blob: 473e09da7d02d3396273221f3094824c91f3b030 [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>
Colin Cross1c441e92013-05-06 23:50:16 +000037#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <asm/io.h>
40#include <asm/mman.h>
Arun Sharma600634972011-07-26 16:09:06 -070041#include <linux/atomic.h>
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -080042#include <linux/proc_fs.h>
43#include <linux/seq_file.h>
Al Viro35280bd2013-02-24 14:52:17 -050044#include <linux/compat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Linus Torvalds1da177e2005-04-16 15:20:36 -070046/*
47 * LOCKING:
48 * There are three level of locking required by epoll :
49 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080050 * 1) epmutex (mutex)
Davide Libenzic7ea7632007-05-15 01:40:47 -070051 * 2) ep->mtx (mutex)
52 * 3) ep->lock (spinlock)
Linus Torvalds1da177e2005-04-16 15:20:36 -070053 *
54 * The acquire order is the one listed above, from 1 to 3.
55 * We need a spinlock (ep->lock) because we manipulate objects
56 * from inside the poll callback, that might be triggered from
57 * a wake_up() that in turn might be called from IRQ context.
58 * So we can't sleep inside the poll callback and hence we need
59 * a spinlock. During the event transfer loop (from kernel to
60 * user space) we could end up sleeping due a copy_to_user(), so
61 * we need a lock that will allow us to sleep. This lock is a
Davide Libenzid47de162007-05-15 01:40:41 -070062 * mutex (ep->mtx). It is acquired during the event transfer loop,
63 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
64 * Then we also need a global mutex to serialize eventpoll_release_file()
65 * and ep_free().
66 * This mutex is acquired by ep_free() during the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -070067 * cleanup path and it is also acquired by eventpoll_release_file()
68 * if a file has been pushed inside an epoll set and it is then
Daniel Balutabf6a41d2011-01-30 23:42:29 +020069 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
Davide Libenzi22bacca2011-02-25 14:44:12 -080070 * It is also acquired when inserting an epoll fd onto another epoll
71 * fd. We do this so that we walk the epoll tree and ensure that this
72 * insertion does not create a cycle of epoll file descriptors, which
73 * could lead to deadlock. We need a global mutex to prevent two
74 * simultaneous inserts (A into B and B into A) from racing and
75 * constructing a cycle without either insert observing that it is
76 * going to.
Nelson Elhaged8805e62011-10-31 17:13:14 -070077 * It is necessary to acquire multiple "ep->mtx"es at once in the
78 * case when one epoll fd is added to another. In this case, we
79 * always acquire the locks in the order of nesting (i.e. after
80 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
81 * before e2->mtx). Since we disallow cycles of epoll file
82 * descriptors, this ensures that the mutexes are well-ordered. In
83 * order to communicate this nesting to lockdep, when walking a tree
84 * of epoll file descriptors, we use the current recursion depth as
85 * the lockdep subkey.
Davide Libenzid47de162007-05-15 01:40:41 -070086 * It is possible to drop the "ep->mtx" and to use the global
87 * mutex "epmutex" (together with "ep->lock") to have it working,
88 * but having "ep->mtx" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080089 * Events that require holding "epmutex" are very rare, while for
Davide Libenzid47de162007-05-15 01:40:41 -070090 * normal operations the epoll private "ep->mtx" will guarantee
91 * a better scalability.
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 */
93
Linus Torvalds1da177e2005-04-16 15:20:36 -070094/* Epoll private bits inside the event mask */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020095#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Davide Libenzi5071f972009-03-31 15:24:10 -070097/* Maximum number of nesting allowed inside epoll sets */
98#define EP_MAX_NESTS 4
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Davide Libenzib6119672006-10-11 01:21:44 -0700100#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
101
Davide Libenzid47de162007-05-15 01:40:41 -0700102#define EP_UNACTIVE_PTR ((void *) -1L)
103
Davide Libenzi7ef99642008-12-01 13:13:55 -0800104#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106struct epoll_filefd {
107 struct file *file;
108 int fd;
Eric Wong39732ca2013-04-30 15:27:38 -0700109} __packed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700112 * Structure used to track possible nested calls, for too deep recursions
113 * and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700115struct nested_call_node {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 struct list_head llink;
Davide Libenzi5071f972009-03-31 15:24:10 -0700117 void *cookie;
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700118 void *ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119};
120
121/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700122 * This structure is used as collector for nested calls, to check for
123 * maximum recursion dept and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700125struct nested_calls {
126 struct list_head tasks_call_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 spinlock_t lock;
128};
129
130/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700132 * have an entry of this type linked to the "rbr" RB tree.
Eric Wong39732ca2013-04-30 15:27:38 -0700133 * Avoid increasing the size of this struct, there can be many thousands
134 * of these on a server and we do not want this to take another cache line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 */
136struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700137 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 struct rb_node rbn;
139
140 /* List header used to link this structure to the eventpoll ready list */
141 struct list_head rdllink;
142
Davide Libenzic7ea7632007-05-15 01:40:47 -0700143 /*
144 * Works together "struct eventpoll"->ovflist in keeping the
145 * single linked chain of items.
146 */
147 struct epitem *next;
148
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 /* The file descriptor information this item refers to */
150 struct epoll_filefd ffd;
151
152 /* Number of active wait queue attached to poll operations */
153 int nwait;
154
155 /* List containing poll wait queues */
156 struct list_head pwqlist;
157
158 /* The "container" of this item */
159 struct eventpoll *ep;
160
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 /* List header used to link this item to the "struct file" items list */
162 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700163
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200164 /* wakeup_source used when EPOLLWAKEUP is set */
Eric Wongeea1d582013-04-30 15:27:39 -0700165 struct wakeup_source __rcu *ws;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200166
Davide Libenzic7ea7632007-05-15 01:40:47 -0700167 /* The structure that describe the interested events and the source fd */
168 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700169};
170
171/*
172 * This structure is stored inside the "private_data" member of the file
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200173 * structure and represents the main data structure for the eventpoll
Davide Libenzid47de162007-05-15 01:40:41 -0700174 * interface.
175 */
176struct eventpoll {
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200177 /* Protect the access to this structure */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700178 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700179
180 /*
181 * This mutex is used to ensure that files are not removed
182 * while epoll is using them. This is held during the event
183 * collection loop, the file cleanup path, the epoll file exit
184 * code and the ctl operations.
185 */
186 struct mutex mtx;
187
188 /* Wait queue used by sys_epoll_wait() */
189 wait_queue_head_t wq;
190
191 /* Wait queue used by file->poll() */
192 wait_queue_head_t poll_wait;
193
194 /* List of ready file descriptors */
195 struct list_head rdllist;
196
Davide Libenzi67647d02007-05-15 01:40:52 -0700197 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700198 struct rb_root rbr;
199
200 /*
201 * This is a single linked list that chains all the "struct epitem" that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300202 * happened while transferring ready events to userspace w/out
Davide Libenzid47de162007-05-15 01:40:41 -0700203 * holding ->lock.
204 */
205 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800206
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200207 /* wakeup_source used when ep_scan_ready_list is running */
208 struct wakeup_source *ws;
209
Davide Libenzi7ef99642008-12-01 13:13:55 -0800210 /* The user that created the eventpoll descriptor */
211 struct user_struct *user;
Jason Baron28d82dc2012-01-12 17:17:43 -0800212
213 struct file *file;
214
215 /* used to optimize loop detection check */
216 int visited;
217 struct list_head visited_list_link;
Davide Libenzid47de162007-05-15 01:40:41 -0700218};
219
220/* Wait structure used by the poll hooks */
221struct eppoll_entry {
222 /* List header used to link this structure to the "struct epitem" */
223 struct list_head llink;
224
225 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700226 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700227
228 /*
229 * Wait queue item that will be linked to the target file wait
230 * queue head.
231 */
232 wait_queue_t wait;
233
234 /* The wait queue head that linked the "wait" wait queue item */
235 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236};
237
238/* Wrapper struct used by poll queueing */
239struct ep_pqueue {
240 poll_table pt;
241 struct epitem *epi;
242};
243
Davide Libenzi5071f972009-03-31 15:24:10 -0700244/* Used by the ep_send_events() function as callback private data */
245struct ep_send_events_data {
246 int maxevents;
247 struct epoll_event __user *events;
248};
249
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800251 * Configuration options available inside /proc/sys/fs/epoll/
252 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800253/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800254static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800255
256/*
Davide Libenzid47de162007-05-15 01:40:41 -0700257 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800259static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
Davide Libenzi22bacca2011-02-25 14:44:12 -0800261/* Used to check for epoll file descriptor inclusion loops */
262static struct nested_calls poll_loop_ncalls;
263
Davide Libenzi5071f972009-03-31 15:24:10 -0700264/* Used for safe wake up implementation */
265static struct nested_calls poll_safewake_ncalls;
266
267/* Used to call file's f_op->poll() under the nested calls boundaries */
268static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800271static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272
273/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800274static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275
Jason Baron28d82dc2012-01-12 17:17:43 -0800276/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
277static LIST_HEAD(visited_list);
278
279/*
280 * List of files with newly added links, where we may need to limit the number
281 * of emanating paths. Protected by the epmutex.
282 */
283static LIST_HEAD(tfile_check_list);
284
Davide Libenzi7ef99642008-12-01 13:13:55 -0800285#ifdef CONFIG_SYSCTL
286
287#include <linux/sysctl.h>
288
Robin Holt52bd19f72011-01-12 17:00:01 -0800289static long zero;
290static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800291
292ctl_table epoll_table[] = {
293 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800294 .procname = "max_user_watches",
295 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800296 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800297 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800298 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800299 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800300 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800301 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800302 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800303};
304#endif /* CONFIG_SYSCTL */
305
Jason Baron28d82dc2012-01-12 17:17:43 -0800306static const struct file_operations eventpoll_fops;
307
308static inline int is_file_epoll(struct file *f)
309{
310 return f->f_op == &eventpoll_fops;
311}
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700312
Davide Libenzi67647d02007-05-15 01:40:52 -0700313/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700314static inline void ep_set_ffd(struct epoll_filefd *ffd,
315 struct file *file, int fd)
316{
317 ffd->file = file;
318 ffd->fd = fd;
319}
320
Davide Libenzi67647d02007-05-15 01:40:52 -0700321/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700322static inline int ep_cmp_ffd(struct epoll_filefd *p1,
323 struct epoll_filefd *p2)
324{
325 return (p1->file > p2->file ? +1:
326 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
327}
328
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700329/* Tells us if the item is currently linked */
330static inline int ep_is_linked(struct list_head *p)
331{
332 return !list_empty(p);
333}
334
Oleg Nesterov971316f2012-02-24 20:07:29 +0100335static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
336{
337 return container_of(p, struct eppoll_entry, wait);
338}
339
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700340/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700341static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700342{
343 return container_of(p, struct eppoll_entry, wait)->base;
344}
345
346/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700347static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700348{
349 return container_of(p, struct ep_pqueue, pt)->epi;
350}
351
352/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700353static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700354{
Andrew Mortona80a6b82012-11-08 15:53:35 -0800355 return op != EPOLL_CTL_DEL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700356}
357
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700359static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360{
Davide Libenzi5071f972009-03-31 15:24:10 -0700361 INIT_LIST_HEAD(&ncalls->tasks_call_list);
362 spin_lock_init(&ncalls->lock);
363}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364
Davide Libenzi5071f972009-03-31 15:24:10 -0700365/**
Davide Libenzi3fb0e582011-03-22 16:34:46 -0700366 * ep_events_available - Checks if ready events might be available.
367 *
368 * @ep: Pointer to the eventpoll context.
369 *
370 * Returns: Returns a value different than zero if ready events are available,
371 * or zero otherwise.
372 */
373static inline int ep_events_available(struct eventpoll *ep)
374{
375 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
376}
377
378/**
Davide Libenzi5071f972009-03-31 15:24:10 -0700379 * ep_call_nested - Perform a bound (possibly) nested call, by checking
380 * that the recursion limit is not exceeded, and that
381 * the same nested call (by the meaning of same cookie) is
382 * no re-entered.
383 *
384 * @ncalls: Pointer to the nested_calls structure to be used for this call.
385 * @max_nests: Maximum number of allowed nesting calls.
386 * @nproc: Nested call core function pointer.
387 * @priv: Opaque data to be passed to the @nproc callback.
388 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700389 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700390 *
391 * Returns: Returns the code returned by the @nproc callback, or -1 if
392 * the maximum recursion limit has been exceeded.
393 */
394static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
395 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700396 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700397{
398 int error, call_nests = 0;
399 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700400 struct list_head *lsthead = &ncalls->tasks_call_list;
401 struct nested_call_node *tncur;
402 struct nested_call_node tnode;
403
404 spin_lock_irqsave(&ncalls->lock, flags);
405
406 /*
407 * Try to see if the current task is already inside this wakeup call.
408 * We use a list here, since the population inside this set is always
409 * very much limited.
410 */
411 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700412 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700413 (tncur->cookie == cookie || ++call_nests > max_nests)) {
414 /*
415 * Ops ... loop detected or maximum nest level reached.
416 * We abort this wake by breaking the cycle itself.
417 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700418 error = -1;
419 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700420 }
421 }
422
423 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700424 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700425 tnode.cookie = cookie;
426 list_add(&tnode.llink, lsthead);
427
428 spin_unlock_irqrestore(&ncalls->lock, flags);
429
430 /* Call the nested function */
431 error = (*nproc)(priv, cookie, call_nests);
432
433 /* Remove the current task from the list */
434 spin_lock_irqsave(&ncalls->lock, flags);
435 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700436out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700437 spin_unlock_irqrestore(&ncalls->lock, flags);
438
439 return error;
440}
441
Steven Rostedt02edc6f2012-03-23 15:02:27 -0700442/*
443 * As described in commit 0ccf831cb lockdep: annotate epoll
444 * the use of wait queues used by epoll is done in a very controlled
445 * manner. Wake ups can nest inside each other, but are never done
446 * with the same locking. For example:
447 *
448 * dfd = socket(...);
449 * efd1 = epoll_create();
450 * efd2 = epoll_create();
451 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
452 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
453 *
454 * When a packet arrives to the device underneath "dfd", the net code will
455 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
456 * callback wakeup entry on that queue, and the wake_up() performed by the
457 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
458 * (efd1) notices that it may have some event ready, so it needs to wake up
459 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
460 * that ends up in another wake_up(), after having checked about the
461 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
462 * avoid stack blasting.
463 *
464 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
465 * this special case of epoll.
466 */
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700467#ifdef CONFIG_DEBUG_LOCK_ALLOC
468static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
469 unsigned long events, int subclass)
470{
471 unsigned long flags;
472
473 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
474 wake_up_locked_poll(wqueue, events);
475 spin_unlock_irqrestore(&wqueue->lock, flags);
476}
477#else
478static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
479 unsigned long events, int subclass)
480{
481 wake_up_poll(wqueue, events);
482}
483#endif
484
Davide Libenzi5071f972009-03-31 15:24:10 -0700485static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
486{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700487 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
488 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700489 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490}
491
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492/*
493 * Perform a safe wake up of the poll wait list. The problem is that
494 * with the new callback'd wake up system, it is possible that the
495 * poll callback is reentered from inside the call to wake_up() done
496 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700497 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 * and we cannot reenter the same wait queue head at all. This will
499 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700500 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700502static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700504 int this_cpu = get_cpu();
505
Davide Libenzi5071f972009-03-31 15:24:10 -0700506 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700507 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
508
509 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510}
511
Oleg Nesterov971316f2012-02-24 20:07:29 +0100512static void ep_remove_wait_queue(struct eppoll_entry *pwq)
513{
514 wait_queue_head_t *whead;
515
516 rcu_read_lock();
517 /* If it is cleared by POLLFREE, it should be rcu-safe */
518 whead = rcu_dereference(pwq->whead);
519 if (whead)
520 remove_wait_queue(whead, &pwq->wait);
521 rcu_read_unlock();
522}
523
Davide Libenzi7699acd2007-05-10 22:23:23 -0700524/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700525 * This function unregisters poll callbacks from the associated file
526 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
527 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700528 */
529static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
530{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700531 struct list_head *lsthead = &epi->pwqlist;
532 struct eppoll_entry *pwq;
533
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700534 while (!list_empty(lsthead)) {
535 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700536
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700537 list_del(&pwq->llink);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100538 ep_remove_wait_queue(pwq);
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700539 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700540 }
541}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542
Eric Wongeea1d582013-04-30 15:27:39 -0700543/* call only when ep->mtx is held */
544static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
545{
546 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
547}
548
549/* call only when ep->mtx is held */
550static inline void ep_pm_stay_awake(struct epitem *epi)
551{
552 struct wakeup_source *ws = ep_wakeup_source(epi);
553
554 if (ws)
555 __pm_stay_awake(ws);
556}
557
558static inline bool ep_has_wakeup_source(struct epitem *epi)
559{
560 return rcu_access_pointer(epi->ws) ? true : false;
561}
562
563/* call when ep->mtx cannot be held (ep_poll_callback) */
564static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
565{
566 struct wakeup_source *ws;
567
568 rcu_read_lock();
569 ws = rcu_dereference(epi->ws);
570 if (ws)
571 __pm_stay_awake(ws);
572 rcu_read_unlock();
573}
574
Davide Libenzi5071f972009-03-31 15:24:10 -0700575/**
576 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
577 * the scan code, to call f_op->poll(). Also allows for
578 * O(NumReady) performance.
579 *
580 * @ep: Pointer to the epoll private data structure.
581 * @sproc: Pointer to the scan callback.
582 * @priv: Private opaque data passed to the @sproc callback.
Nelson Elhaged8805e62011-10-31 17:13:14 -0700583 * @depth: The current depth of recursive f_op->poll calls.
Davide Libenzi5071f972009-03-31 15:24:10 -0700584 *
585 * Returns: The same integer error code returned by the @sproc callback.
586 */
587static int ep_scan_ready_list(struct eventpoll *ep,
588 int (*sproc)(struct eventpoll *,
589 struct list_head *, void *),
Nelson Elhaged8805e62011-10-31 17:13:14 -0700590 void *priv,
591 int depth)
Davide Libenzi5071f972009-03-31 15:24:10 -0700592{
593 int error, pwake = 0;
594 unsigned long flags;
595 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700596 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700597
598 /*
599 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700600 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700601 */
Nelson Elhaged8805e62011-10-31 17:13:14 -0700602 mutex_lock_nested(&ep->mtx, depth);
Davide Libenzi5071f972009-03-31 15:24:10 -0700603
604 /*
605 * Steal the ready list, and re-init the original one to the
606 * empty list. Also, set ep->ovflist to NULL so that events
607 * happening while looping w/out locks, are not lost. We cannot
608 * have the poll callback to queue directly on ep->rdllist,
609 * because we want the "sproc" callback to be able to do it
610 * in a lockless way.
611 */
612 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700613 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700614 ep->ovflist = NULL;
615 spin_unlock_irqrestore(&ep->lock, flags);
616
617 /*
618 * Now call the callback function.
619 */
620 error = (*sproc)(ep, &txlist, priv);
621
622 spin_lock_irqsave(&ep->lock, flags);
623 /*
624 * During the time we spent inside the "sproc" callback, some
625 * other events might have been queued by the poll callback.
626 * We re-insert them inside the main ready-list here.
627 */
628 for (nepi = ep->ovflist; (epi = nepi) != NULL;
629 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
630 /*
631 * We need to check if the item is already in the list.
632 * During the "sproc" callback execution time, items are
633 * queued into ->ovflist but the "txlist" might already
634 * contain them, and the list_splice() below takes care of them.
635 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200636 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700637 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -0700638 ep_pm_stay_awake(epi);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200639 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700640 }
641 /*
642 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
643 * releasing the lock, events will be queued in the normal way inside
644 * ep->rdllist.
645 */
646 ep->ovflist = EP_UNACTIVE_PTR;
647
648 /*
649 * Quickly re-inject items left on "txlist".
650 */
651 list_splice(&txlist, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200652 __pm_relax(ep->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700653
654 if (!list_empty(&ep->rdllist)) {
655 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700656 * Wake up (if active) both the eventpoll wait list and
657 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700658 */
659 if (waitqueue_active(&ep->wq))
660 wake_up_locked(&ep->wq);
661 if (waitqueue_active(&ep->poll_wait))
662 pwake++;
663 }
664 spin_unlock_irqrestore(&ep->lock, flags);
665
666 mutex_unlock(&ep->mtx);
667
668 /* We have to call this outside the lock */
669 if (pwake)
670 ep_poll_safewake(&ep->poll_wait);
671
672 return error;
673}
674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700676 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700677 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700678 */
679static int ep_remove(struct eventpoll *ep, struct epitem *epi)
680{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700681 unsigned long flags;
682 struct file *file = epi->ffd.file;
683
684 /*
685 * Removes poll wait queue hooks. We _have_ to do this without holding
686 * the "ep->lock" otherwise a deadlock might occur. This because of the
687 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
688 * queue head lock when unregistering the wait queue. The wakeup callback
689 * will run by holding the wait queue head lock and will call our callback
690 * that will try to get "ep->lock".
691 */
692 ep_unregister_pollwait(ep, epi);
693
694 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700695 spin_lock(&file->f_lock);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700696 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700697 list_del_init(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700698 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
Davide Libenzicdac75e2008-04-29 00:58:34 -0700700 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701
Davide Libenzic7ea7632007-05-15 01:40:47 -0700702 spin_lock_irqsave(&ep->lock, flags);
703 if (ep_is_linked(&epi->rdllink))
704 list_del_init(&epi->rdllink);
705 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
Eric Wongeea1d582013-04-30 15:27:39 -0700707 wakeup_source_unregister(ep_wakeup_source(epi));
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200708
Davide Libenzi7699acd2007-05-10 22:23:23 -0700709 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700710 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711
Robin Holt52bd19f72011-01-12 17:00:01 -0800712 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800713
Davide Libenzic7ea7632007-05-15 01:40:47 -0700714 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715}
716
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717static void ep_free(struct eventpoll *ep)
718{
719 struct rb_node *rbp;
720 struct epitem *epi;
721
722 /* We need to release all tasks waiting for these file */
723 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700724 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725
726 /*
727 * We need to lock this because we could be hit by
728 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700729 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 * is on the way to be removed and no one has references to it
731 * anymore. The only hit might come from eventpoll_release_file() but
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300732 * holding "epmutex" is sufficient here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800734 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
736 /*
737 * Walks through the whole tree by unregistering poll callbacks.
738 */
739 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
740 epi = rb_entry(rbp, struct epitem, rbn);
741
742 ep_unregister_pollwait(ep, epi);
Eric Dumazet91cf5ab2013-09-11 14:24:06 -0700743 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 }
745
746 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700747 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700749 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 * us during this operation. So we can avoid the lock on "ep->lock".
Eric Wongddf676c2013-04-30 15:27:40 -0700751 * We do not need to lock ep->mtx, either, we only do it to prevent
752 * a lockdep warning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 */
Eric Wongddf676c2013-04-30 15:27:40 -0700754 mutex_lock(&ep->mtx);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700755 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 epi = rb_entry(rbp, struct epitem, rbn);
757 ep_remove(ep, epi);
Eric Dumazet91cf5ab2013-09-11 14:24:06 -0700758 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 }
Eric Wongddf676c2013-04-30 15:27:40 -0700760 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761
Arjan van de Ven144efe32006-03-23 03:00:32 -0800762 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700763 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800764 free_uid(ep->user);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200765 wakeup_source_unregister(ep->ws);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700766 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767}
768
Davide Libenzi7699acd2007-05-10 22:23:23 -0700769static int ep_eventpoll_release(struct inode *inode, struct file *file)
770{
771 struct eventpoll *ep = file->private_data;
772
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700773 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700774 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700775
Davide Libenzi7699acd2007-05-10 22:23:23 -0700776 return 0;
777}
778
Eric Wong450d89e2013-04-30 15:27:42 -0700779static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
780{
781 pt->_key = epi->event.events;
782
783 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
784}
785
Davide Libenzi296e2362009-03-31 15:24:11 -0700786static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
787 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700788{
789 struct epitem *epi, *tmp;
Hans Verkuil626cf232012-03-23 15:02:27 -0700790 poll_table pt;
Davide Libenzi5071f972009-03-31 15:24:10 -0700791
Hans Verkuil626cf232012-03-23 15:02:27 -0700792 init_poll_funcptr(&pt, NULL);
Eric Wong450d89e2013-04-30 15:27:42 -0700793
Davide Libenzi5071f972009-03-31 15:24:10 -0700794 list_for_each_entry_safe(epi, tmp, head, rdllink) {
Eric Wong450d89e2013-04-30 15:27:42 -0700795 if (ep_item_poll(epi, &pt))
Davide Libenzi5071f972009-03-31 15:24:10 -0700796 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700797 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700798 /*
799 * Item has been dropped into the ready list by the poll
800 * callback, but it's not actually ready, as far as
801 * caller requested events goes. We can remove it here.
802 */
Eric Wongeea1d582013-04-30 15:27:39 -0700803 __pm_relax(ep_wakeup_source(epi));
Davide Libenzi5071f972009-03-31 15:24:10 -0700804 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700805 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700806 }
807
808 return 0;
809}
810
811static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
812{
Nelson Elhaged8805e62011-10-31 17:13:14 -0700813 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
Davide Libenzi5071f972009-03-31 15:24:10 -0700814}
815
Davide Libenzi7699acd2007-05-10 22:23:23 -0700816static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
817{
Davide Libenzi5071f972009-03-31 15:24:10 -0700818 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700819 struct eventpoll *ep = file->private_data;
820
821 /* Insert inside our poll wait queue */
822 poll_wait(file, &ep->poll_wait, wait);
823
Davide Libenzi5071f972009-03-31 15:24:10 -0700824 /*
825 * Proceed to find out if wanted events are really available inside
826 * the ready list. This need to be done under ep_call_nested()
827 * supervision, since the call to f_op->poll() done on listed files
828 * could re-enter here.
829 */
830 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700831 ep_poll_readyevents_proc, ep, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700832
Davide Libenzi296e2362009-03-31 15:24:11 -0700833 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700834}
835
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800836#ifdef CONFIG_PROC_FS
837static int ep_show_fdinfo(struct seq_file *m, struct file *f)
838{
839 struct eventpoll *ep = f->private_data;
840 struct rb_node *rbp;
841 int ret = 0;
842
843 mutex_lock(&ep->mtx);
844 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
845 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
846
847 ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
848 epi->ffd.fd, epi->event.events,
849 (long long)epi->event.data);
850 if (ret)
851 break;
852 }
853 mutex_unlock(&ep->mtx);
854
855 return ret;
856}
857#endif
858
Davide Libenzi7699acd2007-05-10 22:23:23 -0700859/* File callbacks that implement the eventpoll file behaviour */
860static const struct file_operations eventpoll_fops = {
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800861#ifdef CONFIG_PROC_FS
862 .show_fdinfo = ep_show_fdinfo,
863#endif
Davide Libenzi7699acd2007-05-10 22:23:23 -0700864 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200865 .poll = ep_eventpoll_poll,
866 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700867};
868
Davide Libenzi7699acd2007-05-10 22:23:23 -0700869/*
870 * This is called from eventpoll_release() to unlink files from the eventpoll
871 * interface. We need to have this facility to cleanup correctly files that are
872 * closed without being removed from the eventpoll interface.
873 */
874void eventpoll_release_file(struct file *file)
875{
876 struct list_head *lsthead = &file->f_ep_links;
877 struct eventpoll *ep;
878 struct epitem *epi;
879
880 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700881 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700882 * necessary. It is not necessary because we're in the "struct file"
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300883 * cleanup path, and this means that no one is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700884 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700885 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700886 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700887 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700888 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700889 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700890 *
891 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700892 */
893 mutex_lock(&epmutex);
894
895 while (!list_empty(lsthead)) {
896 epi = list_first_entry(lsthead, struct epitem, fllink);
897
898 ep = epi->ep;
899 list_del_init(&epi->fllink);
Nelson Elhaged8805e62011-10-31 17:13:14 -0700900 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700901 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700902 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700903 }
904
905 mutex_unlock(&epmutex);
906}
907
908static int ep_alloc(struct eventpoll **pep)
909{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800910 int error;
911 struct user_struct *user;
912 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700913
Davide Libenzi7ef99642008-12-01 13:13:55 -0800914 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800915 error = -ENOMEM;
916 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
917 if (unlikely(!ep))
918 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700919
Davide Libenzic7ea7632007-05-15 01:40:47 -0700920 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700921 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700922 init_waitqueue_head(&ep->wq);
923 init_waitqueue_head(&ep->poll_wait);
924 INIT_LIST_HEAD(&ep->rdllist);
925 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700926 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800927 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700928
929 *pep = ep;
930
Davide Libenzi7699acd2007-05-10 22:23:23 -0700931 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800932
933free_uid:
934 free_uid(user);
935 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700936}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
938/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700939 * Search the file inside the eventpoll tree. The RB tree operations
940 * are protected by the "mtx" mutex, and ep_find() must be called with
941 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 */
943static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
944{
945 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 struct rb_node *rbp;
947 struct epitem *epi, *epir = NULL;
948 struct epoll_filefd ffd;
949
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700950 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 for (rbp = ep->rbr.rb_node; rbp; ) {
952 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700953 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 if (kcmp > 0)
955 rbp = rbp->rb_right;
956 else if (kcmp < 0)
957 rbp = rbp->rb_left;
958 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 epir = epi;
960 break;
961 }
962 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 return epir;
965}
966
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700968 * This is the callback that is passed to the wait queue wakeup
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200969 * mechanism. It is called by the stored file descriptors when they
Davide Libenzi7699acd2007-05-10 22:23:23 -0700970 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700972static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700974 int pwake = 0;
975 unsigned long flags;
976 struct epitem *epi = ep_item_from_wait(wait);
977 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978
Oleg Nesterov971316f2012-02-24 20:07:29 +0100979 if ((unsigned long)key & POLLFREE) {
980 ep_pwq_from_wait(wait)->whead = NULL;
981 /*
982 * whead = NULL above can race with ep_remove_wait_queue()
983 * which can do another remove_wait_queue() after us, so we
984 * can't use __remove_wait_queue(). whead->lock is held by
985 * the caller.
986 */
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100987 list_del_init(&wait->task_list);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100988 }
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100989
Davide Libenzic7ea7632007-05-15 01:40:47 -0700990 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700991
992 /*
993 * If the event mask does not contain any poll(2) event, we consider the
994 * descriptor to be disabled. This condition is likely the effect of the
995 * EPOLLONESHOT bit that disables the descriptor when an event is received,
996 * until the next EPOLL_CTL_MOD will be issued.
997 */
998 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700999 goto out_unlock;
1000
1001 /*
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -07001002 * Check the events coming with the callback. At this stage, not
1003 * every device reports the events in the "key" parameter of the
1004 * callback. We need to be able to handle both cases here, hence the
1005 * test for "key" != NULL before the event match test.
1006 */
1007 if (key && !((unsigned long) key & epi->event.events))
1008 goto out_unlock;
1009
1010 /*
Daniel Balutabf6a41d2011-01-30 23:42:29 +02001011 * If we are transferring events to userspace, we can hold no locks
Davide Libenzid47de162007-05-15 01:40:41 -07001012 * (because we're accessing user memory, and because of linux f_op->poll()
Daniel Balutabf6a41d2011-01-30 23:42:29 +02001013 * semantics). All the events that happen during that period of time are
Davide Libenzid47de162007-05-15 01:40:41 -07001014 * chained in ep->ovflist and requeued later on.
1015 */
1016 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1017 if (epi->next == EP_UNACTIVE_PTR) {
1018 epi->next = ep->ovflist;
1019 ep->ovflist = epi;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001020 if (epi->ws) {
1021 /*
1022 * Activate ep->ws since epi->ws may get
1023 * deactivated at any time.
1024 */
1025 __pm_stay_awake(ep->ws);
1026 }
1027
Davide Libenzid47de162007-05-15 01:40:41 -07001028 }
1029 goto out_unlock;
1030 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001031
1032 /* If this file is already in the ready list we exit soon */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001033 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001034 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001035 ep_pm_stay_awake_rcu(epi);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001036 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001037
Davide Libenzi7699acd2007-05-10 22:23:23 -07001038 /*
1039 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1040 * wait list.
1041 */
1042 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -04001043 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001044 if (waitqueue_active(&ep->poll_wait))
1045 pwake++;
1046
Davide Libenzid47de162007-05-15 01:40:41 -07001047out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001048 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001049
1050 /* We have to call this outside the lock */
1051 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001052 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001053
1054 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055}
1056
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057/*
1058 * This is the callback that is used to add our wait queue to the
1059 * target file wakeup lists.
1060 */
1061static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1062 poll_table *pt)
1063{
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001064 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 struct eppoll_entry *pwq;
1066
Christoph Lametere94b1762006-12-06 20:33:17 -08001067 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1069 pwq->whead = whead;
1070 pwq->base = epi;
1071 add_wait_queue(whead, &pwq->wait);
1072 list_add_tail(&pwq->llink, &epi->pwqlist);
1073 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001074 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 /* We have to signal that an error occurred */
1076 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -07001077 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078}
1079
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1081{
1082 int kcmp;
1083 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1084 struct epitem *epic;
1085
1086 while (*p) {
1087 parent = *p;
1088 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001089 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 if (kcmp > 0)
1091 p = &parent->rb_right;
1092 else
1093 p = &parent->rb_left;
1094 }
1095 rb_link_node(&epi->rbn, parent, p);
1096 rb_insert_color(&epi->rbn, &ep->rbr);
1097}
1098
Andrew Mortona80a6b82012-11-08 15:53:35 -08001099
1100
Jason Baron28d82dc2012-01-12 17:17:43 -08001101#define PATH_ARR_SIZE 5
1102/*
1103 * These are the number paths of length 1 to 5, that we are allowing to emanate
1104 * from a single file of interest. For example, we allow 1000 paths of length
1105 * 1, to emanate from each file of interest. This essentially represents the
1106 * potential wakeup paths, which need to be limited in order to avoid massive
1107 * uncontrolled wakeup storms. The common use case should be a single ep which
1108 * is connected to n file sources. In this case each file source has 1 path
1109 * of length 1. Thus, the numbers below should be more than sufficient. These
1110 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1111 * and delete can't add additional paths. Protected by the epmutex.
1112 */
1113static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1114static int path_count[PATH_ARR_SIZE];
1115
1116static int path_count_inc(int nests)
1117{
Jason Baron93dc6102012-03-16 16:34:03 -04001118 /* Allow an arbitrary number of depth 1 paths */
1119 if (nests == 0)
1120 return 0;
1121
Jason Baron28d82dc2012-01-12 17:17:43 -08001122 if (++path_count[nests] > path_limits[nests])
1123 return -1;
1124 return 0;
1125}
1126
1127static void path_count_init(void)
1128{
1129 int i;
1130
1131 for (i = 0; i < PATH_ARR_SIZE; i++)
1132 path_count[i] = 0;
1133}
1134
1135static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1136{
1137 int error = 0;
1138 struct file *file = priv;
1139 struct file *child_file;
1140 struct epitem *epi;
1141
1142 list_for_each_entry(epi, &file->f_ep_links, fllink) {
1143 child_file = epi->ep->file;
1144 if (is_file_epoll(child_file)) {
1145 if (list_empty(&child_file->f_ep_links)) {
1146 if (path_count_inc(call_nests)) {
1147 error = -1;
1148 break;
1149 }
1150 } else {
1151 error = ep_call_nested(&poll_loop_ncalls,
1152 EP_MAX_NESTS,
1153 reverse_path_check_proc,
1154 child_file, child_file,
1155 current);
1156 }
1157 if (error != 0)
1158 break;
1159 } else {
1160 printk(KERN_ERR "reverse_path_check_proc: "
1161 "file is not an ep!\n");
1162 }
1163 }
1164 return error;
1165}
1166
1167/**
1168 * reverse_path_check - The tfile_check_list is list of file *, which have
1169 * links that are proposed to be newly added. We need to
1170 * make sure that those added links don't add too many
1171 * paths such that we will spend all our time waking up
1172 * eventpoll objects.
1173 *
1174 * Returns: Returns zero if the proposed links don't create too many paths,
1175 * -1 otherwise.
1176 */
1177static int reverse_path_check(void)
1178{
Jason Baron28d82dc2012-01-12 17:17:43 -08001179 int error = 0;
1180 struct file *current_file;
1181
1182 /* let's call this for all tfiles */
1183 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001184 path_count_init();
1185 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1186 reverse_path_check_proc, current_file,
1187 current_file, current);
1188 if (error)
1189 break;
1190 }
1191 return error;
1192}
1193
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001194static int ep_create_wakeup_source(struct epitem *epi)
1195{
1196 const char *name;
Eric Wongeea1d582013-04-30 15:27:39 -07001197 struct wakeup_source *ws;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001198
1199 if (!epi->ep->ws) {
1200 epi->ep->ws = wakeup_source_register("eventpoll");
1201 if (!epi->ep->ws)
1202 return -ENOMEM;
1203 }
1204
1205 name = epi->ffd.file->f_path.dentry->d_name.name;
Eric Wongeea1d582013-04-30 15:27:39 -07001206 ws = wakeup_source_register(name);
1207
1208 if (!ws)
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001209 return -ENOMEM;
Eric Wongeea1d582013-04-30 15:27:39 -07001210 rcu_assign_pointer(epi->ws, ws);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001211
1212 return 0;
1213}
1214
Eric Wongeea1d582013-04-30 15:27:39 -07001215/* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1216static noinline void ep_destroy_wakeup_source(struct epitem *epi)
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001217{
Eric Wongeea1d582013-04-30 15:27:39 -07001218 struct wakeup_source *ws = ep_wakeup_source(epi);
1219
Eric Wongd6d67e72013-04-30 15:27:43 -07001220 RCU_INIT_POINTER(epi->ws, NULL);
Eric Wongeea1d582013-04-30 15:27:39 -07001221
1222 /*
1223 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1224 * used internally by wakeup_source_remove, too (called by
1225 * wakeup_source_unregister), so we cannot use call_rcu
1226 */
1227 synchronize_rcu();
1228 wakeup_source_unregister(ws);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001229}
1230
Davide Libenzic7ea7632007-05-15 01:40:47 -07001231/*
1232 * Must be called with "mtx" held.
1233 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1235 struct file *tfile, int fd)
1236{
1237 int error, revents, pwake = 0;
1238 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -08001239 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 struct epitem *epi;
1241 struct ep_pqueue epq;
1242
Robin Holt52bd19f72011-01-12 17:00:01 -08001243 user_watches = atomic_long_read(&ep->user->epoll_watches);
1244 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001245 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -08001246 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001247 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
1249 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 INIT_LIST_HEAD(&epi->rdllink);
1251 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 INIT_LIST_HEAD(&epi->pwqlist);
1253 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001254 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -07001257 epi->next = EP_UNACTIVE_PTR;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001258 if (epi->event.events & EPOLLWAKEUP) {
1259 error = ep_create_wakeup_source(epi);
1260 if (error)
1261 goto error_create_wakeup_source;
1262 } else {
Eric Wongeea1d582013-04-30 15:27:39 -07001263 RCU_INIT_POINTER(epi->ws, NULL);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001264 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265
1266 /* Initialize the poll table using the queue callback */
1267 epq.epi = epi;
1268 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1269
1270 /*
1271 * Attach the item to the poll hooks and get current event bits.
1272 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -07001273 * been increased by the caller of this function. Note that after
1274 * this operation completes, the poll callback can start hitting
1275 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 */
Eric Wong450d89e2013-04-30 15:27:42 -07001277 revents = ep_item_poll(epi, &epq.pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278
1279 /*
1280 * We have to check if something went wrong during the poll wait queue
1281 * install process. Namely an allocation for a wait queue failed due
1282 * high memory pressure.
1283 */
Davide Libenzi7ef99642008-12-01 13:13:55 -08001284 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001286 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
1288 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -07001289 spin_lock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 list_add_tail(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -07001291 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292
Davide Libenzic7ea7632007-05-15 01:40:47 -07001293 /*
1294 * Add the current item to the RB tree. All RB tree operations are
1295 * protected by "mtx", and ep_insert() is called with "mtx" held.
1296 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 ep_rbtree_insert(ep, epi);
1298
Jason Baron28d82dc2012-01-12 17:17:43 -08001299 /* now check if we've created too many backpaths */
1300 error = -EINVAL;
1301 if (reverse_path_check())
1302 goto error_remove_epi;
1303
Davide Libenzic7ea7632007-05-15 01:40:47 -07001304 /* We have to drop the new item inside our item list to keep track of it */
1305 spin_lock_irqsave(&ep->lock, flags);
1306
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001308 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001310 ep_pm_stay_awake(epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311
1312 /* Notify waiting tasks that events are available */
1313 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -04001314 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 if (waitqueue_active(&ep->poll_wait))
1316 pwake++;
1317 }
1318
Davide Libenzic7ea7632007-05-15 01:40:47 -07001319 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
Robin Holt52bd19f72011-01-12 17:00:01 -08001321 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001322
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 /* We have to call this outside the lock */
1324 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001325 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 return 0;
1328
Jason Baron28d82dc2012-01-12 17:17:43 -08001329error_remove_epi:
1330 spin_lock(&tfile->f_lock);
1331 if (ep_is_linked(&epi->fllink))
1332 list_del_init(&epi->fllink);
1333 spin_unlock(&tfile->f_lock);
1334
1335 rb_erase(&epi->rbn, &ep->rbr);
1336
Davide Libenzi7699acd2007-05-10 22:23:23 -07001337error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 ep_unregister_pollwait(ep, epi);
1339
1340 /*
1341 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -07001342 * allocated wait queue. Note that we don't care about the ep->ovflist
1343 * list, since that is used/cleaned only inside a section bound by "mtx".
1344 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001346 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001347 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001348 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001349 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350
Eric Wongeea1d582013-04-30 15:27:39 -07001351 wakeup_source_unregister(ep_wakeup_source(epi));
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001352
1353error_create_wakeup_source:
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001354 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001355
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 return error;
1357}
1358
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359/*
1360 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001361 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 */
1363static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1364{
1365 int pwake = 0;
1366 unsigned int revents;
Hans Verkuil626cf232012-03-23 15:02:27 -07001367 poll_table pt;
1368
1369 init_poll_funcptr(&pt, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370
1371 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001372 * Set the new event interest mask before calling f_op->poll();
1373 * otherwise we might miss an event that happens between the
1374 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 */
Eric Wong128dd172013-01-01 21:20:27 +00001376 epi->event.events = event->events; /* need barrier below */
Tony Battersbye057e152009-03-31 15:24:15 -07001377 epi->event.data = event->data; /* protected by mtx */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001378 if (epi->event.events & EPOLLWAKEUP) {
Eric Wongeea1d582013-04-30 15:27:39 -07001379 if (!ep_has_wakeup_source(epi))
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001380 ep_create_wakeup_source(epi);
Eric Wongeea1d582013-04-30 15:27:39 -07001381 } else if (ep_has_wakeup_source(epi)) {
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001382 ep_destroy_wakeup_source(epi);
1383 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
1385 /*
Eric Wong128dd172013-01-01 21:20:27 +00001386 * The following barrier has two effects:
1387 *
1388 * 1) Flush epi changes above to other CPUs. This ensures
1389 * we do not miss events from ep_poll_callback if an
1390 * event occurs immediately after we call f_op->poll().
1391 * We need this because we did not take ep->lock while
1392 * changing epi above (but ep_poll_callback does take
1393 * ep->lock).
1394 *
1395 * 2) We also need to ensure we do not miss _past_ events
1396 * when calling f_op->poll(). This barrier also
1397 * pairs with the barrier in wq_has_sleeper (see
1398 * comments for wq_has_sleeper).
1399 *
1400 * This barrier will now guarantee ep_poll_callback or f_op->poll
1401 * (or both) will notice the readiness of an item.
1402 */
1403 smp_mb();
1404
1405 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 * Get current event bits. We can safely use the file* here because
1407 * its usage count has been increased by the caller of this function.
1408 */
Eric Wong450d89e2013-04-30 15:27:42 -07001409 revents = ep_item_poll(epi, &pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001412 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001413 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001415 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001416 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001417 if (!ep_is_linked(&epi->rdllink)) {
1418 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001419 ep_pm_stay_awake(epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420
Davide Libenzic7ea7632007-05-15 01:40:47 -07001421 /* Notify waiting tasks that events are available */
1422 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -04001423 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001424 if (waitqueue_active(&ep->poll_wait))
1425 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 }
Tony Battersbye057e152009-03-31 15:24:15 -07001427 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429
1430 /* We have to call this outside the lock */
1431 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001432 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433
1434 return 0;
1435}
1436
Davide Libenzi296e2362009-03-31 15:24:11 -07001437static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1438 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001439{
1440 struct ep_send_events_data *esed = priv;
1441 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001442 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001443 struct epitem *epi;
1444 struct epoll_event __user *uevent;
Eric Wongeea1d582013-04-30 15:27:39 -07001445 struct wakeup_source *ws;
Hans Verkuil626cf232012-03-23 15:02:27 -07001446 poll_table pt;
1447
1448 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -07001449
Davide Libenzi296e2362009-03-31 15:24:11 -07001450 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001451 * We can loop without lock because we are passed a task private list.
1452 * Items cannot vanish during the loop because ep_scan_ready_list() is
1453 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001454 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001455 for (eventcnt = 0, uevent = esed->events;
1456 !list_empty(head) && eventcnt < esed->maxevents;) {
1457 epi = list_first_entry(head, struct epitem, rdllink);
1458
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001459 /*
1460 * Activate ep->ws before deactivating epi->ws to prevent
1461 * triggering auto-suspend here (in case we reactive epi->ws
1462 * below).
1463 *
1464 * This could be rearranged to delay the deactivation of epi->ws
1465 * instead, but then epi->ws would temporarily be out of sync
1466 * with ep_is_linked().
1467 */
Eric Wongeea1d582013-04-30 15:27:39 -07001468 ws = ep_wakeup_source(epi);
1469 if (ws) {
1470 if (ws->active)
1471 __pm_stay_awake(ep->ws);
1472 __pm_relax(ws);
1473 }
1474
Davide Libenzi5071f972009-03-31 15:24:10 -07001475 list_del_init(&epi->rdllink);
1476
Eric Wong450d89e2013-04-30 15:27:42 -07001477 revents = ep_item_poll(epi, &pt);
Davide Libenzi5071f972009-03-31 15:24:10 -07001478
Davide Libenzi296e2362009-03-31 15:24:11 -07001479 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001480 * If the event mask intersect the caller-requested one,
1481 * deliver the event to userspace. Again, ep_scan_ready_list()
1482 * is holding "mtx", so no operations coming from userspace
1483 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001484 */
1485 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001486 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001487 __put_user(epi->event.data, &uevent->data)) {
1488 list_add(&epi->rdllink, head);
Eric Wongeea1d582013-04-30 15:27:39 -07001489 ep_pm_stay_awake(epi);
Davide Libenzi296e2362009-03-31 15:24:11 -07001490 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001491 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001492 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001493 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001494 if (epi->event.events & EPOLLONESHOT)
1495 epi->event.events &= EP_PRIVATE_BITS;
1496 else if (!(epi->event.events & EPOLLET)) {
1497 /*
1498 * If this file has been added with Level
1499 * Trigger mode, we need to insert back inside
1500 * the ready list, so that the next call to
1501 * epoll_wait() will check again the events
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001502 * availability. At this point, no one can insert
Davide Libenzi296e2362009-03-31 15:24:11 -07001503 * into ep->rdllist besides us. The epoll_ctl()
1504 * callers are locked out by
1505 * ep_scan_ready_list() holding "mtx" and the
1506 * poll callback will queue them in ep->ovflist.
1507 */
1508 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001509 ep_pm_stay_awake(epi);
Davide Libenzi296e2362009-03-31 15:24:11 -07001510 }
1511 }
1512 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001513
1514 return eventcnt;
1515}
1516
Davide Libenzi296e2362009-03-31 15:24:11 -07001517static int ep_send_events(struct eventpoll *ep,
1518 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519{
Davide Libenzi5071f972009-03-31 15:24:10 -07001520 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001521
Davide Libenzi5071f972009-03-31 15:24:10 -07001522 esed.maxevents = maxevents;
1523 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001524
Nelson Elhaged8805e62011-10-31 17:13:14 -07001525 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526}
1527
Eric Dumazet0781b902011-02-01 15:52:35 -08001528static inline struct timespec ep_set_mstimeout(long ms)
1529{
1530 struct timespec now, ts = {
1531 .tv_sec = ms / MSEC_PER_SEC,
1532 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1533 };
1534
1535 ktime_get_ts(&now);
1536 return timespec_add_safe(now, ts);
1537}
1538
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001539/**
1540 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1541 * event buffer.
1542 *
1543 * @ep: Pointer to the eventpoll context.
1544 * @events: Pointer to the userspace buffer where the ready events should be
1545 * stored.
1546 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1547 * @timeout: Maximum timeout for the ready events fetch operation, in
1548 * milliseconds. If the @timeout is zero, the function will not block,
1549 * while if the @timeout is less than zero, the function will block
1550 * until at least one event has been retrieved (or an error
1551 * occurred).
1552 *
1553 * Returns: Returns the number of ready events which have been fetched, or an
1554 * error code, in case of error.
1555 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1557 int maxevents, long timeout)
1558{
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001559 int res = 0, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560 unsigned long flags;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001561 long slack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001563 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001565 if (timeout > 0) {
Eric Dumazet0781b902011-02-01 15:52:35 -08001566 struct timespec end_time = ep_set_mstimeout(timeout);
1567
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001568 slack = select_estimate_accuracy(&end_time);
1569 to = &expires;
1570 *to = timespec_to_ktime(end_time);
1571 } else if (timeout == 0) {
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001572 /*
1573 * Avoid the unnecessary trip to the wait queue loop, if the
1574 * caller specified a non blocking operation.
1575 */
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001576 timed_out = 1;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001577 spin_lock_irqsave(&ep->lock, flags);
1578 goto check_events;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001579 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001581fetch_events:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001582 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001584 if (!ep_events_available(ep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 /*
1586 * We don't have any available event to return to the caller.
1587 * We need to sleep here, and we will be wake up by
1588 * ep_poll_callback() when events will become available.
1589 */
1590 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001591 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592
1593 for (;;) {
1594 /*
1595 * We don't want to sleep if the ep_poll_callback() sends us
1596 * a wakeup in between. That's why we set the task state
1597 * to TASK_INTERRUPTIBLE before doing the checks.
1598 */
1599 set_current_state(TASK_INTERRUPTIBLE);
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001600 if (ep_events_available(ep) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 break;
1602 if (signal_pending(current)) {
1603 res = -EINTR;
1604 break;
1605 }
1606
Davide Libenzic7ea7632007-05-15 01:40:47 -07001607 spin_unlock_irqrestore(&ep->lock, flags);
Colin Cross1c441e92013-05-06 23:50:16 +00001608 if (!freezable_schedule_hrtimeout_range(to, slack,
1609 HRTIMER_MODE_ABS))
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001610 timed_out = 1;
1611
Davide Libenzic7ea7632007-05-15 01:40:47 -07001612 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001614 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615
1616 set_current_state(TASK_RUNNING);
1617 }
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001618check_events:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 /* Is it worth to try to dig for events ? */
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001620 eavail = ep_events_available(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
Davide Libenzic7ea7632007-05-15 01:40:47 -07001622 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
1624 /*
1625 * Try to transfer events to user space. In case we get 0 events and
1626 * there's still timeout left over, we go trying again in search of
1627 * more luck.
1628 */
1629 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001630 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001631 goto fetch_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632
1633 return res;
1634}
1635
Davide Libenzi22bacca2011-02-25 14:44:12 -08001636/**
1637 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1638 * API, to verify that adding an epoll file inside another
1639 * epoll structure, does not violate the constraints, in
1640 * terms of closed loops, or too deep chains (which can
1641 * result in excessive stack usage).
1642 *
1643 * @priv: Pointer to the epoll file to be currently checked.
1644 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1645 * data structure pointer.
1646 * @call_nests: Current dept of the @ep_call_nested() call stack.
1647 *
1648 * Returns: Returns zero if adding the epoll @file inside current epoll
1649 * structure @ep does not violate the constraints, or -1 otherwise.
1650 */
1651static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1652{
1653 int error = 0;
1654 struct file *file = priv;
1655 struct eventpoll *ep = file->private_data;
Jason Baron28d82dc2012-01-12 17:17:43 -08001656 struct eventpoll *ep_tovisit;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001657 struct rb_node *rbp;
1658 struct epitem *epi;
1659
Nelson Elhaged8805e62011-10-31 17:13:14 -07001660 mutex_lock_nested(&ep->mtx, call_nests + 1);
Jason Baron28d82dc2012-01-12 17:17:43 -08001661 ep->visited = 1;
1662 list_add(&ep->visited_list_link, &visited_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001663 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1664 epi = rb_entry(rbp, struct epitem, rbn);
1665 if (unlikely(is_file_epoll(epi->ffd.file))) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001666 ep_tovisit = epi->ffd.file->private_data;
1667 if (ep_tovisit->visited)
1668 continue;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001669 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Jason Baron28d82dc2012-01-12 17:17:43 -08001670 ep_loop_check_proc, epi->ffd.file,
1671 ep_tovisit, current);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001672 if (error != 0)
1673 break;
Jason Baron28d82dc2012-01-12 17:17:43 -08001674 } else {
1675 /*
1676 * If we've reached a file that is not associated with
1677 * an ep, then we need to check if the newly added
1678 * links are going to add too many wakeup paths. We do
1679 * this by adding it to the tfile_check_list, if it's
1680 * not already there, and calling reverse_path_check()
1681 * during ep_insert().
1682 */
1683 if (list_empty(&epi->ffd.file->f_tfile_llink))
1684 list_add(&epi->ffd.file->f_tfile_llink,
1685 &tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001686 }
1687 }
1688 mutex_unlock(&ep->mtx);
1689
1690 return error;
1691}
1692
1693/**
1694 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1695 * another epoll file (represented by @ep) does not create
1696 * closed loops or too deep chains.
1697 *
1698 * @ep: Pointer to the epoll private data structure.
1699 * @file: Pointer to the epoll file to be checked.
1700 *
1701 * Returns: Returns zero if adding the epoll @file inside current epoll
1702 * structure @ep does not violate the constraints, or -1 otherwise.
1703 */
1704static int ep_loop_check(struct eventpoll *ep, struct file *file)
1705{
Jason Baron28d82dc2012-01-12 17:17:43 -08001706 int ret;
1707 struct eventpoll *ep_cur, *ep_next;
1708
1709 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Davide Libenzi22bacca2011-02-25 14:44:12 -08001710 ep_loop_check_proc, file, ep, current);
Jason Baron28d82dc2012-01-12 17:17:43 -08001711 /* clear visited list */
1712 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1713 visited_list_link) {
1714 ep_cur->visited = 0;
1715 list_del(&ep_cur->visited_list_link);
1716 }
1717 return ret;
1718}
1719
1720static void clear_tfile_check_list(void)
1721{
1722 struct file *file;
1723
1724 /* first clear the tfile_check_list */
1725 while (!list_empty(&tfile_check_list)) {
1726 file = list_first_entry(&tfile_check_list, struct file,
1727 f_tfile_llink);
1728 list_del_init(&file->f_tfile_llink);
1729 }
1730 INIT_LIST_HEAD(&tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001731}
1732
Davide Libenzi7699acd2007-05-10 22:23:23 -07001733/*
Andrew Morton523723b2008-08-12 15:09:01 -07001734 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001735 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001736SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001737{
Jason Baron28d82dc2012-01-12 17:17:43 -08001738 int error, fd;
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001739 struct eventpoll *ep = NULL;
Jason Baron28d82dc2012-01-12 17:17:43 -08001740 struct file *file;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001741
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001742 /* Check the EPOLL_* constant for consistency. */
1743 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1744
Davide Libenzi296e2362009-03-31 15:24:11 -07001745 if (flags & ~EPOLL_CLOEXEC)
1746 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001747 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001748 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001749 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001750 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001751 if (error < 0)
1752 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001753 /*
1754 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001755 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001756 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001757 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1758 if (fd < 0) {
1759 error = fd;
1760 goto out_free_ep;
1761 }
1762 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001763 O_RDWR | (flags & O_CLOEXEC));
Jason Baron28d82dc2012-01-12 17:17:43 -08001764 if (IS_ERR(file)) {
1765 error = PTR_ERR(file);
1766 goto out_free_fd;
1767 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001768 ep->file = file;
Al Viro98022742012-08-17 22:42:36 -04001769 fd_install(fd, file);
Jason Baron28d82dc2012-01-12 17:17:43 -08001770 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001771
Jason Baron28d82dc2012-01-12 17:17:43 -08001772out_free_fd:
1773 put_unused_fd(fd);
1774out_free_ep:
1775 ep_free(ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001776 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001777}
1778
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001779SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001780{
Davide Libenzibfe38912009-05-12 13:19:44 -07001781 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001782 return -EINVAL;
1783
1784 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001785}
1786
Davide Libenzi7699acd2007-05-10 22:23:23 -07001787/*
1788 * The following function implements the controller interface for
1789 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001790 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001791 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001792SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1793 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001794{
1795 int error;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001796 int did_lock_epmutex = 0;
Al Viro7e3fb582013-08-30 12:47:21 -04001797 struct fd f, tf;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001798 struct eventpoll *ep;
1799 struct epitem *epi;
1800 struct epoll_event epds;
1801
Davide Libenzi7699acd2007-05-10 22:23:23 -07001802 error = -EFAULT;
1803 if (ep_op_has_event(op) &&
1804 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1805 goto error_return;
1806
Davide Libenzi7699acd2007-05-10 22:23:23 -07001807 error = -EBADF;
Al Viro7e3fb582013-08-30 12:47:21 -04001808 f = fdget(epfd);
1809 if (!f.file)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001810 goto error_return;
1811
1812 /* Get the "struct file *" for the target file */
Al Viro7e3fb582013-08-30 12:47:21 -04001813 tf = fdget(fd);
1814 if (!tf.file)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001815 goto error_fput;
1816
1817 /* The target file descriptor must support poll */
1818 error = -EPERM;
Al Viro7e3fb582013-08-30 12:47:21 -04001819 if (!tf.file->f_op || !tf.file->f_op->poll)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001820 goto error_tgt_fput;
1821
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001822 /* Check if EPOLLWAKEUP is allowed */
Michael Kerriskd9914cf2012-07-17 21:37:27 +02001823 if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
Rafael J. Wysockia8159412012-05-21 21:20:48 +02001824 epds.events &= ~EPOLLWAKEUP;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001825
Davide Libenzi7699acd2007-05-10 22:23:23 -07001826 /*
1827 * We have to check that the file structure underneath the file descriptor
1828 * the user passed to us _is_ an eventpoll file. And also we do not permit
1829 * adding an epoll file descriptor inside itself.
1830 */
1831 error = -EINVAL;
Al Viro7e3fb582013-08-30 12:47:21 -04001832 if (f.file == tf.file || !is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07001833 goto error_tgt_fput;
1834
1835 /*
1836 * At this point it is safe to assume that the "private_data" contains
1837 * our own data structure.
1838 */
Al Viro7e3fb582013-08-30 12:47:21 -04001839 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001840
Davide Libenzi22bacca2011-02-25 14:44:12 -08001841 /*
1842 * When we insert an epoll file descriptor, inside another epoll file
1843 * descriptor, there is the change of creating closed loops, which are
Jason Baron28d82dc2012-01-12 17:17:43 -08001844 * better be handled here, than in more critical paths. While we are
1845 * checking for loops we also determine the list of files reachable
1846 * and hang them on the tfile_check_list, so we can check that we
1847 * haven't created too many possible wakeup paths.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001848 *
Jason Baron28d82dc2012-01-12 17:17:43 -08001849 * We need to hold the epmutex across both ep_insert and ep_remove
1850 * b/c we want to make sure we are looking at a coherent view of
1851 * epoll network.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001852 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001853 if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
Davide Libenzi22bacca2011-02-25 14:44:12 -08001854 mutex_lock(&epmutex);
1855 did_lock_epmutex = 1;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001856 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001857 if (op == EPOLL_CTL_ADD) {
Al Viro7e3fb582013-08-30 12:47:21 -04001858 if (is_file_epoll(tf.file)) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001859 error = -ELOOP;
Al Viro7e3fb582013-08-30 12:47:21 -04001860 if (ep_loop_check(ep, tf.file) != 0) {
Jason Baron13d51802012-04-25 16:01:47 -07001861 clear_tfile_check_list();
Jason Baron28d82dc2012-01-12 17:17:43 -08001862 goto error_tgt_fput;
Jason Baron13d51802012-04-25 16:01:47 -07001863 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001864 } else
Al Viro7e3fb582013-08-30 12:47:21 -04001865 list_add(&tf.file->f_tfile_llink, &tfile_check_list);
Jason Baron28d82dc2012-01-12 17:17:43 -08001866 }
Davide Libenzi22bacca2011-02-25 14:44:12 -08001867
Nelson Elhaged8805e62011-10-31 17:13:14 -07001868 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001869
Davide Libenzi67647d02007-05-15 01:40:52 -07001870 /*
1871 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1872 * above, we can be sure to be able to use the item looked up by
1873 * ep_find() till we release the mutex.
1874 */
Al Viro7e3fb582013-08-30 12:47:21 -04001875 epi = ep_find(ep, tf.file, fd);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001876
1877 error = -EINVAL;
1878 switch (op) {
1879 case EPOLL_CTL_ADD:
1880 if (!epi) {
1881 epds.events |= POLLERR | POLLHUP;
Al Viro7e3fb582013-08-30 12:47:21 -04001882 error = ep_insert(ep, &epds, tf.file, fd);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001883 } else
1884 error = -EEXIST;
Jason Baron28d82dc2012-01-12 17:17:43 -08001885 clear_tfile_check_list();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001886 break;
1887 case EPOLL_CTL_DEL:
1888 if (epi)
1889 error = ep_remove(ep, epi);
1890 else
1891 error = -ENOENT;
1892 break;
1893 case EPOLL_CTL_MOD:
1894 if (epi) {
1895 epds.events |= POLLERR | POLLHUP;
1896 error = ep_modify(ep, epi, &epds);
1897 } else
1898 error = -ENOENT;
1899 break;
1900 }
Davide Libenzid47de162007-05-15 01:40:41 -07001901 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001902
1903error_tgt_fput:
Jason Baron28d82dc2012-01-12 17:17:43 -08001904 if (did_lock_epmutex)
Davide Libenzi22bacca2011-02-25 14:44:12 -08001905 mutex_unlock(&epmutex);
1906
Al Viro7e3fb582013-08-30 12:47:21 -04001907 fdput(tf);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001908error_fput:
Al Viro7e3fb582013-08-30 12:47:21 -04001909 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001910error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001911
1912 return error;
1913}
1914
1915/*
1916 * Implement the event wait interface for the eventpoll file. It is the kernel
1917 * part of the user space epoll_wait(2).
1918 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001919SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1920 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001921{
Al Viro2903ff02012-08-28 12:52:22 -04001922 int error;
1923 struct fd f;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001924 struct eventpoll *ep;
1925
Davide Libenzi7699acd2007-05-10 22:23:23 -07001926 /* The maximum number of event must be greater than zero */
1927 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1928 return -EINVAL;
1929
1930 /* Verify that the area passed by the user is writeable */
Al Viro2903ff02012-08-28 12:52:22 -04001931 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1932 return -EFAULT;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001933
1934 /* Get the "struct file *" for the eventpoll file */
Al Viro2903ff02012-08-28 12:52:22 -04001935 f = fdget(epfd);
1936 if (!f.file)
1937 return -EBADF;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001938
1939 /*
1940 * We have to check that the file structure underneath the fd
1941 * the user passed to us _is_ an eventpoll file.
1942 */
1943 error = -EINVAL;
Al Viro2903ff02012-08-28 12:52:22 -04001944 if (!is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07001945 goto error_fput;
1946
1947 /*
1948 * At this point it is safe to assume that the "private_data" contains
1949 * our own data structure.
1950 */
Al Viro2903ff02012-08-28 12:52:22 -04001951 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001952
1953 /* Time to fish for events ... */
1954 error = ep_poll(ep, events, maxevents, timeout);
1955
1956error_fput:
Al Viro2903ff02012-08-28 12:52:22 -04001957 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001958 return error;
1959}
1960
Davide Libenzi7699acd2007-05-10 22:23:23 -07001961/*
1962 * Implement the event wait interface for the eventpoll file. It is the kernel
1963 * part of the user space epoll_pwait(2).
1964 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001965SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1966 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1967 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001968{
1969 int error;
1970 sigset_t ksigmask, sigsaved;
1971
1972 /*
1973 * If the caller wants a certain signal mask to be set during the wait,
1974 * we apply it here.
1975 */
1976 if (sigmask) {
1977 if (sigsetsize != sizeof(sigset_t))
1978 return -EINVAL;
1979 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1980 return -EFAULT;
Oleg Nesterov77d55912013-07-03 15:08:14 -07001981 sigsaved = current->blocked;
1982 set_current_blocked(&ksigmask);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001983 }
1984
1985 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1986
1987 /*
1988 * If we changed the signal mask, we need to restore the original one.
1989 * In case we've got a signal while waiting, we do not restore the
1990 * signal mask yet, and we allow do_signal() to deliver the signal on
1991 * the way back to userspace, before the signal mask is restored.
1992 */
1993 if (sigmask) {
1994 if (error == -EINTR) {
1995 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001996 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001997 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001998 } else
Oleg Nesterov77d55912013-07-03 15:08:14 -07001999 set_current_blocked(&sigsaved);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002000 }
2001
2002 return error;
2003}
2004
Al Viro35280bd2013-02-24 14:52:17 -05002005#ifdef CONFIG_COMPAT
2006COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2007 struct epoll_event __user *, events,
2008 int, maxevents, int, timeout,
2009 const compat_sigset_t __user *, sigmask,
2010 compat_size_t, sigsetsize)
2011{
2012 long err;
2013 compat_sigset_t csigmask;
2014 sigset_t ksigmask, sigsaved;
2015
2016 /*
2017 * If the caller wants a certain signal mask to be set during the wait,
2018 * we apply it here.
2019 */
2020 if (sigmask) {
2021 if (sigsetsize != sizeof(compat_sigset_t))
2022 return -EINVAL;
2023 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2024 return -EFAULT;
2025 sigset_from_compat(&ksigmask, &csigmask);
Oleg Nesterov77d55912013-07-03 15:08:14 -07002026 sigsaved = current->blocked;
2027 set_current_blocked(&ksigmask);
Al Viro35280bd2013-02-24 14:52:17 -05002028 }
2029
2030 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2031
2032 /*
2033 * If we changed the signal mask, we need to restore the original one.
2034 * In case we've got a signal while waiting, we do not restore the
2035 * signal mask yet, and we allow do_signal() to deliver the signal on
2036 * the way back to userspace, before the signal mask is restored.
2037 */
2038 if (sigmask) {
2039 if (err == -EINTR) {
2040 memcpy(&current->saved_sigmask, &sigsaved,
2041 sizeof(sigsaved));
2042 set_restore_sigmask();
2043 } else
Oleg Nesterov77d55912013-07-03 15:08:14 -07002044 set_current_blocked(&sigsaved);
Al Viro35280bd2013-02-24 14:52:17 -05002045 }
2046
2047 return err;
2048}
2049#endif
2050
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051static int __init eventpoll_init(void)
2052{
Davide Libenzi7ef99642008-12-01 13:13:55 -08002053 struct sysinfo si;
2054
2055 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08002056 /*
2057 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2058 */
2059 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08002060 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08002061 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062
Davide Libenzi22bacca2011-02-25 14:44:12 -08002063 /*
2064 * Initialize the structure used to perform epoll file descriptor
2065 * inclusion loops checks.
2066 */
2067 ep_nested_calls_init(&poll_loop_ncalls);
2068
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07002070 ep_nested_calls_init(&poll_safewake_ncalls);
2071
2072 /* Initialize the structure used to perform file's f_op->poll() calls */
2073 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074
Eric Wong39732ca2013-04-30 15:27:38 -07002075 /*
2076 * We can have many thousands of epitems, so prevent this from
2077 * using an extra cache line on 64-bit (and smaller) CPUs
2078 */
2079 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2080
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 /* Allocates slab cache used to allocate "struct epitem" items */
2082 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07002083 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084
2085 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2086 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07002087 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090}
Davide Libenzicea69242007-05-10 22:23:22 -07002091fs_initcall(eventpoll_init);