blob: 267d0ada45414a6f9cd8e90efeb9714e52e65de9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Davide Libenzi5071f972009-03-31 15:24:10 -07002 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 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/* Epoll private bits inside the event mask */
75#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
76
Davide Libenzi5071f972009-03-31 15:24:10 -070077/* Maximum number of nesting allowed inside epoll sets */
78#define EP_MAX_NESTS 4
Linus Torvalds1da177e2005-04-16 15:20:36 -070079
Davide Libenzib6119672006-10-11 01:21:44 -070080#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
81
Davide Libenzid47de162007-05-15 01:40:41 -070082#define EP_UNACTIVE_PTR ((void *) -1L)
83
Davide Libenzi7ef99642008-12-01 13:13:55 -080084#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
85
Linus Torvalds1da177e2005-04-16 15:20:36 -070086struct epoll_filefd {
87 struct file *file;
88 int fd;
89};
90
91/*
Davide Libenzi5071f972009-03-31 15:24:10 -070092 * Structure used to track possible nested calls, for too deep recursions
93 * and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -070094 */
Davide Libenzi5071f972009-03-31 15:24:10 -070095struct nested_call_node {
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 struct list_head llink;
Davide Libenzi5071f972009-03-31 15:24:10 -070097 void *cookie;
Davide Libenzi3fe4a972009-06-17 16:25:58 -070098 void *ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099};
100
101/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700102 * This structure is used as collector for nested calls, to check for
103 * maximum recursion dept and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700105struct nested_calls {
106 struct list_head tasks_call_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 spinlock_t lock;
108};
109
110/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700112 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 */
114struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700115 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 struct rb_node rbn;
117
118 /* List header used to link this structure to the eventpoll ready list */
119 struct list_head rdllink;
120
Davide Libenzic7ea7632007-05-15 01:40:47 -0700121 /*
122 * Works together "struct eventpoll"->ovflist in keeping the
123 * single linked chain of items.
124 */
125 struct epitem *next;
126
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 /* The file descriptor information this item refers to */
128 struct epoll_filefd ffd;
129
130 /* Number of active wait queue attached to poll operations */
131 int nwait;
132
133 /* List containing poll wait queues */
134 struct list_head pwqlist;
135
136 /* The "container" of this item */
137 struct eventpoll *ep;
138
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 /* List header used to link this item to the "struct file" items list */
140 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700141
Davide Libenzic7ea7632007-05-15 01:40:47 -0700142 /* The structure that describe the interested events and the source fd */
143 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700144};
145
146/*
147 * This structure is stored inside the "private_data" member of the file
148 * structure and rapresent the main data sructure for the eventpoll
149 * interface.
150 */
151struct eventpoll {
152 /* Protect the this structure access */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700153 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700154
155 /*
156 * This mutex is used to ensure that files are not removed
157 * while epoll is using them. This is held during the event
158 * collection loop, the file cleanup path, the epoll file exit
159 * code and the ctl operations.
160 */
161 struct mutex mtx;
162
163 /* Wait queue used by sys_epoll_wait() */
164 wait_queue_head_t wq;
165
166 /* Wait queue used by file->poll() */
167 wait_queue_head_t poll_wait;
168
169 /* List of ready file descriptors */
170 struct list_head rdllist;
171
Davide Libenzi67647d02007-05-15 01:40:52 -0700172 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700173 struct rb_root rbr;
174
175 /*
176 * This is a single linked list that chains all the "struct epitem" that
177 * happened while transfering ready events to userspace w/out
178 * holding ->lock.
179 */
180 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800181
182 /* The user that created the eventpoll descriptor */
183 struct user_struct *user;
Davide Libenzid47de162007-05-15 01:40:41 -0700184};
185
186/* Wait structure used by the poll hooks */
187struct eppoll_entry {
188 /* List header used to link this structure to the "struct epitem" */
189 struct list_head llink;
190
191 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700192 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700193
194 /*
195 * Wait queue item that will be linked to the target file wait
196 * queue head.
197 */
198 wait_queue_t wait;
199
200 /* The wait queue head that linked the "wait" wait queue item */
201 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202};
203
204/* Wrapper struct used by poll queueing */
205struct ep_pqueue {
206 poll_table pt;
207 struct epitem *epi;
208};
209
Davide Libenzi5071f972009-03-31 15:24:10 -0700210/* Used by the ep_send_events() function as callback private data */
211struct ep_send_events_data {
212 int maxevents;
213 struct epoll_event __user *events;
214};
215
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800217 * Configuration options available inside /proc/sys/fs/epoll/
218 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800219/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800220static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800221
222/*
Davide Libenzid47de162007-05-15 01:40:41 -0700223 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800225static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226
Davide Libenzi5071f972009-03-31 15:24:10 -0700227/* Used for safe wake up implementation */
228static struct nested_calls poll_safewake_ncalls;
229
230/* Used to call file's f_op->poll() under the nested calls boundaries */
231static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232
233/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800234static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235
236/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800237static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238
Davide Libenzi7ef99642008-12-01 13:13:55 -0800239#ifdef CONFIG_SYSCTL
240
241#include <linux/sysctl.h>
242
Robin Holt52bd19f72011-01-12 17:00:01 -0800243static long zero;
244static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800245
246ctl_table epoll_table[] = {
247 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800248 .procname = "max_user_watches",
249 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800250 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800251 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800252 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800253 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800254 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800255 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800256 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800257};
258#endif /* CONFIG_SYSCTL */
259
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700260
Davide Libenzi67647d02007-05-15 01:40:52 -0700261/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700262static inline void ep_set_ffd(struct epoll_filefd *ffd,
263 struct file *file, int fd)
264{
265 ffd->file = file;
266 ffd->fd = fd;
267}
268
Davide Libenzi67647d02007-05-15 01:40:52 -0700269/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700270static inline int ep_cmp_ffd(struct epoll_filefd *p1,
271 struct epoll_filefd *p2)
272{
273 return (p1->file > p2->file ? +1:
274 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
275}
276
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700277/* Tells us if the item is currently linked */
278static inline int ep_is_linked(struct list_head *p)
279{
280 return !list_empty(p);
281}
282
283/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700284static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700285{
286 return container_of(p, struct eppoll_entry, wait)->base;
287}
288
289/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700290static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700291{
292 return container_of(p, struct ep_pqueue, pt)->epi;
293}
294
295/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700296static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700297{
298 return op != EPOLL_CTL_DEL;
299}
300
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700302static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303{
Davide Libenzi5071f972009-03-31 15:24:10 -0700304 INIT_LIST_HEAD(&ncalls->tasks_call_list);
305 spin_lock_init(&ncalls->lock);
306}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307
Davide Libenzi5071f972009-03-31 15:24:10 -0700308/**
309 * ep_call_nested - Perform a bound (possibly) nested call, by checking
310 * that the recursion limit is not exceeded, and that
311 * the same nested call (by the meaning of same cookie) is
312 * no re-entered.
313 *
314 * @ncalls: Pointer to the nested_calls structure to be used for this call.
315 * @max_nests: Maximum number of allowed nesting calls.
316 * @nproc: Nested call core function pointer.
317 * @priv: Opaque data to be passed to the @nproc callback.
318 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700319 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700320 *
321 * Returns: Returns the code returned by the @nproc callback, or -1 if
322 * the maximum recursion limit has been exceeded.
323 */
324static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
325 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700326 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700327{
328 int error, call_nests = 0;
329 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700330 struct list_head *lsthead = &ncalls->tasks_call_list;
331 struct nested_call_node *tncur;
332 struct nested_call_node tnode;
333
334 spin_lock_irqsave(&ncalls->lock, flags);
335
336 /*
337 * Try to see if the current task is already inside this wakeup call.
338 * We use a list here, since the population inside this set is always
339 * very much limited.
340 */
341 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700342 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700343 (tncur->cookie == cookie || ++call_nests > max_nests)) {
344 /*
345 * Ops ... loop detected or maximum nest level reached.
346 * We abort this wake by breaking the cycle itself.
347 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700348 error = -1;
349 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700350 }
351 }
352
353 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700354 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700355 tnode.cookie = cookie;
356 list_add(&tnode.llink, lsthead);
357
358 spin_unlock_irqrestore(&ncalls->lock, flags);
359
360 /* Call the nested function */
361 error = (*nproc)(priv, cookie, call_nests);
362
363 /* Remove the current task from the list */
364 spin_lock_irqsave(&ncalls->lock, flags);
365 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700366out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700367 spin_unlock_irqrestore(&ncalls->lock, flags);
368
369 return error;
370}
371
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700372#ifdef CONFIG_DEBUG_LOCK_ALLOC
373static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
374 unsigned long events, int subclass)
375{
376 unsigned long flags;
377
378 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
379 wake_up_locked_poll(wqueue, events);
380 spin_unlock_irqrestore(&wqueue->lock, flags);
381}
382#else
383static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
384 unsigned long events, int subclass)
385{
386 wake_up_poll(wqueue, events);
387}
388#endif
389
Davide Libenzi5071f972009-03-31 15:24:10 -0700390static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
391{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700392 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
393 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700394 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395}
396
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397/*
398 * Perform a safe wake up of the poll wait list. The problem is that
399 * with the new callback'd wake up system, it is possible that the
400 * poll callback is reentered from inside the call to wake_up() done
401 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700402 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 * and we cannot reenter the same wait queue head at all. This will
404 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700405 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700407static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700409 int this_cpu = get_cpu();
410
Davide Libenzi5071f972009-03-31 15:24:10 -0700411 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700412 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
413
414 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415}
416
Davide Libenzi7699acd2007-05-10 22:23:23 -0700417/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700418 * This function unregisters poll callbacks from the associated file
419 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
420 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700421 */
422static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
423{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700424 struct list_head *lsthead = &epi->pwqlist;
425 struct eppoll_entry *pwq;
426
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700427 while (!list_empty(lsthead)) {
428 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700429
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700430 list_del(&pwq->llink);
431 remove_wait_queue(pwq->whead, &pwq->wait);
432 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700433 }
434}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Davide Libenzi5071f972009-03-31 15:24:10 -0700436/**
437 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
438 * the scan code, to call f_op->poll(). Also allows for
439 * O(NumReady) performance.
440 *
441 * @ep: Pointer to the epoll private data structure.
442 * @sproc: Pointer to the scan callback.
443 * @priv: Private opaque data passed to the @sproc callback.
444 *
445 * Returns: The same integer error code returned by the @sproc callback.
446 */
447static int ep_scan_ready_list(struct eventpoll *ep,
448 int (*sproc)(struct eventpoll *,
449 struct list_head *, void *),
450 void *priv)
451{
452 int error, pwake = 0;
453 unsigned long flags;
454 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700455 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700456
457 /*
458 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700459 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700460 */
461 mutex_lock(&ep->mtx);
462
463 /*
464 * Steal the ready list, and re-init the original one to the
465 * empty list. Also, set ep->ovflist to NULL so that events
466 * happening while looping w/out locks, are not lost. We cannot
467 * have the poll callback to queue directly on ep->rdllist,
468 * because we want the "sproc" callback to be able to do it
469 * in a lockless way.
470 */
471 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700472 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700473 ep->ovflist = NULL;
474 spin_unlock_irqrestore(&ep->lock, flags);
475
476 /*
477 * Now call the callback function.
478 */
479 error = (*sproc)(ep, &txlist, priv);
480
481 spin_lock_irqsave(&ep->lock, flags);
482 /*
483 * During the time we spent inside the "sproc" callback, some
484 * other events might have been queued by the poll callback.
485 * We re-insert them inside the main ready-list here.
486 */
487 for (nepi = ep->ovflist; (epi = nepi) != NULL;
488 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
489 /*
490 * We need to check if the item is already in the list.
491 * During the "sproc" callback execution time, items are
492 * queued into ->ovflist but the "txlist" might already
493 * contain them, and the list_splice() below takes care of them.
494 */
495 if (!ep_is_linked(&epi->rdllink))
496 list_add_tail(&epi->rdllink, &ep->rdllist);
497 }
498 /*
499 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
500 * releasing the lock, events will be queued in the normal way inside
501 * ep->rdllist.
502 */
503 ep->ovflist = EP_UNACTIVE_PTR;
504
505 /*
506 * Quickly re-inject items left on "txlist".
507 */
508 list_splice(&txlist, &ep->rdllist);
509
510 if (!list_empty(&ep->rdllist)) {
511 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700512 * Wake up (if active) both the eventpoll wait list and
513 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700514 */
515 if (waitqueue_active(&ep->wq))
516 wake_up_locked(&ep->wq);
517 if (waitqueue_active(&ep->poll_wait))
518 pwake++;
519 }
520 spin_unlock_irqrestore(&ep->lock, flags);
521
522 mutex_unlock(&ep->mtx);
523
524 /* We have to call this outside the lock */
525 if (pwake)
526 ep_poll_safewake(&ep->poll_wait);
527
528 return error;
529}
530
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700532 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700533 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700534 */
535static int ep_remove(struct eventpoll *ep, struct epitem *epi)
536{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700537 unsigned long flags;
538 struct file *file = epi->ffd.file;
539
540 /*
541 * Removes poll wait queue hooks. We _have_ to do this without holding
542 * the "ep->lock" otherwise a deadlock might occur. This because of the
543 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
544 * queue head lock when unregistering the wait queue. The wakeup callback
545 * will run by holding the wait queue head lock and will call our callback
546 * that will try to get "ep->lock".
547 */
548 ep_unregister_pollwait(ep, epi);
549
550 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700551 spin_lock(&file->f_lock);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700552 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700553 list_del_init(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700554 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
Davide Libenzicdac75e2008-04-29 00:58:34 -0700556 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
Davide Libenzic7ea7632007-05-15 01:40:47 -0700558 spin_lock_irqsave(&ep->lock, flags);
559 if (ep_is_linked(&epi->rdllink))
560 list_del_init(&epi->rdllink);
561 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
Davide Libenzi7699acd2007-05-10 22:23:23 -0700563 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700564 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
Robin Holt52bd19f72011-01-12 17:00:01 -0800566 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800567
Davide Libenzic7ea7632007-05-15 01:40:47 -0700568 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569}
570
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571static void ep_free(struct eventpoll *ep)
572{
573 struct rb_node *rbp;
574 struct epitem *epi;
575
576 /* We need to release all tasks waiting for these file */
577 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700578 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
580 /*
581 * We need to lock this because we could be hit by
582 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700583 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 * is on the way to be removed and no one has references to it
585 * anymore. The only hit might come from eventpoll_release_file() but
Arjan van de Ven144efe32006-03-23 03:00:32 -0800586 * holding "epmutex" is sufficent here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800588 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589
590 /*
591 * Walks through the whole tree by unregistering poll callbacks.
592 */
593 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
594 epi = rb_entry(rbp, struct epitem, rbn);
595
596 ep_unregister_pollwait(ep, epi);
597 }
598
599 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700600 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700602 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 * us during this operation. So we can avoid the lock on "ep->lock".
604 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700605 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 epi = rb_entry(rbp, struct epitem, rbn);
607 ep_remove(ep, epi);
608 }
609
Arjan van de Ven144efe32006-03-23 03:00:32 -0800610 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700611 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800612 free_uid(ep->user);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700613 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614}
615
Davide Libenzi7699acd2007-05-10 22:23:23 -0700616static int ep_eventpoll_release(struct inode *inode, struct file *file)
617{
618 struct eventpoll *ep = file->private_data;
619
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700620 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700621 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700622
Davide Libenzi7699acd2007-05-10 22:23:23 -0700623 return 0;
624}
625
Davide Libenzi296e2362009-03-31 15:24:11 -0700626static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
627 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700628{
629 struct epitem *epi, *tmp;
630
631 list_for_each_entry_safe(epi, tmp, head, rdllink) {
632 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
633 epi->event.events)
634 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700635 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700636 /*
637 * Item has been dropped into the ready list by the poll
638 * callback, but it's not actually ready, as far as
639 * caller requested events goes. We can remove it here.
640 */
641 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700642 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700643 }
644
645 return 0;
646}
647
648static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
649{
650 return ep_scan_ready_list(priv, ep_read_events_proc, NULL);
651}
652
Davide Libenzi7699acd2007-05-10 22:23:23 -0700653static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
654{
Davide Libenzi5071f972009-03-31 15:24:10 -0700655 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700656 struct eventpoll *ep = file->private_data;
657
658 /* Insert inside our poll wait queue */
659 poll_wait(file, &ep->poll_wait, wait);
660
Davide Libenzi5071f972009-03-31 15:24:10 -0700661 /*
662 * Proceed to find out if wanted events are really available inside
663 * the ready list. This need to be done under ep_call_nested()
664 * supervision, since the call to f_op->poll() done on listed files
665 * could re-enter here.
666 */
667 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700668 ep_poll_readyevents_proc, ep, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700669
Davide Libenzi296e2362009-03-31 15:24:11 -0700670 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700671}
672
673/* File callbacks that implement the eventpoll file behaviour */
674static const struct file_operations eventpoll_fops = {
675 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200676 .poll = ep_eventpoll_poll,
677 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700678};
679
680/* Fast test to see if the file is an evenpoll file */
681static inline int is_file_epoll(struct file *f)
682{
683 return f->f_op == &eventpoll_fops;
684}
685
686/*
687 * This is called from eventpoll_release() to unlink files from the eventpoll
688 * interface. We need to have this facility to cleanup correctly files that are
689 * closed without being removed from the eventpoll interface.
690 */
691void eventpoll_release_file(struct file *file)
692{
693 struct list_head *lsthead = &file->f_ep_links;
694 struct eventpoll *ep;
695 struct epitem *epi;
696
697 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700698 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700699 * necessary. It is not necessary because we're in the "struct file"
700 * cleanup path, and this means that noone is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700701 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700702 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700703 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700704 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700705 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700706 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700707 *
708 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700709 */
710 mutex_lock(&epmutex);
711
712 while (!list_empty(lsthead)) {
713 epi = list_first_entry(lsthead, struct epitem, fllink);
714
715 ep = epi->ep;
716 list_del_init(&epi->fllink);
Davide Libenzid47de162007-05-15 01:40:41 -0700717 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700718 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700719 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700720 }
721
722 mutex_unlock(&epmutex);
723}
724
725static int ep_alloc(struct eventpoll **pep)
726{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800727 int error;
728 struct user_struct *user;
729 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700730
Davide Libenzi7ef99642008-12-01 13:13:55 -0800731 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800732 error = -ENOMEM;
733 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
734 if (unlikely(!ep))
735 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700736
Davide Libenzic7ea7632007-05-15 01:40:47 -0700737 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700738 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700739 init_waitqueue_head(&ep->wq);
740 init_waitqueue_head(&ep->poll_wait);
741 INIT_LIST_HEAD(&ep->rdllist);
742 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700743 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800744 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700745
746 *pep = ep;
747
Davide Libenzi7699acd2007-05-10 22:23:23 -0700748 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800749
750free_uid:
751 free_uid(user);
752 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700753}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754
755/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700756 * Search the file inside the eventpoll tree. The RB tree operations
757 * are protected by the "mtx" mutex, and ep_find() must be called with
758 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 */
760static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
761{
762 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 struct rb_node *rbp;
764 struct epitem *epi, *epir = NULL;
765 struct epoll_filefd ffd;
766
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700767 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 for (rbp = ep->rbr.rb_node; rbp; ) {
769 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700770 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 if (kcmp > 0)
772 rbp = rbp->rb_right;
773 else if (kcmp < 0)
774 rbp = rbp->rb_left;
775 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 epir = epi;
777 break;
778 }
779 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 return epir;
782}
783
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700785 * This is the callback that is passed to the wait queue wakeup
786 * machanism. It is called by the stored file descriptors when they
787 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700789static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700791 int pwake = 0;
792 unsigned long flags;
793 struct epitem *epi = ep_item_from_wait(wait);
794 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795
Davide Libenzic7ea7632007-05-15 01:40:47 -0700796 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700797
798 /*
799 * If the event mask does not contain any poll(2) event, we consider the
800 * descriptor to be disabled. This condition is likely the effect of the
801 * EPOLLONESHOT bit that disables the descriptor when an event is received,
802 * until the next EPOLL_CTL_MOD will be issued.
803 */
804 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700805 goto out_unlock;
806
807 /*
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700808 * Check the events coming with the callback. At this stage, not
809 * every device reports the events in the "key" parameter of the
810 * callback. We need to be able to handle both cases here, hence the
811 * test for "key" != NULL before the event match test.
812 */
813 if (key && !((unsigned long) key & epi->event.events))
814 goto out_unlock;
815
816 /*
Davide Libenzid47de162007-05-15 01:40:41 -0700817 * If we are trasfering events to userspace, we can hold no locks
818 * (because we're accessing user memory, and because of linux f_op->poll()
819 * semantics). All the events that happens during that period of time are
820 * chained in ep->ovflist and requeued later on.
821 */
822 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
823 if (epi->next == EP_UNACTIVE_PTR) {
824 epi->next = ep->ovflist;
825 ep->ovflist = epi;
826 }
827 goto out_unlock;
828 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700829
830 /* If this file is already in the ready list we exit soon */
Davide Libenzi5071f972009-03-31 15:24:10 -0700831 if (!ep_is_linked(&epi->rdllink))
832 list_add_tail(&epi->rdllink, &ep->rdllist);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700833
Davide Libenzi7699acd2007-05-10 22:23:23 -0700834 /*
835 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
836 * wait list.
837 */
838 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -0400839 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700840 if (waitqueue_active(&ep->poll_wait))
841 pwake++;
842
Davide Libenzid47de162007-05-15 01:40:41 -0700843out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700844 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700845
846 /* We have to call this outside the lock */
847 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -0700848 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700849
850 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851}
852
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853/*
854 * This is the callback that is used to add our wait queue to the
855 * target file wakeup lists.
856 */
857static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
858 poll_table *pt)
859{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700860 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 struct eppoll_entry *pwq;
862
Christoph Lametere94b1762006-12-06 20:33:17 -0800863 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
865 pwq->whead = whead;
866 pwq->base = epi;
867 add_wait_queue(whead, &pwq->wait);
868 list_add_tail(&pwq->llink, &epi->pwqlist);
869 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -0700870 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 /* We have to signal that an error occurred */
872 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -0700873 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874}
875
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
877{
878 int kcmp;
879 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
880 struct epitem *epic;
881
882 while (*p) {
883 parent = *p;
884 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700885 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 if (kcmp > 0)
887 p = &parent->rb_right;
888 else
889 p = &parent->rb_left;
890 }
891 rb_link_node(&epi->rbn, parent, p);
892 rb_insert_color(&epi->rbn, &ep->rbr);
893}
894
Davide Libenzic7ea7632007-05-15 01:40:47 -0700895/*
896 * Must be called with "mtx" held.
897 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
899 struct file *tfile, int fd)
900{
901 int error, revents, pwake = 0;
902 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -0800903 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 struct epitem *epi;
905 struct ep_pqueue epq;
906
Robin Holt52bd19f72011-01-12 17:00:01 -0800907 user_watches = atomic_long_read(&ep->user->epoll_watches);
908 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -0800909 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -0800910 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -0800911 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912
913 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 INIT_LIST_HEAD(&epi->rdllink);
915 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 INIT_LIST_HEAD(&epi->pwqlist);
917 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700918 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -0700921 epi->next = EP_UNACTIVE_PTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922
923 /* Initialize the poll table using the queue callback */
924 epq.epi = epi;
925 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
926
927 /*
928 * Attach the item to the poll hooks and get current event bits.
929 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -0700930 * been increased by the caller of this function. Note that after
931 * this operation completes, the poll callback can start hitting
932 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 */
934 revents = tfile->f_op->poll(tfile, &epq.pt);
935
936 /*
937 * We have to check if something went wrong during the poll wait queue
938 * install process. Namely an allocation for a wait queue failed due
939 * high memory pressure.
940 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800941 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700943 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
945 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -0700946 spin_lock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 list_add_tail(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -0700948 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
Davide Libenzic7ea7632007-05-15 01:40:47 -0700950 /*
951 * Add the current item to the RB tree. All RB tree operations are
952 * protected by "mtx", and ep_insert() is called with "mtx" held.
953 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 ep_rbtree_insert(ep, epi);
955
Davide Libenzic7ea7632007-05-15 01:40:47 -0700956 /* We have to drop the new item inside our item list to keep track of it */
957 spin_lock_irqsave(&ep->lock, flags);
958
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700960 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 list_add_tail(&epi->rdllink, &ep->rdllist);
962
963 /* Notify waiting tasks that events are available */
964 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -0400965 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 if (waitqueue_active(&ep->poll_wait))
967 pwake++;
968 }
969
Davide Libenzic7ea7632007-05-15 01:40:47 -0700970 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Robin Holt52bd19f72011-01-12 17:00:01 -0800972 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800973
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 /* We have to call this outside the lock */
975 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -0700976 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 return 0;
979
Davide Libenzi7699acd2007-05-10 22:23:23 -0700980error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 ep_unregister_pollwait(ep, epi);
982
983 /*
984 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -0700985 * allocated wait queue. Note that we don't care about the ep->ovflist
986 * list, since that is used/cleaned only inside a section bound by "mtx".
987 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700989 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700990 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700991 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -0700992 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700994 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800995
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 return error;
997}
998
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999/*
1000 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001001 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 */
1003static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1004{
1005 int pwake = 0;
1006 unsigned int revents;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007
1008 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001009 * Set the new event interest mask before calling f_op->poll();
1010 * otherwise we might miss an event that happens between the
1011 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 */
1013 epi->event.events = event->events;
Tony Battersbye057e152009-03-31 15:24:15 -07001014 epi->event.data = event->data; /* protected by mtx */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015
1016 /*
1017 * Get current event bits. We can safely use the file* here because
1018 * its usage count has been increased by the caller of this function.
1019 */
1020 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1021
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001023 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001024 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001026 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001027 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001028 if (!ep_is_linked(&epi->rdllink)) {
1029 list_add_tail(&epi->rdllink, &ep->rdllist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030
Davide Libenzic7ea7632007-05-15 01:40:47 -07001031 /* Notify waiting tasks that events are available */
1032 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001033 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001034 if (waitqueue_active(&ep->poll_wait))
1035 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 }
Tony Battersbye057e152009-03-31 15:24:15 -07001037 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039
1040 /* We have to call this outside the lock */
1041 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001042 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043
1044 return 0;
1045}
1046
Davide Libenzi296e2362009-03-31 15:24:11 -07001047static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1048 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001049{
1050 struct ep_send_events_data *esed = priv;
1051 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001052 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001053 struct epitem *epi;
1054 struct epoll_event __user *uevent;
1055
Davide Libenzi296e2362009-03-31 15:24:11 -07001056 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001057 * We can loop without lock because we are passed a task private list.
1058 * Items cannot vanish during the loop because ep_scan_ready_list() is
1059 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001060 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001061 for (eventcnt = 0, uevent = esed->events;
1062 !list_empty(head) && eventcnt < esed->maxevents;) {
1063 epi = list_first_entry(head, struct epitem, rdllink);
1064
1065 list_del_init(&epi->rdllink);
1066
Davide Libenzi296e2362009-03-31 15:24:11 -07001067 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1068 epi->event.events;
Davide Libenzi5071f972009-03-31 15:24:10 -07001069
Davide Libenzi296e2362009-03-31 15:24:11 -07001070 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001071 * If the event mask intersect the caller-requested one,
1072 * deliver the event to userspace. Again, ep_scan_ready_list()
1073 * is holding "mtx", so no operations coming from userspace
1074 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001075 */
1076 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001077 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001078 __put_user(epi->event.data, &uevent->data)) {
1079 list_add(&epi->rdllink, head);
Davide Libenzi296e2362009-03-31 15:24:11 -07001080 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001081 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001082 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001083 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001084 if (epi->event.events & EPOLLONESHOT)
1085 epi->event.events &= EP_PRIVATE_BITS;
1086 else if (!(epi->event.events & EPOLLET)) {
1087 /*
1088 * If this file has been added with Level
1089 * Trigger mode, we need to insert back inside
1090 * the ready list, so that the next call to
1091 * epoll_wait() will check again the events
1092 * availability. At this point, noone can insert
1093 * into ep->rdllist besides us. The epoll_ctl()
1094 * callers are locked out by
1095 * ep_scan_ready_list() holding "mtx" and the
1096 * poll callback will queue them in ep->ovflist.
1097 */
1098 list_add_tail(&epi->rdllink, &ep->rdllist);
1099 }
1100 }
1101 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001102
1103 return eventcnt;
1104}
1105
Davide Libenzi296e2362009-03-31 15:24:11 -07001106static int ep_send_events(struct eventpoll *ep,
1107 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108{
Davide Libenzi5071f972009-03-31 15:24:10 -07001109 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001110
Davide Libenzi5071f972009-03-31 15:24:10 -07001111 esed.maxevents = maxevents;
1112 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001113
Davide Libenzi5071f972009-03-31 15:24:10 -07001114 return ep_scan_ready_list(ep, ep_send_events_proc, &esed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115}
1116
Eric Dumazet0781b902011-02-01 15:52:35 -08001117static inline struct timespec ep_set_mstimeout(long ms)
1118{
1119 struct timespec now, ts = {
1120 .tv_sec = ms / MSEC_PER_SEC,
1121 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1122 };
1123
1124 ktime_get_ts(&now);
1125 return timespec_add_safe(now, ts);
1126}
1127
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1129 int maxevents, long timeout)
1130{
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001131 int res, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 unsigned long flags;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001133 long slack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001135 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001137 if (timeout > 0) {
Eric Dumazet0781b902011-02-01 15:52:35 -08001138 struct timespec end_time = ep_set_mstimeout(timeout);
1139
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001140 slack = select_estimate_accuracy(&end_time);
1141 to = &expires;
1142 *to = timespec_to_ktime(end_time);
1143 } else if (timeout == 0) {
1144 timed_out = 1;
1145 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
1147retry:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001148 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
1150 res = 0;
1151 if (list_empty(&ep->rdllist)) {
1152 /*
1153 * We don't have any available event to return to the caller.
1154 * We need to sleep here, and we will be wake up by
1155 * ep_poll_callback() when events will become available.
1156 */
1157 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001158 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
1160 for (;;) {
1161 /*
1162 * We don't want to sleep if the ep_poll_callback() sends us
1163 * a wakeup in between. That's why we set the task state
1164 * to TASK_INTERRUPTIBLE before doing the checks.
1165 */
1166 set_current_state(TASK_INTERRUPTIBLE);
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001167 if (!list_empty(&ep->rdllist) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 break;
1169 if (signal_pending(current)) {
1170 res = -EINTR;
1171 break;
1172 }
1173
Davide Libenzic7ea7632007-05-15 01:40:47 -07001174 spin_unlock_irqrestore(&ep->lock, flags);
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001175 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1176 timed_out = 1;
1177
Davide Libenzic7ea7632007-05-15 01:40:47 -07001178 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001180 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
1182 set_current_state(TASK_RUNNING);
1183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 /* Is it worth to try to dig for events ? */
Davide Libenzi5071f972009-03-31 15:24:10 -07001185 eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
Davide Libenzic7ea7632007-05-15 01:40:47 -07001187 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
1189 /*
1190 * Try to transfer events to user space. In case we get 0 events and
1191 * there's still timeout left over, we go trying again in search of
1192 * more luck.
1193 */
1194 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001195 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 goto retry;
1197
1198 return res;
1199}
1200
Davide Libenzi7699acd2007-05-10 22:23:23 -07001201/*
Andrew Morton523723b2008-08-12 15:09:01 -07001202 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001203 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001204SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001205{
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001206 int error;
1207 struct eventpoll *ep = NULL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001208
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001209 /* Check the EPOLL_* constant for consistency. */
1210 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1211
Davide Libenzi296e2362009-03-31 15:24:11 -07001212 if (flags & ~EPOLL_CLOEXEC)
1213 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001214 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001215 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001216 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001217 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001218 if (error < 0)
1219 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001220 /*
1221 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001222 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001223 */
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001224 error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001225 O_RDWR | (flags & O_CLOEXEC));
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001226 if (error < 0)
Al Viro2030a422008-02-23 06:46:49 -05001227 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001228
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001229 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001230}
1231
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001232SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001233{
Davide Libenzibfe38912009-05-12 13:19:44 -07001234 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001235 return -EINVAL;
1236
1237 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001238}
1239
Davide Libenzi7699acd2007-05-10 22:23:23 -07001240/*
1241 * The following function implements the controller interface for
1242 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001243 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001244 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001245SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1246 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001247{
1248 int error;
1249 struct file *file, *tfile;
1250 struct eventpoll *ep;
1251 struct epitem *epi;
1252 struct epoll_event epds;
1253
Davide Libenzi7699acd2007-05-10 22:23:23 -07001254 error = -EFAULT;
1255 if (ep_op_has_event(op) &&
1256 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1257 goto error_return;
1258
1259 /* Get the "struct file *" for the eventpoll file */
1260 error = -EBADF;
1261 file = fget(epfd);
1262 if (!file)
1263 goto error_return;
1264
1265 /* Get the "struct file *" for the target file */
1266 tfile = fget(fd);
1267 if (!tfile)
1268 goto error_fput;
1269
1270 /* The target file descriptor must support poll */
1271 error = -EPERM;
1272 if (!tfile->f_op || !tfile->f_op->poll)
1273 goto error_tgt_fput;
1274
1275 /*
1276 * We have to check that the file structure underneath the file descriptor
1277 * the user passed to us _is_ an eventpoll file. And also we do not permit
1278 * adding an epoll file descriptor inside itself.
1279 */
1280 error = -EINVAL;
1281 if (file == tfile || !is_file_epoll(file))
1282 goto error_tgt_fput;
1283
1284 /*
1285 * At this point it is safe to assume that the "private_data" contains
1286 * our own data structure.
1287 */
1288 ep = file->private_data;
1289
Davide Libenzid47de162007-05-15 01:40:41 -07001290 mutex_lock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001291
Davide Libenzi67647d02007-05-15 01:40:52 -07001292 /*
1293 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1294 * above, we can be sure to be able to use the item looked up by
1295 * ep_find() till we release the mutex.
1296 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001297 epi = ep_find(ep, tfile, fd);
1298
1299 error = -EINVAL;
1300 switch (op) {
1301 case EPOLL_CTL_ADD:
1302 if (!epi) {
1303 epds.events |= POLLERR | POLLHUP;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001304 error = ep_insert(ep, &epds, tfile, fd);
1305 } else
1306 error = -EEXIST;
1307 break;
1308 case EPOLL_CTL_DEL:
1309 if (epi)
1310 error = ep_remove(ep, epi);
1311 else
1312 error = -ENOENT;
1313 break;
1314 case EPOLL_CTL_MOD:
1315 if (epi) {
1316 epds.events |= POLLERR | POLLHUP;
1317 error = ep_modify(ep, epi, &epds);
1318 } else
1319 error = -ENOENT;
1320 break;
1321 }
Davide Libenzid47de162007-05-15 01:40:41 -07001322 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001323
1324error_tgt_fput:
1325 fput(tfile);
1326error_fput:
1327 fput(file);
1328error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001329
1330 return error;
1331}
1332
1333/*
1334 * Implement the event wait interface for the eventpoll file. It is the kernel
1335 * part of the user space epoll_wait(2).
1336 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001337SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1338 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001339{
1340 int error;
1341 struct file *file;
1342 struct eventpoll *ep;
1343
Davide Libenzi7699acd2007-05-10 22:23:23 -07001344 /* The maximum number of event must be greater than zero */
1345 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1346 return -EINVAL;
1347
1348 /* Verify that the area passed by the user is writeable */
1349 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1350 error = -EFAULT;
1351 goto error_return;
1352 }
1353
1354 /* Get the "struct file *" for the eventpoll file */
1355 error = -EBADF;
1356 file = fget(epfd);
1357 if (!file)
1358 goto error_return;
1359
1360 /*
1361 * We have to check that the file structure underneath the fd
1362 * the user passed to us _is_ an eventpoll file.
1363 */
1364 error = -EINVAL;
1365 if (!is_file_epoll(file))
1366 goto error_fput;
1367
1368 /*
1369 * At this point it is safe to assume that the "private_data" contains
1370 * our own data structure.
1371 */
1372 ep = file->private_data;
1373
1374 /* Time to fish for events ... */
1375 error = ep_poll(ep, events, maxevents, timeout);
1376
1377error_fput:
1378 fput(file);
1379error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001380
1381 return error;
1382}
1383
Roland McGrathf3de2722008-04-30 00:53:09 -07001384#ifdef HAVE_SET_RESTORE_SIGMASK
Davide Libenzi7699acd2007-05-10 22:23:23 -07001385
1386/*
1387 * Implement the event wait interface for the eventpoll file. It is the kernel
1388 * part of the user space epoll_pwait(2).
1389 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001390SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1391 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1392 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001393{
1394 int error;
1395 sigset_t ksigmask, sigsaved;
1396
1397 /*
1398 * If the caller wants a certain signal mask to be set during the wait,
1399 * we apply it here.
1400 */
1401 if (sigmask) {
1402 if (sigsetsize != sizeof(sigset_t))
1403 return -EINVAL;
1404 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1405 return -EFAULT;
1406 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1407 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1408 }
1409
1410 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1411
1412 /*
1413 * If we changed the signal mask, we need to restore the original one.
1414 * In case we've got a signal while waiting, we do not restore the
1415 * signal mask yet, and we allow do_signal() to deliver the signal on
1416 * the way back to userspace, before the signal mask is restored.
1417 */
1418 if (sigmask) {
1419 if (error == -EINTR) {
1420 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001421 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001422 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001423 } else
1424 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1425 }
1426
1427 return error;
1428}
1429
Roland McGrathf3de2722008-04-30 00:53:09 -07001430#endif /* HAVE_SET_RESTORE_SIGMASK */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001431
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432static int __init eventpoll_init(void)
1433{
Davide Libenzi7ef99642008-12-01 13:13:55 -08001434 struct sysinfo si;
1435
1436 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08001437 /*
1438 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1439 */
1440 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08001441 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08001442 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443
1444 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07001445 ep_nested_calls_init(&poll_safewake_ncalls);
1446
1447 /* Initialize the structure used to perform file's f_op->poll() calls */
1448 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
1450 /* Allocates slab cache used to allocate "struct epitem" items */
1451 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001452 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
1454 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1455 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001456 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459}
Davide Libenzicea69242007-05-10 22:23:22 -07001460fs_initcall(eventpoll_init);