);
388
389 /* Remove the current item from the list of epoll hooks */
390 spin_lock(&file->f_ep_lock);
391 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700392 list_del_init(&epi->fllink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700393 spin_unlock(&file->f_ep_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
Davide Libenzicdac75e2008-04-29 00:58:34 -0700395 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Davide Libenzic7ea7632007-05-15 01:40:47 -0700397 spin_lock_irqsave(&ep->lock, flags);
398 if (ep_is_linked(&epi->rdllink))
399 list_del_init(&epi->rdllink);
400 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
Davide Libenzi7699acd2007-05-10 22:23:23 -0700402 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700403 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
Davide Libenzic7ea7632007-05-15 01:40:47 -0700405 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
406 current, ep, file));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
Davide Libenzic7ea7632007-05-15 01:40:47 -0700408 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409}
410
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411static void ep_free(struct eventpoll *ep)
412{
413 struct rb_node *rbp;
414 struct epitem *epi;
415
416 /* We need to release all tasks waiting for these file */
417 if (waitqueue_active(&ep->poll_wait))
418 ep_poll_safewake(&psw, &ep->poll_wait);
419
420 /*
421 * We need to lock this because we could be hit by
422 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700423 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 * is on the way to be removed and no one has references to it
425 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800426 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800428 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
430 /*
431 * Walks through the whole tree by unregistering poll callbacks.
432 */
433 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
434 epi = rb_entry(rbp, struct epitem, rbn);
435
436 ep_unregister_pollwait(ep, epi);
437 }
438
439 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700440 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700442 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 * us during this operation. So we can avoid the lock on "ep->lock".
444 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700445 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 epi = rb_entry(rbp, struct epitem, rbn);
447 ep_remove(ep, epi);
448 }
449
Arjan van de Ven144efe32006-03-23 03:00:32 -0800450 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700451 mutex_destroy(&ep->mtx);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700452 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453}
454
Davide Libenzi7699acd2007-05-10 22:23:23 -0700455static int ep_eventpoll_release(struct inode *inode, struct file *file)
456{
457 struct eventpoll *ep = file->private_data;
458
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700459 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700460 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700461
462 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
463 return 0;
464}
465
466static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
467{
468 unsigned int pollflags = 0;
469 unsigned long flags;
470 struct eventpoll *ep = file->private_data;
471
472 /* Insert inside our poll wait queue */
473 poll_wait(file, &ep->poll_wait, wait);
474
475 /* Check our condition */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700476 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700477 if (!list_empty(&ep->rdllist))
478 pollflags = POLLIN | POLLRDNORM;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700479 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700480
481 return pollflags;
482}
483
484/* File callbacks that implement the eventpoll file behaviour */
485static const struct file_operations eventpoll_fops = {
486 .release = ep_eventpoll_release,
487 .poll = ep_eventpoll_poll
488};
489
490/* Fast test to see if the file is an evenpoll file */
491static inline int is_file_epoll(struct file *f)
492{
493 return f->f_op == &eventpoll_fops;
494}
495
496/*
497 * This is called from eventpoll_release() to unlink files from the eventpoll
498 * interface. We need to have this facility to cleanup correctly files that are
499 * closed without being removed from the eventpoll interface.
500 */
501void eventpoll_release_file(struct file *file)
502{
503 struct list_head *lsthead = &file->f_ep_links;
504 struct eventpoll *ep;
505 struct epitem *epi;
506
507 /*
508 * We don't want to get "file->f_ep_lock" because it is not
509 * necessary. It is not necessary because we're in the "struct file"
510 * cleanup path, and this means that noone is using this file anymore.
Davide Libenzi67647d02007-05-15 01:40:52 -0700511 * So, for example, epoll_ctl() cannot hit here sicne if we reach this
512 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700513 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700514 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700515 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700516 * from anywhere but ep_free().
517 */
518 mutex_lock(&epmutex);
519
520 while (!list_empty(lsthead)) {
521 epi = list_first_entry(lsthead, struct epitem, fllink);
522
523 ep = epi->ep;
524 list_del_init(&epi->fllink);
Davide Libenzid47de162007-05-15 01:40:41 -0700525 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700526 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700527 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700528 }
529
530 mutex_unlock(&epmutex);
531}
532
533static int ep_alloc(struct eventpoll **pep)
534{
535 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
536
537 if (!ep)
538 return -ENOMEM;
539
Davide Libenzic7ea7632007-05-15 01:40:47 -0700540 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700541 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700542 init_waitqueue_head(&ep->wq);
543 init_waitqueue_head(&ep->poll_wait);
544 INIT_LIST_HEAD(&ep->rdllist);
545 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700546 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700547
548 *pep = ep;
549
550 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
551 current, ep));
552 return 0;
553}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554
555/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700556 * Search the file inside the eventpoll tree. The RB tree operations
557 * are protected by the "mtx" mutex, and ep_find() must be called with
558 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 */
560static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
561{
562 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 struct rb_node *rbp;
564 struct epitem *epi, *epir = NULL;
565 struct epoll_filefd ffd;
566
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700567 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 for (rbp = ep->rbr.rb_node; rbp; ) {
569 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700570 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 if (kcmp > 0)
572 rbp = rbp->rb_right;
573 else if (kcmp < 0)
574 rbp = rbp->rb_left;
575 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 epir = epi;
577 break;
578 }
579 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580
581 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
582 current, file, epir));
583
584 return epir;
585}
586
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700588 * This is the callback that is passed to the wait queue wakeup
589 * machanism. It is called by the stored file descriptors when they
590 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700592static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700594 int pwake = 0;
595 unsigned long flags;
596 struct epitem *epi = ep_item_from_wait(wait);
597 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598
Davide Libenzi7699acd2007-05-10 22:23:23 -0700599 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
600 current, epi->ffd.file, epi, ep));
601
Davide Libenzic7ea7632007-05-15 01:40:47 -0700602 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700603
604 /*
605 * If the event mask does not contain any poll(2) event, we consider the
606 * descriptor to be disabled. This condition is likely the effect of the
607 * EPOLLONESHOT bit that disables the descriptor when an event is received,
608 * until the next EPOLL_CTL_MOD will be issued.
609 */
610 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700611 goto out_unlock;
612
613 /*
614 * If we are trasfering events to userspace, we can hold no locks
615 * (because we're accessing user memory, and because of linux f_op->poll()
616 * semantics). All the events that happens during that period of time are
617 * chained in ep->ovflist and requeued later on.
618 */
619 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
620 if (epi->next == EP_UNACTIVE_PTR) {
621 epi->next = ep->ovflist;
622 ep->ovflist = epi;
623 }
624 goto out_unlock;
625 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700626
627 /* If this file is already in the ready list we exit soon */
628 if (ep_is_linked(&epi->rdllink))
629 goto is_linked;
630
631 list_add_tail(&epi->rdllink, &ep->rdllist);
632
633is_linked:
634 /*
635 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
636 * wait list.
637 */
638 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400639 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700640 if (waitqueue_active(&ep->poll_wait))
641 pwake++;
642
Davide Libenzid47de162007-05-15 01:40:41 -0700643out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700644 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700645
646 /* We have to call this outside the lock */
647 if (pwake)
648 ep_poll_safewake(&psw, &ep->poll_wait);
649
650 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651}
652
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653/*
654 * This is the callback that is used to add our wait queue to the
655 * target file wakeup lists.
656 */
657static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
658 poll_table *pt)
659{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700660 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 struct eppoll_entry *pwq;
662
Christoph Lametere94b1762006-12-06 20:33:17 -0800663 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
665 pwq->whead = whead;
666 pwq->base = epi;
667 add_wait_queue(whead, &pwq->wait);
668 list_add_tail(&pwq->llink, &epi->pwqlist);
669 epi->nwait++;
670 } else {
671 /* We have to signal that an error occurred */
672 epi->nwait = -1;
673 }
674}
675
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
677{
678 int kcmp;
679 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
680 struct epitem *epic;
681
682 while (*p) {
683 parent = *p;
684 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700685 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 if (kcmp > 0)
687 p = &parent->rb_right;
688 else
689 p = &parent->rb_left;
690 }
691 rb_link_node(&epi->rbn, parent, p);
692 rb_insert_color(&epi->rbn, &ep->rbr);
693}
694
Davide Libenzic7ea7632007-05-15 01:40:47 -0700695/*
696 * Must be called with "mtx" held.
697 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
699 struct file *tfile, int fd)
700{
701 int error, revents, pwake = 0;
702 unsigned long flags;
703 struct epitem *epi;
704 struct ep_pqueue epq;
705
706 error = -ENOMEM;
Christoph Lametere94b1762006-12-06 20:33:17 -0800707 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7699acd2007-05-10 22:23:23 -0700708 goto error_return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
710 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 INIT_LIST_HEAD(&epi->rdllink);
712 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 INIT_LIST_HEAD(&epi->pwqlist);
714 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700715 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -0700718 epi->next = EP_UNACTIVE_PTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
720 /* Initialize the poll table using the queue callback */
721 epq.epi = epi;
722 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
723
724 /*
725 * Attach the item to the poll hooks and get current event bits.
726 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -0700727 * been increased by the caller of this function. Note that after
728 * this operation completes, the poll callback can start hitting
729 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 */
731 revents = tfile->f_op->poll(tfile, &epq.pt);
732
733 /*
734 * We have to check if something went wrong during the poll wait queue
735 * install process. Namely an allocation for a wait queue failed due
736 * high memory pressure.
737 */
738 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700739 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740
741 /* Add the current item to the list of active epoll hook for this file */
742 spin_lock(&tfile->f_ep_lock);
743 list_add_tail(&epi->fllink, &tfile->f_ep_links);
744 spin_unlock(&tfile->f_ep_lock);
745
Davide Libenzic7ea7632007-05-15 01:40:47 -0700746 /*
747 * Add the current item to the RB tree. All RB tree operations are
748 * protected by "mtx", and ep_insert() is called with "mtx" held.
749 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 ep_rbtree_insert(ep, epi);
751
Davide Libenzic7ea7632007-05-15 01:40:47 -0700752 /* We have to drop the new item inside our item list to keep track of it */
753 spin_lock_irqsave(&ep->lock, flags);
754
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700756 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 list_add_tail(&epi->rdllink, &ep->rdllist);
758
759 /* Notify waiting tasks that events are available */
760 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400761 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 if (waitqueue_active(&ep->poll_wait))
763 pwake++;
764 }
765
Davide Libenzic7ea7632007-05-15 01:40:47 -0700766 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767
768 /* We have to call this outside the lock */
769 if (pwake)
770 ep_poll_safewake(&psw, &ep->poll_wait);
771
772 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
773 current, ep, tfile, fd));
774
775 return 0;
776
Davide Libenzi7699acd2007-05-10 22:23:23 -0700777error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 ep_unregister_pollwait(ep, epi);
779
780 /*
781 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -0700782 * allocated wait queue. Note that we don't care about the ep->ovflist
783 * list, since that is used/cleaned only inside a section bound by "mtx".
784 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700786 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700787 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700788 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700789 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700791 kmem_cache_free(epi_cache, epi);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700792error_return:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 return error;
794}
795
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796/*
797 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -0700798 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 */
800static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
801{
802 int pwake = 0;
803 unsigned int revents;
804 unsigned long flags;
805
806 /*
807 * Set the new event interest mask before calling f_op->poll(), otherwise
808 * a potential race might occur. In fact if we do this operation inside
809 * the lock, an event might happen between the f_op->poll() call and the
810 * new event set registering.
811 */
812 epi->event.events = event->events;
813
814 /*
815 * Get current event bits. We can safely use the file* here because
816 * its usage count has been increased by the caller of this function.
817 */
818 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
819
Davide Libenzic7ea7632007-05-15 01:40:47 -0700820 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821
822 /* Copy the data member from inside the lock */
823 epi->event.data = event->data;
824
825 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700826 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -0700827 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700829 if (revents & event->events) {
830 if (!ep_is_linked(&epi->rdllink)) {
831 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832
Davide Libenzic7ea7632007-05-15 01:40:47 -0700833 /* Notify waiting tasks that events are available */
834 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400835 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700836 if (waitqueue_active(&ep->poll_wait))
837 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 }
839 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700840 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841
842 /* We have to call this outside the lock */
843 if (pwake)
844 ep_poll_safewake(&psw, &ep->poll_wait);
845
846 return 0;
847}
848
Davide Libenzid47de162007-05-15 01:40:41 -0700849static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
850 int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851{
Davide Libenzi6192bd52007-05-08 00:25:41 -0700852 int eventcnt, error = -EFAULT, pwake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 unsigned int revents;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700854 unsigned long flags;
Davide Libenzid47de162007-05-15 01:40:41 -0700855 struct epitem *epi, *nepi;
856 struct list_head txlist;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700857
Davide Libenzid47de162007-05-15 01:40:41 -0700858 INIT_LIST_HEAD(&txlist);
859
860 /*
861 * We need to lock this because we could be hit by
862 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
863 */
864 mutex_lock(&ep->mtx);
865
866 /*
867 * Steal the ready list, and re-init the original one to the
868 * empty list. Also, set ep->ovflist to NULL so that events
869 * happening while looping w/out locks, are not lost. We cannot
870 * have the poll callback to queue directly on ep->rdllist,
871 * because we are doing it in the loop below, in a lockless way.
872 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700873 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700874 list_splice(&ep->rdllist, &txlist);
875 INIT_LIST_HEAD(&ep->rdllist);
876 ep->ovflist = NULL;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700877 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878
879 /*
880 * We can loop without lock because this is a task private list.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700881 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
Davide Libenzid47de162007-05-15 01:40:41 -0700882 * Items cannot vanish during the loop because we are holding "mtx".
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 */
Davide Libenzid47de162007-05-15 01:40:41 -0700884 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
885 epi = list_first_entry(&txlist, struct epitem, rdllink);
886
887 list_del_init(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888
889 /*
890 * Get the ready file event set. We can safely use the file
Davide Libenzid47de162007-05-15 01:40:41 -0700891 * because we are holding the "mtx" and this will guarantee
892 * that both the file and the item will not vanish.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 */
894 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700895 revents &= epi->event.events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
897 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700898 * Is the event mask intersect the caller-requested one,
899 * deliver the event to userspace. Again, we are holding
Davide Libenzid47de162007-05-15 01:40:41 -0700900 * "mtx", so no operations coming from userspace can change
901 * the item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700903 if (revents) {
904 if (__put_user(revents,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 &events[eventcnt].events) ||
906 __put_user(epi->event.data,
907 &events[eventcnt].data))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700908 goto errxit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 if (epi->event.events & EPOLLONESHOT)
910 epi->event.events &= EP_PRIVATE_BITS;
911 eventcnt++;
912 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700914 * At this point, noone can insert into ep->rdllist besides
915 * us. The epoll_ctl() callers are locked out by us holding
916 * "mtx" and the poll callback will queue them in ep->ovflist.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700918 if (!(epi->event.events & EPOLLET) &&
Davide Libenzid47de162007-05-15 01:40:41 -0700919 (revents & epi->event.events))
920 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 }
Davide Libenzi6192bd52007-05-08 00:25:41 -0700922 error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923
Davide Libenzid47de162007-05-15 01:40:41 -0700924errxit:
925
Davide Libenzic7ea7632007-05-15 01:40:47 -0700926 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700927 /*
928 * During the time we spent in the loop above, some other events
929 * might have been queued by the poll callback. We re-insert them
930 * here (in case they are not already queued, or they're one-shot).
931 */
932 for (nepi = ep->ovflist; (epi = nepi) != NULL;
933 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
934 if (!ep_is_linked(&epi->rdllink) &&
935 (epi->event.events & ~EP_PRIVATE_BITS))
936 list_add_tail(&epi->rdllink, &ep->rdllist);
937 }
938 /*
939 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
940 * releasing the lock, events will be queued in the normal way inside
941 * ep->rdllist.
942 */
943 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700944
945 /*
Davide Libenzi67647d02007-05-15 01:40:52 -0700946 * In case of error in the event-send loop, or in case the number of
947 * ready events exceeds the userspace limit, we need to splice the
948 * "txlist" back inside ep->rdllist.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700949 */
Davide Libenzid47de162007-05-15 01:40:41 -0700950 list_splice(&txlist, &ep->rdllist);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700951
Davide Libenzid47de162007-05-15 01:40:41 -0700952 if (!list_empty(&ep->rdllist)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700954 * Wake up (if active) both the eventpoll wait list and the ->poll()
Davide Libenzi67647d02007-05-15 01:40:52 -0700955 * wait list (delayed after we release the lock).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 */
957 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400958 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 if (waitqueue_active(&ep->poll_wait))
960 pwake++;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700961 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700962 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700963
964 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
966 /* We have to call this outside the lock */
967 if (pwake)
968 ep_poll_safewake(&psw, &ep->poll_wait);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700969
970 return eventcnt == 0 ? error: eventcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971}
972
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
974 int maxevents, long timeout)
975{
976 int res, eavail;
977 unsigned long flags;
978 long jtimeout;
979 wait_queue_t wait;
980
981 /*
982 * Calculate the timeout by checking for the "infinite" value ( -1 )
983 * and the overflow condition. The passed timeout is in milliseconds,
984 * that why (t * HZ) / 1000.
985 */
Davide Libenzie3306dd2005-09-27 21:45:33 -0700986 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
987 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988
989retry:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700990 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991
992 res = 0;
993 if (list_empty(&ep->rdllist)) {
994 /*
995 * We don't have any available event to return to the caller.
996 * We need to sleep here, and we will be wake up by
997 * ep_poll_callback() when events will become available.
998 */
999 init_waitqueue_entry(&wait, current);
Davide Libenzid47de162007-05-15 01:40:41 -07001000 wait.flags |= WQ_FLAG_EXCLUSIVE;
Davide Libenzi3419b232006-06-25 05:48:14 -07001001 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002
1003 for (;;) {
1004 /*
1005 * We don't want to sleep if the ep_poll_callback() sends us
1006 * a wakeup in between. That's why we set the task state
1007 * to TASK_INTERRUPTIBLE before doing the checks.
1008 */
1009 set_current_state(TASK_INTERRUPTIBLE);
1010 if (!list_empty(&ep->rdllist) || !jtimeout)
1011 break;
1012 if (signal_pending(current)) {
1013 res = -EINTR;
1014 break;
1015 }
1016
Davide Libenzic7ea7632007-05-15 01:40:47 -07001017 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 jtimeout = schedule_timeout(jtimeout);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001019 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001021 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022
1023 set_current_state(TASK_RUNNING);
1024 }
1025
1026 /* Is it worth to try to dig for events ? */
1027 eavail = !list_empty(&ep->rdllist);
1028
Davide Libenzic7ea7632007-05-15 01:40:47 -07001029 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030
1031 /*
1032 * Try to transfer events to user space. In case we get 0 events and
1033 * there's still timeout left over, we go trying again in search of
1034 * more luck.
1035 */
1036 if (!res && eavail &&
Davide Libenzid47de162007-05-15 01:40:41 -07001037 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 goto retry;
1039
1040 return res;
1041}
1042
Davide Libenzi7699acd2007-05-10 22:23:23 -07001043/*
Davide Libenzi67647d02007-05-15 01:40:52 -07001044 * It opens an eventpoll file descriptor. The "size" parameter is there
1045 * for historical reasons, when epoll was using an hash instead of an
1046 * RB tree. With the current implementation, the "size" parameter is ignored
1047 * (besides sanity checks).
Davide Libenzi7699acd2007-05-10 22:23:23 -07001048 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001049asmlinkage long sys_epoll_create1(int flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001050{
1051 int error, fd = -1;
1052 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001053
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001054 /* Check the EPOLL_* constant for consistency. */
1055 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1056
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001057 if (flags & ~EPOLL_CLOEXEC)
1058 return -EINVAL;
1059
Davide Libenzi7699acd2007-05-10 22:23:23 -07001060 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001061 current, flags));
Davide Libenzi7699acd2007-05-10 22:23:23 -07001062
1063 /*
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001064 * Create the internal data structure ( "struct eventpoll" ).
Davide Libenzi7699acd2007-05-10 22:23:23 -07001065 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001066 error = ep_alloc(&ep);
1067 if (error < 0) {
Al Viro2030a422008-02-23 06:46:49 -05001068 fd = error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001069 goto error_return;
Al Viro2030a422008-02-23 06:46:49 -05001070 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001071
1072 /*
1073 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001074 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001075 */
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001076 fd = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1077 flags & O_CLOEXEC);
Al Viro2030a422008-02-23 06:46:49 -05001078 if (fd < 0)
1079 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001080
Al Viro2030a422008-02-23 06:46:49 -05001081error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001082 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001083 current, flags, fd));
Davide Libenzi7699acd2007-05-10 22:23:23 -07001084
1085 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001086}
1087
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001088asmlinkage long sys_epoll_create(int size)
1089{
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001090 if (size < 0)
1091 return -EINVAL;
1092
1093 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001094}
1095
Davide Libenzi7699acd2007-05-10 22:23:23 -07001096/*
1097 * The following function implements the controller interface for
1098 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001099 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001100 */
1101asmlinkage long sys_epoll_ctl(int epfd, int op, int fd,
1102 struct epoll_event __user *event)
1103{
1104 int error;
1105 struct file *file, *tfile;
1106 struct eventpoll *ep;
1107 struct epitem *epi;
1108 struct epoll_event epds;
1109
1110 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1111 current, epfd, op, fd, event));
1112
1113 error = -EFAULT;
1114 if (ep_op_has_event(op) &&
1115 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1116 goto error_return;
1117
1118 /* Get the "struct file *" for the eventpoll file */
1119 error = -EBADF;
1120 file = fget(epfd);
1121 if (!file)
1122 goto error_return;
1123
1124 /* Get the "struct file *" for the target file */
1125 tfile = fget(fd);
1126 if (!tfile)
1127 goto error_fput;
1128
1129 /* The target file descriptor must support poll */
1130 error = -EPERM;
1131 if (!tfile->f_op || !tfile->f_op->poll)
1132 goto error_tgt_fput;
1133
1134 /*
1135 * We have to check that the file structure underneath the file descriptor
1136 * the user passed to us _is_ an eventpoll file. And also we do not permit
1137 * adding an epoll file descriptor inside itself.
1138 */
1139 error = -EINVAL;
1140 if (file == tfile || !is_file_epoll(file))
1141 goto error_tgt_fput;
1142
1143 /*
1144 * At this point it is safe to assume that the "private_data" contains
1145 * our own data structure.
1146 */
1147 ep = file->private_data;
1148
Davide Libenzid47de162007-05-15 01:40:41 -07001149 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001150
Davide Libenzi67647d02007-05-15 01:40:52 -07001151 /*
1152 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1153 * above, we can be sure to be able to use the item looked up by
1154 * ep_find() till we release the mutex.
1155 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001156 epi = ep_find(ep, tfile, fd);
1157
1158 error = -EINVAL;
1159 switch (op) {
1160 case EPOLL_CTL_ADD:
1161 if (!epi) {
1162 epds.events |= POLLERR | POLLHUP;
1163
1164 error = ep_insert(ep, &epds, tfile, fd);
1165 } else
1166 error = -EEXIST;
1167 break;
1168 case EPOLL_CTL_DEL:
1169 if (epi)
1170 error = ep_remove(ep, epi);
1171 else
1172 error = -ENOENT;
1173 break;
1174 case EPOLL_CTL_MOD:
1175 if (epi) {
1176 epds.events |= POLLERR | POLLHUP;
1177 error = ep_modify(ep, epi, &epds);
1178 } else
1179 error = -ENOENT;
1180 break;
1181 }
Davide Libenzid47de162007-05-15 01:40:41 -07001182 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001183
1184error_tgt_fput:
1185 fput(tfile);
1186error_fput:
1187 fput(file);
1188error_return:
1189 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1190 current, epfd, op, fd, event, error));
1191
1192 return error;
1193}
1194
1195/*
1196 * Implement the event wait interface for the eventpoll file. It is the kernel
1197 * part of the user space epoll_wait(2).
1198 */
1199asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
1200 int maxevents, int timeout)
1201{
1202 int error;
1203 struct file *file;
1204 struct eventpoll *ep;
1205
1206 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1207 current, epfd, events, maxevents, timeout));
1208
1209 /* The maximum number of event must be greater than zero */
1210 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1211 return -EINVAL;
1212
1213 /* Verify that the area passed by the user is writeable */
1214 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1215 error = -EFAULT;
1216 goto error_return;
1217 }
1218
1219 /* Get the "struct file *" for the eventpoll file */
1220 error = -EBADF;
1221 file = fget(epfd);
1222 if (!file)
1223 goto error_return;
1224
1225 /*
1226 * We have to check that the file structure underneath the fd
1227 * the user passed to us _is_ an eventpoll file.
1228 */
1229 error = -EINVAL;
1230 if (!is_file_epoll(file))
1231 goto error_fput;
1232
1233 /*
1234 * At this point it is safe to assume that the "private_data" contains
1235 * our own data structure.
1236 */
1237 ep = file->private_data;
1238
1239 /* Time to fish for events ... */
1240 error = ep_poll(ep, events, maxevents, timeout);
1241
1242error_fput:
1243 fput(file);
1244error_return:
1245 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1246 current, epfd, events, maxevents, timeout, error));
1247
1248 return error;
1249}
1250
Roland McGrathf3de2722008-04-30 00:53:09 -07001251#ifdef HAVE_SET_RESTORE_SIGMASK
Davide Libenzi7699acd2007-05-10 22:23:23 -07001252
1253/*
1254 * Implement the event wait interface for the eventpoll file. It is the kernel
1255 * part of the user space epoll_pwait(2).
1256 */
1257asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
1258 int maxevents, int timeout, const sigset_t __user *sigmask,
1259 size_t sigsetsize)
1260{
1261 int error;
1262 sigset_t ksigmask, sigsaved;
1263
1264 /*
1265 * If the caller wants a certain signal mask to be set during the wait,
1266 * we apply it here.
1267 */
1268 if (sigmask) {
1269 if (sigsetsize != sizeof(sigset_t))
1270 return -EINVAL;
1271 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1272 return -EFAULT;
1273 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1274 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1275 }
1276
1277 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1278
1279 /*
1280 * If we changed the signal mask, we need to restore the original one.
1281 * In case we've got a signal while waiting, we do not restore the
1282 * signal mask yet, and we allow do_signal() to deliver the signal on
1283 * the way back to userspace, before the signal mask is restored.
1284 */
1285 if (sigmask) {
1286 if (error == -EINTR) {
1287 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001288 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001289 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001290 } else
1291 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1292 }
1293
1294 return error;
1295}
1296
Roland McGrathf3de2722008-04-30 00:53:09 -07001297#endif /* HAVE_SET_RESTORE_SIGMASK */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001298
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299static int __init eventpoll_init(void)
1300{
Arjan van de Ven144efe32006-03-23 03:00:32 -08001301 mutex_init(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302
1303 /* Initialize the structure used to perform safe poll wait head wake ups */
1304 ep_poll_safewake_init(&psw);
1305
1306 /* Allocates slab cache used to allocate "struct epitem" items */
1307 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1308 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001309 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310
1311 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1312 pwq_cache = kmem_cache_create("eventpoll_pwq",
1313 sizeof(struct eppoll_entry), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09001314 EPI_SLAB_DEBUG|SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317}
Davide Libenzicea69242007-05-10 22:23:22 -07001318fs_initcall(eventpoll_init);