blob: da72250ddc1cf2331336a23cbe8239bf880a10cf [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>
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020036#include <linux/device.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/io.h>
39#include <asm/mman.h>
Arun Sharma600634972011-07-26 16:09:06 -070040#include <linux/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
Daniel Balutabf6a41d2011-01-30 23:42:29 +020065 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
Davide Libenzi22bacca2011-02-25 14:44:12 -080066 * It is also acquired when inserting an epoll fd onto another epoll
67 * fd. We do this so that we walk the epoll tree and ensure that this
68 * insertion does not create a cycle of epoll file descriptors, which
69 * could lead to deadlock. We need a global mutex to prevent two
70 * simultaneous inserts (A into B and B into A) from racing and
71 * constructing a cycle without either insert observing that it is
72 * going to.
Nelson Elhaged8805e62011-10-31 17:13:14 -070073 * It is necessary to acquire multiple "ep->mtx"es at once in the
74 * case when one epoll fd is added to another. In this case, we
75 * always acquire the locks in the order of nesting (i.e. after
76 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
77 * before e2->mtx). Since we disallow cycles of epoll file
78 * descriptors, this ensures that the mutexes are well-ordered. In
79 * order to communicate this nesting to lockdep, when walking a tree
80 * of epoll file descriptors, we use the current recursion depth as
81 * the lockdep subkey.
Davide Libenzid47de162007-05-15 01:40:41 -070082 * It is possible to drop the "ep->mtx" and to use the global
83 * mutex "epmutex" (together with "ep->lock") to have it working,
84 * but having "ep->mtx" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080085 * Events that require holding "epmutex" are very rare, while for
Davide Libenzid47de162007-05-15 01:40:41 -070086 * normal operations the epoll private "ep->mtx" will guarantee
87 * a better scalability.
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 */
89
Linus Torvalds1da177e2005-04-16 15:20:36 -070090/* Epoll private bits inside the event mask */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020091#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
Davide Libenzi5071f972009-03-31 15:24:10 -070093/* Maximum number of nesting allowed inside epoll sets */
94#define EP_MAX_NESTS 4
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
Davide Libenzib6119672006-10-11 01:21:44 -070096#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
97
Davide Libenzid47de162007-05-15 01:40:41 -070098#define EP_UNACTIVE_PTR ((void *) -1L)
99
Davide Libenzi7ef99642008-12-01 13:13:55 -0800100#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
101
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102struct epoll_filefd {
103 struct file *file;
104 int fd;
105};
106
107/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700108 * Structure used to track possible nested calls, for too deep recursions
109 * and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700111struct nested_call_node {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 struct list_head llink;
Davide Libenzi5071f972009-03-31 15:24:10 -0700113 void *cookie;
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700114 void *ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115};
116
117/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700118 * This structure is used as collector for nested calls, to check for
119 * maximum recursion dept and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700121struct nested_calls {
122 struct list_head tasks_call_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 spinlock_t lock;
124};
125
126/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700128 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 */
130struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700131 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 struct rb_node rbn;
133
134 /* List header used to link this structure to the eventpoll ready list */
135 struct list_head rdllink;
136
Davide Libenzic7ea7632007-05-15 01:40:47 -0700137 /*
138 * Works together "struct eventpoll"->ovflist in keeping the
139 * single linked chain of items.
140 */
141 struct epitem *next;
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 /* The file descriptor information this item refers to */
144 struct epoll_filefd ffd;
145
146 /* Number of active wait queue attached to poll operations */
147 int nwait;
148
149 /* List containing poll wait queues */
150 struct list_head pwqlist;
151
152 /* The "container" of this item */
153 struct eventpoll *ep;
154
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 /* List header used to link this item to the "struct file" items list */
156 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700157
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200158 /* wakeup_source used when EPOLLWAKEUP is set */
159 struct wakeup_source *ws;
160
Davide Libenzic7ea7632007-05-15 01:40:47 -0700161 /* The structure that describe the interested events and the source fd */
162 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700163};
164
165/*
166 * This structure is stored inside the "private_data" member of the file
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200167 * structure and represents the main data structure for the eventpoll
Davide Libenzid47de162007-05-15 01:40:41 -0700168 * interface.
169 */
170struct eventpoll {
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200171 /* Protect the access to this structure */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700172 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700173
174 /*
175 * This mutex is used to ensure that files are not removed
176 * while epoll is using them. This is held during the event
177 * collection loop, the file cleanup path, the epoll file exit
178 * code and the ctl operations.
179 */
180 struct mutex mtx;
181
182 /* Wait queue used by sys_epoll_wait() */
183 wait_queue_head_t wq;
184
185 /* Wait queue used by file->poll() */
186 wait_queue_head_t poll_wait;
187
188 /* List of ready file descriptors */
189 struct list_head rdllist;
190
Davide Libenzi67647d02007-05-15 01:40:52 -0700191 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700192 struct rb_root rbr;
193
194 /*
195 * This is a single linked list that chains all the "struct epitem" that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300196 * happened while transferring ready events to userspace w/out
Davide Libenzid47de162007-05-15 01:40:41 -0700197 * holding ->lock.
198 */
199 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800200
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200201 /* wakeup_source used when ep_scan_ready_list is running */
202 struct wakeup_source *ws;
203
Davide Libenzi7ef99642008-12-01 13:13:55 -0800204 /* The user that created the eventpoll descriptor */
205 struct user_struct *user;
Jason Baron28d82dc2012-01-12 17:17:43 -0800206
207 struct file *file;
208
209 /* used to optimize loop detection check */
210 int visited;
211 struct list_head visited_list_link;
Davide Libenzid47de162007-05-15 01:40:41 -0700212};
213
214/* Wait structure used by the poll hooks */
215struct eppoll_entry {
216 /* List header used to link this structure to the "struct epitem" */
217 struct list_head llink;
218
219 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700220 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700221
222 /*
223 * Wait queue item that will be linked to the target file wait
224 * queue head.
225 */
226 wait_queue_t wait;
227
228 /* The wait queue head that linked the "wait" wait queue item */
229 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230};
231
232/* Wrapper struct used by poll queueing */
233struct ep_pqueue {
234 poll_table pt;
235 struct epitem *epi;
236};
237
Davide Libenzi5071f972009-03-31 15:24:10 -0700238/* Used by the ep_send_events() function as callback private data */
239struct ep_send_events_data {
240 int maxevents;
241 struct epoll_event __user *events;
242};
243
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800245 * Configuration options available inside /proc/sys/fs/epoll/
246 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800247/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800248static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800249
250/*
Davide Libenzid47de162007-05-15 01:40:41 -0700251 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800253static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254
Davide Libenzi22bacca2011-02-25 14:44:12 -0800255/* Used to check for epoll file descriptor inclusion loops */
256static struct nested_calls poll_loop_ncalls;
257
Davide Libenzi5071f972009-03-31 15:24:10 -0700258/* Used for safe wake up implementation */
259static struct nested_calls poll_safewake_ncalls;
260
261/* Used to call file's f_op->poll() under the nested calls boundaries */
262static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
264/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800265static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
267/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800268static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
Jason Baron28d82dc2012-01-12 17:17:43 -0800270/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
271static LIST_HEAD(visited_list);
272
273/*
274 * List of files with newly added links, where we may need to limit the number
275 * of emanating paths. Protected by the epmutex.
276 */
277static LIST_HEAD(tfile_check_list);
278
Davide Libenzi7ef99642008-12-01 13:13:55 -0800279#ifdef CONFIG_SYSCTL
280
281#include <linux/sysctl.h>
282
Robin Holt52bd19f72011-01-12 17:00:01 -0800283static long zero;
284static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800285
286ctl_table epoll_table[] = {
287 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800288 .procname = "max_user_watches",
289 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800290 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800291 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800292 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800293 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800294 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800295 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800296 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800297};
298#endif /* CONFIG_SYSCTL */
299
Jason Baron28d82dc2012-01-12 17:17:43 -0800300static const struct file_operations eventpoll_fops;
301
302static inline int is_file_epoll(struct file *f)
303{
304 return f->f_op == &eventpoll_fops;
305}
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700306
Davide Libenzi67647d02007-05-15 01:40:52 -0700307/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700308static inline void ep_set_ffd(struct epoll_filefd *ffd,
309 struct file *file, int fd)
310{
311 ffd->file = file;
312 ffd->fd = fd;
313}
314
Davide Libenzi67647d02007-05-15 01:40:52 -0700315/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700316static inline int ep_cmp_ffd(struct epoll_filefd *p1,
317 struct epoll_filefd *p2)
318{
319 return (p1->file > p2->file ? +1:
320 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
321}
322
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700323/* Tells us if the item is currently linked */
324static inline int ep_is_linked(struct list_head *p)
325{
326 return !list_empty(p);
327}
328
Oleg Nesterov971316f2012-02-24 20:07:29 +0100329static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
330{
331 return container_of(p, struct eppoll_entry, wait);
332}
333
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700334/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700335static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700336{
337 return container_of(p, struct eppoll_entry, wait)->base;
338}
339
340/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700341static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700342{
343 return container_of(p, struct ep_pqueue, pt)->epi;
344}
345
346/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700347static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700348{
Paton J. Lewis03a7beb2012-10-04 17:13:39 -0700349 return op == EPOLL_CTL_ADD || op == EPOLL_CTL_MOD;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700350}
351
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700353static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354{
Davide Libenzi5071f972009-03-31 15:24:10 -0700355 INIT_LIST_HEAD(&ncalls->tasks_call_list);
356 spin_lock_init(&ncalls->lock);
357}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Davide Libenzi5071f972009-03-31 15:24:10 -0700359/**
Davide Libenzi3fb0e582011-03-22 16:34:46 -0700360 * ep_events_available - Checks if ready events might be available.
361 *
362 * @ep: Pointer to the eventpoll context.
363 *
364 * Returns: Returns a value different than zero if ready events are available,
365 * or zero otherwise.
366 */
367static inline int ep_events_available(struct eventpoll *ep)
368{
369 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
370}
371
372/**
Davide Libenzi5071f972009-03-31 15:24:10 -0700373 * ep_call_nested - Perform a bound (possibly) nested call, by checking
374 * that the recursion limit is not exceeded, and that
375 * the same nested call (by the meaning of same cookie) is
376 * no re-entered.
377 *
378 * @ncalls: Pointer to the nested_calls structure to be used for this call.
379 * @max_nests: Maximum number of allowed nesting calls.
380 * @nproc: Nested call core function pointer.
381 * @priv: Opaque data to be passed to the @nproc callback.
382 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700383 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700384 *
385 * Returns: Returns the code returned by the @nproc callback, or -1 if
386 * the maximum recursion limit has been exceeded.
387 */
388static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
389 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700390 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700391{
392 int error, call_nests = 0;
393 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700394 struct list_head *lsthead = &ncalls->tasks_call_list;
395 struct nested_call_node *tncur;
396 struct nested_call_node tnode;
397
398 spin_lock_irqsave(&ncalls->lock, flags);
399
400 /*
401 * Try to see if the current task is already inside this wakeup call.
402 * We use a list here, since the population inside this set is always
403 * very much limited.
404 */
405 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700406 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700407 (tncur->cookie == cookie || ++call_nests > max_nests)) {
408 /*
409 * Ops ... loop detected or maximum nest level reached.
410 * We abort this wake by breaking the cycle itself.
411 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700412 error = -1;
413 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700414 }
415 }
416
417 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700418 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700419 tnode.cookie = cookie;
420 list_add(&tnode.llink, lsthead);
421
422 spin_unlock_irqrestore(&ncalls->lock, flags);
423
424 /* Call the nested function */
425 error = (*nproc)(priv, cookie, call_nests);
426
427 /* Remove the current task from the list */
428 spin_lock_irqsave(&ncalls->lock, flags);
429 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700430out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700431 spin_unlock_irqrestore(&ncalls->lock, flags);
432
433 return error;
434}
435
Steven Rostedt02edc6f2012-03-23 15:02:27 -0700436/*
437 * As described in commit 0ccf831cb lockdep: annotate epoll
438 * the use of wait queues used by epoll is done in a very controlled
439 * manner. Wake ups can nest inside each other, but are never done
440 * with the same locking. For example:
441 *
442 * dfd = socket(...);
443 * efd1 = epoll_create();
444 * efd2 = epoll_create();
445 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
446 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
447 *
448 * When a packet arrives to the device underneath "dfd", the net code will
449 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
450 * callback wakeup entry on that queue, and the wake_up() performed by the
451 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
452 * (efd1) notices that it may have some event ready, so it needs to wake up
453 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
454 * that ends up in another wake_up(), after having checked about the
455 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
456 * avoid stack blasting.
457 *
458 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
459 * this special case of epoll.
460 */
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700461#ifdef CONFIG_DEBUG_LOCK_ALLOC
462static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
463 unsigned long events, int subclass)
464{
465 unsigned long flags;
466
467 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
468 wake_up_locked_poll(wqueue, events);
469 spin_unlock_irqrestore(&wqueue->lock, flags);
470}
471#else
472static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
473 unsigned long events, int subclass)
474{
475 wake_up_poll(wqueue, events);
476}
477#endif
478
Davide Libenzi5071f972009-03-31 15:24:10 -0700479static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
480{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700481 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
482 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700483 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484}
485
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486/*
487 * Perform a safe wake up of the poll wait list. The problem is that
488 * with the new callback'd wake up system, it is possible that the
489 * poll callback is reentered from inside the call to wake_up() done
490 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700491 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 * and we cannot reenter the same wait queue head at all. This will
493 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700494 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700496static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700498 int this_cpu = get_cpu();
499
Davide Libenzi5071f972009-03-31 15:24:10 -0700500 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700501 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
502
503 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504}
505
Oleg Nesterov971316f2012-02-24 20:07:29 +0100506static void ep_remove_wait_queue(struct eppoll_entry *pwq)
507{
508 wait_queue_head_t *whead;
509
510 rcu_read_lock();
511 /* If it is cleared by POLLFREE, it should be rcu-safe */
512 whead = rcu_dereference(pwq->whead);
513 if (whead)
514 remove_wait_queue(whead, &pwq->wait);
515 rcu_read_unlock();
516}
517
Davide Libenzi7699acd2007-05-10 22:23:23 -0700518/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700519 * This function unregisters poll callbacks from the associated file
520 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
521 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700522 */
523static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
524{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700525 struct list_head *lsthead = &epi->pwqlist;
526 struct eppoll_entry *pwq;
527
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700528 while (!list_empty(lsthead)) {
529 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700530
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700531 list_del(&pwq->llink);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100532 ep_remove_wait_queue(pwq);
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700533 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700534 }
535}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536
Davide Libenzi5071f972009-03-31 15:24:10 -0700537/**
538 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
539 * the scan code, to call f_op->poll(). Also allows for
540 * O(NumReady) performance.
541 *
542 * @ep: Pointer to the epoll private data structure.
543 * @sproc: Pointer to the scan callback.
544 * @priv: Private opaque data passed to the @sproc callback.
Nelson Elhaged8805e62011-10-31 17:13:14 -0700545 * @depth: The current depth of recursive f_op->poll calls.
Davide Libenzi5071f972009-03-31 15:24:10 -0700546 *
547 * Returns: The same integer error code returned by the @sproc callback.
548 */
549static int ep_scan_ready_list(struct eventpoll *ep,
550 int (*sproc)(struct eventpoll *,
551 struct list_head *, void *),
Nelson Elhaged8805e62011-10-31 17:13:14 -0700552 void *priv,
553 int depth)
Davide Libenzi5071f972009-03-31 15:24:10 -0700554{
555 int error, pwake = 0;
556 unsigned long flags;
557 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700558 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700559
560 /*
561 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700562 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700563 */
Nelson Elhaged8805e62011-10-31 17:13:14 -0700564 mutex_lock_nested(&ep->mtx, depth);
Davide Libenzi5071f972009-03-31 15:24:10 -0700565
566 /*
567 * Steal the ready list, and re-init the original one to the
568 * empty list. Also, set ep->ovflist to NULL so that events
569 * happening while looping w/out locks, are not lost. We cannot
570 * have the poll callback to queue directly on ep->rdllist,
571 * because we want the "sproc" callback to be able to do it
572 * in a lockless way.
573 */
574 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700575 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700576 ep->ovflist = NULL;
577 spin_unlock_irqrestore(&ep->lock, flags);
578
579 /*
580 * Now call the callback function.
581 */
582 error = (*sproc)(ep, &txlist, priv);
583
584 spin_lock_irqsave(&ep->lock, flags);
585 /*
586 * During the time we spent inside the "sproc" callback, some
587 * other events might have been queued by the poll callback.
588 * We re-insert them inside the main ready-list here.
589 */
590 for (nepi = ep->ovflist; (epi = nepi) != NULL;
591 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
592 /*
593 * We need to check if the item is already in the list.
594 * During the "sproc" callback execution time, items are
595 * queued into ->ovflist but the "txlist" might already
596 * contain them, and the list_splice() below takes care of them.
597 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200598 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700599 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200600 __pm_stay_awake(epi->ws);
601 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700602 }
603 /*
604 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
605 * releasing the lock, events will be queued in the normal way inside
606 * ep->rdllist.
607 */
608 ep->ovflist = EP_UNACTIVE_PTR;
609
610 /*
611 * Quickly re-inject items left on "txlist".
612 */
613 list_splice(&txlist, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200614 __pm_relax(ep->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700615
616 if (!list_empty(&ep->rdllist)) {
617 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700618 * Wake up (if active) both the eventpoll wait list and
619 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700620 */
621 if (waitqueue_active(&ep->wq))
622 wake_up_locked(&ep->wq);
623 if (waitqueue_active(&ep->poll_wait))
624 pwake++;
625 }
626 spin_unlock_irqrestore(&ep->lock, flags);
627
628 mutex_unlock(&ep->mtx);
629
630 /* We have to call this outside the lock */
631 if (pwake)
632 ep_poll_safewake(&ep->poll_wait);
633
634 return error;
635}
636
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700638 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700639 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700640 */
641static int ep_remove(struct eventpoll *ep, struct epitem *epi)
642{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700643 unsigned long flags;
644 struct file *file = epi->ffd.file;
645
646 /*
647 * Removes poll wait queue hooks. We _have_ to do this without holding
648 * the "ep->lock" otherwise a deadlock might occur. This because of the
649 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
650 * queue head lock when unregistering the wait queue. The wakeup callback
651 * will run by holding the wait queue head lock and will call our callback
652 * that will try to get "ep->lock".
653 */
654 ep_unregister_pollwait(ep, epi);
655
656 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700657 spin_lock(&file->f_lock);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700658 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700659 list_del_init(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700660 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661
Davide Libenzicdac75e2008-04-29 00:58:34 -0700662 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
Davide Libenzic7ea7632007-05-15 01:40:47 -0700664 spin_lock_irqsave(&ep->lock, flags);
665 if (ep_is_linked(&epi->rdllink))
666 list_del_init(&epi->rdllink);
667 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200669 wakeup_source_unregister(epi->ws);
670
Davide Libenzi7699acd2007-05-10 22:23:23 -0700671 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700672 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
Robin Holt52bd19f72011-01-12 17:00:01 -0800674 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800675
Davide Libenzic7ea7632007-05-15 01:40:47 -0700676 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677}
678
Paton J. Lewis03a7beb2012-10-04 17:13:39 -0700679/*
680 * Disables a "struct epitem" in the eventpoll set. Returns -EBUSY if the item
681 * had no event flags set, indicating that another thread may be currently
682 * handling that item's events (in the case that EPOLLONESHOT was being
683 * used). Otherwise a zero result indicates that the item has been disabled
684 * from receiving events. A disabled item may be re-enabled via
685 * EPOLL_CTL_MOD. Must be called with "mtx" held.
686 */
687static int ep_disable(struct eventpoll *ep, struct epitem *epi)
688{
689 int result = 0;
690 unsigned long flags;
691
692 spin_lock_irqsave(&ep->lock, flags);
693 if (epi->event.events & ~EP_PRIVATE_BITS) {
694 if (ep_is_linked(&epi->rdllink))
695 list_del_init(&epi->rdllink);
696 /* Ensure ep_poll_callback will not add epi back onto ready
697 list: */
698 epi->event.events &= EP_PRIVATE_BITS;
699 }
700 else
701 result = -EBUSY;
702 spin_unlock_irqrestore(&ep->lock, flags);
703
704 return result;
705}
706
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707static void ep_free(struct eventpoll *ep)
708{
709 struct rb_node *rbp;
710 struct epitem *epi;
711
712 /* We need to release all tasks waiting for these file */
713 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700714 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
716 /*
717 * We need to lock this because we could be hit by
718 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700719 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 * is on the way to be removed and no one has references to it
721 * anymore. The only hit might come from eventpoll_release_file() but
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300722 * holding "epmutex" is sufficient here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800724 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725
726 /*
727 * Walks through the whole tree by unregistering poll callbacks.
728 */
729 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
730 epi = rb_entry(rbp, struct epitem, rbn);
731
732 ep_unregister_pollwait(ep, epi);
733 }
734
735 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700736 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700738 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 * us during this operation. So we can avoid the lock on "ep->lock".
740 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700741 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 epi = rb_entry(rbp, struct epitem, rbn);
743 ep_remove(ep, epi);
744 }
745
Arjan van de Ven144efe32006-03-23 03:00:32 -0800746 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700747 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800748 free_uid(ep->user);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200749 wakeup_source_unregister(ep->ws);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700750 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751}
752
Davide Libenzi7699acd2007-05-10 22:23:23 -0700753static int ep_eventpoll_release(struct inode *inode, struct file *file)
754{
755 struct eventpoll *ep = file->private_data;
756
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700757 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700758 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700759
Davide Libenzi7699acd2007-05-10 22:23:23 -0700760 return 0;
761}
762
Davide Libenzi296e2362009-03-31 15:24:11 -0700763static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
764 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700765{
766 struct epitem *epi, *tmp;
Hans Verkuil626cf232012-03-23 15:02:27 -0700767 poll_table pt;
Davide Libenzi5071f972009-03-31 15:24:10 -0700768
Hans Verkuil626cf232012-03-23 15:02:27 -0700769 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -0700770 list_for_each_entry_safe(epi, tmp, head, rdllink) {
Hans Verkuil626cf232012-03-23 15:02:27 -0700771 pt._key = epi->event.events;
772 if (epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
Davide Libenzi5071f972009-03-31 15:24:10 -0700773 epi->event.events)
774 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700775 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700776 /*
777 * Item has been dropped into the ready list by the poll
778 * callback, but it's not actually ready, as far as
779 * caller requested events goes. We can remove it here.
780 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200781 __pm_relax(epi->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700782 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700783 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700784 }
785
786 return 0;
787}
788
789static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
790{
Nelson Elhaged8805e62011-10-31 17:13:14 -0700791 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
Davide Libenzi5071f972009-03-31 15:24:10 -0700792}
793
Davide Libenzi7699acd2007-05-10 22:23:23 -0700794static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
795{
Davide Libenzi5071f972009-03-31 15:24:10 -0700796 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700797 struct eventpoll *ep = file->private_data;
798
799 /* Insert inside our poll wait queue */
800 poll_wait(file, &ep->poll_wait, wait);
801
Davide Libenzi5071f972009-03-31 15:24:10 -0700802 /*
803 * Proceed to find out if wanted events are really available inside
804 * the ready list. This need to be done under ep_call_nested()
805 * supervision, since the call to f_op->poll() done on listed files
806 * could re-enter here.
807 */
808 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700809 ep_poll_readyevents_proc, ep, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700810
Davide Libenzi296e2362009-03-31 15:24:11 -0700811 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700812}
813
814/* File callbacks that implement the eventpoll file behaviour */
815static const struct file_operations eventpoll_fops = {
816 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200817 .poll = ep_eventpoll_poll,
818 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700819};
820
Davide Libenzi7699acd2007-05-10 22:23:23 -0700821/*
822 * This is called from eventpoll_release() to unlink files from the eventpoll
823 * interface. We need to have this facility to cleanup correctly files that are
824 * closed without being removed from the eventpoll interface.
825 */
826void eventpoll_release_file(struct file *file)
827{
828 struct list_head *lsthead = &file->f_ep_links;
829 struct eventpoll *ep;
830 struct epitem *epi;
831
832 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700833 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700834 * necessary. It is not necessary because we're in the "struct file"
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300835 * cleanup path, and this means that no one is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700836 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700837 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700838 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700839 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700840 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700841 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700842 *
843 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700844 */
845 mutex_lock(&epmutex);
846
847 while (!list_empty(lsthead)) {
848 epi = list_first_entry(lsthead, struct epitem, fllink);
849
850 ep = epi->ep;
851 list_del_init(&epi->fllink);
Nelson Elhaged8805e62011-10-31 17:13:14 -0700852 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700853 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700854 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700855 }
856
857 mutex_unlock(&epmutex);
858}
859
860static int ep_alloc(struct eventpoll **pep)
861{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800862 int error;
863 struct user_struct *user;
864 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700865
Davide Libenzi7ef99642008-12-01 13:13:55 -0800866 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800867 error = -ENOMEM;
868 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
869 if (unlikely(!ep))
870 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700871
Davide Libenzic7ea7632007-05-15 01:40:47 -0700872 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700873 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700874 init_waitqueue_head(&ep->wq);
875 init_waitqueue_head(&ep->poll_wait);
876 INIT_LIST_HEAD(&ep->rdllist);
877 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700878 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800879 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700880
881 *pep = ep;
882
Davide Libenzi7699acd2007-05-10 22:23:23 -0700883 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800884
885free_uid:
886 free_uid(user);
887 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700888}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
890/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700891 * Search the file inside the eventpoll tree. The RB tree operations
892 * are protected by the "mtx" mutex, and ep_find() must be called with
893 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 */
895static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
896{
897 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 struct rb_node *rbp;
899 struct epitem *epi, *epir = NULL;
900 struct epoll_filefd ffd;
901
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700902 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 for (rbp = ep->rbr.rb_node; rbp; ) {
904 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700905 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 if (kcmp > 0)
907 rbp = rbp->rb_right;
908 else if (kcmp < 0)
909 rbp = rbp->rb_left;
910 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 epir = epi;
912 break;
913 }
914 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 return epir;
917}
918
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700920 * This is the callback that is passed to the wait queue wakeup
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200921 * mechanism. It is called by the stored file descriptors when they
Davide Libenzi7699acd2007-05-10 22:23:23 -0700922 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700924static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700926 int pwake = 0;
927 unsigned long flags;
928 struct epitem *epi = ep_item_from_wait(wait);
929 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930
Oleg Nesterov971316f2012-02-24 20:07:29 +0100931 if ((unsigned long)key & POLLFREE) {
932 ep_pwq_from_wait(wait)->whead = NULL;
933 /*
934 * whead = NULL above can race with ep_remove_wait_queue()
935 * which can do another remove_wait_queue() after us, so we
936 * can't use __remove_wait_queue(). whead->lock is held by
937 * the caller.
938 */
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100939 list_del_init(&wait->task_list);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100940 }
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100941
Davide Libenzic7ea7632007-05-15 01:40:47 -0700942 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700943
944 /*
945 * If the event mask does not contain any poll(2) event, we consider the
946 * descriptor to be disabled. This condition is likely the effect of the
947 * EPOLLONESHOT bit that disables the descriptor when an event is received,
948 * until the next EPOLL_CTL_MOD will be issued.
949 */
950 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700951 goto out_unlock;
952
953 /*
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700954 * Check the events coming with the callback. At this stage, not
955 * every device reports the events in the "key" parameter of the
956 * callback. We need to be able to handle both cases here, hence the
957 * test for "key" != NULL before the event match test.
958 */
959 if (key && !((unsigned long) key & epi->event.events))
960 goto out_unlock;
961
962 /*
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200963 * If we are transferring events to userspace, we can hold no locks
Davide Libenzid47de162007-05-15 01:40:41 -0700964 * (because we're accessing user memory, and because of linux f_op->poll()
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200965 * semantics). All the events that happen during that period of time are
Davide Libenzid47de162007-05-15 01:40:41 -0700966 * chained in ep->ovflist and requeued later on.
967 */
968 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
969 if (epi->next == EP_UNACTIVE_PTR) {
970 epi->next = ep->ovflist;
971 ep->ovflist = epi;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200972 if (epi->ws) {
973 /*
974 * Activate ep->ws since epi->ws may get
975 * deactivated at any time.
976 */
977 __pm_stay_awake(ep->ws);
978 }
979
Davide Libenzid47de162007-05-15 01:40:41 -0700980 }
981 goto out_unlock;
982 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700983
984 /* If this file is already in the ready list we exit soon */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200985 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700986 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200987 __pm_stay_awake(epi->ws);
988 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700989
Davide Libenzi7699acd2007-05-10 22:23:23 -0700990 /*
991 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
992 * wait list.
993 */
994 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -0400995 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700996 if (waitqueue_active(&ep->poll_wait))
997 pwake++;
998
Davide Libenzid47de162007-05-15 01:40:41 -0700999out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001000 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001001
1002 /* We have to call this outside the lock */
1003 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001004 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001005
1006 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007}
1008
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009/*
1010 * This is the callback that is used to add our wait queue to the
1011 * target file wakeup lists.
1012 */
1013static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1014 poll_table *pt)
1015{
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001016 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 struct eppoll_entry *pwq;
1018
Christoph Lametere94b1762006-12-06 20:33:17 -08001019 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1021 pwq->whead = whead;
1022 pwq->base = epi;
1023 add_wait_queue(whead, &pwq->wait);
1024 list_add_tail(&pwq->llink, &epi->pwqlist);
1025 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001026 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 /* We have to signal that an error occurred */
1028 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -07001029 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030}
1031
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1033{
1034 int kcmp;
1035 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1036 struct epitem *epic;
1037
1038 while (*p) {
1039 parent = *p;
1040 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001041 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 if (kcmp > 0)
1043 p = &parent->rb_right;
1044 else
1045 p = &parent->rb_left;
1046 }
1047 rb_link_node(&epi->rbn, parent, p);
1048 rb_insert_color(&epi->rbn, &ep->rbr);
1049}
1050
Jason Baron28d82dc2012-01-12 17:17:43 -08001051#define PATH_ARR_SIZE 5
1052/*
1053 * These are the number paths of length 1 to 5, that we are allowing to emanate
1054 * from a single file of interest. For example, we allow 1000 paths of length
1055 * 1, to emanate from each file of interest. This essentially represents the
1056 * potential wakeup paths, which need to be limited in order to avoid massive
1057 * uncontrolled wakeup storms. The common use case should be a single ep which
1058 * is connected to n file sources. In this case each file source has 1 path
1059 * of length 1. Thus, the numbers below should be more than sufficient. These
1060 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1061 * and delete can't add additional paths. Protected by the epmutex.
1062 */
1063static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1064static int path_count[PATH_ARR_SIZE];
1065
1066static int path_count_inc(int nests)
1067{
Jason Baron93dc6102012-03-16 16:34:03 -04001068 /* Allow an arbitrary number of depth 1 paths */
1069 if (nests == 0)
1070 return 0;
1071
Jason Baron28d82dc2012-01-12 17:17:43 -08001072 if (++path_count[nests] > path_limits[nests])
1073 return -1;
1074 return 0;
1075}
1076
1077static void path_count_init(void)
1078{
1079 int i;
1080
1081 for (i = 0; i < PATH_ARR_SIZE; i++)
1082 path_count[i] = 0;
1083}
1084
1085static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1086{
1087 int error = 0;
1088 struct file *file = priv;
1089 struct file *child_file;
1090 struct epitem *epi;
1091
1092 list_for_each_entry(epi, &file->f_ep_links, fllink) {
1093 child_file = epi->ep->file;
1094 if (is_file_epoll(child_file)) {
1095 if (list_empty(&child_file->f_ep_links)) {
1096 if (path_count_inc(call_nests)) {
1097 error = -1;
1098 break;
1099 }
1100 } else {
1101 error = ep_call_nested(&poll_loop_ncalls,
1102 EP_MAX_NESTS,
1103 reverse_path_check_proc,
1104 child_file, child_file,
1105 current);
1106 }
1107 if (error != 0)
1108 break;
1109 } else {
1110 printk(KERN_ERR "reverse_path_check_proc: "
1111 "file is not an ep!\n");
1112 }
1113 }
1114 return error;
1115}
1116
1117/**
1118 * reverse_path_check - The tfile_check_list is list of file *, which have
1119 * links that are proposed to be newly added. We need to
1120 * make sure that those added links don't add too many
1121 * paths such that we will spend all our time waking up
1122 * eventpoll objects.
1123 *
1124 * Returns: Returns zero if the proposed links don't create too many paths,
1125 * -1 otherwise.
1126 */
1127static int reverse_path_check(void)
1128{
Jason Baron28d82dc2012-01-12 17:17:43 -08001129 int error = 0;
1130 struct file *current_file;
1131
1132 /* let's call this for all tfiles */
1133 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001134 path_count_init();
1135 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1136 reverse_path_check_proc, current_file,
1137 current_file, current);
1138 if (error)
1139 break;
1140 }
1141 return error;
1142}
1143
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001144static int ep_create_wakeup_source(struct epitem *epi)
1145{
1146 const char *name;
1147
1148 if (!epi->ep->ws) {
1149 epi->ep->ws = wakeup_source_register("eventpoll");
1150 if (!epi->ep->ws)
1151 return -ENOMEM;
1152 }
1153
1154 name = epi->ffd.file->f_path.dentry->d_name.name;
1155 epi->ws = wakeup_source_register(name);
1156 if (!epi->ws)
1157 return -ENOMEM;
1158
1159 return 0;
1160}
1161
1162static void ep_destroy_wakeup_source(struct epitem *epi)
1163{
1164 wakeup_source_unregister(epi->ws);
1165 epi->ws = NULL;
1166}
1167
Davide Libenzic7ea7632007-05-15 01:40:47 -07001168/*
1169 * Must be called with "mtx" held.
1170 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1172 struct file *tfile, int fd)
1173{
1174 int error, revents, pwake = 0;
1175 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -08001176 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 struct epitem *epi;
1178 struct ep_pqueue epq;
1179
Robin Holt52bd19f72011-01-12 17:00:01 -08001180 user_watches = atomic_long_read(&ep->user->epoll_watches);
1181 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001182 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -08001183 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001184 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185
1186 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 INIT_LIST_HEAD(&epi->rdllink);
1188 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 INIT_LIST_HEAD(&epi->pwqlist);
1190 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001191 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -07001194 epi->next = EP_UNACTIVE_PTR;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001195 if (epi->event.events & EPOLLWAKEUP) {
1196 error = ep_create_wakeup_source(epi);
1197 if (error)
1198 goto error_create_wakeup_source;
1199 } else {
1200 epi->ws = NULL;
1201 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202
1203 /* Initialize the poll table using the queue callback */
1204 epq.epi = epi;
1205 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
Hans Verkuil626cf232012-03-23 15:02:27 -07001206 epq.pt._key = event->events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207
1208 /*
1209 * Attach the item to the poll hooks and get current event bits.
1210 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -07001211 * been increased by the caller of this function. Note that after
1212 * this operation completes, the poll callback can start hitting
1213 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 */
1215 revents = tfile->f_op->poll(tfile, &epq.pt);
1216
1217 /*
1218 * We have to check if something went wrong during the poll wait queue
1219 * install process. Namely an allocation for a wait queue failed due
1220 * high memory pressure.
1221 */
Davide Libenzi7ef99642008-12-01 13:13:55 -08001222 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001224 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225
1226 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -07001227 spin_lock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 list_add_tail(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -07001229 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230
Davide Libenzic7ea7632007-05-15 01:40:47 -07001231 /*
1232 * Add the current item to the RB tree. All RB tree operations are
1233 * protected by "mtx", and ep_insert() is called with "mtx" held.
1234 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 ep_rbtree_insert(ep, epi);
1236
Jason Baron28d82dc2012-01-12 17:17:43 -08001237 /* now check if we've created too many backpaths */
1238 error = -EINVAL;
1239 if (reverse_path_check())
1240 goto error_remove_epi;
1241
Davide Libenzic7ea7632007-05-15 01:40:47 -07001242 /* We have to drop the new item inside our item list to keep track of it */
1243 spin_lock_irqsave(&ep->lock, flags);
1244
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001246 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001248 __pm_stay_awake(epi->ws);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249
1250 /* Notify waiting tasks that events are available */
1251 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001252 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 if (waitqueue_active(&ep->poll_wait))
1254 pwake++;
1255 }
1256
Davide Libenzic7ea7632007-05-15 01:40:47 -07001257 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258
Robin Holt52bd19f72011-01-12 17:00:01 -08001259 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 /* We have to call this outside the lock */
1262 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001263 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 return 0;
1266
Jason Baron28d82dc2012-01-12 17:17:43 -08001267error_remove_epi:
1268 spin_lock(&tfile->f_lock);
1269 if (ep_is_linked(&epi->fllink))
1270 list_del_init(&epi->fllink);
1271 spin_unlock(&tfile->f_lock);
1272
1273 rb_erase(&epi->rbn, &ep->rbr);
1274
Davide Libenzi7699acd2007-05-10 22:23:23 -07001275error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 ep_unregister_pollwait(ep, epi);
1277
1278 /*
1279 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -07001280 * allocated wait queue. Note that we don't care about the ep->ovflist
1281 * list, since that is used/cleaned only inside a section bound by "mtx".
1282 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001284 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001285 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001286 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001287 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001289 wakeup_source_unregister(epi->ws);
1290
1291error_create_wakeup_source:
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001292 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001293
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 return error;
1295}
1296
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297/*
1298 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001299 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 */
1301static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1302{
1303 int pwake = 0;
1304 unsigned int revents;
Hans Verkuil626cf232012-03-23 15:02:27 -07001305 poll_table pt;
1306
1307 init_poll_funcptr(&pt, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308
1309 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001310 * Set the new event interest mask before calling f_op->poll();
1311 * otherwise we might miss an event that happens between the
1312 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 */
1314 epi->event.events = event->events;
Hans Verkuil626cf232012-03-23 15:02:27 -07001315 pt._key = event->events;
Tony Battersbye057e152009-03-31 15:24:15 -07001316 epi->event.data = event->data; /* protected by mtx */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001317 if (epi->event.events & EPOLLWAKEUP) {
1318 if (!epi->ws)
1319 ep_create_wakeup_source(epi);
1320 } else if (epi->ws) {
1321 ep_destroy_wakeup_source(epi);
1322 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323
1324 /*
1325 * Get current event bits. We can safely use the file* here because
1326 * its usage count has been increased by the caller of this function.
1327 */
Hans Verkuil626cf232012-03-23 15:02:27 -07001328 revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001331 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001332 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001334 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001335 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001336 if (!ep_is_linked(&epi->rdllink)) {
1337 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001338 __pm_stay_awake(epi->ws);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339
Davide Libenzic7ea7632007-05-15 01:40:47 -07001340 /* Notify waiting tasks that events are available */
1341 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001342 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001343 if (waitqueue_active(&ep->poll_wait))
1344 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 }
Tony Battersbye057e152009-03-31 15:24:15 -07001346 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
1349 /* We have to call this outside the lock */
1350 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001351 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352
1353 return 0;
1354}
1355
Davide Libenzi296e2362009-03-31 15:24:11 -07001356static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1357 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001358{
1359 struct ep_send_events_data *esed = priv;
1360 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001361 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001362 struct epitem *epi;
1363 struct epoll_event __user *uevent;
Hans Verkuil626cf232012-03-23 15:02:27 -07001364 poll_table pt;
1365
1366 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -07001367
Davide Libenzi296e2362009-03-31 15:24:11 -07001368 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001369 * We can loop without lock because we are passed a task private list.
1370 * Items cannot vanish during the loop because ep_scan_ready_list() is
1371 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001372 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001373 for (eventcnt = 0, uevent = esed->events;
1374 !list_empty(head) && eventcnt < esed->maxevents;) {
1375 epi = list_first_entry(head, struct epitem, rdllink);
1376
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001377 /*
1378 * Activate ep->ws before deactivating epi->ws to prevent
1379 * triggering auto-suspend here (in case we reactive epi->ws
1380 * below).
1381 *
1382 * This could be rearranged to delay the deactivation of epi->ws
1383 * instead, but then epi->ws would temporarily be out of sync
1384 * with ep_is_linked().
1385 */
1386 if (epi->ws && epi->ws->active)
1387 __pm_stay_awake(ep->ws);
1388 __pm_relax(epi->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -07001389 list_del_init(&epi->rdllink);
1390
Hans Verkuil626cf232012-03-23 15:02:27 -07001391 pt._key = epi->event.events;
1392 revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
Davide Libenzi296e2362009-03-31 15:24:11 -07001393 epi->event.events;
Davide Libenzi5071f972009-03-31 15:24:10 -07001394
Davide Libenzi296e2362009-03-31 15:24:11 -07001395 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001396 * If the event mask intersect the caller-requested one,
1397 * deliver the event to userspace. Again, ep_scan_ready_list()
1398 * is holding "mtx", so no operations coming from userspace
1399 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001400 */
1401 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001402 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001403 __put_user(epi->event.data, &uevent->data)) {
1404 list_add(&epi->rdllink, head);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001405 __pm_stay_awake(epi->ws);
Davide Libenzi296e2362009-03-31 15:24:11 -07001406 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001407 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001408 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001409 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001410 if (epi->event.events & EPOLLONESHOT)
1411 epi->event.events &= EP_PRIVATE_BITS;
1412 else if (!(epi->event.events & EPOLLET)) {
1413 /*
1414 * If this file has been added with Level
1415 * Trigger mode, we need to insert back inside
1416 * the ready list, so that the next call to
1417 * epoll_wait() will check again the events
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001418 * availability. At this point, no one can insert
Davide Libenzi296e2362009-03-31 15:24:11 -07001419 * into ep->rdllist besides us. The epoll_ctl()
1420 * callers are locked out by
1421 * ep_scan_ready_list() holding "mtx" and the
1422 * poll callback will queue them in ep->ovflist.
1423 */
1424 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001425 __pm_stay_awake(epi->ws);
Davide Libenzi296e2362009-03-31 15:24:11 -07001426 }
1427 }
1428 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001429
1430 return eventcnt;
1431}
1432
Davide Libenzi296e2362009-03-31 15:24:11 -07001433static int ep_send_events(struct eventpoll *ep,
1434 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435{
Davide Libenzi5071f972009-03-31 15:24:10 -07001436 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001437
Davide Libenzi5071f972009-03-31 15:24:10 -07001438 esed.maxevents = maxevents;
1439 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001440
Nelson Elhaged8805e62011-10-31 17:13:14 -07001441 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442}
1443
Eric Dumazet0781b902011-02-01 15:52:35 -08001444static inline struct timespec ep_set_mstimeout(long ms)
1445{
1446 struct timespec now, ts = {
1447 .tv_sec = ms / MSEC_PER_SEC,
1448 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1449 };
1450
1451 ktime_get_ts(&now);
1452 return timespec_add_safe(now, ts);
1453}
1454
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001455/**
1456 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1457 * event buffer.
1458 *
1459 * @ep: Pointer to the eventpoll context.
1460 * @events: Pointer to the userspace buffer where the ready events should be
1461 * stored.
1462 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1463 * @timeout: Maximum timeout for the ready events fetch operation, in
1464 * milliseconds. If the @timeout is zero, the function will not block,
1465 * while if the @timeout is less than zero, the function will block
1466 * until at least one event has been retrieved (or an error
1467 * occurred).
1468 *
1469 * Returns: Returns the number of ready events which have been fetched, or an
1470 * error code, in case of error.
1471 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1473 int maxevents, long timeout)
1474{
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001475 int res = 0, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 unsigned long flags;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001477 long slack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001479 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001481 if (timeout > 0) {
Eric Dumazet0781b902011-02-01 15:52:35 -08001482 struct timespec end_time = ep_set_mstimeout(timeout);
1483
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001484 slack = select_estimate_accuracy(&end_time);
1485 to = &expires;
1486 *to = timespec_to_ktime(end_time);
1487 } else if (timeout == 0) {
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001488 /*
1489 * Avoid the unnecessary trip to the wait queue loop, if the
1490 * caller specified a non blocking operation.
1491 */
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001492 timed_out = 1;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001493 spin_lock_irqsave(&ep->lock, flags);
1494 goto check_events;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001495 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001497fetch_events:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001498 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001500 if (!ep_events_available(ep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 /*
1502 * We don't have any available event to return to the caller.
1503 * We need to sleep here, and we will be wake up by
1504 * ep_poll_callback() when events will become available.
1505 */
1506 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001507 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
1509 for (;;) {
1510 /*
1511 * We don't want to sleep if the ep_poll_callback() sends us
1512 * a wakeup in between. That's why we set the task state
1513 * to TASK_INTERRUPTIBLE before doing the checks.
1514 */
1515 set_current_state(TASK_INTERRUPTIBLE);
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001516 if (ep_events_available(ep) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 break;
1518 if (signal_pending(current)) {
1519 res = -EINTR;
1520 break;
1521 }
1522
Davide Libenzic7ea7632007-05-15 01:40:47 -07001523 spin_unlock_irqrestore(&ep->lock, flags);
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001524 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1525 timed_out = 1;
1526
Davide Libenzic7ea7632007-05-15 01:40:47 -07001527 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001529 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530
1531 set_current_state(TASK_RUNNING);
1532 }
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001533check_events:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 /* Is it worth to try to dig for events ? */
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001535 eavail = ep_events_available(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Davide Libenzic7ea7632007-05-15 01:40:47 -07001537 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538
1539 /*
1540 * Try to transfer events to user space. In case we get 0 events and
1541 * there's still timeout left over, we go trying again in search of
1542 * more luck.
1543 */
1544 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001545 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001546 goto fetch_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547
1548 return res;
1549}
1550
Davide Libenzi22bacca2011-02-25 14:44:12 -08001551/**
1552 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1553 * API, to verify that adding an epoll file inside another
1554 * epoll structure, does not violate the constraints, in
1555 * terms of closed loops, or too deep chains (which can
1556 * result in excessive stack usage).
1557 *
1558 * @priv: Pointer to the epoll file to be currently checked.
1559 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1560 * data structure pointer.
1561 * @call_nests: Current dept of the @ep_call_nested() call stack.
1562 *
1563 * Returns: Returns zero if adding the epoll @file inside current epoll
1564 * structure @ep does not violate the constraints, or -1 otherwise.
1565 */
1566static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1567{
1568 int error = 0;
1569 struct file *file = priv;
1570 struct eventpoll *ep = file->private_data;
Jason Baron28d82dc2012-01-12 17:17:43 -08001571 struct eventpoll *ep_tovisit;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001572 struct rb_node *rbp;
1573 struct epitem *epi;
1574
Nelson Elhaged8805e62011-10-31 17:13:14 -07001575 mutex_lock_nested(&ep->mtx, call_nests + 1);
Jason Baron28d82dc2012-01-12 17:17:43 -08001576 ep->visited = 1;
1577 list_add(&ep->visited_list_link, &visited_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001578 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1579 epi = rb_entry(rbp, struct epitem, rbn);
1580 if (unlikely(is_file_epoll(epi->ffd.file))) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001581 ep_tovisit = epi->ffd.file->private_data;
1582 if (ep_tovisit->visited)
1583 continue;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001584 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Jason Baron28d82dc2012-01-12 17:17:43 -08001585 ep_loop_check_proc, epi->ffd.file,
1586 ep_tovisit, current);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001587 if (error != 0)
1588 break;
Jason Baron28d82dc2012-01-12 17:17:43 -08001589 } else {
1590 /*
1591 * If we've reached a file that is not associated with
1592 * an ep, then we need to check if the newly added
1593 * links are going to add too many wakeup paths. We do
1594 * this by adding it to the tfile_check_list, if it's
1595 * not already there, and calling reverse_path_check()
1596 * during ep_insert().
1597 */
1598 if (list_empty(&epi->ffd.file->f_tfile_llink))
1599 list_add(&epi->ffd.file->f_tfile_llink,
1600 &tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001601 }
1602 }
1603 mutex_unlock(&ep->mtx);
1604
1605 return error;
1606}
1607
1608/**
1609 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1610 * another epoll file (represented by @ep) does not create
1611 * closed loops or too deep chains.
1612 *
1613 * @ep: Pointer to the epoll private data structure.
1614 * @file: Pointer to the epoll file to be checked.
1615 *
1616 * Returns: Returns zero if adding the epoll @file inside current epoll
1617 * structure @ep does not violate the constraints, or -1 otherwise.
1618 */
1619static int ep_loop_check(struct eventpoll *ep, struct file *file)
1620{
Jason Baron28d82dc2012-01-12 17:17:43 -08001621 int ret;
1622 struct eventpoll *ep_cur, *ep_next;
1623
1624 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Davide Libenzi22bacca2011-02-25 14:44:12 -08001625 ep_loop_check_proc, file, ep, current);
Jason Baron28d82dc2012-01-12 17:17:43 -08001626 /* clear visited list */
1627 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1628 visited_list_link) {
1629 ep_cur->visited = 0;
1630 list_del(&ep_cur->visited_list_link);
1631 }
1632 return ret;
1633}
1634
1635static void clear_tfile_check_list(void)
1636{
1637 struct file *file;
1638
1639 /* first clear the tfile_check_list */
1640 while (!list_empty(&tfile_check_list)) {
1641 file = list_first_entry(&tfile_check_list, struct file,
1642 f_tfile_llink);
1643 list_del_init(&file->f_tfile_llink);
1644 }
1645 INIT_LIST_HEAD(&tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001646}
1647
Davide Libenzi7699acd2007-05-10 22:23:23 -07001648/*
Andrew Morton523723b2008-08-12 15:09:01 -07001649 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001650 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001651SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001652{
Jason Baron28d82dc2012-01-12 17:17:43 -08001653 int error, fd;
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001654 struct eventpoll *ep = NULL;
Jason Baron28d82dc2012-01-12 17:17:43 -08001655 struct file *file;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001656
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001657 /* Check the EPOLL_* constant for consistency. */
1658 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1659
Davide Libenzi296e2362009-03-31 15:24:11 -07001660 if (flags & ~EPOLL_CLOEXEC)
1661 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001662 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001663 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001664 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001665 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001666 if (error < 0)
1667 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001668 /*
1669 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001670 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001671 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001672 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1673 if (fd < 0) {
1674 error = fd;
1675 goto out_free_ep;
1676 }
1677 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001678 O_RDWR | (flags & O_CLOEXEC));
Jason Baron28d82dc2012-01-12 17:17:43 -08001679 if (IS_ERR(file)) {
1680 error = PTR_ERR(file);
1681 goto out_free_fd;
1682 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001683 ep->file = file;
Al Viro98022742012-08-17 22:42:36 -04001684 fd_install(fd, file);
Jason Baron28d82dc2012-01-12 17:17:43 -08001685 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001686
Jason Baron28d82dc2012-01-12 17:17:43 -08001687out_free_fd:
1688 put_unused_fd(fd);
1689out_free_ep:
1690 ep_free(ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001691 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001692}
1693
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001694SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001695{
Davide Libenzibfe38912009-05-12 13:19:44 -07001696 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001697 return -EINVAL;
1698
1699 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001700}
1701
Davide Libenzi7699acd2007-05-10 22:23:23 -07001702/*
1703 * The following function implements the controller interface for
1704 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001705 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001706 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001707SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1708 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001709{
1710 int error;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001711 int did_lock_epmutex = 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001712 struct file *file, *tfile;
1713 struct eventpoll *ep;
1714 struct epitem *epi;
1715 struct epoll_event epds;
1716
Davide Libenzi7699acd2007-05-10 22:23:23 -07001717 error = -EFAULT;
1718 if (ep_op_has_event(op) &&
1719 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1720 goto error_return;
1721
1722 /* Get the "struct file *" for the eventpoll file */
1723 error = -EBADF;
1724 file = fget(epfd);
1725 if (!file)
1726 goto error_return;
1727
1728 /* Get the "struct file *" for the target file */
1729 tfile = fget(fd);
1730 if (!tfile)
1731 goto error_fput;
1732
1733 /* The target file descriptor must support poll */
1734 error = -EPERM;
1735 if (!tfile->f_op || !tfile->f_op->poll)
1736 goto error_tgt_fput;
1737
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001738 /* Check if EPOLLWAKEUP is allowed */
Michael Kerriskd9914cf2012-07-17 21:37:27 +02001739 if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
Rafael J. Wysockia8159412012-05-21 21:20:48 +02001740 epds.events &= ~EPOLLWAKEUP;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001741
Davide Libenzi7699acd2007-05-10 22:23:23 -07001742 /*
1743 * We have to check that the file structure underneath the file descriptor
1744 * the user passed to us _is_ an eventpoll file. And also we do not permit
1745 * adding an epoll file descriptor inside itself.
1746 */
1747 error = -EINVAL;
1748 if (file == tfile || !is_file_epoll(file))
1749 goto error_tgt_fput;
1750
1751 /*
1752 * At this point it is safe to assume that the "private_data" contains
1753 * our own data structure.
1754 */
1755 ep = file->private_data;
1756
Davide Libenzi22bacca2011-02-25 14:44:12 -08001757 /*
1758 * When we insert an epoll file descriptor, inside another epoll file
1759 * descriptor, there is the change of creating closed loops, which are
Jason Baron28d82dc2012-01-12 17:17:43 -08001760 * better be handled here, than in more critical paths. While we are
1761 * checking for loops we also determine the list of files reachable
1762 * and hang them on the tfile_check_list, so we can check that we
1763 * haven't created too many possible wakeup paths.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001764 *
Jason Baron28d82dc2012-01-12 17:17:43 -08001765 * We need to hold the epmutex across both ep_insert and ep_remove
1766 * b/c we want to make sure we are looking at a coherent view of
1767 * epoll network.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001768 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001769 if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
Davide Libenzi22bacca2011-02-25 14:44:12 -08001770 mutex_lock(&epmutex);
1771 did_lock_epmutex = 1;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001772 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001773 if (op == EPOLL_CTL_ADD) {
1774 if (is_file_epoll(tfile)) {
1775 error = -ELOOP;
Jason Baron13d51802012-04-25 16:01:47 -07001776 if (ep_loop_check(ep, tfile) != 0) {
1777 clear_tfile_check_list();
Jason Baron28d82dc2012-01-12 17:17:43 -08001778 goto error_tgt_fput;
Jason Baron13d51802012-04-25 16:01:47 -07001779 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001780 } else
1781 list_add(&tfile->f_tfile_llink, &tfile_check_list);
1782 }
Davide Libenzi22bacca2011-02-25 14:44:12 -08001783
Nelson Elhaged8805e62011-10-31 17:13:14 -07001784 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001785
Davide Libenzi67647d02007-05-15 01:40:52 -07001786 /*
1787 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1788 * above, we can be sure to be able to use the item looked up by
1789 * ep_find() till we release the mutex.
1790 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001791 epi = ep_find(ep, tfile, fd);
1792
1793 error = -EINVAL;
1794 switch (op) {
1795 case EPOLL_CTL_ADD:
1796 if (!epi) {
1797 epds.events |= POLLERR | POLLHUP;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001798 error = ep_insert(ep, &epds, tfile, fd);
1799 } else
1800 error = -EEXIST;
Jason Baron28d82dc2012-01-12 17:17:43 -08001801 clear_tfile_check_list();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001802 break;
1803 case EPOLL_CTL_DEL:
1804 if (epi)
1805 error = ep_remove(ep, epi);
1806 else
1807 error = -ENOENT;
1808 break;
1809 case EPOLL_CTL_MOD:
1810 if (epi) {
1811 epds.events |= POLLERR | POLLHUP;
1812 error = ep_modify(ep, epi, &epds);
1813 } else
1814 error = -ENOENT;
1815 break;
Paton J. Lewis03a7beb2012-10-04 17:13:39 -07001816 case EPOLL_CTL_DISABLE:
1817 if (epi)
1818 error = ep_disable(ep, epi);
1819 else
1820 error = -ENOENT;
1821 break;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001822 }
Davide Libenzid47de162007-05-15 01:40:41 -07001823 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001824
1825error_tgt_fput:
Jason Baron28d82dc2012-01-12 17:17:43 -08001826 if (did_lock_epmutex)
Davide Libenzi22bacca2011-02-25 14:44:12 -08001827 mutex_unlock(&epmutex);
1828
Davide Libenzi7699acd2007-05-10 22:23:23 -07001829 fput(tfile);
1830error_fput:
1831 fput(file);
1832error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001833
1834 return error;
1835}
1836
1837/*
1838 * Implement the event wait interface for the eventpoll file. It is the kernel
1839 * part of the user space epoll_wait(2).
1840 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001841SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1842 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001843{
Al Viro2903ff02012-08-28 12:52:22 -04001844 int error;
1845 struct fd f;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001846 struct eventpoll *ep;
1847
Davide Libenzi7699acd2007-05-10 22:23:23 -07001848 /* The maximum number of event must be greater than zero */
1849 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1850 return -EINVAL;
1851
1852 /* Verify that the area passed by the user is writeable */
Al Viro2903ff02012-08-28 12:52:22 -04001853 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1854 return -EFAULT;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001855
1856 /* Get the "struct file *" for the eventpoll file */
Al Viro2903ff02012-08-28 12:52:22 -04001857 f = fdget(epfd);
1858 if (!f.file)
1859 return -EBADF;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001860
1861 /*
1862 * We have to check that the file structure underneath the fd
1863 * the user passed to us _is_ an eventpoll file.
1864 */
1865 error = -EINVAL;
Al Viro2903ff02012-08-28 12:52:22 -04001866 if (!is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07001867 goto error_fput;
1868
1869 /*
1870 * At this point it is safe to assume that the "private_data" contains
1871 * our own data structure.
1872 */
Al Viro2903ff02012-08-28 12:52:22 -04001873 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001874
1875 /* Time to fish for events ... */
1876 error = ep_poll(ep, events, maxevents, timeout);
1877
1878error_fput:
Al Viro2903ff02012-08-28 12:52:22 -04001879 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001880 return error;
1881}
1882
Davide Libenzi7699acd2007-05-10 22:23:23 -07001883/*
1884 * Implement the event wait interface for the eventpoll file. It is the kernel
1885 * part of the user space epoll_pwait(2).
1886 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001887SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1888 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1889 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001890{
1891 int error;
1892 sigset_t ksigmask, sigsaved;
1893
1894 /*
1895 * If the caller wants a certain signal mask to be set during the wait,
1896 * we apply it here.
1897 */
1898 if (sigmask) {
1899 if (sigsetsize != sizeof(sigset_t))
1900 return -EINVAL;
1901 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1902 return -EFAULT;
1903 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1904 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1905 }
1906
1907 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1908
1909 /*
1910 * If we changed the signal mask, we need to restore the original one.
1911 * In case we've got a signal while waiting, we do not restore the
1912 * signal mask yet, and we allow do_signal() to deliver the signal on
1913 * the way back to userspace, before the signal mask is restored.
1914 */
1915 if (sigmask) {
1916 if (error == -EINTR) {
1917 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001918 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001919 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001920 } else
1921 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1922 }
1923
1924 return error;
1925}
1926
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927static int __init eventpoll_init(void)
1928{
Davide Libenzi7ef99642008-12-01 13:13:55 -08001929 struct sysinfo si;
1930
1931 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08001932 /*
1933 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1934 */
1935 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08001936 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08001937 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
Davide Libenzi22bacca2011-02-25 14:44:12 -08001939 /*
1940 * Initialize the structure used to perform epoll file descriptor
1941 * inclusion loops checks.
1942 */
1943 ep_nested_calls_init(&poll_loop_ncalls);
1944
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07001946 ep_nested_calls_init(&poll_safewake_ncalls);
1947
1948 /* Initialize the structure used to perform file's f_op->poll() calls */
1949 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950
1951 /* Allocates slab cache used to allocate "struct epitem" items */
1952 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001953 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
1955 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1956 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001957 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960}
Davide Libenzicea69242007-05-10 22:23:22 -07001961fs_initcall(eventpoll_init);