blob: 5b96ba77e62309515334a508a9bee83bfb6ab08f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Davide Libenzi5071f972009-03-31 15:24:10 -07002 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * Davide Libenzi <davidel@xmailserver.org>
11 *
12 */
13
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/init.h>
15#include <linux/kernel.h>
16#include <linux/sched.h>
17#include <linux/fs.h>
18#include <linux/file.h>
19#include <linux/signal.h>
20#include <linux/errno.h>
21#include <linux/mm.h>
22#include <linux/slab.h>
23#include <linux/poll.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/string.h>
25#include <linux/list.h>
26#include <linux/hash.h>
27#include <linux/spinlock.h>
28#include <linux/syscalls.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/rbtree.h>
30#include <linux/wait.h>
31#include <linux/eventpoll.h>
32#include <linux/mount.h>
33#include <linux/bitops.h>
Arjan van de Ven144efe32006-03-23 03:00:32 -080034#include <linux/mutex.h>
Davide Libenzida66f7c2007-05-10 22:23:21 -070035#include <linux/anon_inodes.h>
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020036#include <linux/device.h>
Colin Crosseb5d5502013-05-06 23:50:16 +000037#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <asm/io.h>
40#include <asm/mman.h>
Arun Sharma600634972011-07-26 16:09:06 -070041#include <linux/atomic.h>
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -080042#include <linux/proc_fs.h>
43#include <linux/seq_file.h>
Al Viro35280bd2013-02-24 14:52:17 -050044#include <linux/compat.h>
Jason Baronae10b2b2013-11-12 15:10:16 -080045#include <linux/rculist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Linus Torvalds1da177e2005-04-16 15:20:36 -070047/*
48 * LOCKING:
49 * There are three level of locking required by epoll :
50 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080051 * 1) epmutex (mutex)
Davide Libenzic7ea7632007-05-15 01:40:47 -070052 * 2) ep->mtx (mutex)
53 * 3) ep->lock (spinlock)
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 *
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
Davide Libenzid47de162007-05-15 01:40:41 -070063 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
66 * and ep_free().
67 * This mutex is acquired by ep_free() during the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -070068 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
Daniel Balutabf6a41d2011-01-30 23:42:29 +020070 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
Davide Libenzi22bacca2011-02-25 14:44:12 -080071 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
77 * going to.
Nelson Elhaged8805e62011-10-31 17:13:14 -070078 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
86 * the lockdep subkey.
Davide Libenzid47de162007-05-15 01:40:41 -070087 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080090 * Events that require holding "epmutex" are very rare, while for
Davide Libenzid47de162007-05-15 01:40:41 -070091 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
Linus Torvalds1da177e2005-04-16 15:20:36 -070093 */
94
Linus Torvalds1da177e2005-04-16 15:20:36 -070095/* Epoll private bits inside the event mask */
Jason Barondf0108c2016-01-20 14:59:24 -080096#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Jason Baronb6a515c2016-02-05 15:37:04 -080098#define EPOLLINOUT_BITS (POLLIN | POLLOUT)
99
100#define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | POLLERR | POLLHUP | \
101 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
102
Davide Libenzi5071f972009-03-31 15:24:10 -0700103/* Maximum number of nesting allowed inside epoll sets */
104#define EP_MAX_NESTS 4
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Davide Libenzib6119672006-10-11 01:21:44 -0700106#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
107
Davide Libenzid47de162007-05-15 01:40:41 -0700108#define EP_UNACTIVE_PTR ((void *) -1L)
109
Davide Libenzi7ef99642008-12-01 13:13:55 -0800110#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
111
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112struct epoll_filefd {
113 struct file *file;
114 int fd;
Eric Wong39732ca2013-04-30 15:27:38 -0700115} __packed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
117/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700118 * Structure used to track possible nested calls, for too deep recursions
119 * and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700121struct nested_call_node {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 struct list_head llink;
Davide Libenzi5071f972009-03-31 15:24:10 -0700123 void *cookie;
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700124 void *ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125};
126
127/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700128 * This structure is used as collector for nested calls, to check for
129 * maximum recursion dept and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700131struct nested_calls {
132 struct list_head tasks_call_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 spinlock_t lock;
134};
135
136/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700138 * have an entry of this type linked to the "rbr" RB tree.
Eric Wong39732ca2013-04-30 15:27:38 -0700139 * Avoid increasing the size of this struct, there can be many thousands
140 * of these on a server and we do not want this to take another cache line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 */
142struct epitem {
Jason Baronae10b2b2013-11-12 15:10:16 -0800143 union {
144 /* RB tree node links this structure to the eventpoll RB tree */
145 struct rb_node rbn;
146 /* Used to free the struct epitem */
147 struct rcu_head rcu;
148 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
150 /* List header used to link this structure to the eventpoll ready list */
151 struct list_head rdllink;
152
Davide Libenzic7ea7632007-05-15 01:40:47 -0700153 /*
154 * Works together "struct eventpoll"->ovflist in keeping the
155 * single linked chain of items.
156 */
157 struct epitem *next;
158
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 /* The file descriptor information this item refers to */
160 struct epoll_filefd ffd;
161
162 /* Number of active wait queue attached to poll operations */
163 int nwait;
164
165 /* List containing poll wait queues */
166 struct list_head pwqlist;
167
168 /* The "container" of this item */
169 struct eventpoll *ep;
170
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 /* List header used to link this item to the "struct file" items list */
172 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700173
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200174 /* wakeup_source used when EPOLLWAKEUP is set */
Eric Wongeea1d582013-04-30 15:27:39 -0700175 struct wakeup_source __rcu *ws;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200176
Davide Libenzic7ea7632007-05-15 01:40:47 -0700177 /* The structure that describe the interested events and the source fd */
178 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700179};
180
181/*
182 * This structure is stored inside the "private_data" member of the file
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200183 * structure and represents the main data structure for the eventpoll
Davide Libenzid47de162007-05-15 01:40:41 -0700184 * interface.
185 */
186struct eventpoll {
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200187 /* Protect the access to this structure */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700188 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700189
190 /*
191 * This mutex is used to ensure that files are not removed
192 * while epoll is using them. This is held during the event
193 * collection loop, the file cleanup path, the epoll file exit
194 * code and the ctl operations.
195 */
196 struct mutex mtx;
197
198 /* Wait queue used by sys_epoll_wait() */
199 wait_queue_head_t wq;
200
201 /* Wait queue used by file->poll() */
202 wait_queue_head_t poll_wait;
203
204 /* List of ready file descriptors */
205 struct list_head rdllist;
206
Davide Libenzi67647d02007-05-15 01:40:52 -0700207 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700208 struct rb_root rbr;
209
210 /*
211 * This is a single linked list that chains all the "struct epitem" that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300212 * happened while transferring ready events to userspace w/out
Davide Libenzid47de162007-05-15 01:40:41 -0700213 * holding ->lock.
214 */
215 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800216
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200217 /* wakeup_source used when ep_scan_ready_list is running */
218 struct wakeup_source *ws;
219
Davide Libenzi7ef99642008-12-01 13:13:55 -0800220 /* The user that created the eventpoll descriptor */
221 struct user_struct *user;
Jason Baron28d82dc2012-01-12 17:17:43 -0800222
223 struct file *file;
224
225 /* used to optimize loop detection check */
226 int visited;
227 struct list_head visited_list_link;
Davide Libenzid47de162007-05-15 01:40:41 -0700228};
229
230/* Wait structure used by the poll hooks */
231struct eppoll_entry {
232 /* List header used to link this structure to the "struct epitem" */
233 struct list_head llink;
234
235 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700236 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700237
238 /*
239 * Wait queue item that will be linked to the target file wait
240 * queue head.
241 */
242 wait_queue_t wait;
243
244 /* The wait queue head that linked the "wait" wait queue item */
245 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246};
247
248/* Wrapper struct used by poll queueing */
249struct ep_pqueue {
250 poll_table pt;
251 struct epitem *epi;
252};
253
Davide Libenzi5071f972009-03-31 15:24:10 -0700254/* Used by the ep_send_events() function as callback private data */
255struct ep_send_events_data {
256 int maxevents;
257 struct epoll_event __user *events;
258};
259
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800261 * Configuration options available inside /proc/sys/fs/epoll/
262 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800263/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800264static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800265
266/*
Davide Libenzid47de162007-05-15 01:40:41 -0700267 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800269static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270
Davide Libenzi22bacca2011-02-25 14:44:12 -0800271/* Used to check for epoll file descriptor inclusion loops */
272static struct nested_calls poll_loop_ncalls;
273
Davide Libenzi5071f972009-03-31 15:24:10 -0700274/* Used for safe wake up implementation */
275static struct nested_calls poll_safewake_ncalls;
276
277/* Used to call file's f_op->poll() under the nested calls boundaries */
278static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
280/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800281static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
283/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800284static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
Jason Baron28d82dc2012-01-12 17:17:43 -0800286/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
287static LIST_HEAD(visited_list);
288
289/*
290 * List of files with newly added links, where we may need to limit the number
291 * of emanating paths. Protected by the epmutex.
292 */
293static LIST_HEAD(tfile_check_list);
294
Davide Libenzi7ef99642008-12-01 13:13:55 -0800295#ifdef CONFIG_SYSCTL
296
297#include <linux/sysctl.h>
298
Robin Holt52bd19f72011-01-12 17:00:01 -0800299static long zero;
300static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800301
Joe Perches1f7e0612014-06-06 14:38:05 -0700302struct ctl_table epoll_table[] = {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800303 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800304 .procname = "max_user_watches",
305 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800306 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800307 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800308 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800309 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800310 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800311 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800312 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800313};
314#endif /* CONFIG_SYSCTL */
315
Jason Baron28d82dc2012-01-12 17:17:43 -0800316static const struct file_operations eventpoll_fops;
317
318static inline int is_file_epoll(struct file *f)
319{
320 return f->f_op == &eventpoll_fops;
321}
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700322
Davide Libenzi67647d02007-05-15 01:40:52 -0700323/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700324static inline void ep_set_ffd(struct epoll_filefd *ffd,
325 struct file *file, int fd)
326{
327 ffd->file = file;
328 ffd->fd = fd;
329}
330
Davide Libenzi67647d02007-05-15 01:40:52 -0700331/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700332static inline int ep_cmp_ffd(struct epoll_filefd *p1,
333 struct epoll_filefd *p2)
334{
335 return (p1->file > p2->file ? +1:
336 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
337}
338
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700339/* Tells us if the item is currently linked */
340static inline int ep_is_linked(struct list_head *p)
341{
342 return !list_empty(p);
343}
344
Oleg Nesterov971316f2012-02-24 20:07:29 +0100345static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
346{
347 return container_of(p, struct eppoll_entry, wait);
348}
349
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700350/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700351static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700352{
353 return container_of(p, struct eppoll_entry, wait)->base;
354}
355
356/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700357static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700358{
359 return container_of(p, struct ep_pqueue, pt)->epi;
360}
361
362/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700363static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700364{
Andrew Mortona80a6b82012-11-08 15:53:35 -0800365 return op != EPOLL_CTL_DEL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700366}
367
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700369static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370{
Davide Libenzi5071f972009-03-31 15:24:10 -0700371 INIT_LIST_HEAD(&ncalls->tasks_call_list);
372 spin_lock_init(&ncalls->lock);
373}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
Davide Libenzi5071f972009-03-31 15:24:10 -0700375/**
Davide Libenzi3fb0e582011-03-22 16:34:46 -0700376 * ep_events_available - Checks if ready events might be available.
377 *
378 * @ep: Pointer to the eventpoll context.
379 *
380 * Returns: Returns a value different than zero if ready events are available,
381 * or zero otherwise.
382 */
383static inline int ep_events_available(struct eventpoll *ep)
384{
385 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
386}
387
388/**
Davide Libenzi5071f972009-03-31 15:24:10 -0700389 * ep_call_nested - Perform a bound (possibly) nested call, by checking
390 * that the recursion limit is not exceeded, and that
391 * the same nested call (by the meaning of same cookie) is
392 * no re-entered.
393 *
394 * @ncalls: Pointer to the nested_calls structure to be used for this call.
395 * @max_nests: Maximum number of allowed nesting calls.
396 * @nproc: Nested call core function pointer.
397 * @priv: Opaque data to be passed to the @nproc callback.
398 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700399 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700400 *
401 * Returns: Returns the code returned by the @nproc callback, or -1 if
402 * the maximum recursion limit has been exceeded.
403 */
404static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
405 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700406 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700407{
408 int error, call_nests = 0;
409 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700410 struct list_head *lsthead = &ncalls->tasks_call_list;
411 struct nested_call_node *tncur;
412 struct nested_call_node tnode;
413
414 spin_lock_irqsave(&ncalls->lock, flags);
415
416 /*
417 * Try to see if the current task is already inside this wakeup call.
418 * We use a list here, since the population inside this set is always
419 * very much limited.
420 */
421 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700422 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700423 (tncur->cookie == cookie || ++call_nests > max_nests)) {
424 /*
425 * Ops ... loop detected or maximum nest level reached.
426 * We abort this wake by breaking the cycle itself.
427 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700428 error = -1;
429 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700430 }
431 }
432
433 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700434 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700435 tnode.cookie = cookie;
436 list_add(&tnode.llink, lsthead);
437
438 spin_unlock_irqrestore(&ncalls->lock, flags);
439
440 /* Call the nested function */
441 error = (*nproc)(priv, cookie, call_nests);
442
443 /* Remove the current task from the list */
444 spin_lock_irqsave(&ncalls->lock, flags);
445 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700446out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700447 spin_unlock_irqrestore(&ncalls->lock, flags);
448
449 return error;
450}
451
Steven Rostedt02edc6f2012-03-23 15:02:27 -0700452/*
453 * As described in commit 0ccf831cb lockdep: annotate epoll
454 * the use of wait queues used by epoll is done in a very controlled
455 * manner. Wake ups can nest inside each other, but are never done
456 * with the same locking. For example:
457 *
458 * dfd = socket(...);
459 * efd1 = epoll_create();
460 * efd2 = epoll_create();
461 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
462 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
463 *
464 * When a packet arrives to the device underneath "dfd", the net code will
465 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
466 * callback wakeup entry on that queue, and the wake_up() performed by the
467 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
468 * (efd1) notices that it may have some event ready, so it needs to wake up
469 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
470 * that ends up in another wake_up(), after having checked about the
471 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
472 * avoid stack blasting.
473 *
474 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
475 * this special case of epoll.
476 */
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700477#ifdef CONFIG_DEBUG_LOCK_ALLOC
478static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
479 unsigned long events, int subclass)
480{
481 unsigned long flags;
482
483 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
484 wake_up_locked_poll(wqueue, events);
485 spin_unlock_irqrestore(&wqueue->lock, flags);
486}
487#else
488static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
489 unsigned long events, int subclass)
490{
491 wake_up_poll(wqueue, events);
492}
493#endif
494
Davide Libenzi5071f972009-03-31 15:24:10 -0700495static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
496{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700497 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
498 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700499 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500}
501
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502/*
503 * Perform a safe wake up of the poll wait list. The problem is that
504 * with the new callback'd wake up system, it is possible that the
505 * poll callback is reentered from inside the call to wake_up() done
506 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700507 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 * and we cannot reenter the same wait queue head at all. This will
509 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700510 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700512static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700514 int this_cpu = get_cpu();
515
Davide Libenzi5071f972009-03-31 15:24:10 -0700516 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700517 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
518
519 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520}
521
Oleg Nesterov971316f2012-02-24 20:07:29 +0100522static void ep_remove_wait_queue(struct eppoll_entry *pwq)
523{
524 wait_queue_head_t *whead;
525
526 rcu_read_lock();
Oleg Nesterovd325f1f2017-09-01 18:55:33 +0200527 /*
528 * If it is cleared by POLLFREE, it should be rcu-safe.
529 * If we read NULL we need a barrier paired with
530 * smp_store_release() in ep_poll_callback(), otherwise
531 * we rely on whead->lock.
532 */
533 whead = smp_load_acquire(&pwq->whead);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100534 if (whead)
535 remove_wait_queue(whead, &pwq->wait);
536 rcu_read_unlock();
537}
538
Davide Libenzi7699acd2007-05-10 22:23:23 -0700539/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700540 * This function unregisters poll callbacks from the associated file
541 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
542 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700543 */
544static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
545{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700546 struct list_head *lsthead = &epi->pwqlist;
547 struct eppoll_entry *pwq;
548
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700549 while (!list_empty(lsthead)) {
550 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700551
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700552 list_del(&pwq->llink);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100553 ep_remove_wait_queue(pwq);
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700554 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700555 }
556}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
Eric Wongeea1d582013-04-30 15:27:39 -0700558/* call only when ep->mtx is held */
559static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
560{
561 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
562}
563
564/* call only when ep->mtx is held */
565static inline void ep_pm_stay_awake(struct epitem *epi)
566{
567 struct wakeup_source *ws = ep_wakeup_source(epi);
568
569 if (ws)
570 __pm_stay_awake(ws);
571}
572
573static inline bool ep_has_wakeup_source(struct epitem *epi)
574{
575 return rcu_access_pointer(epi->ws) ? true : false;
576}
577
578/* call when ep->mtx cannot be held (ep_poll_callback) */
579static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
580{
581 struct wakeup_source *ws;
582
583 rcu_read_lock();
584 ws = rcu_dereference(epi->ws);
585 if (ws)
586 __pm_stay_awake(ws);
587 rcu_read_unlock();
588}
589
Davide Libenzi5071f972009-03-31 15:24:10 -0700590/**
591 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
592 * the scan code, to call f_op->poll(). Also allows for
593 * O(NumReady) performance.
594 *
595 * @ep: Pointer to the epoll private data structure.
596 * @sproc: Pointer to the scan callback.
597 * @priv: Private opaque data passed to the @sproc callback.
Nelson Elhaged8805e62011-10-31 17:13:14 -0700598 * @depth: The current depth of recursive f_op->poll calls.
Jason Baron67347fe2013-11-12 15:10:18 -0800599 * @ep_locked: caller already holds ep->mtx
Davide Libenzi5071f972009-03-31 15:24:10 -0700600 *
601 * Returns: The same integer error code returned by the @sproc callback.
602 */
603static int ep_scan_ready_list(struct eventpoll *ep,
604 int (*sproc)(struct eventpoll *,
605 struct list_head *, void *),
Jason Baron67347fe2013-11-12 15:10:18 -0800606 void *priv, int depth, bool ep_locked)
Davide Libenzi5071f972009-03-31 15:24:10 -0700607{
608 int error, pwake = 0;
609 unsigned long flags;
610 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700611 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700612
613 /*
614 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700615 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700616 */
Jason Baron67347fe2013-11-12 15:10:18 -0800617
618 if (!ep_locked)
619 mutex_lock_nested(&ep->mtx, depth);
Davide Libenzi5071f972009-03-31 15:24:10 -0700620
621 /*
622 * Steal the ready list, and re-init the original one to the
623 * empty list. Also, set ep->ovflist to NULL so that events
624 * happening while looping w/out locks, are not lost. We cannot
625 * have the poll callback to queue directly on ep->rdllist,
626 * because we want the "sproc" callback to be able to do it
627 * in a lockless way.
628 */
629 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700630 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700631 ep->ovflist = NULL;
632 spin_unlock_irqrestore(&ep->lock, flags);
633
634 /*
635 * Now call the callback function.
636 */
637 error = (*sproc)(ep, &txlist, priv);
638
639 spin_lock_irqsave(&ep->lock, flags);
640 /*
641 * During the time we spent inside the "sproc" callback, some
642 * other events might have been queued by the poll callback.
643 * We re-insert them inside the main ready-list here.
644 */
645 for (nepi = ep->ovflist; (epi = nepi) != NULL;
646 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
647 /*
648 * We need to check if the item is already in the list.
649 * During the "sproc" callback execution time, items are
650 * queued into ->ovflist but the "txlist" might already
651 * contain them, and the list_splice() below takes care of them.
652 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200653 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700654 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -0700655 ep_pm_stay_awake(epi);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200656 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700657 }
658 /*
659 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
660 * releasing the lock, events will be queued in the normal way inside
661 * ep->rdllist.
662 */
663 ep->ovflist = EP_UNACTIVE_PTR;
664
665 /*
666 * Quickly re-inject items left on "txlist".
667 */
668 list_splice(&txlist, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200669 __pm_relax(ep->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700670
671 if (!list_empty(&ep->rdllist)) {
672 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700673 * Wake up (if active) both the eventpoll wait list and
674 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700675 */
676 if (waitqueue_active(&ep->wq))
677 wake_up_locked(&ep->wq);
678 if (waitqueue_active(&ep->poll_wait))
679 pwake++;
680 }
681 spin_unlock_irqrestore(&ep->lock, flags);
682
Jason Baron67347fe2013-11-12 15:10:18 -0800683 if (!ep_locked)
684 mutex_unlock(&ep->mtx);
Davide Libenzi5071f972009-03-31 15:24:10 -0700685
686 /* We have to call this outside the lock */
687 if (pwake)
688 ep_poll_safewake(&ep->poll_wait);
689
690 return error;
691}
692
Jason Baronae10b2b2013-11-12 15:10:16 -0800693static void epi_rcu_free(struct rcu_head *head)
694{
695 struct epitem *epi = container_of(head, struct epitem, rcu);
696 kmem_cache_free(epi_cache, epi);
697}
698
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700700 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700701 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700702 */
703static int ep_remove(struct eventpoll *ep, struct epitem *epi)
704{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700705 unsigned long flags;
706 struct file *file = epi->ffd.file;
707
708 /*
709 * Removes poll wait queue hooks. We _have_ to do this without holding
710 * the "ep->lock" otherwise a deadlock might occur. This because of the
711 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
712 * queue head lock when unregistering the wait queue. The wakeup callback
713 * will run by holding the wait queue head lock and will call our callback
714 * that will try to get "ep->lock".
715 */
716 ep_unregister_pollwait(ep, epi);
717
718 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700719 spin_lock(&file->f_lock);
Jason Baronae10b2b2013-11-12 15:10:16 -0800720 list_del_rcu(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700721 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
Davide Libenzicdac75e2008-04-29 00:58:34 -0700723 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
Davide Libenzic7ea7632007-05-15 01:40:47 -0700725 spin_lock_irqsave(&ep->lock, flags);
726 if (ep_is_linked(&epi->rdllink))
727 list_del_init(&epi->rdllink);
728 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729
Eric Wongeea1d582013-04-30 15:27:39 -0700730 wakeup_source_unregister(ep_wakeup_source(epi));
Jason Baronae10b2b2013-11-12 15:10:16 -0800731 /*
732 * At this point it is safe to free the eventpoll item. Use the union
733 * field epi->rcu, since we are trying to minimize the size of
734 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
735 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
736 * use of the rbn field.
737 */
738 call_rcu(&epi->rcu, epi_rcu_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739
Robin Holt52bd19f72011-01-12 17:00:01 -0800740 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800741
Davide Libenzic7ea7632007-05-15 01:40:47 -0700742 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743}
744
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745static void ep_free(struct eventpoll *ep)
746{
747 struct rb_node *rbp;
748 struct epitem *epi;
749
750 /* We need to release all tasks waiting for these file */
751 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700752 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753
754 /*
755 * We need to lock this because we could be hit by
756 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700757 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 * is on the way to be removed and no one has references to it
759 * anymore. The only hit might come from eventpoll_release_file() but
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300760 * holding "epmutex" is sufficient here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800762 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
764 /*
765 * Walks through the whole tree by unregistering poll callbacks.
766 */
767 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
768 epi = rb_entry(rbp, struct epitem, rbn);
769
770 ep_unregister_pollwait(ep, epi);
Eric Dumazet91cf5ab2013-09-11 14:24:06 -0700771 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 }
773
774 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700775 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700777 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 * us during this operation. So we can avoid the lock on "ep->lock".
Eric Wongddf676c2013-04-30 15:27:40 -0700779 * We do not need to lock ep->mtx, either, we only do it to prevent
780 * a lockdep warning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 */
Eric Wongddf676c2013-04-30 15:27:40 -0700782 mutex_lock(&ep->mtx);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700783 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 epi = rb_entry(rbp, struct epitem, rbn);
785 ep_remove(ep, epi);
Eric Dumazet91cf5ab2013-09-11 14:24:06 -0700786 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 }
Eric Wongddf676c2013-04-30 15:27:40 -0700788 mutex_unlock(&ep->mtx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
Arjan van de Ven144efe32006-03-23 03:00:32 -0800790 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700791 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800792 free_uid(ep->user);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200793 wakeup_source_unregister(ep->ws);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700794 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
Davide Libenzi7699acd2007-05-10 22:23:23 -0700797static int ep_eventpoll_release(struct inode *inode, struct file *file)
798{
799 struct eventpoll *ep = file->private_data;
800
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700801 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700802 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700803
Davide Libenzi7699acd2007-05-10 22:23:23 -0700804 return 0;
805}
806
Eric Wong450d89e2013-04-30 15:27:42 -0700807static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
808{
809 pt->_key = epi->event.events;
810
811 return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
812}
813
Davide Libenzi296e2362009-03-31 15:24:11 -0700814static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
815 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700816{
817 struct epitem *epi, *tmp;
Hans Verkuil626cf232012-03-23 15:02:27 -0700818 poll_table pt;
Davide Libenzi5071f972009-03-31 15:24:10 -0700819
Hans Verkuil626cf232012-03-23 15:02:27 -0700820 init_poll_funcptr(&pt, NULL);
Eric Wong450d89e2013-04-30 15:27:42 -0700821
Davide Libenzi5071f972009-03-31 15:24:10 -0700822 list_for_each_entry_safe(epi, tmp, head, rdllink) {
Eric Wong450d89e2013-04-30 15:27:42 -0700823 if (ep_item_poll(epi, &pt))
Davide Libenzi5071f972009-03-31 15:24:10 -0700824 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700825 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700826 /*
827 * Item has been dropped into the ready list by the poll
828 * callback, but it's not actually ready, as far as
829 * caller requested events goes. We can remove it here.
830 */
Eric Wongeea1d582013-04-30 15:27:39 -0700831 __pm_relax(ep_wakeup_source(epi));
Davide Libenzi5071f972009-03-31 15:24:10 -0700832 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700833 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700834 }
835
836 return 0;
837}
838
Jason Baron67347fe2013-11-12 15:10:18 -0800839static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
840 poll_table *pt);
841
842struct readyevents_arg {
843 struct eventpoll *ep;
844 bool locked;
845};
846
Davide Libenzi5071f972009-03-31 15:24:10 -0700847static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
848{
Jason Baron67347fe2013-11-12 15:10:18 -0800849 struct readyevents_arg *arg = priv;
850
851 return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
852 call_nests + 1, arg->locked);
Davide Libenzi5071f972009-03-31 15:24:10 -0700853}
854
Davide Libenzi7699acd2007-05-10 22:23:23 -0700855static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
856{
Davide Libenzi5071f972009-03-31 15:24:10 -0700857 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700858 struct eventpoll *ep = file->private_data;
Jason Baron67347fe2013-11-12 15:10:18 -0800859 struct readyevents_arg arg;
860
861 /*
862 * During ep_insert() we already hold the ep->mtx for the tfile.
863 * Prevent re-aquisition.
864 */
865 arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
866 arg.ep = ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700867
868 /* Insert inside our poll wait queue */
869 poll_wait(file, &ep->poll_wait, wait);
870
Davide Libenzi5071f972009-03-31 15:24:10 -0700871 /*
872 * Proceed to find out if wanted events are really available inside
873 * the ready list. This need to be done under ep_call_nested()
874 * supervision, since the call to f_op->poll() done on listed files
875 * could re-enter here.
876 */
877 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Jason Baron67347fe2013-11-12 15:10:18 -0800878 ep_poll_readyevents_proc, &arg, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700879
Davide Libenzi296e2362009-03-31 15:24:11 -0700880 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700881}
882
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800883#ifdef CONFIG_PROC_FS
Joe Perchesa3816ab2014-09-29 16:08:25 -0700884static void ep_show_fdinfo(struct seq_file *m, struct file *f)
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800885{
886 struct eventpoll *ep = f->private_data;
887 struct rb_node *rbp;
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800888
889 mutex_lock(&ep->mtx);
890 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
891 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
892
Joe Perchesa3816ab2014-09-29 16:08:25 -0700893 seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
894 epi->ffd.fd, epi->event.events,
895 (long long)epi->event.data);
896 if (seq_has_overflowed(m))
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800897 break;
898 }
899 mutex_unlock(&ep->mtx);
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800900}
901#endif
902
Davide Libenzi7699acd2007-05-10 22:23:23 -0700903/* File callbacks that implement the eventpoll file behaviour */
904static const struct file_operations eventpoll_fops = {
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800905#ifdef CONFIG_PROC_FS
906 .show_fdinfo = ep_show_fdinfo,
907#endif
Davide Libenzi7699acd2007-05-10 22:23:23 -0700908 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200909 .poll = ep_eventpoll_poll,
910 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700911};
912
Davide Libenzi7699acd2007-05-10 22:23:23 -0700913/*
914 * This is called from eventpoll_release() to unlink files from the eventpoll
915 * interface. We need to have this facility to cleanup correctly files that are
916 * closed without being removed from the eventpoll interface.
917 */
918void eventpoll_release_file(struct file *file)
919{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700920 struct eventpoll *ep;
Konstantin Khlebnikovebe06182014-06-17 06:58:05 +0400921 struct epitem *epi, *next;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700922
923 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700924 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700925 * necessary. It is not necessary because we're in the "struct file"
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300926 * cleanup path, and this means that no one is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700927 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700928 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700929 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700930 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700931 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700932 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700933 *
934 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700935 */
936 mutex_lock(&epmutex);
Konstantin Khlebnikovebe06182014-06-17 06:58:05 +0400937 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
Davide Libenzi7699acd2007-05-10 22:23:23 -0700938 ep = epi->ep;
Nelson Elhaged8805e62011-10-31 17:13:14 -0700939 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700940 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700941 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700942 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700943 mutex_unlock(&epmutex);
944}
945
946static int ep_alloc(struct eventpoll **pep)
947{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800948 int error;
949 struct user_struct *user;
950 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700951
Davide Libenzi7ef99642008-12-01 13:13:55 -0800952 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800953 error = -ENOMEM;
954 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
955 if (unlikely(!ep))
956 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700957
Davide Libenzic7ea7632007-05-15 01:40:47 -0700958 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700959 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700960 init_waitqueue_head(&ep->wq);
961 init_waitqueue_head(&ep->poll_wait);
962 INIT_LIST_HEAD(&ep->rdllist);
963 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700964 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800965 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700966
967 *pep = ep;
968
Davide Libenzi7699acd2007-05-10 22:23:23 -0700969 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800970
971free_uid:
972 free_uid(user);
973 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700974}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
976/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700977 * Search the file inside the eventpoll tree. The RB tree operations
978 * are protected by the "mtx" mutex, and ep_find() must be called with
979 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 */
981static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
982{
983 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 struct rb_node *rbp;
985 struct epitem *epi, *epir = NULL;
986 struct epoll_filefd ffd;
987
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700988 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 for (rbp = ep->rbr.rb_node; rbp; ) {
990 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700991 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 if (kcmp > 0)
993 rbp = rbp->rb_right;
994 else if (kcmp < 0)
995 rbp = rbp->rb_left;
996 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 epir = epi;
998 break;
999 }
1000 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 return epir;
1003}
1004
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005/*
Davide Libenzi7699acd2007-05-10 22:23:23 -07001006 * This is the callback that is passed to the wait queue wakeup
Daniel Balutabf6a41d2011-01-30 23:42:29 +02001007 * mechanism. It is called by the stored file descriptors when they
Davide Libenzi7699acd2007-05-10 22:23:23 -07001008 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001010static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011{
Davide Libenzi7699acd2007-05-10 22:23:23 -07001012 int pwake = 0;
1013 unsigned long flags;
1014 struct epitem *epi = ep_item_from_wait(wait);
1015 struct eventpoll *ep = epi->ep;
Jason Barondf0108c2016-01-20 14:59:24 -08001016 int ewake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
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)) {
Jason Baronb6a515c2016-02-05 15:37:04 -08001071 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1072 !((unsigned long)key & POLLFREE)) {
1073 switch ((unsigned long)key & EPOLLINOUT_BITS) {
1074 case POLLIN:
1075 if (epi->event.events & POLLIN)
1076 ewake = 1;
1077 break;
1078 case POLLOUT:
1079 if (epi->event.events & POLLOUT)
1080 ewake = 1;
1081 break;
1082 case 0:
1083 ewake = 1;
1084 break;
1085 }
1086 }
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001087 wake_up_locked(&ep->wq);
Jason Barondf0108c2016-01-20 14:59:24 -08001088 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001089 if (waitqueue_active(&ep->poll_wait))
1090 pwake++;
1091
Davide Libenzid47de162007-05-15 01:40:41 -07001092out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001093 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001094
1095 /* We have to call this outside the lock */
1096 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001097 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001098
Oleg Nesterovd325f1f2017-09-01 18:55:33 +02001099 if (!(epi->event.events & EPOLLEXCLUSIVE))
1100 ewake = 1;
Jason Barondf0108c2016-01-20 14:59:24 -08001101
Oleg Nesterovd325f1f2017-09-01 18:55:33 +02001102 if ((unsigned long)key & POLLFREE) {
1103 /*
1104 * If we race with ep_remove_wait_queue() it can miss
1105 * ->whead = NULL and do another remove_wait_queue() after
1106 * us, so we can't use __remove_wait_queue().
1107 */
1108 list_del_init(&wait->task_list);
1109 /*
1110 * ->whead != NULL protects us from the race with ep_free()
1111 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1112 * held by the caller. Once we nullify it, nothing protects
1113 * ep/epi or even wait.
1114 */
1115 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1116 }
1117
1118 return ewake;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119}
1120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121/*
1122 * This is the callback that is used to add our wait queue to the
1123 * target file wakeup lists.
1124 */
1125static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1126 poll_table *pt)
1127{
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001128 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 struct eppoll_entry *pwq;
1130
Christoph Lametere94b1762006-12-06 20:33:17 -08001131 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1133 pwq->whead = whead;
1134 pwq->base = epi;
Jason Barondf0108c2016-01-20 14:59:24 -08001135 if (epi->event.events & EPOLLEXCLUSIVE)
1136 add_wait_queue_exclusive(whead, &pwq->wait);
1137 else
1138 add_wait_queue(whead, &pwq->wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 list_add_tail(&pwq->llink, &epi->pwqlist);
1140 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001141 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 /* We have to signal that an error occurred */
1143 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -07001144 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145}
1146
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1148{
1149 int kcmp;
1150 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1151 struct epitem *epic;
1152
1153 while (*p) {
1154 parent = *p;
1155 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001156 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 if (kcmp > 0)
1158 p = &parent->rb_right;
1159 else
1160 p = &parent->rb_left;
1161 }
1162 rb_link_node(&epi->rbn, parent, p);
1163 rb_insert_color(&epi->rbn, &ep->rbr);
1164}
1165
Andrew Mortona80a6b82012-11-08 15:53:35 -08001166
1167
Jason Baron28d82dc2012-01-12 17:17:43 -08001168#define PATH_ARR_SIZE 5
1169/*
1170 * These are the number paths of length 1 to 5, that we are allowing to emanate
1171 * from a single file of interest. For example, we allow 1000 paths of length
1172 * 1, to emanate from each file of interest. This essentially represents the
1173 * potential wakeup paths, which need to be limited in order to avoid massive
1174 * uncontrolled wakeup storms. The common use case should be a single ep which
1175 * is connected to n file sources. In this case each file source has 1 path
1176 * of length 1. Thus, the numbers below should be more than sufficient. These
1177 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1178 * and delete can't add additional paths. Protected by the epmutex.
1179 */
1180static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1181static int path_count[PATH_ARR_SIZE];
1182
1183static int path_count_inc(int nests)
1184{
Jason Baron93dc6102012-03-16 16:34:03 -04001185 /* Allow an arbitrary number of depth 1 paths */
1186 if (nests == 0)
1187 return 0;
1188
Jason Baron28d82dc2012-01-12 17:17:43 -08001189 if (++path_count[nests] > path_limits[nests])
1190 return -1;
1191 return 0;
1192}
1193
1194static void path_count_init(void)
1195{
1196 int i;
1197
1198 for (i = 0; i < PATH_ARR_SIZE; i++)
1199 path_count[i] = 0;
1200}
1201
1202static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1203{
1204 int error = 0;
1205 struct file *file = priv;
1206 struct file *child_file;
1207 struct epitem *epi;
1208
Jason Baronae10b2b2013-11-12 15:10:16 -08001209 /* CTL_DEL can remove links here, but that can't increase our count */
1210 rcu_read_lock();
1211 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001212 child_file = epi->ep->file;
1213 if (is_file_epoll(child_file)) {
1214 if (list_empty(&child_file->f_ep_links)) {
1215 if (path_count_inc(call_nests)) {
1216 error = -1;
1217 break;
1218 }
1219 } else {
1220 error = ep_call_nested(&poll_loop_ncalls,
1221 EP_MAX_NESTS,
1222 reverse_path_check_proc,
1223 child_file, child_file,
1224 current);
1225 }
1226 if (error != 0)
1227 break;
1228 } else {
1229 printk(KERN_ERR "reverse_path_check_proc: "
1230 "file is not an ep!\n");
1231 }
1232 }
Jason Baronae10b2b2013-11-12 15:10:16 -08001233 rcu_read_unlock();
Jason Baron28d82dc2012-01-12 17:17:43 -08001234 return error;
1235}
1236
1237/**
1238 * reverse_path_check - The tfile_check_list is list of file *, which have
1239 * links that are proposed to be newly added. We need to
1240 * make sure that those added links don't add too many
1241 * paths such that we will spend all our time waking up
1242 * eventpoll objects.
1243 *
1244 * Returns: Returns zero if the proposed links don't create too many paths,
1245 * -1 otherwise.
1246 */
1247static int reverse_path_check(void)
1248{
Jason Baron28d82dc2012-01-12 17:17:43 -08001249 int error = 0;
1250 struct file *current_file;
1251
1252 /* let's call this for all tfiles */
1253 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001254 path_count_init();
1255 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1256 reverse_path_check_proc, current_file,
1257 current_file, current);
1258 if (error)
1259 break;
1260 }
1261 return error;
1262}
1263
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001264static int ep_create_wakeup_source(struct epitem *epi)
1265{
1266 const char *name;
Eric Wongeea1d582013-04-30 15:27:39 -07001267 struct wakeup_source *ws;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001268
1269 if (!epi->ep->ws) {
1270 epi->ep->ws = wakeup_source_register("eventpoll");
1271 if (!epi->ep->ws)
1272 return -ENOMEM;
1273 }
1274
1275 name = epi->ffd.file->f_path.dentry->d_name.name;
Eric Wongeea1d582013-04-30 15:27:39 -07001276 ws = wakeup_source_register(name);
1277
1278 if (!ws)
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001279 return -ENOMEM;
Eric Wongeea1d582013-04-30 15:27:39 -07001280 rcu_assign_pointer(epi->ws, ws);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001281
1282 return 0;
1283}
1284
Eric Wongeea1d582013-04-30 15:27:39 -07001285/* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1286static noinline void ep_destroy_wakeup_source(struct epitem *epi)
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001287{
Eric Wongeea1d582013-04-30 15:27:39 -07001288 struct wakeup_source *ws = ep_wakeup_source(epi);
1289
Eric Wongd6d67e72013-04-30 15:27:43 -07001290 RCU_INIT_POINTER(epi->ws, NULL);
Eric Wongeea1d582013-04-30 15:27:39 -07001291
1292 /*
1293 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1294 * used internally by wakeup_source_remove, too (called by
1295 * wakeup_source_unregister), so we cannot use call_rcu
1296 */
1297 synchronize_rcu();
1298 wakeup_source_unregister(ws);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001299}
1300
Davide Libenzic7ea7632007-05-15 01:40:47 -07001301/*
1302 * Must be called with "mtx" held.
1303 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
Jason Baron67347fe2013-11-12 15:10:18 -08001305 struct file *tfile, int fd, int full_check)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306{
1307 int error, revents, pwake = 0;
1308 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -08001309 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 struct epitem *epi;
1311 struct ep_pqueue epq;
1312
Robin Holt52bd19f72011-01-12 17:00:01 -08001313 user_watches = atomic_long_read(&ep->user->epoll_watches);
1314 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001315 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -08001316 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001317 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318
1319 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 INIT_LIST_HEAD(&epi->rdllink);
1321 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 INIT_LIST_HEAD(&epi->pwqlist);
1323 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001324 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -07001327 epi->next = EP_UNACTIVE_PTR;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001328 if (epi->event.events & EPOLLWAKEUP) {
1329 error = ep_create_wakeup_source(epi);
1330 if (error)
1331 goto error_create_wakeup_source;
1332 } else {
Eric Wongeea1d582013-04-30 15:27:39 -07001333 RCU_INIT_POINTER(epi->ws, NULL);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001334 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335
1336 /* Initialize the poll table using the queue callback */
1337 epq.epi = epi;
1338 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1339
1340 /*
1341 * Attach the item to the poll hooks and get current event bits.
1342 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -07001343 * been increased by the caller of this function. Note that after
1344 * this operation completes, the poll callback can start hitting
1345 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 */
Eric Wong450d89e2013-04-30 15:27:42 -07001347 revents = ep_item_poll(epi, &epq.pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
1349 /*
1350 * We have to check if something went wrong during the poll wait queue
1351 * install process. Namely an allocation for a wait queue failed due
1352 * high memory pressure.
1353 */
Davide Libenzi7ef99642008-12-01 13:13:55 -08001354 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001356 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357
1358 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -07001359 spin_lock(&tfile->f_lock);
Jason Baronae10b2b2013-11-12 15:10:16 -08001360 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -07001361 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
Davide Libenzic7ea7632007-05-15 01:40:47 -07001363 /*
1364 * Add the current item to the RB tree. All RB tree operations are
1365 * protected by "mtx", and ep_insert() is called with "mtx" held.
1366 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 ep_rbtree_insert(ep, epi);
1368
Jason Baron28d82dc2012-01-12 17:17:43 -08001369 /* now check if we've created too many backpaths */
1370 error = -EINVAL;
Jason Baron67347fe2013-11-12 15:10:18 -08001371 if (full_check && reverse_path_check())
Jason Baron28d82dc2012-01-12 17:17:43 -08001372 goto error_remove_epi;
1373
Davide Libenzic7ea7632007-05-15 01:40:47 -07001374 /* We have to drop the new item inside our item list to keep track of it */
1375 spin_lock_irqsave(&ep->lock, flags);
1376
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001378 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001380 ep_pm_stay_awake(epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381
1382 /* Notify waiting tasks that events are available */
1383 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001384 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 if (waitqueue_active(&ep->poll_wait))
1386 pwake++;
1387 }
1388
Davide Libenzic7ea7632007-05-15 01:40:47 -07001389 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390
Robin Holt52bd19f72011-01-12 17:00:01 -08001391 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001392
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 /* We have to call this outside the lock */
1394 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001395 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 return 0;
1398
Jason Baron28d82dc2012-01-12 17:17:43 -08001399error_remove_epi:
1400 spin_lock(&tfile->f_lock);
Jason Baronae10b2b2013-11-12 15:10:16 -08001401 list_del_rcu(&epi->fllink);
Jason Baron28d82dc2012-01-12 17:17:43 -08001402 spin_unlock(&tfile->f_lock);
1403
1404 rb_erase(&epi->rbn, &ep->rbr);
1405
Davide Libenzi7699acd2007-05-10 22:23:23 -07001406error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 ep_unregister_pollwait(ep, epi);
1408
1409 /*
1410 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -07001411 * allocated wait queue. Note that we don't care about the ep->ovflist
1412 * list, since that is used/cleaned only inside a section bound by "mtx".
1413 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001415 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001416 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001417 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001418 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419
Eric Wongeea1d582013-04-30 15:27:39 -07001420 wakeup_source_unregister(ep_wakeup_source(epi));
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001421
1422error_create_wakeup_source:
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001423 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001424
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 return error;
1426}
1427
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428/*
1429 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001430 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 */
1432static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1433{
1434 int pwake = 0;
1435 unsigned int revents;
Hans Verkuil626cf232012-03-23 15:02:27 -07001436 poll_table pt;
1437
1438 init_poll_funcptr(&pt, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439
1440 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001441 * Set the new event interest mask before calling f_op->poll();
1442 * otherwise we might miss an event that happens between the
1443 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 */
Eric Wong128dd172013-01-01 21:20:27 +00001445 epi->event.events = event->events; /* need barrier below */
Tony Battersbye057e152009-03-31 15:24:15 -07001446 epi->event.data = event->data; /* protected by mtx */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001447 if (epi->event.events & EPOLLWAKEUP) {
Eric Wongeea1d582013-04-30 15:27:39 -07001448 if (!ep_has_wakeup_source(epi))
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001449 ep_create_wakeup_source(epi);
Eric Wongeea1d582013-04-30 15:27:39 -07001450 } else if (ep_has_wakeup_source(epi)) {
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001451 ep_destroy_wakeup_source(epi);
1452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
1454 /*
Eric Wong128dd172013-01-01 21:20:27 +00001455 * The following barrier has two effects:
1456 *
1457 * 1) Flush epi changes above to other CPUs. This ensures
1458 * we do not miss events from ep_poll_callback if an
1459 * event occurs immediately after we call f_op->poll().
1460 * We need this because we did not take ep->lock while
1461 * changing epi above (but ep_poll_callback does take
1462 * ep->lock).
1463 *
1464 * 2) We also need to ensure we do not miss _past_ events
1465 * when calling f_op->poll(). This barrier also
1466 * pairs with the barrier in wq_has_sleeper (see
1467 * comments for wq_has_sleeper).
1468 *
1469 * This barrier will now guarantee ep_poll_callback or f_op->poll
1470 * (or both) will notice the readiness of an item.
1471 */
1472 smp_mb();
1473
1474 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 * Get current event bits. We can safely use the file* here because
1476 * its usage count has been increased by the caller of this function.
1477 */
Eric Wong450d89e2013-04-30 15:27:42 -07001478 revents = ep_item_poll(epi, &pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001481 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001482 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001484 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001485 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001486 if (!ep_is_linked(&epi->rdllink)) {
1487 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001488 ep_pm_stay_awake(epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
Davide Libenzic7ea7632007-05-15 01:40:47 -07001490 /* Notify waiting tasks that events are available */
1491 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001492 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001493 if (waitqueue_active(&ep->poll_wait))
1494 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 }
Tony Battersbye057e152009-03-31 15:24:15 -07001496 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
1499 /* We have to call this outside the lock */
1500 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001501 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
1503 return 0;
1504}
1505
Davide Libenzi296e2362009-03-31 15:24:11 -07001506static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1507 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001508{
1509 struct ep_send_events_data *esed = priv;
1510 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001511 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001512 struct epitem *epi;
1513 struct epoll_event __user *uevent;
Eric Wongeea1d582013-04-30 15:27:39 -07001514 struct wakeup_source *ws;
Hans Verkuil626cf232012-03-23 15:02:27 -07001515 poll_table pt;
1516
1517 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -07001518
Davide Libenzi296e2362009-03-31 15:24:11 -07001519 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001520 * We can loop without lock because we are passed a task private list.
1521 * Items cannot vanish during the loop because ep_scan_ready_list() is
1522 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001523 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001524 for (eventcnt = 0, uevent = esed->events;
1525 !list_empty(head) && eventcnt < esed->maxevents;) {
1526 epi = list_first_entry(head, struct epitem, rdllink);
1527
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001528 /*
1529 * Activate ep->ws before deactivating epi->ws to prevent
1530 * triggering auto-suspend here (in case we reactive epi->ws
1531 * below).
1532 *
1533 * This could be rearranged to delay the deactivation of epi->ws
1534 * instead, but then epi->ws would temporarily be out of sync
1535 * with ep_is_linked().
1536 */
Eric Wongeea1d582013-04-30 15:27:39 -07001537 ws = ep_wakeup_source(epi);
1538 if (ws) {
1539 if (ws->active)
1540 __pm_stay_awake(ep->ws);
1541 __pm_relax(ws);
1542 }
1543
Davide Libenzi5071f972009-03-31 15:24:10 -07001544 list_del_init(&epi->rdllink);
1545
Eric Wong450d89e2013-04-30 15:27:42 -07001546 revents = ep_item_poll(epi, &pt);
Davide Libenzi5071f972009-03-31 15:24:10 -07001547
Davide Libenzi296e2362009-03-31 15:24:11 -07001548 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001549 * If the event mask intersect the caller-requested one,
1550 * deliver the event to userspace. Again, ep_scan_ready_list()
1551 * is holding "mtx", so no operations coming from userspace
1552 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001553 */
1554 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001555 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001556 __put_user(epi->event.data, &uevent->data)) {
1557 list_add(&epi->rdllink, head);
Eric Wongeea1d582013-04-30 15:27:39 -07001558 ep_pm_stay_awake(epi);
Davide Libenzi296e2362009-03-31 15:24:11 -07001559 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001560 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001561 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001562 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001563 if (epi->event.events & EPOLLONESHOT)
1564 epi->event.events &= EP_PRIVATE_BITS;
1565 else if (!(epi->event.events & EPOLLET)) {
1566 /*
1567 * If this file has been added with Level
1568 * Trigger mode, we need to insert back inside
1569 * the ready list, so that the next call to
1570 * epoll_wait() will check again the events
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001571 * availability. At this point, no one can insert
Davide Libenzi296e2362009-03-31 15:24:11 -07001572 * into ep->rdllist besides us. The epoll_ctl()
1573 * callers are locked out by
1574 * ep_scan_ready_list() holding "mtx" and the
1575 * poll callback will queue them in ep->ovflist.
1576 */
1577 list_add_tail(&epi->rdllink, &ep->rdllist);
Eric Wongeea1d582013-04-30 15:27:39 -07001578 ep_pm_stay_awake(epi);
Davide Libenzi296e2362009-03-31 15:24:11 -07001579 }
1580 }
1581 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001582
1583 return eventcnt;
1584}
1585
Davide Libenzi296e2362009-03-31 15:24:11 -07001586static int ep_send_events(struct eventpoll *ep,
1587 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588{
Davide Libenzi5071f972009-03-31 15:24:10 -07001589 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001590
Davide Libenzi5071f972009-03-31 15:24:10 -07001591 esed.maxevents = maxevents;
1592 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001593
Jason Baron67347fe2013-11-12 15:10:18 -08001594 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595}
1596
Deepa Dinamani766b9f92016-05-19 17:09:05 -07001597static inline struct timespec64 ep_set_mstimeout(long ms)
Eric Dumazet0781b902011-02-01 15:52:35 -08001598{
Deepa Dinamani766b9f92016-05-19 17:09:05 -07001599 struct timespec64 now, ts = {
Eric Dumazet0781b902011-02-01 15:52:35 -08001600 .tv_sec = ms / MSEC_PER_SEC,
1601 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1602 };
1603
Deepa Dinamani766b9f92016-05-19 17:09:05 -07001604 ktime_get_ts64(&now);
1605 return timespec64_add_safe(now, ts);
Eric Dumazet0781b902011-02-01 15:52:35 -08001606}
1607
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001608/**
1609 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1610 * event buffer.
1611 *
1612 * @ep: Pointer to the eventpoll context.
1613 * @events: Pointer to the userspace buffer where the ready events should be
1614 * stored.
1615 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1616 * @timeout: Maximum timeout for the ready events fetch operation, in
1617 * milliseconds. If the @timeout is zero, the function will not block,
1618 * while if the @timeout is less than zero, the function will block
1619 * until at least one event has been retrieved (or an error
1620 * occurred).
1621 *
1622 * Returns: Returns the number of ready events which have been fetched, or an
1623 * error code, in case of error.
1624 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1626 int maxevents, long timeout)
1627{
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001628 int res = 0, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 unsigned long flags;
John Stultzda8b44d2016-03-17 14:20:51 -07001630 u64 slack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001632 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001634 if (timeout > 0) {
Deepa Dinamani766b9f92016-05-19 17:09:05 -07001635 struct timespec64 end_time = ep_set_mstimeout(timeout);
Eric Dumazet0781b902011-02-01 15:52:35 -08001636
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001637 slack = select_estimate_accuracy(&end_time);
1638 to = &expires;
Deepa Dinamani766b9f92016-05-19 17:09:05 -07001639 *to = timespec64_to_ktime(end_time);
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001640 } else if (timeout == 0) {
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001641 /*
1642 * Avoid the unnecessary trip to the wait queue loop, if the
1643 * caller specified a non blocking operation.
1644 */
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001645 timed_out = 1;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001646 spin_lock_irqsave(&ep->lock, flags);
1647 goto check_events;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001650fetch_events:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001651 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001653 if (!ep_events_available(ep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 /*
1655 * We don't have any available event to return to the caller.
1656 * We need to sleep here, and we will be wake up by
1657 * ep_poll_callback() when events will become available.
1658 */
1659 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001660 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
1662 for (;;) {
1663 /*
1664 * We don't want to sleep if the ep_poll_callback() sends us
1665 * a wakeup in between. That's why we set the task state
1666 * to TASK_INTERRUPTIBLE before doing the checks.
1667 */
1668 set_current_state(TASK_INTERRUPTIBLE);
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001669 if (ep_events_available(ep) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 break;
1671 if (signal_pending(current)) {
1672 res = -EINTR;
1673 break;
1674 }
1675
Davide Libenzic7ea7632007-05-15 01:40:47 -07001676 spin_unlock_irqrestore(&ep->lock, flags);
Colin Crosseb5d5502013-05-06 23:50:16 +00001677 if (!freezable_schedule_hrtimeout_range(to, slack,
1678 HRTIMER_MODE_ABS))
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001679 timed_out = 1;
1680
Davide Libenzic7ea7632007-05-15 01:40:47 -07001681 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683
Davidlohr Bueso4d5755b2015-02-13 14:39:08 -08001684 __remove_wait_queue(&ep->wq, &wait);
1685 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 }
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001687check_events:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 /* Is it worth to try to dig for events ? */
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001689 eavail = ep_events_available(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690
Davide Libenzic7ea7632007-05-15 01:40:47 -07001691 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692
1693 /*
1694 * Try to transfer events to user space. In case we get 0 events and
1695 * there's still timeout left over, we go trying again in search of
1696 * more luck.
1697 */
1698 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001699 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001700 goto fetch_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
1702 return res;
1703}
1704
Davide Libenzi22bacca2011-02-25 14:44:12 -08001705/**
1706 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1707 * API, to verify that adding an epoll file inside another
1708 * epoll structure, does not violate the constraints, in
1709 * terms of closed loops, or too deep chains (which can
1710 * result in excessive stack usage).
1711 *
1712 * @priv: Pointer to the epoll file to be currently checked.
1713 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1714 * data structure pointer.
1715 * @call_nests: Current dept of the @ep_call_nested() call stack.
1716 *
1717 * Returns: Returns zero if adding the epoll @file inside current epoll
1718 * structure @ep does not violate the constraints, or -1 otherwise.
1719 */
1720static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1721{
1722 int error = 0;
1723 struct file *file = priv;
1724 struct eventpoll *ep = file->private_data;
Jason Baron28d82dc2012-01-12 17:17:43 -08001725 struct eventpoll *ep_tovisit;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001726 struct rb_node *rbp;
1727 struct epitem *epi;
1728
Nelson Elhaged8805e62011-10-31 17:13:14 -07001729 mutex_lock_nested(&ep->mtx, call_nests + 1);
Jason Baron28d82dc2012-01-12 17:17:43 -08001730 ep->visited = 1;
1731 list_add(&ep->visited_list_link, &visited_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001732 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1733 epi = rb_entry(rbp, struct epitem, rbn);
1734 if (unlikely(is_file_epoll(epi->ffd.file))) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001735 ep_tovisit = epi->ffd.file->private_data;
1736 if (ep_tovisit->visited)
1737 continue;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001738 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Jason Baron28d82dc2012-01-12 17:17:43 -08001739 ep_loop_check_proc, epi->ffd.file,
1740 ep_tovisit, current);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001741 if (error != 0)
1742 break;
Jason Baron28d82dc2012-01-12 17:17:43 -08001743 } else {
1744 /*
1745 * If we've reached a file that is not associated with
1746 * an ep, then we need to check if the newly added
1747 * links are going to add too many wakeup paths. We do
1748 * this by adding it to the tfile_check_list, if it's
1749 * not already there, and calling reverse_path_check()
1750 * during ep_insert().
1751 */
1752 if (list_empty(&epi->ffd.file->f_tfile_llink))
1753 list_add(&epi->ffd.file->f_tfile_llink,
1754 &tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001755 }
1756 }
1757 mutex_unlock(&ep->mtx);
1758
1759 return error;
1760}
1761
1762/**
1763 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1764 * another epoll file (represented by @ep) does not create
1765 * closed loops or too deep chains.
1766 *
1767 * @ep: Pointer to the epoll private data structure.
1768 * @file: Pointer to the epoll file to be checked.
1769 *
1770 * Returns: Returns zero if adding the epoll @file inside current epoll
1771 * structure @ep does not violate the constraints, or -1 otherwise.
1772 */
1773static int ep_loop_check(struct eventpoll *ep, struct file *file)
1774{
Jason Baron28d82dc2012-01-12 17:17:43 -08001775 int ret;
1776 struct eventpoll *ep_cur, *ep_next;
1777
1778 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Davide Libenzi22bacca2011-02-25 14:44:12 -08001779 ep_loop_check_proc, file, ep, current);
Jason Baron28d82dc2012-01-12 17:17:43 -08001780 /* clear visited list */
1781 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1782 visited_list_link) {
1783 ep_cur->visited = 0;
1784 list_del(&ep_cur->visited_list_link);
1785 }
1786 return ret;
1787}
1788
1789static void clear_tfile_check_list(void)
1790{
1791 struct file *file;
1792
1793 /* first clear the tfile_check_list */
1794 while (!list_empty(&tfile_check_list)) {
1795 file = list_first_entry(&tfile_check_list, struct file,
1796 f_tfile_llink);
1797 list_del_init(&file->f_tfile_llink);
1798 }
1799 INIT_LIST_HEAD(&tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001800}
1801
Davide Libenzi7699acd2007-05-10 22:23:23 -07001802/*
Andrew Morton523723b2008-08-12 15:09:01 -07001803 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001804 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001805SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001806{
Jason Baron28d82dc2012-01-12 17:17:43 -08001807 int error, fd;
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001808 struct eventpoll *ep = NULL;
Jason Baron28d82dc2012-01-12 17:17:43 -08001809 struct file *file;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001810
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001811 /* Check the EPOLL_* constant for consistency. */
1812 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1813
Davide Libenzi296e2362009-03-31 15:24:11 -07001814 if (flags & ~EPOLL_CLOEXEC)
1815 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001816 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001817 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001818 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001819 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001820 if (error < 0)
1821 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001822 /*
1823 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001824 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001825 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001826 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1827 if (fd < 0) {
1828 error = fd;
1829 goto out_free_ep;
1830 }
1831 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001832 O_RDWR | (flags & O_CLOEXEC));
Jason Baron28d82dc2012-01-12 17:17:43 -08001833 if (IS_ERR(file)) {
1834 error = PTR_ERR(file);
1835 goto out_free_fd;
1836 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001837 ep->file = file;
Al Viro98022742012-08-17 22:42:36 -04001838 fd_install(fd, file);
Jason Baron28d82dc2012-01-12 17:17:43 -08001839 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001840
Jason Baron28d82dc2012-01-12 17:17:43 -08001841out_free_fd:
1842 put_unused_fd(fd);
1843out_free_ep:
1844 ep_free(ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001845 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001846}
1847
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001848SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001849{
Davide Libenzibfe38912009-05-12 13:19:44 -07001850 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001851 return -EINVAL;
1852
1853 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001854}
1855
Davide Libenzi7699acd2007-05-10 22:23:23 -07001856/*
1857 * The following function implements the controller interface for
1858 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001859 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001860 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001861SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1862 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001863{
1864 int error;
Jason Baron67347fe2013-11-12 15:10:18 -08001865 int full_check = 0;
Al Viro7e3fb582013-08-30 12:47:21 -04001866 struct fd f, tf;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001867 struct eventpoll *ep;
1868 struct epitem *epi;
1869 struct epoll_event epds;
Jason Baron67347fe2013-11-12 15:10:18 -08001870 struct eventpoll *tep = NULL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001871
Davide Libenzi7699acd2007-05-10 22:23:23 -07001872 error = -EFAULT;
1873 if (ep_op_has_event(op) &&
1874 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1875 goto error_return;
1876
Davide Libenzi7699acd2007-05-10 22:23:23 -07001877 error = -EBADF;
Al Viro7e3fb582013-08-30 12:47:21 -04001878 f = fdget(epfd);
1879 if (!f.file)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001880 goto error_return;
1881
1882 /* Get the "struct file *" for the target file */
Al Viro7e3fb582013-08-30 12:47:21 -04001883 tf = fdget(fd);
1884 if (!tf.file)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001885 goto error_fput;
1886
1887 /* The target file descriptor must support poll */
1888 error = -EPERM;
Al Viro72c2d532013-09-22 16:27:52 -04001889 if (!tf.file->f_op->poll)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001890 goto error_tgt_fput;
1891
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001892 /* Check if EPOLLWAKEUP is allowed */
Nicolas Ioossc680e412014-09-09 14:50:51 -07001893 if (ep_op_has_event(op))
1894 ep_take_care_of_epollwakeup(&epds);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001895
Davide Libenzi7699acd2007-05-10 22:23:23 -07001896 /*
1897 * We have to check that the file structure underneath the file descriptor
1898 * the user passed to us _is_ an eventpoll file. And also we do not permit
1899 * adding an epoll file descriptor inside itself.
1900 */
1901 error = -EINVAL;
Al Viro7e3fb582013-08-30 12:47:21 -04001902 if (f.file == tf.file || !is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07001903 goto error_tgt_fput;
1904
1905 /*
Jason Barondf0108c2016-01-20 14:59:24 -08001906 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
1907 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
1908 * Also, we do not currently supported nested exclusive wakeups.
1909 */
Jason Baronb6a515c2016-02-05 15:37:04 -08001910 if (epds.events & EPOLLEXCLUSIVE) {
1911 if (op == EPOLL_CTL_MOD)
1912 goto error_tgt_fput;
1913 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
1914 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
1915 goto error_tgt_fput;
1916 }
Jason Barondf0108c2016-01-20 14:59:24 -08001917
1918 /*
Davide Libenzi7699acd2007-05-10 22:23:23 -07001919 * At this point it is safe to assume that the "private_data" contains
1920 * our own data structure.
1921 */
Al Viro7e3fb582013-08-30 12:47:21 -04001922 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001923
Davide Libenzi22bacca2011-02-25 14:44:12 -08001924 /*
1925 * When we insert an epoll file descriptor, inside another epoll file
1926 * descriptor, there is the change of creating closed loops, which are
Jason Baron28d82dc2012-01-12 17:17:43 -08001927 * better be handled here, than in more critical paths. While we are
1928 * checking for loops we also determine the list of files reachable
1929 * and hang them on the tfile_check_list, so we can check that we
1930 * haven't created too many possible wakeup paths.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001931 *
Jason Baron67347fe2013-11-12 15:10:18 -08001932 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1933 * the epoll file descriptor is attaching directly to a wakeup source,
1934 * unless the epoll file descriptor is nested. The purpose of taking the
1935 * 'epmutex' on add is to prevent complex toplogies such as loops and
1936 * deep wakeup paths from forming in parallel through multiple
1937 * EPOLL_CTL_ADD operations.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001938 */
Nelson Elhaged8805e62011-10-31 17:13:14 -07001939 mutex_lock_nested(&ep->mtx, 0);
Jason Baron67347fe2013-11-12 15:10:18 -08001940 if (op == EPOLL_CTL_ADD) {
1941 if (!list_empty(&f.file->f_ep_links) ||
1942 is_file_epoll(tf.file)) {
1943 full_check = 1;
1944 mutex_unlock(&ep->mtx);
1945 mutex_lock(&epmutex);
1946 if (is_file_epoll(tf.file)) {
1947 error = -ELOOP;
1948 if (ep_loop_check(ep, tf.file) != 0) {
1949 clear_tfile_check_list();
1950 goto error_tgt_fput;
1951 }
1952 } else
1953 list_add(&tf.file->f_tfile_llink,
1954 &tfile_check_list);
1955 mutex_lock_nested(&ep->mtx, 0);
1956 if (is_file_epoll(tf.file)) {
1957 tep = tf.file->private_data;
1958 mutex_lock_nested(&tep->mtx, 1);
1959 }
1960 }
1961 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001962
Davide Libenzi67647d02007-05-15 01:40:52 -07001963 /*
1964 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1965 * above, we can be sure to be able to use the item looked up by
1966 * ep_find() till we release the mutex.
1967 */
Al Viro7e3fb582013-08-30 12:47:21 -04001968 epi = ep_find(ep, tf.file, fd);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001969
1970 error = -EINVAL;
1971 switch (op) {
1972 case EPOLL_CTL_ADD:
1973 if (!epi) {
1974 epds.events |= POLLERR | POLLHUP;
Jason Baron67347fe2013-11-12 15:10:18 -08001975 error = ep_insert(ep, &epds, tf.file, fd, full_check);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001976 } else
1977 error = -EEXIST;
Jason Baron67347fe2013-11-12 15:10:18 -08001978 if (full_check)
1979 clear_tfile_check_list();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001980 break;
1981 case EPOLL_CTL_DEL:
1982 if (epi)
1983 error = ep_remove(ep, epi);
1984 else
1985 error = -ENOENT;
1986 break;
1987 case EPOLL_CTL_MOD:
1988 if (epi) {
Jason Baronb6a515c2016-02-05 15:37:04 -08001989 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
1990 epds.events |= POLLERR | POLLHUP;
1991 error = ep_modify(ep, epi, &epds);
1992 }
Davide Libenzi7699acd2007-05-10 22:23:23 -07001993 } else
1994 error = -ENOENT;
1995 break;
1996 }
Jason Baron67347fe2013-11-12 15:10:18 -08001997 if (tep != NULL)
1998 mutex_unlock(&tep->mtx);
Davide Libenzid47de162007-05-15 01:40:41 -07001999 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002000
2001error_tgt_fput:
Jason Baron67347fe2013-11-12 15:10:18 -08002002 if (full_check)
Davide Libenzi22bacca2011-02-25 14:44:12 -08002003 mutex_unlock(&epmutex);
2004
Al Viro7e3fb582013-08-30 12:47:21 -04002005 fdput(tf);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002006error_fput:
Al Viro7e3fb582013-08-30 12:47:21 -04002007 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002008error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07002009
2010 return error;
2011}
2012
2013/*
2014 * Implement the event wait interface for the eventpoll file. It is the kernel
2015 * part of the user space epoll_wait(2).
2016 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01002017SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2018 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07002019{
Al Viro2903ff02012-08-28 12:52:22 -04002020 int error;
2021 struct fd f;
Davide Libenzi7699acd2007-05-10 22:23:23 -07002022 struct eventpoll *ep;
2023
Davide Libenzi7699acd2007-05-10 22:23:23 -07002024 /* The maximum number of event must be greater than zero */
2025 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2026 return -EINVAL;
2027
2028 /* Verify that the area passed by the user is writeable */
Al Viro2903ff02012-08-28 12:52:22 -04002029 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2030 return -EFAULT;
Davide Libenzi7699acd2007-05-10 22:23:23 -07002031
2032 /* Get the "struct file *" for the eventpoll file */
Al Viro2903ff02012-08-28 12:52:22 -04002033 f = fdget(epfd);
2034 if (!f.file)
2035 return -EBADF;
Davide Libenzi7699acd2007-05-10 22:23:23 -07002036
2037 /*
2038 * We have to check that the file structure underneath the fd
2039 * the user passed to us _is_ an eventpoll file.
2040 */
2041 error = -EINVAL;
Al Viro2903ff02012-08-28 12:52:22 -04002042 if (!is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07002043 goto error_fput;
2044
2045 /*
2046 * At this point it is safe to assume that the "private_data" contains
2047 * our own data structure.
2048 */
Al Viro2903ff02012-08-28 12:52:22 -04002049 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07002050
2051 /* Time to fish for events ... */
2052 error = ep_poll(ep, events, maxevents, timeout);
2053
2054error_fput:
Al Viro2903ff02012-08-28 12:52:22 -04002055 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002056 return error;
2057}
2058
Davide Libenzi7699acd2007-05-10 22:23:23 -07002059/*
2060 * Implement the event wait interface for the eventpoll file. It is the kernel
2061 * part of the user space epoll_pwait(2).
2062 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01002063SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2064 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2065 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07002066{
2067 int error;
2068 sigset_t ksigmask, sigsaved;
2069
2070 /*
2071 * If the caller wants a certain signal mask to be set during the wait,
2072 * we apply it here.
2073 */
2074 if (sigmask) {
2075 if (sigsetsize != sizeof(sigset_t))
2076 return -EINVAL;
2077 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2078 return -EFAULT;
Oleg Nesterov77d55912013-07-03 15:08:14 -07002079 sigsaved = current->blocked;
2080 set_current_blocked(&ksigmask);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002081 }
2082
2083 error = sys_epoll_wait(epfd, events, maxevents, timeout);
2084
2085 /*
2086 * If we changed the signal mask, we need to restore the original one.
2087 * In case we've got a signal while waiting, we do not restore the
2088 * signal mask yet, and we allow do_signal() to deliver the signal on
2089 * the way back to userspace, before the signal mask is restored.
2090 */
2091 if (sigmask) {
2092 if (error == -EINTR) {
2093 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07002094 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07002095 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07002096 } else
Oleg Nesterov77d55912013-07-03 15:08:14 -07002097 set_current_blocked(&sigsaved);
Davide Libenzi7699acd2007-05-10 22:23:23 -07002098 }
2099
2100 return error;
2101}
2102
Al Viro35280bd2013-02-24 14:52:17 -05002103#ifdef CONFIG_COMPAT
2104COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2105 struct epoll_event __user *, events,
2106 int, maxevents, int, timeout,
2107 const compat_sigset_t __user *, sigmask,
2108 compat_size_t, sigsetsize)
2109{
2110 long err;
2111 compat_sigset_t csigmask;
2112 sigset_t ksigmask, sigsaved;
2113
2114 /*
2115 * If the caller wants a certain signal mask to be set during the wait,
2116 * we apply it here.
2117 */
2118 if (sigmask) {
2119 if (sigsetsize != sizeof(compat_sigset_t))
2120 return -EINVAL;
2121 if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2122 return -EFAULT;
2123 sigset_from_compat(&ksigmask, &csigmask);
Oleg Nesterov77d55912013-07-03 15:08:14 -07002124 sigsaved = current->blocked;
2125 set_current_blocked(&ksigmask);
Al Viro35280bd2013-02-24 14:52:17 -05002126 }
2127
2128 err = sys_epoll_wait(epfd, events, maxevents, timeout);
2129
2130 /*
2131 * If we changed the signal mask, we need to restore the original one.
2132 * In case we've got a signal while waiting, we do not restore the
2133 * signal mask yet, and we allow do_signal() to deliver the signal on
2134 * the way back to userspace, before the signal mask is restored.
2135 */
2136 if (sigmask) {
2137 if (err == -EINTR) {
2138 memcpy(&current->saved_sigmask, &sigsaved,
2139 sizeof(sigsaved));
2140 set_restore_sigmask();
2141 } else
Oleg Nesterov77d55912013-07-03 15:08:14 -07002142 set_current_blocked(&sigsaved);
Al Viro35280bd2013-02-24 14:52:17 -05002143 }
2144
2145 return err;
2146}
2147#endif
2148
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149static int __init eventpoll_init(void)
2150{
Davide Libenzi7ef99642008-12-01 13:13:55 -08002151 struct sysinfo si;
2152
2153 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08002154 /*
2155 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2156 */
2157 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08002158 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08002159 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160
Davide Libenzi22bacca2011-02-25 14:44:12 -08002161 /*
2162 * Initialize the structure used to perform epoll file descriptor
2163 * inclusion loops checks.
2164 */
2165 ep_nested_calls_init(&poll_loop_ncalls);
2166
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07002168 ep_nested_calls_init(&poll_safewake_ncalls);
2169
2170 /* Initialize the structure used to perform file's f_op->poll() calls */
2171 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
Eric Wong39732ca2013-04-30 15:27:38 -07002173 /*
2174 * We can have many thousands of epitems, so prevent this from
2175 * using an extra cache line on 64-bit (and smaller) CPUs
2176 */
2177 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2178
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 /* Allocates slab cache used to allocate "struct epitem" items */
2180 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07002181 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182
2183 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2184 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07002185 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188}
Davide Libenzicea69242007-05-10 22:23:22 -07002189fs_initcall(eventpoll_init);