blob: 34f68f3a069a1b7d4924dc8be718927f9e6c791d [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
Davide Libenzi67647d02007-05-15 01:40:52 -0700260/* Special initialization for the RB tree node to detect linkage */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700261static inline void ep_rb_initnode(struct rb_node *n)
262{
David Woodhousec5698822006-04-21 13:17:24 +0100263 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700264}
265
Davide Libenzi67647d02007-05-15 01:40:52 -0700266/* Removes a node from the RB tree and marks it for a fast is-linked check */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700267static inline void ep_rb_erase(struct rb_node *n, struct rb_root *r)
268{
269 rb_erase(n, r);
David Woodhousec5698822006-04-21 13:17:24 +0100270 rb_set_parent(n, n);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700271}
272
Davide Libenzi67647d02007-05-15 01:40:52 -0700273/* Fast check to verify that the item is linked to the main RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700274static inline int ep_rb_linked(struct rb_node *n)
275{
David Woodhousec5698822006-04-21 13:17:24 +0100276 return rb_parent(n) != n;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700277}
278
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700279/* Tells us if the item is currently linked */
280static inline int ep_is_linked(struct list_head *p)
281{
282 return !list_empty(p);
283}
284
285/* Get the "struct epitem" from a wait queue pointer */
286static inline struct epitem * ep_item_from_wait(wait_queue_t *p)
287{
288 return container_of(p, struct eppoll_entry, wait)->base;
289}
290
291/* Get the "struct epitem" from an epoll queue wrapper */
292static inline struct epitem * ep_item_from_epqueue(poll_table *p)
293{
294 return container_of(p, struct ep_pqueue, pt)->epi;
295}
296
297/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700298static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700299{
300 return op != EPOLL_CTL_DEL;
301}
302
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303/* Initialize the poll safe wake up structure */
304static void ep_poll_safewake_init(struct poll_safewake *psw)
305{
306
307 INIT_LIST_HEAD(&psw->wake_task_list);
308 spin_lock_init(&psw->lock);
309}
310
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311/*
312 * Perform a safe wake up of the poll wait list. The problem is that
313 * with the new callback'd wake up system, it is possible that the
314 * poll callback is reentered from inside the call to wake_up() done
315 * on the poll wait queue head. The rule is that we cannot reenter the
316 * wake up code from the same task more than EP_MAX_POLLWAKE_NESTS times,
317 * and we cannot reenter the same wait queue head at all. This will
318 * enable to have a hierarchy of epoll file descriptor of no more than
319 * EP_MAX_POLLWAKE_NESTS deep. We need the irq version of the spin lock
320 * because this one gets called by the poll callback, that in turn is called
321 * from inside a wake_up(), that might be called from irq context.
322 */
323static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
324{
325 int wake_nests = 0;
326 unsigned long flags;
Ingo Molnar36c8b582006-07-03 00:25:41 -0700327 struct task_struct *this_task = current;
Matthias Kaehlckeb70c3942007-10-18 23:39:56 -0700328 struct list_head *lsthead = &psw->wake_task_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 struct wake_task_node *tncur;
330 struct wake_task_node tnode;
331
332 spin_lock_irqsave(&psw->lock, flags);
333
334 /* Try to see if the current task is already inside this wakeup call */
Matthias Kaehlckeb70c3942007-10-18 23:39:56 -0700335 list_for_each_entry(tncur, lsthead, llink) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336
337 if (tncur->wq == wq ||
338 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
339 /*
340 * Ops ... loop detected or maximum nest level reached.
341 * We abort this wake by breaking the cycle itself.
342 */
343 spin_unlock_irqrestore(&psw->lock, flags);
344 return;
345 }
346 }
347
348 /* Add the current task to the list */
349 tnode.task = this_task;
350 tnode.wq = wq;
351 list_add(&tnode.llink, lsthead);
352
353 spin_unlock_irqrestore(&psw->lock, flags);
354
355 /* Do really wake up now */
356 wake_up(wq);
357
358 /* Remove the current task from the list */
359 spin_lock_irqsave(&psw->lock, flags);
360 list_del(&tnode.llink);
361 spin_unlock_irqrestore(&psw->lock, flags);
362}
363
Davide Libenzi7699acd2007-05-10 22:23:23 -0700364/*
365 * This function unregister poll callbacks from the associated file descriptor.
366 * Since this must be called without holding "ep->lock" the atomic exchange trick
367 * will protect us from multiple unregister.
368 */
369static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
370{
371 int nwait;
372 struct list_head *lsthead = &epi->pwqlist;
373 struct eppoll_entry *pwq;
374
375 /* This is called without locks, so we need the atomic exchange */
376 nwait = xchg(&epi->nwait, 0);
377
378 if (nwait) {
379 while (!list_empty(lsthead)) {
380 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
381
382 list_del_init(&pwq->llink);
383 remove_wait_queue(pwq->whead, &pwq->wait);
384 kmem_cache_free(pwq_cache, pwq);
385 }
386 }
387}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700390 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700391 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700392 */
393static int ep_remove(struct eventpoll *ep, struct epitem *epi)
394{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700395 unsigned long flags;
396 struct file *file = epi->ffd.file;
397
398 /*
399 * Removes poll wait queue hooks. We _have_ to do this without holding
400 * the "ep->lock" otherwise a deadlock might occur. This because of the
401 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
402 * queue head lock when unregistering the wait queue. The wakeup callback
403 * will run by holding the wait queue head lock and will call our callback
404 * that will try to get "ep->lock".
405 */
406 ep_unregister_pollwait(ep, epi);
407
408 /* Remove the current item from the list of epoll hooks */
409 spin_lock(&file->f_ep_lock);
410 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700411 list_del_init(&epi->fllink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700412 spin_unlock(&file->f_ep_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
Davide Libenzic7ea7632007-05-15 01:40:47 -0700414 if (ep_rb_linked(&epi->rbn))
415 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Davide Libenzic7ea7632007-05-15 01:40:47 -0700417 spin_lock_irqsave(&ep->lock, flags);
418 if (ep_is_linked(&epi->rdllink))
419 list_del_init(&epi->rdllink);
420 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421
Davide Libenzi7699acd2007-05-10 22:23:23 -0700422 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700423 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424
Davide Libenzic7ea7632007-05-15 01:40:47 -0700425 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
426 current, ep, file));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427
Davide Libenzic7ea7632007-05-15 01:40:47 -0700428 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429}
430
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431static void ep_free(struct eventpoll *ep)
432{
433 struct rb_node *rbp;
434 struct epitem *epi;
435
436 /* We need to release all tasks waiting for these file */
437 if (waitqueue_active(&ep->poll_wait))
438 ep_poll_safewake(&psw, &ep->poll_wait);
439
440 /*
441 * We need to lock this because we could be hit by
442 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700443 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 * is on the way to be removed and no one has references to it
445 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800446 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800448 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
450 /*
451 * Walks through the whole tree by unregistering poll callbacks.
452 */
453 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
454 epi = rb_entry(rbp, struct epitem, rbn);
455
456 ep_unregister_pollwait(ep, epi);
457 }
458
459 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700460 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700462 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 * us during this operation. So we can avoid the lock on "ep->lock".
464 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700465 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 epi = rb_entry(rbp, struct epitem, rbn);
467 ep_remove(ep, epi);
468 }
469
Arjan van de Ven144efe32006-03-23 03:00:32 -0800470 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700471 mutex_destroy(&ep->mtx);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700472 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
Davide Libenzi7699acd2007-05-10 22:23:23 -0700475static int ep_eventpoll_release(struct inode *inode, struct file *file)
476{
477 struct eventpoll *ep = file->private_data;
478
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700479 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700480 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700481
482 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
483 return 0;
484}
485
486static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
487{
488 unsigned int pollflags = 0;
489 unsigned long flags;
490 struct eventpoll *ep = file->private_data;
491
492 /* Insert inside our poll wait queue */
493 poll_wait(file, &ep->poll_wait, wait);
494
495 /* Check our condition */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700496 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700497 if (!list_empty(&ep->rdllist))
498 pollflags = POLLIN | POLLRDNORM;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700499 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700500
501 return pollflags;
502}
503
504/* File callbacks that implement the eventpoll file behaviour */
505static const struct file_operations eventpoll_fops = {
506 .release = ep_eventpoll_release,
507 .poll = ep_eventpoll_poll
508};
509
510/* Fast test to see if the file is an evenpoll file */
511static inline int is_file_epoll(struct file *f)
512{
513 return f->f_op == &eventpoll_fops;
514}
515
516/*
517 * This is called from eventpoll_release() to unlink files from the eventpoll
518 * interface. We need to have this facility to cleanup correctly files that are
519 * closed without being removed from the eventpoll interface.
520 */
521void eventpoll_release_file(struct file *file)
522{
523 struct list_head *lsthead = &file->f_ep_links;
524 struct eventpoll *ep;
525 struct epitem *epi;
526
527 /*
528 * We don't want to get "file->f_ep_lock" because it is not
529 * necessary. It is not necessary because we're in the "struct file"
530 * cleanup path, and this means that noone is using this file anymore.
Davide Libenzi67647d02007-05-15 01:40:52 -0700531 * So, for example, epoll_ctl() cannot hit here sicne if we reach this
532 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700533 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700534 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700535 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700536 * from anywhere but ep_free().
537 */
538 mutex_lock(&epmutex);
539
540 while (!list_empty(lsthead)) {
541 epi = list_first_entry(lsthead, struct epitem, fllink);
542
543 ep = epi->ep;
544 list_del_init(&epi->fllink);
Davide Libenzid47de162007-05-15 01:40:41 -0700545 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700546 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700547 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700548 }
549
550 mutex_unlock(&epmutex);
551}
552
553static int ep_alloc(struct eventpoll **pep)
554{
555 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
556
557 if (!ep)
558 return -ENOMEM;
559
Davide Libenzic7ea7632007-05-15 01:40:47 -0700560 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700561 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700562 init_waitqueue_head(&ep->wq);
563 init_waitqueue_head(&ep->poll_wait);
564 INIT_LIST_HEAD(&ep->rdllist);
565 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700566 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700567
568 *pep = ep;
569
570 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
571 current, ep));
572 return 0;
573}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574
575/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700576 * Search the file inside the eventpoll tree. The RB tree operations
577 * are protected by the "mtx" mutex, and ep_find() must be called with
578 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 */
580static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
581{
582 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 struct rb_node *rbp;
584 struct epitem *epi, *epir = NULL;
585 struct epoll_filefd ffd;
586
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700587 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 for (rbp = ep->rbr.rb_node; rbp; ) {
589 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700590 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 if (kcmp > 0)
592 rbp = rbp->rb_right;
593 else if (kcmp < 0)
594 rbp = rbp->rb_left;
595 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 epir = epi;
597 break;
598 }
599 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600
601 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
602 current, file, epir));
603
604 return epir;
605}
606
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700608 * This is the callback that is passed to the wait queue wakeup
609 * machanism. It is called by the stored file descriptors when they
610 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700612static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700614 int pwake = 0;
615 unsigned long flags;
616 struct epitem *epi = ep_item_from_wait(wait);
617 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
Davide Libenzi7699acd2007-05-10 22:23:23 -0700619 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
620 current, epi->ffd.file, epi, ep));
621
Davide Libenzic7ea7632007-05-15 01:40:47 -0700622 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700623
624 /*
625 * If the event mask does not contain any poll(2) event, we consider the
626 * descriptor to be disabled. This condition is likely the effect of the
627 * EPOLLONESHOT bit that disables the descriptor when an event is received,
628 * until the next EPOLL_CTL_MOD will be issued.
629 */
630 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700631 goto out_unlock;
632
633 /*
634 * If we are trasfering events to userspace, we can hold no locks
635 * (because we're accessing user memory, and because of linux f_op->poll()
636 * semantics). All the events that happens during that period of time are
637 * chained in ep->ovflist and requeued later on.
638 */
639 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
640 if (epi->next == EP_UNACTIVE_PTR) {
641 epi->next = ep->ovflist;
642 ep->ovflist = epi;
643 }
644 goto out_unlock;
645 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700646
647 /* If this file is already in the ready list we exit soon */
648 if (ep_is_linked(&epi->rdllink))
649 goto is_linked;
650
651 list_add_tail(&epi->rdllink, &ep->rdllist);
652
653is_linked:
654 /*
655 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
656 * wait list.
657 */
658 if (waitqueue_active(&ep->wq))
659 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
660 TASK_INTERRUPTIBLE);
661 if (waitqueue_active(&ep->poll_wait))
662 pwake++;
663
Davide Libenzid47de162007-05-15 01:40:41 -0700664out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700665 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700666
667 /* We have to call this outside the lock */
668 if (pwake)
669 ep_poll_safewake(&psw, &ep->poll_wait);
670
671 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672}
673
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674/*
675 * This is the callback that is used to add our wait queue to the
676 * target file wakeup lists.
677 */
678static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
679 poll_table *pt)
680{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700681 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 struct eppoll_entry *pwq;
683
Christoph Lametere94b1762006-12-06 20:33:17 -0800684 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
686 pwq->whead = whead;
687 pwq->base = epi;
688 add_wait_queue(whead, &pwq->wait);
689 list_add_tail(&pwq->llink, &epi->pwqlist);
690 epi->nwait++;
691 } else {
692 /* We have to signal that an error occurred */
693 epi->nwait = -1;
694 }
695}
696
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
698{
699 int kcmp;
700 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
701 struct epitem *epic;
702
703 while (*p) {
704 parent = *p;
705 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700706 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 if (kcmp > 0)
708 p = &parent->rb_right;
709 else
710 p = &parent->rb_left;
711 }
712 rb_link_node(&epi->rbn, parent, p);
713 rb_insert_color(&epi->rbn, &ep->rbr);
714}
715
Davide Libenzic7ea7632007-05-15 01:40:47 -0700716/*
717 * Must be called with "mtx" held.
718 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
720 struct file *tfile, int fd)
721{
722 int error, revents, pwake = 0;
723 unsigned long flags;
724 struct epitem *epi;
725 struct ep_pqueue epq;
726
727 error = -ENOMEM;
Christoph Lametere94b1762006-12-06 20:33:17 -0800728 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7699acd2007-05-10 22:23:23 -0700729 goto error_return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730
731 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700732 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 INIT_LIST_HEAD(&epi->rdllink);
734 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 INIT_LIST_HEAD(&epi->pwqlist);
736 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700737 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -0700740 epi->next = EP_UNACTIVE_PTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741
742 /* Initialize the poll table using the queue callback */
743 epq.epi = epi;
744 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
745
746 /*
747 * Attach the item to the poll hooks and get current event bits.
748 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -0700749 * been increased by the caller of this function. Note that after
750 * this operation completes, the poll callback can start hitting
751 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 */
753 revents = tfile->f_op->poll(tfile, &epq.pt);
754
755 /*
756 * We have to check if something went wrong during the poll wait queue
757 * install process. Namely an allocation for a wait queue failed due
758 * high memory pressure.
759 */
760 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700761 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
763 /* Add the current item to the list of active epoll hook for this file */
764 spin_lock(&tfile->f_ep_lock);
765 list_add_tail(&epi->fllink, &tfile->f_ep_links);
766 spin_unlock(&tfile->f_ep_lock);
767
Davide Libenzic7ea7632007-05-15 01:40:47 -0700768 /*
769 * Add the current item to the RB tree. All RB tree operations are
770 * protected by "mtx", and ep_insert() is called with "mtx" held.
771 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 ep_rbtree_insert(ep, epi);
773
Davide Libenzic7ea7632007-05-15 01:40:47 -0700774 /* We have to drop the new item inside our item list to keep track of it */
775 spin_lock_irqsave(&ep->lock, flags);
776
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700778 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 list_add_tail(&epi->rdllink, &ep->rdllist);
780
781 /* Notify waiting tasks that events are available */
782 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -0700783 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 if (waitqueue_active(&ep->poll_wait))
785 pwake++;
786 }
787
Davide Libenzic7ea7632007-05-15 01:40:47 -0700788 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
790 /* We have to call this outside the lock */
791 if (pwake)
792 ep_poll_safewake(&psw, &ep->poll_wait);
793
794 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
795 current, ep, tfile, fd));
796
797 return 0;
798
Davide Libenzi7699acd2007-05-10 22:23:23 -0700799error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 ep_unregister_pollwait(ep, epi);
801
802 /*
803 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -0700804 * allocated wait queue. Note that we don't care about the ep->ovflist
805 * list, since that is used/cleaned only inside a section bound by "mtx".
806 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700808 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700809 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700810 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700811 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700813 kmem_cache_free(epi_cache, epi);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700814error_return:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 return error;
816}
817
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818/*
819 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -0700820 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 */
822static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
823{
824 int pwake = 0;
825 unsigned int revents;
826 unsigned long flags;
827
828 /*
829 * Set the new event interest mask before calling f_op->poll(), otherwise
830 * a potential race might occur. In fact if we do this operation inside
831 * the lock, an event might happen between the f_op->poll() call and the
832 * new event set registering.
833 */
834 epi->event.events = event->events;
835
836 /*
837 * Get current event bits. We can safely use the file* here because
838 * its usage count has been increased by the caller of this function.
839 */
840 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
841
Davide Libenzic7ea7632007-05-15 01:40:47 -0700842 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843
844 /* Copy the data member from inside the lock */
845 epi->event.data = event->data;
846
847 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700848 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -0700849 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700851 if (revents & event->events) {
852 if (!ep_is_linked(&epi->rdllink)) {
853 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854
Davide Libenzic7ea7632007-05-15 01:40:47 -0700855 /* Notify waiting tasks that events are available */
856 if (waitqueue_active(&ep->wq))
857 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
858 TASK_INTERRUPTIBLE);
859 if (waitqueue_active(&ep->poll_wait))
860 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 }
862 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700863 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864
865 /* We have to call this outside the lock */
866 if (pwake)
867 ep_poll_safewake(&psw, &ep->poll_wait);
868
869 return 0;
870}
871
Davide Libenzid47de162007-05-15 01:40:41 -0700872static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
873 int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874{
Davide Libenzi6192bd52007-05-08 00:25:41 -0700875 int eventcnt, error = -EFAULT, pwake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 unsigned int revents;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700877 unsigned long flags;
Davide Libenzid47de162007-05-15 01:40:41 -0700878 struct epitem *epi, *nepi;
879 struct list_head txlist;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700880
Davide Libenzid47de162007-05-15 01:40:41 -0700881 INIT_LIST_HEAD(&txlist);
882
883 /*
884 * We need to lock this because we could be hit by
885 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
886 */
887 mutex_lock(&ep->mtx);
888
889 /*
890 * Steal the ready list, and re-init the original one to the
891 * empty list. Also, set ep->ovflist to NULL so that events
892 * happening while looping w/out locks, are not lost. We cannot
893 * have the poll callback to queue directly on ep->rdllist,
894 * because we are doing it in the loop below, in a lockless way.
895 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700896 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700897 list_splice(&ep->rdllist, &txlist);
898 INIT_LIST_HEAD(&ep->rdllist);
899 ep->ovflist = NULL;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700900 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901
902 /*
903 * We can loop without lock because this is a task private list.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700904 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
Davide Libenzid47de162007-05-15 01:40:41 -0700905 * Items cannot vanish during the loop because we are holding "mtx".
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 */
Davide Libenzid47de162007-05-15 01:40:41 -0700907 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
908 epi = list_first_entry(&txlist, struct epitem, rdllink);
909
910 list_del_init(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
912 /*
913 * Get the ready file event set. We can safely use the file
Davide Libenzid47de162007-05-15 01:40:41 -0700914 * because we are holding the "mtx" and this will guarantee
915 * that both the file and the item will not vanish.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 */
917 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700918 revents &= epi->event.events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919
920 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700921 * Is the event mask intersect the caller-requested one,
922 * deliver the event to userspace. Again, we are holding
Davide Libenzid47de162007-05-15 01:40:41 -0700923 * "mtx", so no operations coming from userspace can change
924 * the item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700926 if (revents) {
927 if (__put_user(revents,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 &events[eventcnt].events) ||
929 __put_user(epi->event.data,
930 &events[eventcnt].data))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700931 goto errxit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 if (epi->event.events & EPOLLONESHOT)
933 epi->event.events &= EP_PRIVATE_BITS;
934 eventcnt++;
935 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700937 * At this point, noone can insert into ep->rdllist besides
938 * us. The epoll_ctl() callers are locked out by us holding
939 * "mtx" and the poll callback will queue them in ep->ovflist.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700941 if (!(epi->event.events & EPOLLET) &&
Davide Libenzid47de162007-05-15 01:40:41 -0700942 (revents & epi->event.events))
943 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 }
Davide Libenzi6192bd52007-05-08 00:25:41 -0700945 error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
Davide Libenzid47de162007-05-15 01:40:41 -0700947errxit:
948
Davide Libenzic7ea7632007-05-15 01:40:47 -0700949 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700950 /*
951 * During the time we spent in the loop above, some other events
952 * might have been queued by the poll callback. We re-insert them
953 * here (in case they are not already queued, or they're one-shot).
954 */
955 for (nepi = ep->ovflist; (epi = nepi) != NULL;
956 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
957 if (!ep_is_linked(&epi->rdllink) &&
958 (epi->event.events & ~EP_PRIVATE_BITS))
959 list_add_tail(&epi->rdllink, &ep->rdllist);
960 }
961 /*
962 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
963 * releasing the lock, events will be queued in the normal way inside
964 * ep->rdllist.
965 */
966 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700967
968 /*
Davide Libenzi67647d02007-05-15 01:40:52 -0700969 * In case of error in the event-send loop, or in case the number of
970 * ready events exceeds the userspace limit, we need to splice the
971 * "txlist" back inside ep->rdllist.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700972 */
Davide Libenzid47de162007-05-15 01:40:41 -0700973 list_splice(&txlist, &ep->rdllist);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700974
Davide Libenzid47de162007-05-15 01:40:41 -0700975 if (!list_empty(&ep->rdllist)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700977 * Wake up (if active) both the eventpoll wait list and the ->poll()
Davide Libenzi67647d02007-05-15 01:40:52 -0700978 * wait list (delayed after we release the lock).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 */
980 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -0700981 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
982 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 if (waitqueue_active(&ep->poll_wait))
984 pwake++;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700985 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700986 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700987
988 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989
990 /* We have to call this outside the lock */
991 if (pwake)
992 ep_poll_safewake(&psw, &ep->poll_wait);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700993
994 return eventcnt == 0 ? error: eventcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995}
996
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
998 int maxevents, long timeout)
999{
1000 int res, eavail;
1001 unsigned long flags;
1002 long jtimeout;
1003 wait_queue_t wait;
1004
1005 /*
1006 * Calculate the timeout by checking for the "infinite" value ( -1 )
1007 * and the overflow condition. The passed timeout is in milliseconds,
1008 * that why (t * HZ) / 1000.
1009 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001010 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1011 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012
1013retry:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001014 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015
1016 res = 0;
1017 if (list_empty(&ep->rdllist)) {
1018 /*
1019 * We don't have any available event to return to the caller.
1020 * We need to sleep here, and we will be wake up by
1021 * ep_poll_callback() when events will become available.
1022 */
1023 init_waitqueue_entry(&wait, current);
Davide Libenzid47de162007-05-15 01:40:41 -07001024 wait.flags |= WQ_FLAG_EXCLUSIVE;
Davide Libenzi3419b232006-06-25 05:48:14 -07001025 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
1027 for (;;) {
1028 /*
1029 * We don't want to sleep if the ep_poll_callback() sends us
1030 * a wakeup in between. That's why we set the task state
1031 * to TASK_INTERRUPTIBLE before doing the checks.
1032 */
1033 set_current_state(TASK_INTERRUPTIBLE);
1034 if (!list_empty(&ep->rdllist) || !jtimeout)
1035 break;
1036 if (signal_pending(current)) {
1037 res = -EINTR;
1038 break;
1039 }
1040
Davide Libenzic7ea7632007-05-15 01:40:47 -07001041 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 jtimeout = schedule_timeout(jtimeout);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001043 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001045 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046
1047 set_current_state(TASK_RUNNING);
1048 }
1049
1050 /* Is it worth to try to dig for events ? */
1051 eavail = !list_empty(&ep->rdllist);
1052
Davide Libenzic7ea7632007-05-15 01:40:47 -07001053 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054
1055 /*
1056 * Try to transfer events to user space. In case we get 0 events and
1057 * there's still timeout left over, we go trying again in search of
1058 * more luck.
1059 */
1060 if (!res && eavail &&
Davide Libenzid47de162007-05-15 01:40:41 -07001061 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 goto retry;
1063
1064 return res;
1065}
1066
Davide Libenzi7699acd2007-05-10 22:23:23 -07001067/*
Davide Libenzi67647d02007-05-15 01:40:52 -07001068 * It opens an eventpoll file descriptor. The "size" parameter is there
1069 * for historical reasons, when epoll was using an hash instead of an
1070 * RB tree. With the current implementation, the "size" parameter is ignored
1071 * (besides sanity checks).
Davide Libenzi7699acd2007-05-10 22:23:23 -07001072 */
1073asmlinkage long sys_epoll_create(int size)
1074{
1075 int error, fd = -1;
1076 struct eventpoll *ep;
1077 struct inode *inode;
1078 struct file *file;
1079
1080 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
1081 current, size));
1082
1083 /*
1084 * Sanity check on the size parameter, and create the internal data
1085 * structure ( "struct eventpoll" ).
1086 */
1087 error = -EINVAL;
1088 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
1089 goto error_return;
1090
1091 /*
1092 * Creates all the items needed to setup an eventpoll file. That is,
1093 * a file structure, and inode and a free file descriptor.
1094 */
1095 error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]",
1096 &eventpoll_fops, ep);
1097 if (error)
1098 goto error_free;
1099
1100 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1101 current, size, fd));
1102
1103 return fd;
1104
1105error_free:
1106 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001107error_return:
1108 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1109 current, size, error));
1110 return error;
1111}
1112
1113/*
1114 * The following function implements the controller interface for
1115 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001116 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001117 */
1118asmlinkage long sys_epoll_ctl(int epfd, int op, int fd,
1119 struct epoll_event __user *event)
1120{
1121 int error;
1122 struct file *file, *tfile;
1123 struct eventpoll *ep;
1124 struct epitem *epi;
1125 struct epoll_event epds;
1126
1127 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1128 current, epfd, op, fd, event));
1129
1130 error = -EFAULT;
1131 if (ep_op_has_event(op) &&
1132 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1133 goto error_return;
1134
1135 /* Get the "struct file *" for the eventpoll file */
1136 error = -EBADF;
1137 file = fget(epfd);
1138 if (!file)
1139 goto error_return;
1140
1141 /* Get the "struct file *" for the target file */
1142 tfile = fget(fd);
1143 if (!tfile)
1144 goto error_fput;
1145
1146 /* The target file descriptor must support poll */
1147 error = -EPERM;
1148 if (!tfile->f_op || !tfile->f_op->poll)
1149 goto error_tgt_fput;
1150
1151 /*
1152 * We have to check that the file structure underneath the file descriptor
1153 * the user passed to us _is_ an eventpoll file. And also we do not permit
1154 * adding an epoll file descriptor inside itself.
1155 */
1156 error = -EINVAL;
1157 if (file == tfile || !is_file_epoll(file))
1158 goto error_tgt_fput;
1159
1160 /*
1161 * At this point it is safe to assume that the "private_data" contains
1162 * our own data structure.
1163 */
1164 ep = file->private_data;
1165
Davide Libenzid47de162007-05-15 01:40:41 -07001166 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001167
Davide Libenzi67647d02007-05-15 01:40:52 -07001168 /*
1169 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1170 * above, we can be sure to be able to use the item looked up by
1171 * ep_find() till we release the mutex.
1172 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001173 epi = ep_find(ep, tfile, fd);
1174
1175 error = -EINVAL;
1176 switch (op) {
1177 case EPOLL_CTL_ADD:
1178 if (!epi) {
1179 epds.events |= POLLERR | POLLHUP;
1180
1181 error = ep_insert(ep, &epds, tfile, fd);
1182 } else
1183 error = -EEXIST;
1184 break;
1185 case EPOLL_CTL_DEL:
1186 if (epi)
1187 error = ep_remove(ep, epi);
1188 else
1189 error = -ENOENT;
1190 break;
1191 case EPOLL_CTL_MOD:
1192 if (epi) {
1193 epds.events |= POLLERR | POLLHUP;
1194 error = ep_modify(ep, epi, &epds);
1195 } else
1196 error = -ENOENT;
1197 break;
1198 }
Davide Libenzid47de162007-05-15 01:40:41 -07001199 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001200
1201error_tgt_fput:
1202 fput(tfile);
1203error_fput:
1204 fput(file);
1205error_return:
1206 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1207 current, epfd, op, fd, event, error));
1208
1209 return error;
1210}
1211
1212/*
1213 * Implement the event wait interface for the eventpoll file. It is the kernel
1214 * part of the user space epoll_wait(2).
1215 */
1216asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
1217 int maxevents, int timeout)
1218{
1219 int error;
1220 struct file *file;
1221 struct eventpoll *ep;
1222
1223 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1224 current, epfd, events, maxevents, timeout));
1225
1226 /* The maximum number of event must be greater than zero */
1227 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1228 return -EINVAL;
1229
1230 /* Verify that the area passed by the user is writeable */
1231 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1232 error = -EFAULT;
1233 goto error_return;
1234 }
1235
1236 /* Get the "struct file *" for the eventpoll file */
1237 error = -EBADF;
1238 file = fget(epfd);
1239 if (!file)
1240 goto error_return;
1241
1242 /*
1243 * We have to check that the file structure underneath the fd
1244 * the user passed to us _is_ an eventpoll file.
1245 */
1246 error = -EINVAL;
1247 if (!is_file_epoll(file))
1248 goto error_fput;
1249
1250 /*
1251 * At this point it is safe to assume that the "private_data" contains
1252 * our own data structure.
1253 */
1254 ep = file->private_data;
1255
1256 /* Time to fish for events ... */
1257 error = ep_poll(ep, events, maxevents, timeout);
1258
1259error_fput:
1260 fput(file);
1261error_return:
1262 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1263 current, epfd, events, maxevents, timeout, error));
1264
1265 return error;
1266}
1267
1268#ifdef TIF_RESTORE_SIGMASK
1269
1270/*
1271 * Implement the event wait interface for the eventpoll file. It is the kernel
1272 * part of the user space epoll_pwait(2).
1273 */
1274asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
1275 int maxevents, int timeout, const sigset_t __user *sigmask,
1276 size_t sigsetsize)
1277{
1278 int error;
1279 sigset_t ksigmask, sigsaved;
1280
1281 /*
1282 * If the caller wants a certain signal mask to be set during the wait,
1283 * we apply it here.
1284 */
1285 if (sigmask) {
1286 if (sigsetsize != sizeof(sigset_t))
1287 return -EINVAL;
1288 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1289 return -EFAULT;
1290 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1291 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1292 }
1293
1294 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1295
1296 /*
1297 * If we changed the signal mask, we need to restore the original one.
1298 * In case we've got a signal while waiting, we do not restore the
1299 * signal mask yet, and we allow do_signal() to deliver the signal on
1300 * the way back to userspace, before the signal mask is restored.
1301 */
1302 if (sigmask) {
1303 if (error == -EINTR) {
1304 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001305 sizeof(sigsaved));
Davide Libenzi7699acd2007-05-10 22:23:23 -07001306 set_thread_flag(TIF_RESTORE_SIGMASK);
1307 } else
1308 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1309 }
1310
1311 return error;
1312}
1313
1314#endif /* #ifdef TIF_RESTORE_SIGMASK */
1315
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316static int __init eventpoll_init(void)
1317{
Arjan van de Ven144efe32006-03-23 03:00:32 -08001318 mutex_init(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319
1320 /* Initialize the structure used to perform safe poll wait head wake ups */
1321 ep_poll_safewake_init(&psw);
1322
1323 /* Allocates slab cache used to allocate "struct epitem" items */
1324 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1325 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001326 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327
1328 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1329 pwq_cache = kmem_cache_create("eventpoll_pwq",
1330 sizeof(struct eppoll_entry), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09001331 EPI_SLAB_DEBUG|SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334}
Davide Libenzicea69242007-05-10 22:23:22 -07001335fs_initcall(eventpoll_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336