blob: 0b73cd45a06d6a05a2b37231a39a51cf1b8a8e4e [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;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 struct list_head *lsthead = &psw->wake_task_list, *lnk;
329 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 */
335 list_for_each(lnk, lsthead) {
336 tncur = list_entry(lnk, struct wake_task_node, llink);
337
338 if (tncur->wq == wq ||
339 (tncur->task == this_task && ++wake_nests > EP_MAX_POLLWAKE_NESTS)) {
340 /*
341 * Ops ... loop detected or maximum nest level reached.
342 * We abort this wake by breaking the cycle itself.
343 */
344 spin_unlock_irqrestore(&psw->lock, flags);
345 return;
346 }
347 }
348
349 /* Add the current task to the list */
350 tnode.task = this_task;
351 tnode.wq = wq;
352 list_add(&tnode.llink, lsthead);
353
354 spin_unlock_irqrestore(&psw->lock, flags);
355
356 /* Do really wake up now */
357 wake_up(wq);
358
359 /* Remove the current task from the list */
360 spin_lock_irqsave(&psw->lock, flags);
361 list_del(&tnode.llink);
362 spin_unlock_irqrestore(&psw->lock, flags);
363}
364
Davide Libenzi7699acd2007-05-10 22:23:23 -0700365/*
366 * This function unregister poll callbacks from the associated file descriptor.
367 * Since this must be called without holding "ep->lock" the atomic exchange trick
368 * will protect us from multiple unregister.
369 */
370static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
371{
372 int nwait;
373 struct list_head *lsthead = &epi->pwqlist;
374 struct eppoll_entry *pwq;
375
376 /* This is called without locks, so we need the atomic exchange */
377 nwait = xchg(&epi->nwait, 0);
378
379 if (nwait) {
380 while (!list_empty(lsthead)) {
381 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
382
383 list_del_init(&pwq->llink);
384 remove_wait_queue(pwq->whead, &pwq->wait);
385 kmem_cache_free(pwq_cache, pwq);
386 }
387 }
388}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700391 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700392 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700393 */
394static int ep_remove(struct eventpoll *ep, struct epitem *epi)
395{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700396 unsigned long flags;
397 struct file *file = epi->ffd.file;
398
399 /*
400 * Removes poll wait queue hooks. We _have_ to do this without holding
401 * the "ep->lock" otherwise a deadlock might occur. This because of the
402 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
403 * queue head lock when unregistering the wait queue. The wakeup callback
404 * will run by holding the wait queue head lock and will call our callback
405 * that will try to get "ep->lock".
406 */
407 ep_unregister_pollwait(ep, epi);
408
409 /* Remove the current item from the list of epoll hooks */
410 spin_lock(&file->f_ep_lock);
411 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700412 list_del_init(&epi->fllink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700413 spin_unlock(&file->f_ep_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Davide Libenzic7ea7632007-05-15 01:40:47 -0700415 if (ep_rb_linked(&epi->rbn))
416 ep_rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
Davide Libenzic7ea7632007-05-15 01:40:47 -0700418 spin_lock_irqsave(&ep->lock, flags);
419 if (ep_is_linked(&epi->rdllink))
420 list_del_init(&epi->rdllink);
421 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422
Davide Libenzi7699acd2007-05-10 22:23:23 -0700423 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700424 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
Davide Libenzic7ea7632007-05-15 01:40:47 -0700426 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_remove(%p, %p)\n",
427 current, ep, file));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
Davide Libenzic7ea7632007-05-15 01:40:47 -0700429 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430}
431
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432static void ep_free(struct eventpoll *ep)
433{
434 struct rb_node *rbp;
435 struct epitem *epi;
436
437 /* We need to release all tasks waiting for these file */
438 if (waitqueue_active(&ep->poll_wait))
439 ep_poll_safewake(&psw, &ep->poll_wait);
440
441 /*
442 * We need to lock this because we could be hit by
443 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700444 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 * is on the way to be removed and no one has references to it
446 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800447 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800449 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
451 /*
452 * Walks through the whole tree by unregistering poll callbacks.
453 */
454 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
455 epi = rb_entry(rbp, struct epitem, rbn);
456
457 ep_unregister_pollwait(ep, epi);
458 }
459
460 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700461 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700463 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 * us during this operation. So we can avoid the lock on "ep->lock".
465 */
466 while ((rbp = rb_first(&ep->rbr)) != 0) {
467 epi = rb_entry(rbp, struct epitem, rbn);
468 ep_remove(ep, epi);
469 }
470
Arjan van de Ven144efe32006-03-23 03:00:32 -0800471 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700472 mutex_destroy(&ep->mtx);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700473 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474}
475
Davide Libenzi7699acd2007-05-10 22:23:23 -0700476static int ep_eventpoll_release(struct inode *inode, struct file *file)
477{
478 struct eventpoll *ep = file->private_data;
479
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700480 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700481 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700482
483 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: close() ep=%p\n", current, ep));
484 return 0;
485}
486
487static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
488{
489 unsigned int pollflags = 0;
490 unsigned long flags;
491 struct eventpoll *ep = file->private_data;
492
493 /* Insert inside our poll wait queue */
494 poll_wait(file, &ep->poll_wait, wait);
495
496 /* Check our condition */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700497 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700498 if (!list_empty(&ep->rdllist))
499 pollflags = POLLIN | POLLRDNORM;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700500 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700501
502 return pollflags;
503}
504
505/* File callbacks that implement the eventpoll file behaviour */
506static const struct file_operations eventpoll_fops = {
507 .release = ep_eventpoll_release,
508 .poll = ep_eventpoll_poll
509};
510
511/* Fast test to see if the file is an evenpoll file */
512static inline int is_file_epoll(struct file *f)
513{
514 return f->f_op == &eventpoll_fops;
515}
516
517/*
518 * This is called from eventpoll_release() to unlink files from the eventpoll
519 * interface. We need to have this facility to cleanup correctly files that are
520 * closed without being removed from the eventpoll interface.
521 */
522void eventpoll_release_file(struct file *file)
523{
524 struct list_head *lsthead = &file->f_ep_links;
525 struct eventpoll *ep;
526 struct epitem *epi;
527
528 /*
529 * We don't want to get "file->f_ep_lock" because it is not
530 * necessary. It is not necessary because we're in the "struct file"
531 * cleanup path, and this means that noone is using this file anymore.
Davide Libenzi67647d02007-05-15 01:40:52 -0700532 * So, for example, epoll_ctl() cannot hit here sicne if we reach this
533 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700534 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700535 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700536 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700537 * from anywhere but ep_free().
538 */
539 mutex_lock(&epmutex);
540
541 while (!list_empty(lsthead)) {
542 epi = list_first_entry(lsthead, struct epitem, fllink);
543
544 ep = epi->ep;
545 list_del_init(&epi->fllink);
Davide Libenzid47de162007-05-15 01:40:41 -0700546 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700547 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700548 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700549 }
550
551 mutex_unlock(&epmutex);
552}
553
554static int ep_alloc(struct eventpoll **pep)
555{
556 struct eventpoll *ep = kzalloc(sizeof(*ep), GFP_KERNEL);
557
558 if (!ep)
559 return -ENOMEM;
560
Davide Libenzic7ea7632007-05-15 01:40:47 -0700561 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700562 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700563 init_waitqueue_head(&ep->wq);
564 init_waitqueue_head(&ep->poll_wait);
565 INIT_LIST_HEAD(&ep->rdllist);
566 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700567 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700568
569 *pep = ep;
570
571 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_alloc() ep=%p\n",
572 current, ep));
573 return 0;
574}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
576/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700577 * Search the file inside the eventpoll tree. The RB tree operations
578 * are protected by the "mtx" mutex, and ep_find() must be called with
579 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 */
581static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
582{
583 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 struct rb_node *rbp;
585 struct epitem *epi, *epir = NULL;
586 struct epoll_filefd ffd;
587
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700588 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 for (rbp = ep->rbr.rb_node; rbp; ) {
590 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700591 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 if (kcmp > 0)
593 rbp = rbp->rb_right;
594 else if (kcmp < 0)
595 rbp = rbp->rb_left;
596 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 epir = epi;
598 break;
599 }
600 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601
602 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_find(%p) -> %p\n",
603 current, file, epir));
604
605 return epir;
606}
607
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700609 * This is the callback that is passed to the wait queue wakeup
610 * machanism. It is called by the stored file descriptors when they
611 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700613static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700615 int pwake = 0;
616 unsigned long flags;
617 struct epitem *epi = ep_item_from_wait(wait);
618 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Davide Libenzi7699acd2007-05-10 22:23:23 -0700620 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: poll_callback(%p) epi=%p ep=%p\n",
621 current, epi->ffd.file, epi, ep));
622
Davide Libenzic7ea7632007-05-15 01:40:47 -0700623 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700624
625 /*
626 * If the event mask does not contain any poll(2) event, we consider the
627 * descriptor to be disabled. This condition is likely the effect of the
628 * EPOLLONESHOT bit that disables the descriptor when an event is received,
629 * until the next EPOLL_CTL_MOD will be issued.
630 */
631 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700632 goto out_unlock;
633
634 /*
635 * If we are trasfering events to userspace, we can hold no locks
636 * (because we're accessing user memory, and because of linux f_op->poll()
637 * semantics). All the events that happens during that period of time are
638 * chained in ep->ovflist and requeued later on.
639 */
640 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
641 if (epi->next == EP_UNACTIVE_PTR) {
642 epi->next = ep->ovflist;
643 ep->ovflist = epi;
644 }
645 goto out_unlock;
646 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700647
648 /* If this file is already in the ready list we exit soon */
649 if (ep_is_linked(&epi->rdllink))
650 goto is_linked;
651
652 list_add_tail(&epi->rdllink, &ep->rdllist);
653
654is_linked:
655 /*
656 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
657 * wait list.
658 */
659 if (waitqueue_active(&ep->wq))
660 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
661 TASK_INTERRUPTIBLE);
662 if (waitqueue_active(&ep->poll_wait))
663 pwake++;
664
Davide Libenzid47de162007-05-15 01:40:41 -0700665out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700666 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700667
668 /* We have to call this outside the lock */
669 if (pwake)
670 ep_poll_safewake(&psw, &ep->poll_wait);
671
672 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673}
674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675/*
676 * This is the callback that is used to add our wait queue to the
677 * target file wakeup lists.
678 */
679static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
680 poll_table *pt)
681{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700682 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 struct eppoll_entry *pwq;
684
Christoph Lametere94b1762006-12-06 20:33:17 -0800685 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
687 pwq->whead = whead;
688 pwq->base = epi;
689 add_wait_queue(whead, &pwq->wait);
690 list_add_tail(&pwq->llink, &epi->pwqlist);
691 epi->nwait++;
692 } else {
693 /* We have to signal that an error occurred */
694 epi->nwait = -1;
695 }
696}
697
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
699{
700 int kcmp;
701 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
702 struct epitem *epic;
703
704 while (*p) {
705 parent = *p;
706 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700707 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 if (kcmp > 0)
709 p = &parent->rb_right;
710 else
711 p = &parent->rb_left;
712 }
713 rb_link_node(&epi->rbn, parent, p);
714 rb_insert_color(&epi->rbn, &ep->rbr);
715}
716
Davide Libenzic7ea7632007-05-15 01:40:47 -0700717/*
718 * Must be called with "mtx" held.
719 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
721 struct file *tfile, int fd)
722{
723 int error, revents, pwake = 0;
724 unsigned long flags;
725 struct epitem *epi;
726 struct ep_pqueue epq;
727
728 error = -ENOMEM;
Christoph Lametere94b1762006-12-06 20:33:17 -0800729 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7699acd2007-05-10 22:23:23 -0700730 goto error_return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731
732 /* Item initialization follow here ... */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700733 ep_rb_initnode(&epi->rbn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 INIT_LIST_HEAD(&epi->rdllink);
735 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 INIT_LIST_HEAD(&epi->pwqlist);
737 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700738 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -0700741 epi->next = EP_UNACTIVE_PTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742
743 /* Initialize the poll table using the queue callback */
744 epq.epi = epi;
745 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
746
747 /*
748 * Attach the item to the poll hooks and get current event bits.
749 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -0700750 * been increased by the caller of this function. Note that after
751 * this operation completes, the poll callback can start hitting
752 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 */
754 revents = tfile->f_op->poll(tfile, &epq.pt);
755
756 /*
757 * We have to check if something went wrong during the poll wait queue
758 * install process. Namely an allocation for a wait queue failed due
759 * high memory pressure.
760 */
761 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700762 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
764 /* Add the current item to the list of active epoll hook for this file */
765 spin_lock(&tfile->f_ep_lock);
766 list_add_tail(&epi->fllink, &tfile->f_ep_links);
767 spin_unlock(&tfile->f_ep_lock);
768
Davide Libenzic7ea7632007-05-15 01:40:47 -0700769 /*
770 * Add the current item to the RB tree. All RB tree operations are
771 * protected by "mtx", and ep_insert() is called with "mtx" held.
772 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 ep_rbtree_insert(ep, epi);
774
Davide Libenzic7ea7632007-05-15 01:40:47 -0700775 /* We have to drop the new item inside our item list to keep track of it */
776 spin_lock_irqsave(&ep->lock, flags);
777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700779 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 list_add_tail(&epi->rdllink, &ep->rdllist);
781
782 /* Notify waiting tasks that events are available */
783 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -0700784 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 if (waitqueue_active(&ep->poll_wait))
786 pwake++;
787 }
788
Davide Libenzic7ea7632007-05-15 01:40:47 -0700789 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
791 /* We have to call this outside the lock */
792 if (pwake)
793 ep_poll_safewake(&psw, &ep->poll_wait);
794
795 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: ep_insert(%p, %p, %d)\n",
796 current, ep, tfile, fd));
797
798 return 0;
799
Davide Libenzi7699acd2007-05-10 22:23:23 -0700800error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 ep_unregister_pollwait(ep, epi);
802
803 /*
804 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -0700805 * allocated wait queue. Note that we don't care about the ep->ovflist
806 * list, since that is used/cleaned only inside a section bound by "mtx".
807 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700809 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700810 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700811 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700812 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700814 kmem_cache_free(epi_cache, epi);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700815error_return:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 return error;
817}
818
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819/*
820 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -0700821 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 */
823static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
824{
825 int pwake = 0;
826 unsigned int revents;
827 unsigned long flags;
828
829 /*
830 * Set the new event interest mask before calling f_op->poll(), otherwise
831 * a potential race might occur. In fact if we do this operation inside
832 * the lock, an event might happen between the f_op->poll() call and the
833 * new event set registering.
834 */
835 epi->event.events = event->events;
836
837 /*
838 * Get current event bits. We can safely use the file* here because
839 * its usage count has been increased by the caller of this function.
840 */
841 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
842
Davide Libenzic7ea7632007-05-15 01:40:47 -0700843 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
845 /* Copy the data member from inside the lock */
846 epi->event.data = event->data;
847
848 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700849 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -0700850 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700852 if (revents & event->events) {
853 if (!ep_is_linked(&epi->rdllink)) {
854 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855
Davide Libenzic7ea7632007-05-15 01:40:47 -0700856 /* Notify waiting tasks that events are available */
857 if (waitqueue_active(&ep->wq))
858 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
859 TASK_INTERRUPTIBLE);
860 if (waitqueue_active(&ep->poll_wait))
861 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 }
863 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700864 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865
866 /* We have to call this outside the lock */
867 if (pwake)
868 ep_poll_safewake(&psw, &ep->poll_wait);
869
870 return 0;
871}
872
Davide Libenzid47de162007-05-15 01:40:41 -0700873static int ep_send_events(struct eventpoll *ep, struct epoll_event __user *events,
874 int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875{
Davide Libenzi6192bd52007-05-08 00:25:41 -0700876 int eventcnt, error = -EFAULT, pwake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 unsigned int revents;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700878 unsigned long flags;
Davide Libenzid47de162007-05-15 01:40:41 -0700879 struct epitem *epi, *nepi;
880 struct list_head txlist;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700881
Davide Libenzid47de162007-05-15 01:40:41 -0700882 INIT_LIST_HEAD(&txlist);
883
884 /*
885 * We need to lock this because we could be hit by
886 * eventpoll_release_file() and epoll_ctl(EPOLL_CTL_DEL).
887 */
888 mutex_lock(&ep->mtx);
889
890 /*
891 * Steal the ready list, and re-init the original one to the
892 * empty list. Also, set ep->ovflist to NULL so that events
893 * happening while looping w/out locks, are not lost. We cannot
894 * have the poll callback to queue directly on ep->rdllist,
895 * because we are doing it in the loop below, in a lockless way.
896 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700897 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700898 list_splice(&ep->rdllist, &txlist);
899 INIT_LIST_HEAD(&ep->rdllist);
900 ep->ovflist = NULL;
Davide Libenzic7ea7632007-05-15 01:40:47 -0700901 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902
903 /*
904 * We can loop without lock because this is a task private list.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700905 * We just splice'd out the ep->rdllist in ep_collect_ready_items().
Davide Libenzid47de162007-05-15 01:40:41 -0700906 * Items cannot vanish during the loop because we are holding "mtx".
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 */
Davide Libenzid47de162007-05-15 01:40:41 -0700908 for (eventcnt = 0; !list_empty(&txlist) && eventcnt < maxevents;) {
909 epi = list_first_entry(&txlist, struct epitem, rdllink);
910
911 list_del_init(&epi->rdllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912
913 /*
914 * Get the ready file event set. We can safely use the file
Davide Libenzid47de162007-05-15 01:40:41 -0700915 * because we are holding the "mtx" and this will guarantee
916 * that both the file and the item will not vanish.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 */
918 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700919 revents &= epi->event.events;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
921 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700922 * Is the event mask intersect the caller-requested one,
923 * deliver the event to userspace. Again, we are holding
Davide Libenzid47de162007-05-15 01:40:41 -0700924 * "mtx", so no operations coming from userspace can change
925 * the item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700927 if (revents) {
928 if (__put_user(revents,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 &events[eventcnt].events) ||
930 __put_user(epi->event.data,
931 &events[eventcnt].data))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700932 goto errxit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 if (epi->event.events & EPOLLONESHOT)
934 epi->event.events &= EP_PRIVATE_BITS;
935 eventcnt++;
936 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700938 * At this point, noone can insert into ep->rdllist besides
939 * us. The epoll_ctl() callers are locked out by us holding
940 * "mtx" and the poll callback will queue them in ep->ovflist.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700942 if (!(epi->event.events & EPOLLET) &&
Davide Libenzid47de162007-05-15 01:40:41 -0700943 (revents & epi->event.events))
944 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 }
Davide Libenzi6192bd52007-05-08 00:25:41 -0700946 error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
Davide Libenzid47de162007-05-15 01:40:41 -0700948errxit:
949
Davide Libenzic7ea7632007-05-15 01:40:47 -0700950 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700951 /*
952 * During the time we spent in the loop above, some other events
953 * might have been queued by the poll callback. We re-insert them
954 * here (in case they are not already queued, or they're one-shot).
955 */
956 for (nepi = ep->ovflist; (epi = nepi) != NULL;
957 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
958 if (!ep_is_linked(&epi->rdllink) &&
959 (epi->event.events & ~EP_PRIVATE_BITS))
960 list_add_tail(&epi->rdllink, &ep->rdllist);
961 }
962 /*
963 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
964 * releasing the lock, events will be queued in the normal way inside
965 * ep->rdllist.
966 */
967 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700968
969 /*
Davide Libenzi67647d02007-05-15 01:40:52 -0700970 * In case of error in the event-send loop, or in case the number of
971 * ready events exceeds the userspace limit, we need to splice the
972 * "txlist" back inside ep->rdllist.
Davide Libenzi6192bd52007-05-08 00:25:41 -0700973 */
Davide Libenzid47de162007-05-15 01:40:41 -0700974 list_splice(&txlist, &ep->rdllist);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700975
Davide Libenzid47de162007-05-15 01:40:41 -0700976 if (!list_empty(&ep->rdllist)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700978 * Wake up (if active) both the eventpoll wait list and the ->poll()
Davide Libenzi67647d02007-05-15 01:40:52 -0700979 * wait list (delayed after we release the lock).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 */
981 if (waitqueue_active(&ep->wq))
Davide Libenzi3419b232006-06-25 05:48:14 -0700982 __wake_up_locked(&ep->wq, TASK_UNINTERRUPTIBLE |
983 TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 if (waitqueue_active(&ep->poll_wait))
985 pwake++;
Davide Libenzi6192bd52007-05-08 00:25:41 -0700986 }
Davide Libenzic7ea7632007-05-15 01:40:47 -0700987 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzid47de162007-05-15 01:40:41 -0700988
989 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990
991 /* We have to call this outside the lock */
992 if (pwake)
993 ep_poll_safewake(&psw, &ep->poll_wait);
Davide Libenzi6192bd52007-05-08 00:25:41 -0700994
995 return eventcnt == 0 ? error: eventcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996}
997
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
999 int maxevents, long timeout)
1000{
1001 int res, eavail;
1002 unsigned long flags;
1003 long jtimeout;
1004 wait_queue_t wait;
1005
1006 /*
1007 * Calculate the timeout by checking for the "infinite" value ( -1 )
1008 * and the overflow condition. The passed timeout is in milliseconds,
1009 * that why (t * HZ) / 1000.
1010 */
Davide Libenzie3306dd2005-09-27 21:45:33 -07001011 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1012 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013
1014retry:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001015 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
1017 res = 0;
1018 if (list_empty(&ep->rdllist)) {
1019 /*
1020 * We don't have any available event to return to the caller.
1021 * We need to sleep here, and we will be wake up by
1022 * ep_poll_callback() when events will become available.
1023 */
1024 init_waitqueue_entry(&wait, current);
Davide Libenzid47de162007-05-15 01:40:41 -07001025 wait.flags |= WQ_FLAG_EXCLUSIVE;
Davide Libenzi3419b232006-06-25 05:48:14 -07001026 __add_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027
1028 for (;;) {
1029 /*
1030 * We don't want to sleep if the ep_poll_callback() sends us
1031 * a wakeup in between. That's why we set the task state
1032 * to TASK_INTERRUPTIBLE before doing the checks.
1033 */
1034 set_current_state(TASK_INTERRUPTIBLE);
1035 if (!list_empty(&ep->rdllist) || !jtimeout)
1036 break;
1037 if (signal_pending(current)) {
1038 res = -EINTR;
1039 break;
1040 }
1041
Davide Libenzic7ea7632007-05-15 01:40:47 -07001042 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 jtimeout = schedule_timeout(jtimeout);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001044 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001046 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047
1048 set_current_state(TASK_RUNNING);
1049 }
1050
1051 /* Is it worth to try to dig for events ? */
1052 eavail = !list_empty(&ep->rdllist);
1053
Davide Libenzic7ea7632007-05-15 01:40:47 -07001054 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
1056 /*
1057 * Try to transfer events to user space. In case we get 0 events and
1058 * there's still timeout left over, we go trying again in search of
1059 * more luck.
1060 */
1061 if (!res && eavail &&
Davide Libenzid47de162007-05-15 01:40:41 -07001062 !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 goto retry;
1064
1065 return res;
1066}
1067
Davide Libenzi7699acd2007-05-10 22:23:23 -07001068/*
Davide Libenzi67647d02007-05-15 01:40:52 -07001069 * It opens an eventpoll file descriptor. The "size" parameter is there
1070 * for historical reasons, when epoll was using an hash instead of an
1071 * RB tree. With the current implementation, the "size" parameter is ignored
1072 * (besides sanity checks).
Davide Libenzi7699acd2007-05-10 22:23:23 -07001073 */
1074asmlinkage long sys_epoll_create(int size)
1075{
1076 int error, fd = -1;
1077 struct eventpoll *ep;
1078 struct inode *inode;
1079 struct file *file;
1080
1081 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d)\n",
1082 current, size));
1083
1084 /*
1085 * Sanity check on the size parameter, and create the internal data
1086 * structure ( "struct eventpoll" ).
1087 */
1088 error = -EINVAL;
1089 if (size <= 0 || (error = ep_alloc(&ep)) != 0)
1090 goto error_return;
1091
1092 /*
1093 * Creates all the items needed to setup an eventpoll file. That is,
1094 * a file structure, and inode and a free file descriptor.
1095 */
1096 error = anon_inode_getfd(&fd, &inode, &file, "[eventpoll]",
1097 &eventpoll_fops, ep);
1098 if (error)
1099 goto error_free;
1100
1101 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1102 current, size, fd));
1103
1104 return fd;
1105
1106error_free:
1107 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001108error_return:
1109 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_create(%d) = %d\n",
1110 current, size, error));
1111 return error;
1112}
1113
1114/*
1115 * The following function implements the controller interface for
1116 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001117 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001118 */
1119asmlinkage long sys_epoll_ctl(int epfd, int op, int fd,
1120 struct epoll_event __user *event)
1121{
1122 int error;
1123 struct file *file, *tfile;
1124 struct eventpoll *ep;
1125 struct epitem *epi;
1126 struct epoll_event epds;
1127
1128 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p)\n",
1129 current, epfd, op, fd, event));
1130
1131 error = -EFAULT;
1132 if (ep_op_has_event(op) &&
1133 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1134 goto error_return;
1135
1136 /* Get the "struct file *" for the eventpoll file */
1137 error = -EBADF;
1138 file = fget(epfd);
1139 if (!file)
1140 goto error_return;
1141
1142 /* Get the "struct file *" for the target file */
1143 tfile = fget(fd);
1144 if (!tfile)
1145 goto error_fput;
1146
1147 /* The target file descriptor must support poll */
1148 error = -EPERM;
1149 if (!tfile->f_op || !tfile->f_op->poll)
1150 goto error_tgt_fput;
1151
1152 /*
1153 * We have to check that the file structure underneath the file descriptor
1154 * the user passed to us _is_ an eventpoll file. And also we do not permit
1155 * adding an epoll file descriptor inside itself.
1156 */
1157 error = -EINVAL;
1158 if (file == tfile || !is_file_epoll(file))
1159 goto error_tgt_fput;
1160
1161 /*
1162 * At this point it is safe to assume that the "private_data" contains
1163 * our own data structure.
1164 */
1165 ep = file->private_data;
1166
Davide Libenzid47de162007-05-15 01:40:41 -07001167 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001168
Davide Libenzi67647d02007-05-15 01:40:52 -07001169 /*
1170 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1171 * above, we can be sure to be able to use the item looked up by
1172 * ep_find() till we release the mutex.
1173 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001174 epi = ep_find(ep, tfile, fd);
1175
1176 error = -EINVAL;
1177 switch (op) {
1178 case EPOLL_CTL_ADD:
1179 if (!epi) {
1180 epds.events |= POLLERR | POLLHUP;
1181
1182 error = ep_insert(ep, &epds, tfile, fd);
1183 } else
1184 error = -EEXIST;
1185 break;
1186 case EPOLL_CTL_DEL:
1187 if (epi)
1188 error = ep_remove(ep, epi);
1189 else
1190 error = -ENOENT;
1191 break;
1192 case EPOLL_CTL_MOD:
1193 if (epi) {
1194 epds.events |= POLLERR | POLLHUP;
1195 error = ep_modify(ep, epi, &epds);
1196 } else
1197 error = -ENOENT;
1198 break;
1199 }
Davide Libenzid47de162007-05-15 01:40:41 -07001200 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001201
1202error_tgt_fput:
1203 fput(tfile);
1204error_fput:
1205 fput(file);
1206error_return:
1207 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_ctl(%d, %d, %d, %p) = %d\n",
1208 current, epfd, op, fd, event, error));
1209
1210 return error;
1211}
1212
1213/*
1214 * Implement the event wait interface for the eventpoll file. It is the kernel
1215 * part of the user space epoll_wait(2).
1216 */
1217asmlinkage long sys_epoll_wait(int epfd, struct epoll_event __user *events,
1218 int maxevents, int timeout)
1219{
1220 int error;
1221 struct file *file;
1222 struct eventpoll *ep;
1223
1224 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d)\n",
1225 current, epfd, events, maxevents, timeout));
1226
1227 /* The maximum number of event must be greater than zero */
1228 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1229 return -EINVAL;
1230
1231 /* Verify that the area passed by the user is writeable */
1232 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1233 error = -EFAULT;
1234 goto error_return;
1235 }
1236
1237 /* Get the "struct file *" for the eventpoll file */
1238 error = -EBADF;
1239 file = fget(epfd);
1240 if (!file)
1241 goto error_return;
1242
1243 /*
1244 * We have to check that the file structure underneath the fd
1245 * the user passed to us _is_ an eventpoll file.
1246 */
1247 error = -EINVAL;
1248 if (!is_file_epoll(file))
1249 goto error_fput;
1250
1251 /*
1252 * At this point it is safe to assume that the "private_data" contains
1253 * our own data structure.
1254 */
1255 ep = file->private_data;
1256
1257 /* Time to fish for events ... */
1258 error = ep_poll(ep, events, maxevents, timeout);
1259
1260error_fput:
1261 fput(file);
1262error_return:
1263 DNPRINTK(3, (KERN_INFO "[%p] eventpoll: sys_epoll_wait(%d, %p, %d, %d) = %d\n",
1264 current, epfd, events, maxevents, timeout, error));
1265
1266 return error;
1267}
1268
1269#ifdef TIF_RESTORE_SIGMASK
1270
1271/*
1272 * Implement the event wait interface for the eventpoll file. It is the kernel
1273 * part of the user space epoll_pwait(2).
1274 */
1275asmlinkage long sys_epoll_pwait(int epfd, struct epoll_event __user *events,
1276 int maxevents, int timeout, const sigset_t __user *sigmask,
1277 size_t sigsetsize)
1278{
1279 int error;
1280 sigset_t ksigmask, sigsaved;
1281
1282 /*
1283 * If the caller wants a certain signal mask to be set during the wait,
1284 * we apply it here.
1285 */
1286 if (sigmask) {
1287 if (sigsetsize != sizeof(sigset_t))
1288 return -EINVAL;
1289 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1290 return -EFAULT;
1291 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1292 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1293 }
1294
1295 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1296
1297 /*
1298 * If we changed the signal mask, we need to restore the original one.
1299 * In case we've got a signal while waiting, we do not restore the
1300 * signal mask yet, and we allow do_signal() to deliver the signal on
1301 * the way back to userspace, before the signal mask is restored.
1302 */
1303 if (sigmask) {
1304 if (error == -EINTR) {
1305 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001306 sizeof(sigsaved));
Davide Libenzi7699acd2007-05-10 22:23:23 -07001307 set_thread_flag(TIF_RESTORE_SIGMASK);
1308 } else
1309 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1310 }
1311
1312 return error;
1313}
1314
1315#endif /* #ifdef TIF_RESTORE_SIGMASK */
1316
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317static int __init eventpoll_init(void)
1318{
Arjan van de Ven144efe32006-03-23 03:00:32 -08001319 mutex_init(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
1321 /* Initialize the structure used to perform safe poll wait head wake ups */
1322 ep_poll_safewake_init(&psw);
1323
1324 /* Allocates slab cache used to allocate "struct epitem" items */
1325 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1326 0, SLAB_HWCACHE_ALIGN|EPI_SLAB_DEBUG|SLAB_PANIC,
1327 NULL, NULL);
1328
1329 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1330 pwq_cache = kmem_cache_create("eventpoll_pwq",
1331 sizeof(struct eppoll_entry), 0,
1332 EPI_SLAB_DEBUG|SLAB_PANIC, NULL, NULL);
1333
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335}
Davide Libenzicea69242007-05-10 22:23:22 -07001336fs_initcall(eventpoll_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337