blob: ae1dbcf47e979d48b67ee83e1ec413e4d119a48b [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Davide Libenzi5071f972009-03-31 15:24:10 -07002 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * Davide Libenzi <davidel@xmailserver.org>
11 *
12 */
13
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/init.h>
15#include <linux/kernel.h>
16#include <linux/sched.h>
17#include <linux/fs.h>
18#include <linux/file.h>
19#include <linux/signal.h>
20#include <linux/errno.h>
21#include <linux/mm.h>
22#include <linux/slab.h>
23#include <linux/poll.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/string.h>
25#include <linux/list.h>
26#include <linux/hash.h>
27#include <linux/spinlock.h>
28#include <linux/syscalls.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/rbtree.h>
30#include <linux/wait.h>
31#include <linux/eventpoll.h>
32#include <linux/mount.h>
33#include <linux/bitops.h>
Arjan van de Ven144efe32006-03-23 03:00:32 -080034#include <linux/mutex.h>
Davide Libenzida66f7c2007-05-10 22:23:21 -070035#include <linux/anon_inodes.h>
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020036#include <linux/device.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/io.h>
39#include <asm/mman.h>
Arun Sharma600634972011-07-26 16:09:06 -070040#include <linux/atomic.h>
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -080041#include <linux/proc_fs.h>
42#include <linux/seq_file.h>
Al Viro35280bd2013-02-24 14:52:17 -050043#include <linux/compat.h>
Jason Baronae10b2b2013-11-12 15:10:16 -080044#include <linux/rculist.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 */
Jason Barondf0108c2016-01-20 14:59:24 -080095#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
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 {
Jason Baronae10b2b2013-11-12 15:10:16 -0800137 union {
138 /* RB tree node links this structure to the eventpoll RB tree */
139 struct rb_node rbn;
140 /* Used to free the struct epitem */
141 struct rcu_head rcu;
142 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
144 /* List header used to link this structure to the eventpoll ready list */
145 struct list_head rdllink;
146
Davide Libenzic7ea7632007-05-15 01:40:47 -0700147 /*
148 * Works together "struct eventpoll"->ovflist in keeping the
149 * single linked chain of items.
150 */
151 struct epitem *next;
152
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 /* The file descriptor information this item refers to */
154 struct epoll_filefd ffd;
155
156 /* Number of active wait queue attached to poll operations */
157 int nwait;
158
159 /* List containing poll wait queues */
160 struct list_head pwqlist;
161
162 /* The "container" of this item */
163 struct eventpoll *ep;
164
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 /* List header used to link this item to the "struct file" items list */
166 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700167
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200168 /* wakeup_source used when EPOLLWAKEUP is set */
Eric Wongeea1d582013-04-30 15:27:39 -0700169 struct wakeup_source __rcu *ws;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200170
Davide Libenzic7ea7632007-05-15 01:40:47 -0700171 /* The structure that describe the interested events and the source fd */
172 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700173};
174
175/*
176 * This structure is stored inside the "private_data" member of the file
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200177 * structure and represents the main data structure for the eventpoll
Davide Libenzid47de162007-05-15 01:40:41 -0700178 * interface.
179 */
180struct eventpoll {
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200181 /* Protect the access to this structure */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700182 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700183
184 /*
185 * This mutex is used to ensure that files are not removed
186 * while epoll is using them. This is held during the event
187 * collection loop, the file cleanup path, the epoll file exit
188 * code and the ctl operations.
189 */
190 struct mutex mtx;
191
192 /* Wait queue used by sys_epoll_wait() */
193 wait_queue_head_t wq;
194
195 /* Wait queue used by file->poll() */
196 wait_queue_head_t poll_wait;
197
198 /* List of ready file descriptors */
199 struct list_head rdllist;
200
Davide Libenzi67647d02007-05-15 01:40:52 -0700201 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700202 struct rb_root rbr;
203
204 /*
205 * This is a single linked list that chains all the "struct epitem" that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300206 * happened while transferring ready events to userspace w/out
Davide Libenzid47de162007-05-15 01:40:41 -0700207 * holding ->lock.
208 */
209 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800210
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200211 /* wakeup_source used when ep_scan_ready_list is running */
212 struct wakeup_source *ws;
213
Davide Libenzi7ef99642008-12-01 13:13:55 -0800214 /* The user that created the eventpoll descriptor */
215 struct user_struct *user;
Jason Baron28d82dc2012-01-12 17:17:43 -0800216
217 struct file *file;
218
219 /* used to optimize loop detection check */
220 int visited;
221 struct list_head visited_list_link;
Davide Libenzid47de162007-05-15 01:40:41 -0700222};
223
224/* Wait structure used by the poll hooks */
225struct eppoll_entry {
226 /* List header used to link this structure to the "struct epitem" */
227 struct list_head llink;
228
229 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700230 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700231
232 /*
233 * Wait queue item that will be linked to the target file wait
234 * queue head.
235 */
236 wait_queue_t wait;
237
238 /* The wait queue head that linked the "wait" wait queue item */
239 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240};
241
242/* Wrapper struct used by poll queueing */
243struct ep_pqueue {
244 poll_table pt;
245 struct epitem *epi;
246};
247
Davide Libenzi5071f972009-03-31 15:24:10 -0700248/* Used by the ep_send_events() function as callback private data */
249struct ep_send_events_data {
250 int maxevents;
251 struct epoll_event __user *events;
252};
253
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800255 * Configuration options available inside /proc/sys/fs/epoll/
256 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800257/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800258static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800259
260/*
Davide Libenzid47de162007-05-15 01:40:41 -0700261 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800263static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264
Davide Libenzi22bacca2011-02-25 14:44:12 -0800265/* Used to check for epoll file descriptor inclusion loops */
266static struct nested_calls poll_loop_ncalls;
267
Davide Libenzi5071f972009-03-31 15:24:10 -0700268/* Used for safe wake up implementation */
269static struct nested_calls poll_safewake_ncalls;
270
271/* Used to call file's f_op->poll() under the nested calls boundaries */
272static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800275static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276
277/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800278static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Jason Baron28d82dc2012-01-12 17:17:43 -0800280/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
281static LIST_HEAD(visited_list);
282
283/*
284 * List of files with newly added links, where we may need to limit the number
285 * of emanating paths. Protected by the epmutex.
286 */
287static LIST_HEAD(tfile_check_list);
288
Davide Libenzi7ef99642008-12-01 13:13:55 -0800289#ifdef CONFIG_SYSCTL
290
291#include <linux/sysctl.h>
292
Robin Holt52bd19f72011-01-12 17:00:01 -0800293static long zero;
294static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800295
Joe Perches1f7e0612014-06-06 14:38:05 -0700296struct ctl_table epoll_table[] = {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800297 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800298 .procname = "max_user_watches",
299 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800300 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800301 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800302 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800303 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800304 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800305 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800306 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800307};
308#endif /* CONFIG_SYSCTL */
309
Jason Baron28d82dc2012-01-12 17:17:43 -0800310static const struct file_operations eventpoll_fops;
311
312static inline int is_file_epoll(struct file *f)
313{
314 return f->f_op == &eventpoll_fops;
315}
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700316
Davide Libenzi67647d02007-05-15 01:40:52 -0700317/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700318static inline void ep_set_ffd(struct epoll_filefd *ffd,
319 struct file *file, int fd)
320{
321 ffd->file = file;
322 ffd->fd = fd;
323}
324
Davide Libenzi67647d02007-05-15 01:40:52 -0700325/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700326static inline int ep_cmp_ffd(struct epoll_filefd *p1,
327 struct epoll_filefd *p2)
328{
329 return (p1->file > p2->file ? +1:
330 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
331}
332
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700333/* Tells us if the item is currently linked */
334static inline int ep_is_linked(struct list_head *p)
335{
336 return !list_empty(p);
337}
338
Oleg Nesterov971316f2012-02-24 20:07:29 +0100339static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
340{
341 return container_of(p, struct eppoll_entry, wait);
342}
343
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700344/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700345static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700346{
347 return container_of(p, struct eppoll_entry, wait)->base;
348}
349
350/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700351static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700352{
353 return container_of(p, struct ep_pqueue, pt)->epi;
354}
355
356/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700357static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700358{
Andrew Mortona80a6b82012-11-08 15:53:35 -0800359 return op != EPOLL_CTL_DEL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700360}
361
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700363static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
Davide Libenzi5071f972009-03-31 15:24:10 -0700365 INIT_LIST_HEAD(&ncalls->tasks_call_list);
366 spin_lock_init(&ncalls->lock);
367}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368
Davide Libenzi5071f972009-03-31 15:24:10 -0700369/**
Davide Libenzi3fb0e582011-03-22 16:34:46 -0700370 * ep_events_available - Checks if ready events might be available.
371 *
372 * @ep: Pointer to the eventpoll context.
373 *
374 * Returns: Returns a value different than zero if ready events are available,
375 * or zero otherwise.
376 */
377static inline int ep_events_available(struct eventpoll *ep)
378{
379 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
380}
381
382/**
Davide Libenzi5071f972009-03-31 15:24:10 -0700383 * ep_call_nested - Perform a bound (possibly) nested call, by checking
384 * that the recursion limit is not exceeded, and that
385 * the same nested call (by the meaning of same cookie) is
386 * no re-entered.
387 *
388 * @ncalls: Pointer to the nested_calls structure to be used for this call.
389 * @max_nests: Maximum number of allowed nesting calls.
390 * @nproc: Nested call core function pointer.
391 * @priv: Opaque data to be passed to the @nproc callback.
392 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700393 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700394 *
395 * Returns: Returns the code returned by the @nproc callback, or -1 if
396 * the maximum recursion limit has been exceeded.
397 */
398static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
399 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700400 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700401{
402 int error, call_nests = 0;
403 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700404 struct list_head *lsthead = &ncalls->tasks_call_list;
405 struct nested_call_node *tncur;
406 struct nested_call_node tnode;
407
408 spin_lock_irqsave(&ncalls->lock, flags);
409
410 /*
411 * Try to see if the current task is already inside this wakeup call.
412 * We use a list here, since the population inside this set is always
413 * very much limited.
414 */
415 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700416 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700417 (tncur->cookie == cookie || ++call_nests > max_nests)) {
418 /*
419 * Ops ... loop detected or maximum nest level reached.
420 * We abort this wake by breaking the cycle itself.
421 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700422 error = -1;
423 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700424 }
425 }
426
427 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700428 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700429 tnode.cookie = cookie;
430 list_add(&tnode.llink, lsthead);
431
432 spin_unlock_irqrestore(&ncalls->lock, flags);
433
434 /* Call the nested function */
435 error = (*nproc)(priv, cookie, call_nests);
436
437 /* Remove the current task from the list */
438 spin_lock_irqsave(&ncalls->lock, flags);
439 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700440out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700441 spin_unlock_irqrestore(&ncalls->lock, flags);
442
443 return error;
444}
445
Steven Rostedt02edc6f2012-03-23 15:02:27 -0700446/*
447 * As described in commit 0ccf831cb lockdep: annotate epoll
448 * the use of wait queues used by epoll is done in a very controlled
449 * manner. Wake ups can nest inside each other, but are never done
450 * with the same locking. For example:
451 *
452 * dfd = socket(...);
453 * efd1 = epoll_create();
454 * efd2 = epoll_create();
455 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
456 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
457 *
458 * When a packet arrives to the device underneath "dfd", the net code will
459 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
460 * callback wakeup entry on that queue, and the wake_up() performed by the
461 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
462 * (efd1) notices that it may have some event ready, so it needs to wake up
463 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
464 * that ends up in another wake_up(), after having checked about the
465 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
466 * avoid stack blasting.
467 *
468 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
469 * this special case of epoll.
470 */
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700471#ifdef CONFIG_DEBUG_LOCK_ALLOC
472static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
473 unsigned long events, int subclass)
474{
475 unsigned long flags;
476
477 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
478 wake_up_locked_poll(wqueue, events);
479 spin_unlock_irqrestore(&wqueue->lock, flags);
480}
481#else
482static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
483 unsigned long events, int subclass)
484{
485 wake_up_poll(wqueue, events);
486}
487#endif
488
Davide Libenzi5071f972009-03-31 15:24:10 -0700489static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
490{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700491 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
492 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700493 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494}
495
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496/*
497 * Perform a safe wake up of the poll wait list. The problem is that
498 * with the new callback'd wake up system, it is possible that the
499 * poll callback is reentered from inside the call to wake_up() done
500 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700501 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 * and we cannot reenter the same wait queue head at all. This will
503 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700504 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700506static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700508 int this_cpu = get_cpu();
509
Davide Libenzi5071f972009-03-31 15:24:10 -0700510 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700511 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
512
513 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514}
515
Oleg Nesterov971316f2012-02-24 20:07:29 +0100516static void ep_remove_wait_queue(struct eppoll_entry *pwq)
517{
518 wait_queue_head_t *whead;
519
520 rcu_read_lock();
521 /* If it is cleared by POLLFREE, it should be rcu-safe */
522 whead = rcu_dereference(pwq->whead);
523 if (whead)
524 remove_wait_queue(whead, &pwq->wait);
525 rcu_read_unlock();
526}
527
Davide Libenzi7699acd2007-05-10 22:23:23 -0700528/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700529 * This function unregisters poll callbacks from the associated file
530 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
531 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700532 */
533static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
534{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700535 struct list_head *lsthead = &epi->pwqlist;
536 struct eppoll_entry *pwq;
537
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700538 while (!list_empty(lsthead)) {
539 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700540
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700541 list_del(&pwq->llink);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100542 ep_remove_wait_queue(pwq);
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700543 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700544 }
545}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546
Eric Wongeea1d582013-04-30 15:27:39 -0700547/* call only when ep->mtx is held */
548static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
549{
550 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
551}
552
553/* call only when ep->mtx is held */
554static inline void ep_pm_stay_awake(struct epitem *epi)
555{
556 struct wakeup_source *ws = ep_wakeup_source(epi);
557
558 if (ws)
559 __pm_stay_awake(ws);
560}
561
562static inline bool ep_has_wakeup_source(struct epitem *epi)
563{
564 return rcu_access_pointer(epi->ws) ? true : false;
565}
566
567/* call when ep->mtx cannot be held (ep_poll_callback) */
568static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
569{
570 struct wakeup_source *ws;
571
572 rcu_read_lock();
573 ws = rcu_dereference(epi->ws);
574 if (ws)
575 __pm_stay_awake(ws);
576 rcu_read_unlock();
577}
578
Davide Libenzi5071f972009-03-31 15:24:10 -0700579/**
580 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
581 * the scan code, to call f_op->poll(). Also allows for
582 * O(NumReady) performance.
583 *
584 * @ep: Pointer to the epoll private data structure.
585 * @sproc: Pointer to the scan callback.
586 * @priv: Private opaque data passed to the @sproc callback.
Nelson Elhaged8805e62011-10-31 17:13:14 -0700587 * @depth: The current depth of recursive f_op->poll calls.
Jason Baron67347fe2013-11-12 15:10:18 -0800588 * @ep_locked: caller already holds ep->mtx
Davide Libenzi5071f972009-03-31 15:24:10 -0700589 *
590 * Returns: The same integer error code returned by the @sproc callback.
591 */
592static int ep_scan_ready_list(struct eventpoll *ep,
593 int (*sproc)(struct eventpoll *,
594 struct list_head *, void *),
Jason Baron67347fe2013-11-12 15:10:18 -0800595 void *priv, int depth, bool ep_locked)
Davide Libenzi5071f972009-03-31 15:24:10 -0700596{
597 int error, pwake = 0;
598 unsigned long flags;
599 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700600 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700601
602 /*
603 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700604 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700605 */
Jason Baron67347fe2013-11-12 15:10:18 -0800606
607 if (!ep_locked)
608 mutex_lock_nested(&ep->mtx, depth);
Davide Libenzi5071f972009-03-31 15:24:10 -0700609
610 /*
611 * Steal the ready list, and re-init the original one to the
612 * empty list. Also, set ep->ovflist to NULL so that events
613 * happening while looping w/out locks, are not lost. We cannot
614 * have the poll callback to queue directly on ep->rdllist,
615 * because we want the "sproc" callback to be able to do it
616 * in a lockless way.
617 */
618 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700619 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700620 ep->ovflist = NULL;
621 spin_unlock_irqrestore(&ep->lock, flags);
622
623 /*
624 * Now call the callback function.
625 */
626 error = (*sproc)(ep, &txlist, priv);
627
628 spin_lock_irqsave(&ep->lock, flags);
629 /*
630 * During the time we spent inside the "sproc" callback, some
631 * other events might have been queued by the poll callback.
632 * We re-insert them inside the main ready-list here.
633 */
634 for (nepi = ep->ovflist; (epi = nepi) != NULL;
635 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
636 /*
637 * We need to check if the item is already in the list.
638 * During the "sproc" callback execution time, items are
639 * queued into ->ovflist but the "txlist" might already
640 * contain them, and the list_splice() below takes care of them.
641 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200642 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700643 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -0700644 ep_pm_stay_awake(epi);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200645 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700646 }
647 /*
648 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
649 * releasing the lock, events will be queued in the normal way inside
650 * ep->rdllist.
651 */
652 ep->ovflist = EP_UNACTIVE_PTR;
653
654 /*
655 * Quickly re-inject items left on "txlist".
656 */
657 list_splice(&txlist, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200658 __pm_relax(ep->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700659
660 if (!list_empty(&ep->rdllist)) {
661 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700662 * Wake up (if active) both the eventpoll wait list and
663 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700664 */
665 if (waitqueue_active(&ep->wq))
666 wake_up_locked(&ep->wq);
667 if (waitqueue_active(&ep->poll_wait))
668 pwake++;
669 }
670 spin_unlock_irqrestore(&ep->lock, flags);
671
Jason Baron67347fe2013-11-12 15:10:18 -0800672 if (!ep_locked)
673 mutex_unlock(&ep->mtx);
Davide Libenzi5071f972009-03-31 15:24:10 -0700674
675 /* We have to call this outside the lock */
676 if (pwake)
677 ep_poll_safewake(&ep->poll_wait);
678
679 return error;
680}
681
Jason Baronae10b2b2013-11-12 15:10:16 -0800682static void epi_rcu_free(struct rcu_head *head)
683{
684 struct epitem *epi = container_of(head, struct epitem, rcu);
685 kmem_cache_free(epi_cache, epi);
686}
687
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700689 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700690 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700691 */
692static int ep_remove(struct eventpoll *ep, struct epitem *epi)
693{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700694 unsigned long flags;
695 struct file *file = epi->ffd.file;
696
697 /*
698 * Removes poll wait queue hooks. We _have_ to do this without holding
699 * the "ep->lock" otherwise a deadlock might occur. This because of the
700 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
701 * queue head lock when unregistering the wait queue. The wakeup callback
702 * will run by holding the wait queue head lock and will call our callback
703 * that will try to get "ep->lock".
704 */
705 ep_unregister_pollwait(ep, epi);
706
707 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700708 spin_lock(&file->f_lock);
Jason Baronae10b2b2013-11-12 15:10:16 -0800709 list_del_rcu(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700710 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711
Davide Libenzicdac75e2008-04-29 00:58:34 -0700712 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713
Davide Libenzic7ea7632007-05-15 01:40:47 -0700714 spin_lock_irqsave(&ep->lock, flags);
715 if (ep_is_linked(&epi->rdllink))
716 list_del_init(&epi->rdllink);
717 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
Eric Wongeea1d582013-04-30 15:27:39 -0700719 wakeup_source_unregister(ep_wakeup_source(epi));
Jason Baronae10b2b2013-11-12 15:10:16 -0800720 /*
721 * At this point it is safe to free the eventpoll item. Use the union
722 * field epi->rcu, since we are trying to minimize the size of
723 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
724 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
725 * use of the rbn field.
726 */
727 call_rcu(&epi->rcu, epi_rcu_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728
Robin Holt52bd19f72011-01-12 17:00:01 -0800729 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800730
Davide Libenzic7ea7632007-05-15 01:40:47 -0700731 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732}
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734static void ep_free(struct eventpoll *ep)
735{
736 struct rb_node *rbp;
737 struct epitem *epi;
738
739 /* We need to release all tasks waiting for these file */
740 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700741 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742
743 /*
744 * We need to lock this because we could be hit by
745 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700746 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 * is on the way to be removed and no one has references to it
748 * anymore. The only hit might come from eventpoll_release_file() but
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300749 * holding "epmutex" is sufficient here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800751 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752
753 /*
754 * Walks through the whole tree by unregistering poll callbacks.
755 */
756 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
757 epi = rb_entry(rbp, struct epitem, rbn);
758
759 ep_unregister_pollwait(ep, epi);
Eric Dumazet91cf5ab2013-09-11 14:24:06 -0700760 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 }
762
763 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700764 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700766 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 * us during this operation. So we can avoid the lock on "ep->lock".
Eric Wongddf676c2013-04-30 15:27:40 -0700768 * We do not need to lock ep->mtx, either, we only do it to prevent
769 * a lockdep warning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 */
Eric Wongddf676c2013-04-30 15:27:40 -0700771 mutex_lock(&ep->mtx);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700772 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 epi = rb_entry(rbp, struct epitem, rbn);
774 ep_remove(ep, epi);
Eric Dumazet91cf5ab2013-09-11 14:24:06 -0700775 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 }
Eric Wongddf676c2013-04-30 15:27:40 -0700777 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778
Arjan van de Ven144efe32006-03-23 03:00:32 -0800779 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700780 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800781 free_uid(ep->user);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200782 wakeup_source_unregister(ep->ws);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700783 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784}
785
Davide Libenzi7699acd2007-05-10 22:23:23 -0700786static int ep_eventpoll_release(struct inode *inode, struct file *file)
787{
788 struct eventpoll *ep = file->private_data;
789
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700790 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700791 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700792
Davide Libenzi7699acd2007-05-10 22:23:23 -0700793 return 0;
794}
795
Eric Wong450d89e2013-04-30 15:27:42 -0700796static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
797{
798 pt->_key = epi->event.events;
799
800 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
801}
802
Davide Libenzi296e2362009-03-31 15:24:11 -0700803static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
804 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700805{
806 struct epitem *epi, *tmp;
Hans Verkuil626cf232012-03-23 15:02:27 -0700807 poll_table pt;
Davide Libenzi5071f972009-03-31 15:24:10 -0700808
Hans Verkuil626cf232012-03-23 15:02:27 -0700809 init_poll_funcptr(&pt, NULL);
Eric Wong450d89e2013-04-30 15:27:42 -0700810
Davide Libenzi5071f972009-03-31 15:24:10 -0700811 list_for_each_entry_safe(epi, tmp, head, rdllink) {
Eric Wong450d89e2013-04-30 15:27:42 -0700812 if (ep_item_poll(epi, &pt))
Davide Libenzi5071f972009-03-31 15:24:10 -0700813 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700814 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700815 /*
816 * Item has been dropped into the ready list by the poll
817 * callback, but it's not actually ready, as far as
818 * caller requested events goes. We can remove it here.
819 */
Eric Wongeea1d582013-04-30 15:27:39 -0700820 __pm_relax(ep_wakeup_source(epi));
Davide Libenzi5071f972009-03-31 15:24:10 -0700821 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700822 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700823 }
824
825 return 0;
826}
827
Jason Baron67347fe2013-11-12 15:10:18 -0800828static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
829 poll_table *pt);
830
831struct readyevents_arg {
832 struct eventpoll *ep;
833 bool locked;
834};
835
Davide Libenzi5071f972009-03-31 15:24:10 -0700836static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
837{
Jason Baron67347fe2013-11-12 15:10:18 -0800838 struct readyevents_arg *arg = priv;
839
840 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
841 call_nests + 1, arg->locked);
Davide Libenzi5071f972009-03-31 15:24:10 -0700842}
843
Davide Libenzi7699acd2007-05-10 22:23:23 -0700844static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
845{
Davide Libenzi5071f972009-03-31 15:24:10 -0700846 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700847 struct eventpoll *ep = file->private_data;
Jason Baron67347fe2013-11-12 15:10:18 -0800848 struct readyevents_arg arg;
849
850 /*
851 * During ep_insert() we already hold the ep->mtx for the tfile.
852 * Prevent re-aquisition.
853 */
854 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
855 arg.ep = ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700856
857 /* Insert inside our poll wait queue */
858 poll_wait(file, &ep->poll_wait, wait);
859
Davide Libenzi5071f972009-03-31 15:24:10 -0700860 /*
861 * Proceed to find out if wanted events are really available inside
862 * the ready list. This need to be done under ep_call_nested()
863 * supervision, since the call to f_op->poll() done on listed files
864 * could re-enter here.
865 */
866 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Jason Baron67347fe2013-11-12 15:10:18 -0800867 ep_poll_readyevents_proc, &arg, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700868
Davide Libenzi296e2362009-03-31 15:24:11 -0700869 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700870}
871
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800872#ifdef CONFIG_PROC_FS
Joe Perchesa3816ab2014-09-29 16:08:25 -0700873static void ep_show_fdinfo(struct seq_file *m, struct file *f)
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800874{
875 struct eventpoll *ep = f->private_data;
876 struct rb_node *rbp;
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800877
878 mutex_lock(&ep->mtx);
879 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
880 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
881
Joe Perchesa3816ab2014-09-29 16:08:25 -0700882 seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
883 epi->ffd.fd, epi->event.events,
884 (long long)epi->event.data);
885 if (seq_has_overflowed(m))
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800886 break;
887 }
888 mutex_unlock(&ep->mtx);
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800889}
890#endif
891
Davide Libenzi7699acd2007-05-10 22:23:23 -0700892/* File callbacks that implement the eventpoll file behaviour */
893static const struct file_operations eventpoll_fops = {
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800894#ifdef CONFIG_PROC_FS
895 .show_fdinfo = ep_show_fdinfo,
896#endif
Davide Libenzi7699acd2007-05-10 22:23:23 -0700897 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200898 .poll = ep_eventpoll_poll,
899 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700900};
901
Davide Libenzi7699acd2007-05-10 22:23:23 -0700902/*
903 * This is called from eventpoll_release() to unlink files from the eventpoll
904 * interface. We need to have this facility to cleanup correctly files that are
905 * closed without being removed from the eventpoll interface.
906 */
907void eventpoll_release_file(struct file *file)
908{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700909 struct eventpoll *ep;
Konstantin Khlebnikovebe06182014-06-17 06:58:05 +0400910 struct epitem *epi, *next;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700911
912 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700913 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700914 * necessary. It is not necessary because we're in the "struct file"
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300915 * cleanup path, and this means that no one is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700916 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700917 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700918 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700919 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700920 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700921 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700922 *
923 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700924 */
925 mutex_lock(&epmutex);
Konstantin Khlebnikovebe06182014-06-17 06:58:05 +0400926 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
Davide Libenzi7699acd2007-05-10 22:23:23 -0700927 ep = epi->ep;
Nelson Elhaged8805e62011-10-31 17:13:14 -0700928 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700929 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700930 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700931 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700932 mutex_unlock(&epmutex);
933}
934
935static int ep_alloc(struct eventpoll **pep)
936{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800937 int error;
938 struct user_struct *user;
939 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700940
Davide Libenzi7ef99642008-12-01 13:13:55 -0800941 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800942 error = -ENOMEM;
943 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
944 if (unlikely(!ep))
945 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700946
Davide Libenzic7ea7632007-05-15 01:40:47 -0700947 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700948 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700949 init_waitqueue_head(&ep->wq);
950 init_waitqueue_head(&ep->poll_wait);
951 INIT_LIST_HEAD(&ep->rdllist);
952 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700953 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800954 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700955
956 *pep = ep;
957
Davide Libenzi7699acd2007-05-10 22:23:23 -0700958 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800959
960free_uid:
961 free_uid(user);
962 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700963}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
965/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700966 * Search the file inside the eventpoll tree. The RB tree operations
967 * are protected by the "mtx" mutex, and ep_find() must be called with
968 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 */
970static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
971{
972 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 struct rb_node *rbp;
974 struct epitem *epi, *epir = NULL;
975 struct epoll_filefd ffd;
976
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700977 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 for (rbp = ep->rbr.rb_node; rbp; ) {
979 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700980 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 if (kcmp > 0)
982 rbp = rbp->rb_right;
983 else if (kcmp < 0)
984 rbp = rbp->rb_left;
985 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 epir = epi;
987 break;
988 }
989 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 return epir;
992}
993
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700995 * This is the callback that is passed to the wait queue wakeup
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200996 * mechanism. It is called by the stored file descriptors when they
Davide Libenzi7699acd2007-05-10 22:23:23 -0700997 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700999static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000{
Davide Libenzi7699acd2007-05-10 22:23:23 -07001001 int pwake = 0;
1002 unsigned long flags;
1003 struct epitem *epi = ep_item_from_wait(wait);
1004 struct eventpoll *ep = epi->ep;
Jason Barondf0108c2016-01-20 14:59:24 -08001005 int ewake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
Oleg Nesterov971316f2012-02-24 20:07:29 +01001007 if ((unsigned long)key & POLLFREE) {
1008 ep_pwq_from_wait(wait)->whead = NULL;
1009 /*
1010 * whead = NULL above can race with ep_remove_wait_queue()
1011 * which can do another remove_wait_queue() after us, so we
1012 * can't use __remove_wait_queue(). whead->lock is held by
1013 * the caller.
1014 */
Oleg Nesterovd80e7312012-02-24 20:07:11 +01001015 list_del_init(&wait->task_list);
Oleg Nesterov971316f2012-02-24 20:07:29 +01001016 }
Oleg Nesterovd80e7312012-02-24 20:07:11 +01001017
Davide Libenzic7ea7632007-05-15 01:40:47 -07001018 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001019
1020 /*
1021 * If the event mask does not contain any poll(2) event, we consider the
1022 * descriptor to be disabled. This condition is likely the effect of the
1023 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1024 * until the next EPOLL_CTL_MOD will be issued.
1025 */
1026 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -07001027 goto out_unlock;
1028
1029 /*
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -07001030 * Check the events coming with the callback. At this stage, not
1031 * every device reports the events in the "key" parameter of the
1032 * callback. We need to be able to handle both cases here, hence the
1033 * test for "key" != NULL before the event match test.
1034 */
1035 if (key && !((unsigned long) key & epi->event.events))
1036 goto out_unlock;
1037
1038 /*
Daniel Balutabf6a41d2011-01-30 23:42:29 +02001039 * If we are transferring events to userspace, we can hold no locks
Davide Libenzid47de162007-05-15 01:40:41 -07001040 * (because we're accessing user memory, and because of linux f_op->poll()
Daniel Balutabf6a41d2011-01-30 23:42:29 +02001041 * semantics). All the events that happen during that period of time are
Davide Libenzid47de162007-05-15 01:40:41 -07001042 * chained in ep->ovflist and requeued later on.
1043 */
1044 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1045 if (epi->next == EP_UNACTIVE_PTR) {
1046 epi->next = ep->ovflist;
1047 ep->ovflist = epi;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001048 if (epi->ws) {
1049 /*
1050 * Activate ep->ws since epi->ws may get
1051 * deactivated at any time.
1052 */
1053 __pm_stay_awake(ep->ws);
1054 }
1055
Davide Libenzid47de162007-05-15 01:40:41 -07001056 }
1057 goto out_unlock;
1058 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001059
1060 /* If this file is already in the ready list we exit soon */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001061 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001062 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001063 ep_pm_stay_awake_rcu(epi);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001064 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001065
Davide Libenzi7699acd2007-05-10 22:23:23 -07001066 /*
1067 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1068 * wait list.
1069 */
Jason Barondf0108c2016-01-20 14:59:24 -08001070 if (waitqueue_active(&ep->wq)) {
1071 ewake = 1;
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001072 wake_up_locked(&ep->wq);
Jason Barondf0108c2016-01-20 14:59:24 -08001073 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001074 if (waitqueue_active(&ep->poll_wait))
1075 pwake++;
1076
Davide Libenzid47de162007-05-15 01:40:41 -07001077out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001078 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001079
1080 /* We have to call this outside the lock */
1081 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001082 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001083
Jason Barondf0108c2016-01-20 14:59:24 -08001084 if (epi->event.events & EPOLLEXCLUSIVE)
1085 return ewake;
1086
Davide Libenzi7699acd2007-05-10 22:23:23 -07001087 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088}
1089
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090/*
1091 * This is the callback that is used to add our wait queue to the
1092 * target file wakeup lists.
1093 */
1094static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1095 poll_table *pt)
1096{
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001097 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 struct eppoll_entry *pwq;
1099
Christoph Lametere94b1762006-12-06 20:33:17 -08001100 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1102 pwq->whead = whead;
1103 pwq->base = epi;
Jason Barondf0108c2016-01-20 14:59:24 -08001104 if (epi->event.events & EPOLLEXCLUSIVE)
1105 add_wait_queue_exclusive(whead, &pwq->wait);
1106 else
1107 add_wait_queue(whead, &pwq->wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 list_add_tail(&pwq->llink, &epi->pwqlist);
1109 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001110 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 /* We have to signal that an error occurred */
1112 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -07001113 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114}
1115
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1117{
1118 int kcmp;
1119 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1120 struct epitem *epic;
1121
1122 while (*p) {
1123 parent = *p;
1124 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001125 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 if (kcmp > 0)
1127 p = &parent->rb_right;
1128 else
1129 p = &parent->rb_left;
1130 }
1131 rb_link_node(&epi->rbn, parent, p);
1132 rb_insert_color(&epi->rbn, &ep->rbr);
1133}
1134
Andrew Mortona80a6b82012-11-08 15:53:35 -08001135
1136
Jason Baron28d82dc2012-01-12 17:17:43 -08001137#define PATH_ARR_SIZE 5
1138/*
1139 * These are the number paths of length 1 to 5, that we are allowing to emanate
1140 * from a single file of interest. For example, we allow 1000 paths of length
1141 * 1, to emanate from each file of interest. This essentially represents the
1142 * potential wakeup paths, which need to be limited in order to avoid massive
1143 * uncontrolled wakeup storms. The common use case should be a single ep which
1144 * is connected to n file sources. In this case each file source has 1 path
1145 * of length 1. Thus, the numbers below should be more than sufficient. These
1146 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1147 * and delete can't add additional paths. Protected by the epmutex.
1148 */
1149static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1150static int path_count[PATH_ARR_SIZE];
1151
1152static int path_count_inc(int nests)
1153{
Jason Baron93dc6102012-03-16 16:34:03 -04001154 /* Allow an arbitrary number of depth 1 paths */
1155 if (nests == 0)
1156 return 0;
1157
Jason Baron28d82dc2012-01-12 17:17:43 -08001158 if (++path_count[nests] > path_limits[nests])
1159 return -1;
1160 return 0;
1161}
1162
1163static void path_count_init(void)
1164{
1165 int i;
1166
1167 for (i = 0; i < PATH_ARR_SIZE; i++)
1168 path_count[i] = 0;
1169}
1170
1171static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1172{
1173 int error = 0;
1174 struct file *file = priv;
1175 struct file *child_file;
1176 struct epitem *epi;
1177
Jason Baronae10b2b2013-11-12 15:10:16 -08001178 /* CTL_DEL can remove links here, but that can't increase our count */
1179 rcu_read_lock();
1180 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001181 child_file = epi->ep->file;
1182 if (is_file_epoll(child_file)) {
1183 if (list_empty(&child_file->f_ep_links)) {
1184 if (path_count_inc(call_nests)) {
1185 error = -1;
1186 break;
1187 }
1188 } else {
1189 error = ep_call_nested(&poll_loop_ncalls,
1190 EP_MAX_NESTS,
1191 reverse_path_check_proc,
1192 child_file, child_file,
1193 current);
1194 }
1195 if (error != 0)
1196 break;
1197 } else {
1198 printk(KERN_ERR "reverse_path_check_proc: "
1199 "file is not an ep!\n");
1200 }
1201 }
Jason Baronae10b2b2013-11-12 15:10:16 -08001202 rcu_read_unlock();
Jason Baron28d82dc2012-01-12 17:17:43 -08001203 return error;
1204}
1205
1206/**
1207 * reverse_path_check - The tfile_check_list is list of file *, which have
1208 * links that are proposed to be newly added. We need to
1209 * make sure that those added links don't add too many
1210 * paths such that we will spend all our time waking up
1211 * eventpoll objects.
1212 *
1213 * Returns: Returns zero if the proposed links don't create too many paths,
1214 * -1 otherwise.
1215 */
1216static int reverse_path_check(void)
1217{
Jason Baron28d82dc2012-01-12 17:17:43 -08001218 int error = 0;
1219 struct file *current_file;
1220
1221 /* let's call this for all tfiles */
1222 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001223 path_count_init();
1224 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1225 reverse_path_check_proc, current_file,
1226 current_file, current);
1227 if (error)
1228 break;
1229 }
1230 return error;
1231}
1232
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001233static int ep_create_wakeup_source(struct epitem *epi)
1234{
1235 const char *name;
Eric Wongeea1d582013-04-30 15:27:39 -07001236 struct wakeup_source *ws;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001237
1238 if (!epi->ep->ws) {
1239 epi->ep->ws = wakeup_source_register("eventpoll");
1240 if (!epi->ep->ws)
1241 return -ENOMEM;
1242 }
1243
1244 name = epi->ffd.file->f_path.dentry->d_name.name;
Eric Wongeea1d582013-04-30 15:27:39 -07001245 ws = wakeup_source_register(name);
1246
1247 if (!ws)
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001248 return -ENOMEM;
Eric Wongeea1d582013-04-30 15:27:39 -07001249 rcu_assign_pointer(epi->ws, ws);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001250
1251 return 0;
1252}
1253
Eric Wongeea1d582013-04-30 15:27:39 -07001254/* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1255static noinline void ep_destroy_wakeup_source(struct epitem *epi)
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001256{
Eric Wongeea1d582013-04-30 15:27:39 -07001257 struct wakeup_source *ws = ep_wakeup_source(epi);
1258
Eric Wongd6d67e72013-04-30 15:27:43 -07001259 RCU_INIT_POINTER(epi->ws, NULL);
Eric Wongeea1d582013-04-30 15:27:39 -07001260
1261 /*
1262 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1263 * used internally by wakeup_source_remove, too (called by
1264 * wakeup_source_unregister), so we cannot use call_rcu
1265 */
1266 synchronize_rcu();
1267 wakeup_source_unregister(ws);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001268}
1269
Davide Libenzic7ea7632007-05-15 01:40:47 -07001270/*
1271 * Must be called with "mtx" held.
1272 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
Jason Baron67347fe2013-11-12 15:10:18 -08001274 struct file *tfile, int fd, int full_check)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275{
1276 int error, revents, pwake = 0;
1277 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -08001278 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 struct epitem *epi;
1280 struct ep_pqueue epq;
1281
Robin Holt52bd19f72011-01-12 17:00:01 -08001282 user_watches = atomic_long_read(&ep->user->epoll_watches);
1283 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001284 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -08001285 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001286 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
1288 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 INIT_LIST_HEAD(&epi->rdllink);
1290 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 INIT_LIST_HEAD(&epi->pwqlist);
1292 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001293 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -07001296 epi->next = EP_UNACTIVE_PTR;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001297 if (epi->event.events & EPOLLWAKEUP) {
1298 error = ep_create_wakeup_source(epi);
1299 if (error)
1300 goto error_create_wakeup_source;
1301 } else {
Eric Wongeea1d582013-04-30 15:27:39 -07001302 RCU_INIT_POINTER(epi->ws, NULL);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001303 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304
1305 /* Initialize the poll table using the queue callback */
1306 epq.epi = epi;
1307 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1308
1309 /*
1310 * Attach the item to the poll hooks and get current event bits.
1311 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -07001312 * been increased by the caller of this function. Note that after
1313 * this operation completes, the poll callback can start hitting
1314 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 */
Eric Wong450d89e2013-04-30 15:27:42 -07001316 revents = ep_item_poll(epi, &epq.pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317
1318 /*
1319 * We have to check if something went wrong during the poll wait queue
1320 * install process. Namely an allocation for a wait queue failed due
1321 * high memory pressure.
1322 */
Davide Libenzi7ef99642008-12-01 13:13:55 -08001323 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001325 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
1327 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -07001328 spin_lock(&tfile->f_lock);
Jason Baronae10b2b2013-11-12 15:10:16 -08001329 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -07001330 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331
Davide Libenzic7ea7632007-05-15 01:40:47 -07001332 /*
1333 * Add the current item to the RB tree. All RB tree operations are
1334 * protected by "mtx", and ep_insert() is called with "mtx" held.
1335 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 ep_rbtree_insert(ep, epi);
1337
Jason Baron28d82dc2012-01-12 17:17:43 -08001338 /* now check if we've created too many backpaths */
1339 error = -EINVAL;
Jason Baron67347fe2013-11-12 15:10:18 -08001340 if (full_check && reverse_path_check())
Jason Baron28d82dc2012-01-12 17:17:43 -08001341 goto error_remove_epi;
1342
Davide Libenzic7ea7632007-05-15 01:40:47 -07001343 /* We have to drop the new item inside our item list to keep track of it */
1344 spin_lock_irqsave(&ep->lock, flags);
1345
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001347 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001349 ep_pm_stay_awake(epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350
1351 /* Notify waiting tasks that events are available */
1352 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001353 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 if (waitqueue_active(&ep->poll_wait))
1355 pwake++;
1356 }
1357
Davide Libenzic7ea7632007-05-15 01:40:47 -07001358 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
Robin Holt52bd19f72011-01-12 17:00:01 -08001360 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001361
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 /* We have to call this outside the lock */
1363 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001364 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 return 0;
1367
Jason Baron28d82dc2012-01-12 17:17:43 -08001368error_remove_epi:
1369 spin_lock(&tfile->f_lock);
Jason Baronae10b2b2013-11-12 15:10:16 -08001370 list_del_rcu(&epi->fllink);
Jason Baron28d82dc2012-01-12 17:17:43 -08001371 spin_unlock(&tfile->f_lock);
1372
1373 rb_erase(&epi->rbn, &ep->rbr);
1374
Davide Libenzi7699acd2007-05-10 22:23:23 -07001375error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 ep_unregister_pollwait(ep, epi);
1377
1378 /*
1379 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -07001380 * allocated wait queue. Note that we don't care about the ep->ovflist
1381 * list, since that is used/cleaned only inside a section bound by "mtx".
1382 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001384 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001385 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001386 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001387 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
Eric Wongeea1d582013-04-30 15:27:39 -07001389 wakeup_source_unregister(ep_wakeup_source(epi));
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001390
1391error_create_wakeup_source:
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001392 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 return error;
1395}
1396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397/*
1398 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001399 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 */
1401static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1402{
1403 int pwake = 0;
1404 unsigned int revents;
Hans Verkuil626cf232012-03-23 15:02:27 -07001405 poll_table pt;
1406
1407 init_poll_funcptr(&pt, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408
1409 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001410 * Set the new event interest mask before calling f_op->poll();
1411 * otherwise we might miss an event that happens between the
1412 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 */
Eric Wong128dd172013-01-01 21:20:27 +00001414 epi->event.events = event->events; /* need barrier below */
Tony Battersbye057e152009-03-31 15:24:15 -07001415 epi->event.data = event->data; /* protected by mtx */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001416 if (epi->event.events & EPOLLWAKEUP) {
Eric Wongeea1d582013-04-30 15:27:39 -07001417 if (!ep_has_wakeup_source(epi))
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001418 ep_create_wakeup_source(epi);
Eric Wongeea1d582013-04-30 15:27:39 -07001419 } else if (ep_has_wakeup_source(epi)) {
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001420 ep_destroy_wakeup_source(epi);
1421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422
1423 /*
Eric Wong128dd172013-01-01 21:20:27 +00001424 * The following barrier has two effects:
1425 *
1426 * 1) Flush epi changes above to other CPUs. This ensures
1427 * we do not miss events from ep_poll_callback if an
1428 * event occurs immediately after we call f_op->poll().
1429 * We need this because we did not take ep->lock while
1430 * changing epi above (but ep_poll_callback does take
1431 * ep->lock).
1432 *
1433 * 2) We also need to ensure we do not miss _past_ events
1434 * when calling f_op->poll(). This barrier also
1435 * pairs with the barrier in wq_has_sleeper (see
1436 * comments for wq_has_sleeper).
1437 *
1438 * This barrier will now guarantee ep_poll_callback or f_op->poll
1439 * (or both) will notice the readiness of an item.
1440 */
1441 smp_mb();
1442
1443 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 * Get current event bits. We can safely use the file* here because
1445 * its usage count has been increased by the caller of this function.
1446 */
Eric Wong450d89e2013-04-30 15:27:42 -07001447 revents = ep_item_poll(epi, &pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001450 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001451 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001453 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001454 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001455 if (!ep_is_linked(&epi->rdllink)) {
1456 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001457 ep_pm_stay_awake(epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458
Davide Libenzic7ea7632007-05-15 01:40:47 -07001459 /* Notify waiting tasks that events are available */
1460 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001461 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001462 if (waitqueue_active(&ep->poll_wait))
1463 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 }
Tony Battersbye057e152009-03-31 15:24:15 -07001465 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467
1468 /* We have to call this outside the lock */
1469 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001470 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471
1472 return 0;
1473}
1474
Davide Libenzi296e2362009-03-31 15:24:11 -07001475static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1476 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001477{
1478 struct ep_send_events_data *esed = priv;
1479 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001480 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001481 struct epitem *epi;
1482 struct epoll_event __user *uevent;
Eric Wongeea1d582013-04-30 15:27:39 -07001483 struct wakeup_source *ws;
Hans Verkuil626cf232012-03-23 15:02:27 -07001484 poll_table pt;
1485
1486 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -07001487
Davide Libenzi296e2362009-03-31 15:24:11 -07001488 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001489 * We can loop without lock because we are passed a task private list.
1490 * Items cannot vanish during the loop because ep_scan_ready_list() is
1491 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001492 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001493 for (eventcnt = 0, uevent = esed->events;
1494 !list_empty(head) && eventcnt < esed->maxevents;) {
1495 epi = list_first_entry(head, struct epitem, rdllink);
1496
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001497 /*
1498 * Activate ep->ws before deactivating epi->ws to prevent
1499 * triggering auto-suspend here (in case we reactive epi->ws
1500 * below).
1501 *
1502 * This could be rearranged to delay the deactivation of epi->ws
1503 * instead, but then epi->ws would temporarily be out of sync
1504 * with ep_is_linked().
1505 */
Eric Wongeea1d582013-04-30 15:27:39 -07001506 ws = ep_wakeup_source(epi);
1507 if (ws) {
1508 if (ws->active)
1509 __pm_stay_awake(ep->ws);
1510 __pm_relax(ws);
1511 }
1512
Davide Libenzi5071f972009-03-31 15:24:10 -07001513 list_del_init(&epi->rdllink);
1514
Eric Wong450d89e2013-04-30 15:27:42 -07001515 revents = ep_item_poll(epi, &pt);
Davide Libenzi5071f972009-03-31 15:24:10 -07001516
Davide Libenzi296e2362009-03-31 15:24:11 -07001517 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001518 * If the event mask intersect the caller-requested one,
1519 * deliver the event to userspace. Again, ep_scan_ready_list()
1520 * is holding "mtx", so no operations coming from userspace
1521 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001522 */
1523 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001524 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001525 __put_user(epi->event.data, &uevent->data)) {
1526 list_add(&epi->rdllink, head);
Eric Wongeea1d582013-04-30 15:27:39 -07001527 ep_pm_stay_awake(epi);
Davide Libenzi296e2362009-03-31 15:24:11 -07001528 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001529 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001530 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001531 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001532 if (epi->event.events & EPOLLONESHOT)
1533 epi->event.events &= EP_PRIVATE_BITS;
1534 else if (!(epi->event.events & EPOLLET)) {
1535 /*
1536 * If this file has been added with Level
1537 * Trigger mode, we need to insert back inside
1538 * the ready list, so that the next call to
1539 * epoll_wait() will check again the events
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001540 * availability. At this point, no one can insert
Davide Libenzi296e2362009-03-31 15:24:11 -07001541 * into ep->rdllist besides us. The epoll_ctl()
1542 * callers are locked out by
1543 * ep_scan_ready_list() holding "mtx" and the
1544 * poll callback will queue them in ep->ovflist.
1545 */
1546 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001547 ep_pm_stay_awake(epi);
Davide Libenzi296e2362009-03-31 15:24:11 -07001548 }
1549 }
1550 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001551
1552 return eventcnt;
1553}
1554
Davide Libenzi296e2362009-03-31 15:24:11 -07001555static int ep_send_events(struct eventpoll *ep,
1556 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557{
Davide Libenzi5071f972009-03-31 15:24:10 -07001558 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001559
Davide Libenzi5071f972009-03-31 15:24:10 -07001560 esed.maxevents = maxevents;
1561 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001562
Jason Baron67347fe2013-11-12 15:10:18 -08001563 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564}
1565
Eric Dumazet0781b902011-02-01 15:52:35 -08001566static inline struct timespec ep_set_mstimeout(long ms)
1567{
1568 struct timespec now, ts = {
1569 .tv_sec = ms / MSEC_PER_SEC,
1570 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1571 };
1572
1573 ktime_get_ts(&now);
1574 return timespec_add_safe(now, ts);
1575}
1576
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001577/**
1578 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1579 * event buffer.
1580 *
1581 * @ep: Pointer to the eventpoll context.
1582 * @events: Pointer to the userspace buffer where the ready events should be
1583 * stored.
1584 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1585 * @timeout: Maximum timeout for the ready events fetch operation, in
1586 * milliseconds. If the @timeout is zero, the function will not block,
1587 * while if the @timeout is less than zero, the function will block
1588 * until at least one event has been retrieved (or an error
1589 * occurred).
1590 *
1591 * Returns: Returns the number of ready events which have been fetched, or an
1592 * error code, in case of error.
1593 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1595 int maxevents, long timeout)
1596{
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001597 int res = 0, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 unsigned long flags;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001599 long slack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001601 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001603 if (timeout > 0) {
Eric Dumazet0781b902011-02-01 15:52:35 -08001604 struct timespec end_time = ep_set_mstimeout(timeout);
1605
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001606 slack = select_estimate_accuracy(&end_time);
1607 to = &expires;
1608 *to = timespec_to_ktime(end_time);
1609 } else if (timeout == 0) {
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001610 /*
1611 * Avoid the unnecessary trip to the wait queue loop, if the
1612 * caller specified a non blocking operation.
1613 */
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001614 timed_out = 1;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001615 spin_lock_irqsave(&ep->lock, flags);
1616 goto check_events;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001617 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001619fetch_events:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001620 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001622 if (!ep_events_available(ep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 /*
1624 * We don't have any available event to return to the caller.
1625 * We need to sleep here, and we will be wake up by
1626 * ep_poll_callback() when events will become available.
1627 */
1628 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001629 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630
1631 for (;;) {
1632 /*
1633 * We don't want to sleep if the ep_poll_callback() sends us
1634 * a wakeup in between. That's why we set the task state
1635 * to TASK_INTERRUPTIBLE before doing the checks.
1636 */
1637 set_current_state(TASK_INTERRUPTIBLE);
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001638 if (ep_events_available(ep) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 break;
1640 if (signal_pending(current)) {
1641 res = -EINTR;
1642 break;
1643 }
1644
Davide Libenzic7ea7632007-05-15 01:40:47 -07001645 spin_unlock_irqrestore(&ep->lock, flags);
Rafael J. Wysockic5118512013-10-29 13:12:56 +01001646 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001647 timed_out = 1;
1648
Davide Libenzic7ea7632007-05-15 01:40:47 -07001649 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651
Davidlohr Bueso4d5755b2015-02-13 14:39:08 -08001652 __remove_wait_queue(&ep->wq, &wait);
1653 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 }
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001655check_events:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 /* Is it worth to try to dig for events ? */
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001657 eavail = ep_events_available(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658
Davide Libenzic7ea7632007-05-15 01:40:47 -07001659 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660
1661 /*
1662 * Try to transfer events to user space. In case we get 0 events and
1663 * there's still timeout left over, we go trying again in search of
1664 * more luck.
1665 */
1666 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001667 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001668 goto fetch_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669
1670 return res;
1671}
1672
Davide Libenzi22bacca2011-02-25 14:44:12 -08001673/**
1674 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1675 * API, to verify that adding an epoll file inside another
1676 * epoll structure, does not violate the constraints, in
1677 * terms of closed loops, or too deep chains (which can
1678 * result in excessive stack usage).
1679 *
1680 * @priv: Pointer to the epoll file to be currently checked.
1681 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1682 * data structure pointer.
1683 * @call_nests: Current dept of the @ep_call_nested() call stack.
1684 *
1685 * Returns: Returns zero if adding the epoll @file inside current epoll
1686 * structure @ep does not violate the constraints, or -1 otherwise.
1687 */
1688static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1689{
1690 int error = 0;
1691 struct file *file = priv;
1692 struct eventpoll *ep = file->private_data;
Jason Baron28d82dc2012-01-12 17:17:43 -08001693 struct eventpoll *ep_tovisit;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001694 struct rb_node *rbp;
1695 struct epitem *epi;
1696
Nelson Elhaged8805e62011-10-31 17:13:14 -07001697 mutex_lock_nested(&ep->mtx, call_nests + 1);
Jason Baron28d82dc2012-01-12 17:17:43 -08001698 ep->visited = 1;
1699 list_add(&ep->visited_list_link, &visited_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001700 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1701 epi = rb_entry(rbp, struct epitem, rbn);
1702 if (unlikely(is_file_epoll(epi->ffd.file))) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001703 ep_tovisit = epi->ffd.file->private_data;
1704 if (ep_tovisit->visited)
1705 continue;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001706 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Jason Baron28d82dc2012-01-12 17:17:43 -08001707 ep_loop_check_proc, epi->ffd.file,
1708 ep_tovisit, current);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001709 if (error != 0)
1710 break;
Jason Baron28d82dc2012-01-12 17:17:43 -08001711 } else {
1712 /*
1713 * If we've reached a file that is not associated with
1714 * an ep, then we need to check if the newly added
1715 * links are going to add too many wakeup paths. We do
1716 * this by adding it to the tfile_check_list, if it's
1717 * not already there, and calling reverse_path_check()
1718 * during ep_insert().
1719 */
1720 if (list_empty(&epi->ffd.file->f_tfile_llink))
1721 list_add(&epi->ffd.file->f_tfile_llink,
1722 &tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001723 }
1724 }
1725 mutex_unlock(&ep->mtx);
1726
1727 return error;
1728}
1729
1730/**
1731 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1732 * another epoll file (represented by @ep) does not create
1733 * closed loops or too deep chains.
1734 *
1735 * @ep: Pointer to the epoll private data structure.
1736 * @file: Pointer to the epoll file to be checked.
1737 *
1738 * Returns: Returns zero if adding the epoll @file inside current epoll
1739 * structure @ep does not violate the constraints, or -1 otherwise.
1740 */
1741static int ep_loop_check(struct eventpoll *ep, struct file *file)
1742{
Jason Baron28d82dc2012-01-12 17:17:43 -08001743 int ret;
1744 struct eventpoll *ep_cur, *ep_next;
1745
1746 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Davide Libenzi22bacca2011-02-25 14:44:12 -08001747 ep_loop_check_proc, file, ep, current);
Jason Baron28d82dc2012-01-12 17:17:43 -08001748 /* clear visited list */
1749 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1750 visited_list_link) {
1751 ep_cur->visited = 0;
1752 list_del(&ep_cur->visited_list_link);
1753 }
1754 return ret;
1755}
1756
1757static void clear_tfile_check_list(void)
1758{
1759 struct file *file;
1760
1761 /* first clear the tfile_check_list */
1762 while (!list_empty(&tfile_check_list)) {
1763 file = list_first_entry(&tfile_check_list, struct file,
1764 f_tfile_llink);
1765 list_del_init(&file->f_tfile_llink);
1766 }
1767 INIT_LIST_HEAD(&tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001768}
1769
Davide Libenzi7699acd2007-05-10 22:23:23 -07001770/*
Andrew Morton523723b2008-08-12 15:09:01 -07001771 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001772 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001773SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001774{
Jason Baron28d82dc2012-01-12 17:17:43 -08001775 int error, fd;
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001776 struct eventpoll *ep = NULL;
Jason Baron28d82dc2012-01-12 17:17:43 -08001777 struct file *file;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001778
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001779 /* Check the EPOLL_* constant for consistency. */
1780 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1781
Davide Libenzi296e2362009-03-31 15:24:11 -07001782 if (flags & ~EPOLL_CLOEXEC)
1783 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001784 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001785 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001786 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001787 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001788 if (error < 0)
1789 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001790 /*
1791 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001792 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001793 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001794 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1795 if (fd < 0) {
1796 error = fd;
1797 goto out_free_ep;
1798 }
1799 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001800 O_RDWR | (flags & O_CLOEXEC));
Jason Baron28d82dc2012-01-12 17:17:43 -08001801 if (IS_ERR(file)) {
1802 error = PTR_ERR(file);
1803 goto out_free_fd;
1804 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001805 ep->file = file;
Al Viro98022742012-08-17 22:42:36 -04001806 fd_install(fd, file);
Jason Baron28d82dc2012-01-12 17:17:43 -08001807 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001808
Jason Baron28d82dc2012-01-12 17:17:43 -08001809out_free_fd:
1810 put_unused_fd(fd);
1811out_free_ep:
1812 ep_free(ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001813 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001814}
1815
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001816SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001817{
Davide Libenzibfe38912009-05-12 13:19:44 -07001818 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001819 return -EINVAL;
1820
1821 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001822}
1823
Davide Libenzi7699acd2007-05-10 22:23:23 -07001824/*
1825 * The following function implements the controller interface for
1826 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001827 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001828 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001829SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1830 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001831{
1832 int error;
Jason Baron67347fe2013-11-12 15:10:18 -08001833 int full_check = 0;
Al Viro7e3fb582013-08-30 12:47:21 -04001834 struct fd f, tf;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001835 struct eventpoll *ep;
1836 struct epitem *epi;
1837 struct epoll_event epds;
Jason Baron67347fe2013-11-12 15:10:18 -08001838 struct eventpoll *tep = NULL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001839
Davide Libenzi7699acd2007-05-10 22:23:23 -07001840 error = -EFAULT;
1841 if (ep_op_has_event(op) &&
1842 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1843 goto error_return;
1844
Davide Libenzi7699acd2007-05-10 22:23:23 -07001845 error = -EBADF;
Al Viro7e3fb582013-08-30 12:47:21 -04001846 f = fdget(epfd);
1847 if (!f.file)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001848 goto error_return;
1849
1850 /* Get the "struct file *" for the target file */
Al Viro7e3fb582013-08-30 12:47:21 -04001851 tf = fdget(fd);
1852 if (!tf.file)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001853 goto error_fput;
1854
1855 /* The target file descriptor must support poll */
1856 error = -EPERM;
Al Viro72c2d532013-09-22 16:27:52 -04001857 if (!tf.file->f_op->poll)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001858 goto error_tgt_fput;
1859
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001860 /* Check if EPOLLWAKEUP is allowed */
Nicolas Ioossc680e412014-09-09 14:50:51 -07001861 if (ep_op_has_event(op))
1862 ep_take_care_of_epollwakeup(&epds);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001863
Davide Libenzi7699acd2007-05-10 22:23:23 -07001864 /*
1865 * We have to check that the file structure underneath the file descriptor
1866 * the user passed to us _is_ an eventpoll file. And also we do not permit
1867 * adding an epoll file descriptor inside itself.
1868 */
1869 error = -EINVAL;
Al Viro7e3fb582013-08-30 12:47:21 -04001870 if (f.file == tf.file || !is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07001871 goto error_tgt_fput;
1872
1873 /*
Jason Barondf0108c2016-01-20 14:59:24 -08001874 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
1875 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
1876 * Also, we do not currently supported nested exclusive wakeups.
1877 */
1878 if ((epds.events & EPOLLEXCLUSIVE) && (op == EPOLL_CTL_MOD ||
1879 (op == EPOLL_CTL_ADD && is_file_epoll(tf.file))))
1880 goto error_tgt_fput;
1881
1882 /*
Davide Libenzi7699acd2007-05-10 22:23:23 -07001883 * At this point it is safe to assume that the "private_data" contains
1884 * our own data structure.
1885 */
Al Viro7e3fb582013-08-30 12:47:21 -04001886 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001887
Davide Libenzi22bacca2011-02-25 14:44:12 -08001888 /*
1889 * When we insert an epoll file descriptor, inside another epoll file
1890 * descriptor, there is the change of creating closed loops, which are
Jason Baron28d82dc2012-01-12 17:17:43 -08001891 * better be handled here, than in more critical paths. While we are
1892 * checking for loops we also determine the list of files reachable
1893 * and hang them on the tfile_check_list, so we can check that we
1894 * haven't created too many possible wakeup paths.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001895 *
Jason Baron67347fe2013-11-12 15:10:18 -08001896 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1897 * the epoll file descriptor is attaching directly to a wakeup source,
1898 * unless the epoll file descriptor is nested. The purpose of taking the
1899 * 'epmutex' on add is to prevent complex toplogies such as loops and
1900 * deep wakeup paths from forming in parallel through multiple
1901 * EPOLL_CTL_ADD operations.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001902 */
Nelson Elhaged8805e62011-10-31 17:13:14 -07001903 mutex_lock_nested(&ep->mtx, 0);
Jason Baron67347fe2013-11-12 15:10:18 -08001904 if (op == EPOLL_CTL_ADD) {
1905 if (!list_empty(&f.file->f_ep_links) ||
1906 is_file_epoll(tf.file)) {
1907 full_check = 1;
1908 mutex_unlock(&ep->mtx);
1909 mutex_lock(&epmutex);
1910 if (is_file_epoll(tf.file)) {
1911 error = -ELOOP;
1912 if (ep_loop_check(ep, tf.file) != 0) {
1913 clear_tfile_check_list();
1914 goto error_tgt_fput;
1915 }
1916 } else
1917 list_add(&tf.file->f_tfile_llink,
1918 &tfile_check_list);
1919 mutex_lock_nested(&ep->mtx, 0);
1920 if (is_file_epoll(tf.file)) {
1921 tep = tf.file->private_data;
1922 mutex_lock_nested(&tep->mtx, 1);
1923 }
1924 }
1925 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001926
Davide Libenzi67647d02007-05-15 01:40:52 -07001927 /*
1928 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1929 * above, we can be sure to be able to use the item looked up by
1930 * ep_find() till we release the mutex.
1931 */
Al Viro7e3fb582013-08-30 12:47:21 -04001932 epi = ep_find(ep, tf.file, fd);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001933
1934 error = -EINVAL;
1935 switch (op) {
1936 case EPOLL_CTL_ADD:
1937 if (!epi) {
1938 epds.events |= POLLERR | POLLHUP;
Jason Baron67347fe2013-11-12 15:10:18 -08001939 error = ep_insert(ep, &epds, tf.file, fd, full_check);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001940 } else
1941 error = -EEXIST;
Jason Baron67347fe2013-11-12 15:10:18 -08001942 if (full_check)
1943 clear_tfile_check_list();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001944 break;
1945 case EPOLL_CTL_DEL:
1946 if (epi)
1947 error = ep_remove(ep, epi);
1948 else
1949 error = -ENOENT;
1950 break;
1951 case EPOLL_CTL_MOD:
1952 if (epi) {
1953 epds.events |= POLLERR | POLLHUP;
1954 error = ep_modify(ep, epi, &epds);
1955 } else
1956 error = -ENOENT;
1957 break;
1958 }
Jason Baron67347fe2013-11-12 15:10:18 -08001959 if (tep != NULL)
1960 mutex_unlock(&tep->mtx);
Davide Libenzid47de162007-05-15 01:40:41 -07001961 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001962
1963error_tgt_fput:
Jason Baron67347fe2013-11-12 15:10:18 -08001964 if (full_check)
Davide Libenzi22bacca2011-02-25 14:44:12 -08001965 mutex_unlock(&epmutex);
1966
Al Viro7e3fb582013-08-30 12:47:21 -04001967 fdput(tf);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001968error_fput:
Al Viro7e3fb582013-08-30 12:47:21 -04001969 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001970error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001971
1972 return error;
1973}
1974
1975/*
1976 * Implement the event wait interface for the eventpoll file. It is the kernel
1977 * part of the user space epoll_wait(2).
1978 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001979SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1980 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001981{
Al Viro2903ff02012-08-28 12:52:22 -04001982 int error;
1983 struct fd f;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001984 struct eventpoll *ep;
1985
Davide Libenzi7699acd2007-05-10 22:23:23 -07001986 /* The maximum number of event must be greater than zero */
1987 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1988 return -EINVAL;
1989
1990 /* Verify that the area passed by the user is writeable */
Al Viro2903ff02012-08-28 12:52:22 -04001991 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1992 return -EFAULT;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001993
1994 /* Get the "struct file *" for the eventpoll file */
Al Viro2903ff02012-08-28 12:52:22 -04001995 f = fdget(epfd);
1996 if (!f.file)
1997 return -EBADF;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001998
1999 /*
2000 * We have to check that the file structure underneath the fd
2001 * the user passed to us _is_ an eventpoll file.
2002 */
2003 error = -EINVAL;
Al Viro2903ff02012-08-28 12:52:22 -04002004 if (!is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07002005 goto error_fput;
2006
2007 /*
2008 * At this point it is safe to assume that the "private_data" contains
2009 * our own data structure.
2010 */
Al Viro2903ff02012-08-28 12:52:22 -04002011 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07002012
2013 /* Time to fish for events ... */
2014 error = ep_poll(ep, events, maxevents, timeout);
2015
2016error_fput:
Al Viro2903ff02012-08-28 12:52:22 -04002017 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002018 return error;
2019}
2020
Davide Libenzi7699acd2007-05-10 22:23:23 -07002021/*
2022 * Implement the event wait interface for the eventpoll file. It is the kernel
2023 * part of the user space epoll_pwait(2).
2024 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01002025SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2026 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2027 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07002028{
2029 int error;
2030 sigset_t ksigmask, sigsaved;
2031
2032 /*
2033 * If the caller wants a certain signal mask to be set during the wait,
2034 * we apply it here.
2035 */
2036 if (sigmask) {
2037 if (sigsetsize != sizeof(sigset_t))
2038 return -EINVAL;
2039 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2040 return -EFAULT;
Oleg Nesterov77d55912013-07-03 15:08:14 -07002041 sigsaved = current->blocked;
2042 set_current_blocked(&ksigmask);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002043 }
2044
2045 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2046
2047 /*
2048 * If we changed the signal mask, we need to restore the original one.
2049 * In case we've got a signal while waiting, we do not restore the
2050 * signal mask yet, and we allow do_signal() to deliver the signal on
2051 * the way back to userspace, before the signal mask is restored.
2052 */
2053 if (sigmask) {
2054 if (error == -EINTR) {
2055 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07002056 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07002057 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07002058 } else
Oleg Nesterov77d55912013-07-03 15:08:14 -07002059 set_current_blocked(&sigsaved);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002060 }
2061
2062 return error;
2063}
2064
Al Viro35280bd2013-02-24 14:52:17 -05002065#ifdef CONFIG_COMPAT
2066COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2067 struct epoll_event __user *, events,
2068 int, maxevents, int, timeout,
2069 const compat_sigset_t __user *, sigmask,
2070 compat_size_t, sigsetsize)
2071{
2072 long err;
2073 compat_sigset_t csigmask;
2074 sigset_t ksigmask, sigsaved;
2075
2076 /*
2077 * If the caller wants a certain signal mask to be set during the wait,
2078 * we apply it here.
2079 */
2080 if (sigmask) {
2081 if (sigsetsize != sizeof(compat_sigset_t))
2082 return -EINVAL;
2083 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2084 return -EFAULT;
2085 sigset_from_compat(&ksigmask, &csigmask);
Oleg Nesterov77d55912013-07-03 15:08:14 -07002086 sigsaved = current->blocked;
2087 set_current_blocked(&ksigmask);
Al Viro35280bd2013-02-24 14:52:17 -05002088 }
2089
2090 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2091
2092 /*
2093 * If we changed the signal mask, we need to restore the original one.
2094 * In case we've got a signal while waiting, we do not restore the
2095 * signal mask yet, and we allow do_signal() to deliver the signal on
2096 * the way back to userspace, before the signal mask is restored.
2097 */
2098 if (sigmask) {
2099 if (err == -EINTR) {
2100 memcpy(&current->saved_sigmask, &sigsaved,
2101 sizeof(sigsaved));
2102 set_restore_sigmask();
2103 } else
Oleg Nesterov77d55912013-07-03 15:08:14 -07002104 set_current_blocked(&sigsaved);
Al Viro35280bd2013-02-24 14:52:17 -05002105 }
2106
2107 return err;
2108}
2109#endif
2110
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111static int __init eventpoll_init(void)
2112{
Davide Libenzi7ef99642008-12-01 13:13:55 -08002113 struct sysinfo si;
2114
2115 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08002116 /*
2117 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2118 */
2119 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08002120 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08002121 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122
Davide Libenzi22bacca2011-02-25 14:44:12 -08002123 /*
2124 * Initialize the structure used to perform epoll file descriptor
2125 * inclusion loops checks.
2126 */
2127 ep_nested_calls_init(&poll_loop_ncalls);
2128
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07002130 ep_nested_calls_init(&poll_safewake_ncalls);
2131
2132 /* Initialize the structure used to perform file's f_op->poll() calls */
2133 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134
Eric Wong39732ca2013-04-30 15:27:38 -07002135 /*
2136 * We can have many thousands of epitems, so prevent this from
2137 * using an extra cache line on 64-bit (and smaller) CPUs
2138 */
2139 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2140
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 /* Allocates slab cache used to allocate "struct epitem" items */
2142 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07002143 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144
2145 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2146 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07002147 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150}
Davide Libenzicea69242007-05-10 22:23:22 -07002151fs_initcall(eventpoll_init);