blob: a12820c08a0f98d8ddad7173bb086ed49d55711d [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 Hjønnevåg0a907c62012-05-01 21:33:34 +020036#include <linux/device.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include <asm/uaccess.h>
Colin Cross3bd702c2013-05-06 23:50:16 +000038#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <asm/io.h>
40#include <asm/mman.h>
Arun Sharma60063492011-07-26 16:09:06 -070041#include <linux/atomic.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
Linus Torvalds1da177e2005-04-16 15:20:36 -070043/*
44 * LOCKING:
45 * There are three level of locking required by epoll :
46 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080047 * 1) epmutex (mutex)
Davide Libenzic7ea7632007-05-15 01:40:47 -070048 * 2) ep->mtx (mutex)
49 * 3) ep->lock (spinlock)
Linus Torvalds1da177e2005-04-16 15:20:36 -070050 *
51 * The acquire order is the one listed above, from 1 to 3.
52 * We need a spinlock (ep->lock) because we manipulate objects
53 * from inside the poll callback, that might be triggered from
54 * a wake_up() that in turn might be called from IRQ context.
55 * So we can't sleep inside the poll callback and hence we need
56 * a spinlock. During the event transfer loop (from kernel to
57 * user space) we could end up sleeping due a copy_to_user(), so
58 * we need a lock that will allow us to sleep. This lock is a
Davide Libenzid47de162007-05-15 01:40:41 -070059 * mutex (ep->mtx). It is acquired during the event transfer loop,
60 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
61 * Then we also need a global mutex to serialize eventpoll_release_file()
62 * and ep_free().
63 * This mutex is acquired by ep_free() during the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 * cleanup path and it is also acquired by eventpoll_release_file()
65 * if a file has been pushed inside an epoll set and it is then
Daniel Balutabf6a41d2011-01-30 23:42:29 +020066 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
Davide Libenzi22bacca2011-02-25 14:44:12 -080067 * It is also acquired when inserting an epoll fd onto another epoll
68 * fd. We do this so that we walk the epoll tree and ensure that this
69 * insertion does not create a cycle of epoll file descriptors, which
70 * could lead to deadlock. We need a global mutex to prevent two
71 * simultaneous inserts (A into B and B into A) from racing and
72 * constructing a cycle without either insert observing that it is
73 * going to.
Nelson Elhaged8805e62011-10-31 17:13:14 -070074 * It is necessary to acquire multiple "ep->mtx"es at once in the
75 * case when one epoll fd is added to another. In this case, we
76 * always acquire the locks in the order of nesting (i.e. after
77 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
78 * before e2->mtx). Since we disallow cycles of epoll file
79 * descriptors, this ensures that the mutexes are well-ordered. In
80 * order to communicate this nesting to lockdep, when walking a tree
81 * of epoll file descriptors, we use the current recursion depth as
82 * the lockdep subkey.
Davide Libenzid47de162007-05-15 01:40:41 -070083 * It is possible to drop the "ep->mtx" and to use the global
84 * mutex "epmutex" (together with "ep->lock") to have it working,
85 * but having "ep->mtx" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080086 * Events that require holding "epmutex" are very rare, while for
Davide Libenzid47de162007-05-15 01:40:41 -070087 * normal operations the epoll private "ep->mtx" will guarantee
88 * a better scalability.
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 */
90
Linus Torvalds1da177e2005-04-16 15:20:36 -070091/* Epoll private bits inside the event mask */
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +020092#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
Linus Torvalds1da177e2005-04-16 15:20:36 -070093
Davide Libenzi5071f972009-03-31 15:24:10 -070094/* Maximum number of nesting allowed inside epoll sets */
95#define EP_MAX_NESTS 4
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Davide Libenzib6119672006-10-11 01:21:44 -070097#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
98
Davide Libenzid47de162007-05-15 01:40:41 -070099#define EP_UNACTIVE_PTR ((void *) -1L)
100
Davide Libenzi7ef99642008-12-01 13:13:55 -0800101#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103struct epoll_filefd {
104 struct file *file;
105 int fd;
106};
107
108/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700109 * Structure used to track possible nested calls, for too deep recursions
110 * and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700112struct nested_call_node {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 struct list_head llink;
Davide Libenzi5071f972009-03-31 15:24:10 -0700114 void *cookie;
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700115 void *ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116};
117
118/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700119 * This structure is used as collector for nested calls, to check for
120 * maximum recursion dept and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700122struct nested_calls {
123 struct list_head tasks_call_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 spinlock_t lock;
125};
126
127/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700129 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 */
131struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700132 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 struct rb_node rbn;
134
135 /* List header used to link this structure to the eventpoll ready list */
136 struct list_head rdllink;
137
Davide Libenzic7ea7632007-05-15 01:40:47 -0700138 /*
139 * Works together "struct eventpoll"->ovflist in keeping the
140 * single linked chain of items.
141 */
142 struct epitem *next;
143
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* The file descriptor information this item refers to */
145 struct epoll_filefd ffd;
146
147 /* Number of active wait queue attached to poll operations */
148 int nwait;
149
150 /* List containing poll wait queues */
151 struct list_head pwqlist;
152
153 /* The "container" of this item */
154 struct eventpoll *ep;
155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 /* List header used to link this item to the "struct file" items list */
157 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700158
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200159 /* wakeup_source used when EPOLLWAKEUP is set */
160 struct wakeup_source *ws;
161
Davide Libenzic7ea7632007-05-15 01:40:47 -0700162 /* The structure that describe the interested events and the source fd */
163 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700164};
165
166/*
167 * This structure is stored inside the "private_data" member of the file
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200168 * structure and represents the main data structure for the eventpoll
Davide Libenzid47de162007-05-15 01:40:41 -0700169 * interface.
170 */
171struct eventpoll {
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200172 /* Protect the access to this structure */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700173 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700174
175 /*
176 * This mutex is used to ensure that files are not removed
177 * while epoll is using them. This is held during the event
178 * collection loop, the file cleanup path, the epoll file exit
179 * code and the ctl operations.
180 */
181 struct mutex mtx;
182
183 /* Wait queue used by sys_epoll_wait() */
184 wait_queue_head_t wq;
185
186 /* Wait queue used by file->poll() */
187 wait_queue_head_t poll_wait;
188
189 /* List of ready file descriptors */
190 struct list_head rdllist;
191
Davide Libenzi67647d02007-05-15 01:40:52 -0700192 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700193 struct rb_root rbr;
194
195 /*
196 * This is a single linked list that chains all the "struct epitem" that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300197 * happened while transferring ready events to userspace w/out
Davide Libenzid47de162007-05-15 01:40:41 -0700198 * holding ->lock.
199 */
200 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800201
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200202 /* wakeup_source used when ep_scan_ready_list is running */
203 struct wakeup_source *ws;
204
Davide Libenzi7ef99642008-12-01 13:13:55 -0800205 /* The user that created the eventpoll descriptor */
206 struct user_struct *user;
Jason Baron28d82dc2012-01-12 17:17:43 -0800207
208 struct file *file;
209
210 /* used to optimize loop detection check */
211 int visited;
212 struct list_head visited_list_link;
Davide Libenzid47de162007-05-15 01:40:41 -0700213};
214
215/* Wait structure used by the poll hooks */
216struct eppoll_entry {
217 /* List header used to link this structure to the "struct epitem" */
218 struct list_head llink;
219
220 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700221 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700222
223 /*
224 * Wait queue item that will be linked to the target file wait
225 * queue head.
226 */
227 wait_queue_t wait;
228
229 /* The wait queue head that linked the "wait" wait queue item */
230 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231};
232
233/* Wrapper struct used by poll queueing */
234struct ep_pqueue {
235 poll_table pt;
236 struct epitem *epi;
237};
238
Davide Libenzi5071f972009-03-31 15:24:10 -0700239/* Used by the ep_send_events() function as callback private data */
240struct ep_send_events_data {
241 int maxevents;
242 struct epoll_event __user *events;
243};
244
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800246 * Configuration options available inside /proc/sys/fs/epoll/
247 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800248/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800249static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800250
251/*
Davide Libenzid47de162007-05-15 01:40:41 -0700252 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800254static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255
Davide Libenzi22bacca2011-02-25 14:44:12 -0800256/* Used to check for epoll file descriptor inclusion loops */
257static struct nested_calls poll_loop_ncalls;
258
Davide Libenzi5071f972009-03-31 15:24:10 -0700259/* Used for safe wake up implementation */
260static struct nested_calls poll_safewake_ncalls;
261
262/* Used to call file's f_op->poll() under the nested calls boundaries */
263static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264
265/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800266static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
268/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800269static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270
Jason Baron28d82dc2012-01-12 17:17:43 -0800271/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
272static LIST_HEAD(visited_list);
273
274/*
275 * List of files with newly added links, where we may need to limit the number
276 * of emanating paths. Protected by the epmutex.
277 */
278static LIST_HEAD(tfile_check_list);
279
Davide Libenzi7ef99642008-12-01 13:13:55 -0800280#ifdef CONFIG_SYSCTL
281
282#include <linux/sysctl.h>
283
Robin Holt52bd19f72011-01-12 17:00:01 -0800284static long zero;
285static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800286
287ctl_table epoll_table[] = {
288 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800289 .procname = "max_user_watches",
290 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800291 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800292 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800293 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800294 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800295 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800296 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800297 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800298};
299#endif /* CONFIG_SYSCTL */
300
Jason Baron28d82dc2012-01-12 17:17:43 -0800301static const struct file_operations eventpoll_fops;
302
303static inline int is_file_epoll(struct file *f)
304{
305 return f->f_op == &eventpoll_fops;
306}
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700307
Davide Libenzi67647d02007-05-15 01:40:52 -0700308/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700309static inline void ep_set_ffd(struct epoll_filefd *ffd,
310 struct file *file, int fd)
311{
312 ffd->file = file;
313 ffd->fd = fd;
314}
315
Davide Libenzi67647d02007-05-15 01:40:52 -0700316/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700317static inline int ep_cmp_ffd(struct epoll_filefd *p1,
318 struct epoll_filefd *p2)
319{
320 return (p1->file > p2->file ? +1:
321 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
322}
323
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700324/* Tells us if the item is currently linked */
325static inline int ep_is_linked(struct list_head *p)
326{
327 return !list_empty(p);
328}
329
Oleg Nesterov971316f2012-02-24 20:07:29 +0100330static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
331{
332 return container_of(p, struct eppoll_entry, wait);
333}
334
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700335/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700336static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700337{
338 return container_of(p, struct eppoll_entry, wait)->base;
339}
340
341/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700342static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700343{
344 return container_of(p, struct ep_pqueue, pt)->epi;
345}
346
347/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700348static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700349{
350 return op != EPOLL_CTL_DEL;
351}
352
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700354static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355{
Davide Libenzi5071f972009-03-31 15:24:10 -0700356 INIT_LIST_HEAD(&ncalls->tasks_call_list);
357 spin_lock_init(&ncalls->lock);
358}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
Davide Libenzi5071f972009-03-31 15:24:10 -0700360/**
Davide Libenzi3fb0e582011-03-22 16:34:46 -0700361 * ep_events_available - Checks if ready events might be available.
362 *
363 * @ep: Pointer to the eventpoll context.
364 *
365 * Returns: Returns a value different than zero if ready events are available,
366 * or zero otherwise.
367 */
368static inline int ep_events_available(struct eventpoll *ep)
369{
370 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
371}
372
373/**
Davide Libenzi5071f972009-03-31 15:24:10 -0700374 * ep_call_nested - Perform a bound (possibly) nested call, by checking
375 * that the recursion limit is not exceeded, and that
376 * the same nested call (by the meaning of same cookie) is
377 * no re-entered.
378 *
379 * @ncalls: Pointer to the nested_calls structure to be used for this call.
380 * @max_nests: Maximum number of allowed nesting calls.
381 * @nproc: Nested call core function pointer.
382 * @priv: Opaque data to be passed to the @nproc callback.
383 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700384 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700385 *
386 * Returns: Returns the code returned by the @nproc callback, or -1 if
387 * the maximum recursion limit has been exceeded.
388 */
389static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
390 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700391 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700392{
393 int error, call_nests = 0;
394 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700395 struct list_head *lsthead = &ncalls->tasks_call_list;
396 struct nested_call_node *tncur;
397 struct nested_call_node tnode;
398
399 spin_lock_irqsave(&ncalls->lock, flags);
400
401 /*
402 * Try to see if the current task is already inside this wakeup call.
403 * We use a list here, since the population inside this set is always
404 * very much limited.
405 */
406 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700407 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700408 (tncur->cookie == cookie || ++call_nests > max_nests)) {
409 /*
410 * Ops ... loop detected or maximum nest level reached.
411 * We abort this wake by breaking the cycle itself.
412 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700413 error = -1;
414 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700415 }
416 }
417
418 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700419 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700420 tnode.cookie = cookie;
421 list_add(&tnode.llink, lsthead);
422
423 spin_unlock_irqrestore(&ncalls->lock, flags);
424
425 /* Call the nested function */
426 error = (*nproc)(priv, cookie, call_nests);
427
428 /* Remove the current task from the list */
429 spin_lock_irqsave(&ncalls->lock, flags);
430 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700431out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700432 spin_unlock_irqrestore(&ncalls->lock, flags);
433
434 return error;
435}
436
Steven Rostedt02edc6f2012-03-23 15:02:27 -0700437/*
438 * As described in commit 0ccf831cb lockdep: annotate epoll
439 * the use of wait queues used by epoll is done in a very controlled
440 * manner. Wake ups can nest inside each other, but are never done
441 * with the same locking. For example:
442 *
443 * dfd = socket(...);
444 * efd1 = epoll_create();
445 * efd2 = epoll_create();
446 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
447 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
448 *
449 * When a packet arrives to the device underneath "dfd", the net code will
450 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
451 * callback wakeup entry on that queue, and the wake_up() performed by the
452 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
453 * (efd1) notices that it may have some event ready, so it needs to wake up
454 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
455 * that ends up in another wake_up(), after having checked about the
456 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
457 * avoid stack blasting.
458 *
459 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
460 * this special case of epoll.
461 */
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700462#ifdef CONFIG_DEBUG_LOCK_ALLOC
463static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
464 unsigned long events, int subclass)
465{
466 unsigned long flags;
467
468 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
469 wake_up_locked_poll(wqueue, events);
470 spin_unlock_irqrestore(&wqueue->lock, flags);
471}
472#else
473static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
474 unsigned long events, int subclass)
475{
476 wake_up_poll(wqueue, events);
477}
478#endif
479
Davide Libenzi5071f972009-03-31 15:24:10 -0700480static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
481{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700482 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
483 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700484 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485}
486
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487/*
488 * Perform a safe wake up of the poll wait list. The problem is that
489 * with the new callback'd wake up system, it is possible that the
490 * poll callback is reentered from inside the call to wake_up() done
491 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700492 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 * and we cannot reenter the same wait queue head at all. This will
494 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700495 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700497static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700499 int this_cpu = get_cpu();
500
Davide Libenzi5071f972009-03-31 15:24:10 -0700501 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700502 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
503
504 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505}
506
Oleg Nesterov971316f2012-02-24 20:07:29 +0100507static void ep_remove_wait_queue(struct eppoll_entry *pwq)
508{
509 wait_queue_head_t *whead;
510
511 rcu_read_lock();
512 /* If it is cleared by POLLFREE, it should be rcu-safe */
513 whead = rcu_dereference(pwq->whead);
514 if (whead)
515 remove_wait_queue(whead, &pwq->wait);
516 rcu_read_unlock();
517}
518
Davide Libenzi7699acd2007-05-10 22:23:23 -0700519/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700520 * This function unregisters poll callbacks from the associated file
521 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
522 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700523 */
524static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
525{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700526 struct list_head *lsthead = &epi->pwqlist;
527 struct eppoll_entry *pwq;
528
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700529 while (!list_empty(lsthead)) {
530 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700531
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700532 list_del(&pwq->llink);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100533 ep_remove_wait_queue(pwq);
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700534 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700535 }
536}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537
Davide Libenzi5071f972009-03-31 15:24:10 -0700538/**
539 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
540 * the scan code, to call f_op->poll(). Also allows for
541 * O(NumReady) performance.
542 *
543 * @ep: Pointer to the epoll private data structure.
544 * @sproc: Pointer to the scan callback.
545 * @priv: Private opaque data passed to the @sproc callback.
Nelson Elhaged8805e62011-10-31 17:13:14 -0700546 * @depth: The current depth of recursive f_op->poll calls.
Davide Libenzi5071f972009-03-31 15:24:10 -0700547 *
548 * Returns: The same integer error code returned by the @sproc callback.
549 */
550static int ep_scan_ready_list(struct eventpoll *ep,
551 int (*sproc)(struct eventpoll *,
552 struct list_head *, void *),
Nelson Elhaged8805e62011-10-31 17:13:14 -0700553 void *priv,
554 int depth)
Davide Libenzi5071f972009-03-31 15:24:10 -0700555{
556 int error, pwake = 0;
557 unsigned long flags;
558 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700559 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700560
561 /*
562 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700563 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700564 */
Nelson Elhaged8805e62011-10-31 17:13:14 -0700565 mutex_lock_nested(&ep->mtx, depth);
Davide Libenzi5071f972009-03-31 15:24:10 -0700566
567 /*
568 * Steal the ready list, and re-init the original one to the
569 * empty list. Also, set ep->ovflist to NULL so that events
570 * happening while looping w/out locks, are not lost. We cannot
571 * have the poll callback to queue directly on ep->rdllist,
572 * because we want the "sproc" callback to be able to do it
573 * in a lockless way.
574 */
575 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700576 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700577 ep->ovflist = NULL;
578 spin_unlock_irqrestore(&ep->lock, flags);
579
580 /*
581 * Now call the callback function.
582 */
583 error = (*sproc)(ep, &txlist, priv);
584
585 spin_lock_irqsave(&ep->lock, flags);
586 /*
587 * During the time we spent inside the "sproc" callback, some
588 * other events might have been queued by the poll callback.
589 * We re-insert them inside the main ready-list here.
590 */
591 for (nepi = ep->ovflist; (epi = nepi) != NULL;
592 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
593 /*
594 * We need to check if the item is already in the list.
595 * During the "sproc" callback execution time, items are
596 * queued into ->ovflist but the "txlist" might already
597 * contain them, and the list_splice() below takes care of them.
598 */
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200599 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700600 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200601 __pm_stay_awake(epi->ws);
602 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700603 }
604 /*
605 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
606 * releasing the lock, events will be queued in the normal way inside
607 * ep->rdllist.
608 */
609 ep->ovflist = EP_UNACTIVE_PTR;
610
611 /*
612 * Quickly re-inject items left on "txlist".
613 */
614 list_splice(&txlist, &ep->rdllist);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200615 __pm_relax(ep->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700616
617 if (!list_empty(&ep->rdllist)) {
618 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700619 * Wake up (if active) both the eventpoll wait list and
620 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700621 */
622 if (waitqueue_active(&ep->wq))
623 wake_up_locked(&ep->wq);
624 if (waitqueue_active(&ep->poll_wait))
625 pwake++;
626 }
627 spin_unlock_irqrestore(&ep->lock, flags);
628
629 mutex_unlock(&ep->mtx);
630
631 /* We have to call this outside the lock */
632 if (pwake)
633 ep_poll_safewake(&ep->poll_wait);
634
635 return error;
636}
637
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700639 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700640 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700641 */
642static int ep_remove(struct eventpoll *ep, struct epitem *epi)
643{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700644 unsigned long flags;
645 struct file *file = epi->ffd.file;
646
647 /*
648 * Removes poll wait queue hooks. We _have_ to do this without holding
649 * the "ep->lock" otherwise a deadlock might occur. This because of the
650 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
651 * queue head lock when unregistering the wait queue. The wakeup callback
652 * will run by holding the wait queue head lock and will call our callback
653 * that will try to get "ep->lock".
654 */
655 ep_unregister_pollwait(ep, epi);
656
657 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700658 spin_lock(&file->f_lock);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700659 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700660 list_del_init(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700661 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Davide Libenzicdac75e2008-04-29 00:58:34 -0700663 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
Davide Libenzic7ea7632007-05-15 01:40:47 -0700665 spin_lock_irqsave(&ep->lock, flags);
666 if (ep_is_linked(&epi->rdllink))
667 list_del_init(&epi->rdllink);
668 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200670 wakeup_source_unregister(epi->ws);
671
Davide Libenzi7699acd2007-05-10 22:23:23 -0700672 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700673 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
Robin Holt52bd19f72011-01-12 17:00:01 -0800675 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800676
Davide Libenzic7ea7632007-05-15 01:40:47 -0700677 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678}
679
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680static void ep_free(struct eventpoll *ep)
681{
682 struct rb_node *rbp;
683 struct epitem *epi;
684
685 /* We need to release all tasks waiting for these file */
686 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700687 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
689 /*
690 * We need to lock this because we could be hit by
691 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700692 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 * is on the way to be removed and no one has references to it
694 * anymore. The only hit might come from eventpoll_release_file() but
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300695 * holding "epmutex" is sufficient here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800697 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
699 /*
700 * Walks through the whole tree by unregistering poll callbacks.
701 */
702 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
703 epi = rb_entry(rbp, struct epitem, rbn);
704
705 ep_unregister_pollwait(ep, epi);
706 }
707
708 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700709 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700711 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 * us during this operation. So we can avoid the lock on "ep->lock".
713 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700714 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 epi = rb_entry(rbp, struct epitem, rbn);
716 ep_remove(ep, epi);
717 }
718
Arjan van de Ven144efe32006-03-23 03:00:32 -0800719 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700720 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800721 free_uid(ep->user);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200722 wakeup_source_unregister(ep->ws);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700723 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724}
725
Davide Libenzi7699acd2007-05-10 22:23:23 -0700726static int ep_eventpoll_release(struct inode *inode, struct file *file)
727{
728 struct eventpoll *ep = file->private_data;
729
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700730 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700731 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700732
Davide Libenzi7699acd2007-05-10 22:23:23 -0700733 return 0;
734}
735
Davide Libenzi296e2362009-03-31 15:24:11 -0700736static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
737 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700738{
739 struct epitem *epi, *tmp;
Hans Verkuil626cf232012-03-23 15:02:27 -0700740 poll_table pt;
Davide Libenzi5071f972009-03-31 15:24:10 -0700741
Hans Verkuil626cf232012-03-23 15:02:27 -0700742 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -0700743 list_for_each_entry_safe(epi, tmp, head, rdllink) {
Hans Verkuil626cf232012-03-23 15:02:27 -0700744 pt._key = epi->event.events;
745 if (epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
Davide Libenzi5071f972009-03-31 15:24:10 -0700746 epi->event.events)
747 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700748 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700749 /*
750 * Item has been dropped into the ready list by the poll
751 * callback, but it's not actually ready, as far as
752 * caller requested events goes. We can remove it here.
753 */
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200754 __pm_relax(epi->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700755 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700756 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700757 }
758
759 return 0;
760}
761
762static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
763{
Nelson Elhaged8805e62011-10-31 17:13:14 -0700764 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
Davide Libenzi5071f972009-03-31 15:24:10 -0700765}
766
Davide Libenzi7699acd2007-05-10 22:23:23 -0700767static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
768{
Davide Libenzi5071f972009-03-31 15:24:10 -0700769 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700770 struct eventpoll *ep = file->private_data;
771
772 /* Insert inside our poll wait queue */
773 poll_wait(file, &ep->poll_wait, wait);
774
Davide Libenzi5071f972009-03-31 15:24:10 -0700775 /*
776 * Proceed to find out if wanted events are really available inside
777 * the ready list. This need to be done under ep_call_nested()
778 * supervision, since the call to f_op->poll() done on listed files
779 * could re-enter here.
780 */
781 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700782 ep_poll_readyevents_proc, ep, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700783
Davide Libenzi296e2362009-03-31 15:24:11 -0700784 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700785}
786
787/* File callbacks that implement the eventpoll file behaviour */
788static const struct file_operations eventpoll_fops = {
789 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200790 .poll = ep_eventpoll_poll,
791 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700792};
793
Davide Libenzi7699acd2007-05-10 22:23:23 -0700794/*
795 * This is called from eventpoll_release() to unlink files from the eventpoll
796 * interface. We need to have this facility to cleanup correctly files that are
797 * closed without being removed from the eventpoll interface.
798 */
799void eventpoll_release_file(struct file *file)
800{
801 struct list_head *lsthead = &file->f_ep_links;
802 struct eventpoll *ep;
803 struct epitem *epi;
804
805 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700806 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700807 * necessary. It is not necessary because we're in the "struct file"
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300808 * cleanup path, and this means that no one is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700809 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700810 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700811 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700812 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700813 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700814 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700815 *
816 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700817 */
818 mutex_lock(&epmutex);
819
820 while (!list_empty(lsthead)) {
821 epi = list_first_entry(lsthead, struct epitem, fllink);
822
823 ep = epi->ep;
824 list_del_init(&epi->fllink);
Nelson Elhaged8805e62011-10-31 17:13:14 -0700825 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700826 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700827 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700828 }
829
830 mutex_unlock(&epmutex);
831}
832
833static int ep_alloc(struct eventpoll **pep)
834{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800835 int error;
836 struct user_struct *user;
837 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700838
Davide Libenzi7ef99642008-12-01 13:13:55 -0800839 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800840 error = -ENOMEM;
841 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
842 if (unlikely(!ep))
843 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700844
Davide Libenzic7ea7632007-05-15 01:40:47 -0700845 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700846 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700847 init_waitqueue_head(&ep->wq);
848 init_waitqueue_head(&ep->poll_wait);
849 INIT_LIST_HEAD(&ep->rdllist);
850 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700851 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800852 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700853
854 *pep = ep;
855
Davide Libenzi7699acd2007-05-10 22:23:23 -0700856 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800857
858free_uid:
859 free_uid(user);
860 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700861}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
863/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700864 * Search the file inside the eventpoll tree. The RB tree operations
865 * are protected by the "mtx" mutex, and ep_find() must be called with
866 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 */
868static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
869{
870 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 struct rb_node *rbp;
872 struct epitem *epi, *epir = NULL;
873 struct epoll_filefd ffd;
874
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700875 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 for (rbp = ep->rbr.rb_node; rbp; ) {
877 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700878 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 if (kcmp > 0)
880 rbp = rbp->rb_right;
881 else if (kcmp < 0)
882 rbp = rbp->rb_left;
883 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 epir = epi;
885 break;
886 }
887 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 return epir;
890}
891
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700893 * This is the callback that is passed to the wait queue wakeup
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200894 * mechanism. It is called by the stored file descriptors when they
Davide Libenzi7699acd2007-05-10 22:23:23 -0700895 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700897static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700899 int pwake = 0;
900 unsigned long flags;
901 struct epitem *epi = ep_item_from_wait(wait);
902 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903
Oleg Nesterov971316f2012-02-24 20:07:29 +0100904 if ((unsigned long)key & POLLFREE) {
905 ep_pwq_from_wait(wait)->whead = NULL;
906 /*
907 * whead = NULL above can race with ep_remove_wait_queue()
908 * which can do another remove_wait_queue() after us, so we
909 * can't use __remove_wait_queue(). whead->lock is held by
910 * the caller.
911 */
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100912 list_del_init(&wait->task_list);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100913 }
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100914
Davide Libenzic7ea7632007-05-15 01:40:47 -0700915 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700916
917 /*
918 * If the event mask does not contain any poll(2) event, we consider the
919 * descriptor to be disabled. This condition is likely the effect of the
920 * EPOLLONESHOT bit that disables the descriptor when an event is received,
921 * until the next EPOLL_CTL_MOD will be issued.
922 */
923 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700924 goto out_unlock;
925
926 /*
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700927 * Check the events coming with the callback. At this stage, not
928 * every device reports the events in the "key" parameter of the
929 * callback. We need to be able to handle both cases here, hence the
930 * test for "key" != NULL before the event match test.
931 */
932 if (key && !((unsigned long) key & epi->event.events))
933 goto out_unlock;
934
935 /*
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200936 * If we are transferring events to userspace, we can hold no locks
Davide Libenzid47de162007-05-15 01:40:41 -0700937 * (because we're accessing user memory, and because of linux f_op->poll()
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200938 * semantics). All the events that happen during that period of time are
Davide Libenzid47de162007-05-15 01:40:41 -0700939 * chained in ep->ovflist and requeued later on.
940 */
941 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
942 if (epi->next == EP_UNACTIVE_PTR) {
943 epi->next = ep->ovflist;
944 ep->ovflist = epi;
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200945 if (epi->ws) {
946 /*
947 * Activate ep->ws since epi->ws may get
948 * deactivated at any time.
949 */
950 __pm_stay_awake(ep->ws);
951 }
952
Davide Libenzid47de162007-05-15 01:40:41 -0700953 }
954 goto out_unlock;
955 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700956
957 /* If this file is already in the ready list we exit soon */
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200958 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700959 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +0200960 __pm_stay_awake(epi->ws);
961 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700962
Davide Libenzi7699acd2007-05-10 22:23:23 -0700963 /*
964 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
965 * wait list.
966 */
967 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -0400968 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700969 if (waitqueue_active(&ep->poll_wait))
970 pwake++;
971
Davide Libenzid47de162007-05-15 01:40:41 -0700972out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -0700973 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700974
975 /* We have to call this outside the lock */
976 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -0700977 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700978
979 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980}
981
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982/*
983 * This is the callback that is used to add our wait queue to the
984 * target file wakeup lists.
985 */
986static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
987 poll_table *pt)
988{
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700989 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 struct eppoll_entry *pwq;
991
Christoph Lametere94b1762006-12-06 20:33:17 -0800992 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
994 pwq->whead = whead;
995 pwq->base = epi;
996 add_wait_queue(whead, &pwq->wait);
997 list_add_tail(&pwq->llink, &epi->pwqlist);
998 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -0700999 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 /* We have to signal that an error occurred */
1001 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -07001002 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003}
1004
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1006{
1007 int kcmp;
1008 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1009 struct epitem *epic;
1010
1011 while (*p) {
1012 parent = *p;
1013 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001014 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 if (kcmp > 0)
1016 p = &parent->rb_right;
1017 else
1018 p = &parent->rb_left;
1019 }
1020 rb_link_node(&epi->rbn, parent, p);
1021 rb_insert_color(&epi->rbn, &ep->rbr);
1022}
1023
Jason Baron28d82dc2012-01-12 17:17:43 -08001024
1025
1026#define PATH_ARR_SIZE 5
1027/*
1028 * These are the number paths of length 1 to 5, that we are allowing to emanate
1029 * from a single file of interest. For example, we allow 1000 paths of length
1030 * 1, to emanate from each file of interest. This essentially represents the
1031 * potential wakeup paths, which need to be limited in order to avoid massive
1032 * uncontrolled wakeup storms. The common use case should be a single ep which
1033 * is connected to n file sources. In this case each file source has 1 path
1034 * of length 1. Thus, the numbers below should be more than sufficient. These
1035 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1036 * and delete can't add additional paths. Protected by the epmutex.
1037 */
1038static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1039static int path_count[PATH_ARR_SIZE];
1040
1041static int path_count_inc(int nests)
1042{
Jason Baron93dc6102012-03-16 16:34:03 -04001043 /* Allow an arbitrary number of depth 1 paths */
1044 if (nests == 0)
1045 return 0;
1046
Jason Baron28d82dc2012-01-12 17:17:43 -08001047 if (++path_count[nests] > path_limits[nests])
1048 return -1;
1049 return 0;
1050}
1051
1052static void path_count_init(void)
1053{
1054 int i;
1055
1056 for (i = 0; i < PATH_ARR_SIZE; i++)
1057 path_count[i] = 0;
1058}
1059
1060static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1061{
1062 int error = 0;
1063 struct file *file = priv;
1064 struct file *child_file;
1065 struct epitem *epi;
1066
1067 list_for_each_entry(epi, &file->f_ep_links, fllink) {
1068 child_file = epi->ep->file;
1069 if (is_file_epoll(child_file)) {
1070 if (list_empty(&child_file->f_ep_links)) {
1071 if (path_count_inc(call_nests)) {
1072 error = -1;
1073 break;
1074 }
1075 } else {
1076 error = ep_call_nested(&poll_loop_ncalls,
1077 EP_MAX_NESTS,
1078 reverse_path_check_proc,
1079 child_file, child_file,
1080 current);
1081 }
1082 if (error != 0)
1083 break;
1084 } else {
1085 printk(KERN_ERR "reverse_path_check_proc: "
1086 "file is not an ep!\n");
1087 }
1088 }
1089 return error;
1090}
1091
1092/**
1093 * reverse_path_check - The tfile_check_list is list of file *, which have
1094 * links that are proposed to be newly added. We need to
1095 * make sure that those added links don't add too many
1096 * paths such that we will spend all our time waking up
1097 * eventpoll objects.
1098 *
1099 * Returns: Returns zero if the proposed links don't create too many paths,
1100 * -1 otherwise.
1101 */
1102static int reverse_path_check(void)
1103{
Jason Baron28d82dc2012-01-12 17:17:43 -08001104 int error = 0;
1105 struct file *current_file;
1106
1107 /* let's call this for all tfiles */
1108 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001109 path_count_init();
1110 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1111 reverse_path_check_proc, current_file,
1112 current_file, current);
1113 if (error)
1114 break;
1115 }
1116 return error;
1117}
1118
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001119static int ep_create_wakeup_source(struct epitem *epi)
1120{
1121 const char *name;
1122
1123 if (!epi->ep->ws) {
1124 epi->ep->ws = wakeup_source_register("eventpoll");
1125 if (!epi->ep->ws)
1126 return -ENOMEM;
1127 }
1128
1129 name = epi->ffd.file->f_path.dentry->d_name.name;
1130 epi->ws = wakeup_source_register(name);
1131 if (!epi->ws)
1132 return -ENOMEM;
1133
1134 return 0;
1135}
1136
1137static void ep_destroy_wakeup_source(struct epitem *epi)
1138{
1139 wakeup_source_unregister(epi->ws);
1140 epi->ws = NULL;
1141}
1142
Davide Libenzic7ea7632007-05-15 01:40:47 -07001143/*
1144 * Must be called with "mtx" held.
1145 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1147 struct file *tfile, int fd)
1148{
1149 int error, revents, pwake = 0;
1150 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -08001151 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 struct epitem *epi;
1153 struct ep_pqueue epq;
1154
Robin Holt52bd19f72011-01-12 17:00:01 -08001155 user_watches = atomic_long_read(&ep->user->epoll_watches);
1156 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001157 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -08001158 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001159 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
1161 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 INIT_LIST_HEAD(&epi->rdllink);
1163 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 INIT_LIST_HEAD(&epi->pwqlist);
1165 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001166 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -07001169 epi->next = EP_UNACTIVE_PTR;
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001170 if (epi->event.events & EPOLLWAKEUP) {
1171 error = ep_create_wakeup_source(epi);
1172 if (error)
1173 goto error_create_wakeup_source;
1174 } else {
1175 epi->ws = NULL;
1176 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
1178 /* Initialize the poll table using the queue callback */
1179 epq.epi = epi;
1180 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
Hans Verkuil626cf232012-03-23 15:02:27 -07001181 epq.pt._key = event->events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182
1183 /*
1184 * Attach the item to the poll hooks and get current event bits.
1185 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -07001186 * been increased by the caller of this function. Note that after
1187 * this operation completes, the poll callback can start hitting
1188 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 */
1190 revents = tfile->f_op->poll(tfile, &epq.pt);
1191
1192 /*
1193 * We have to check if something went wrong during the poll wait queue
1194 * install process. Namely an allocation for a wait queue failed due
1195 * high memory pressure.
1196 */
Davide Libenzi7ef99642008-12-01 13:13:55 -08001197 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001199 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
1201 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -07001202 spin_lock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 list_add_tail(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -07001204 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205
Davide Libenzic7ea7632007-05-15 01:40:47 -07001206 /*
1207 * Add the current item to the RB tree. All RB tree operations are
1208 * protected by "mtx", and ep_insert() is called with "mtx" held.
1209 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 ep_rbtree_insert(ep, epi);
1211
Jason Baron28d82dc2012-01-12 17:17:43 -08001212 /* now check if we've created too many backpaths */
1213 error = -EINVAL;
1214 if (reverse_path_check())
1215 goto error_remove_epi;
1216
Davide Libenzic7ea7632007-05-15 01:40:47 -07001217 /* We have to drop the new item inside our item list to keep track of it */
1218 spin_lock_irqsave(&ep->lock, flags);
1219
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001221 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001223 __pm_stay_awake(epi->ws);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224
1225 /* Notify waiting tasks that events are available */
1226 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001227 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 if (waitqueue_active(&ep->poll_wait))
1229 pwake++;
1230 }
1231
Davide Libenzic7ea7632007-05-15 01:40:47 -07001232 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233
Robin Holt52bd19f72011-01-12 17:00:01 -08001234 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001235
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 /* We have to call this outside the lock */
1237 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001238 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 return 0;
1241
Jason Baron28d82dc2012-01-12 17:17:43 -08001242error_remove_epi:
1243 spin_lock(&tfile->f_lock);
1244 if (ep_is_linked(&epi->fllink))
1245 list_del_init(&epi->fllink);
1246 spin_unlock(&tfile->f_lock);
1247
1248 rb_erase(&epi->rbn, &ep->rbr);
1249
Davide Libenzi7699acd2007-05-10 22:23:23 -07001250error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 ep_unregister_pollwait(ep, epi);
1252
1253 /*
1254 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -07001255 * allocated wait queue. Note that we don't care about the ep->ovflist
1256 * list, since that is used/cleaned only inside a section bound by "mtx".
1257 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001259 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001260 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001261 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001262 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001264 wakeup_source_unregister(epi->ws);
1265
1266error_create_wakeup_source:
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001267 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001268
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 return error;
1270}
1271
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272/*
1273 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001274 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 */
1276static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1277{
1278 int pwake = 0;
1279 unsigned int revents;
Hans Verkuil626cf232012-03-23 15:02:27 -07001280 poll_table pt;
1281
1282 init_poll_funcptr(&pt, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283
1284 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001285 * Set the new event interest mask before calling f_op->poll();
1286 * otherwise we might miss an event that happens between the
1287 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 */
1289 epi->event.events = event->events;
Hans Verkuil626cf232012-03-23 15:02:27 -07001290 pt._key = event->events;
Tony Battersbye057e152009-03-31 15:24:15 -07001291 epi->event.data = event->data; /* protected by mtx */
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001292 if (epi->event.events & EPOLLWAKEUP) {
1293 if (!epi->ws)
1294 ep_create_wakeup_source(epi);
1295 } else if (epi->ws) {
1296 ep_destroy_wakeup_source(epi);
1297 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
1299 /*
Eric Wong37689292013-01-01 21:20:27 +00001300 * The following barrier has two effects:
1301 *
1302 * 1) Flush epi changes above to other CPUs. This ensures
1303 * we do not miss events from ep_poll_callback if an
1304 * event occurs immediately after we call f_op->poll().
1305 * We need this because we did not take ep->lock while
1306 * changing epi above (but ep_poll_callback does take
1307 * ep->lock).
1308 *
1309 * 2) We also need to ensure we do not miss _past_ events
1310 * when calling f_op->poll(). This barrier also
1311 * pairs with the barrier in wq_has_sleeper (see
1312 * comments for wq_has_sleeper).
1313 *
1314 * This barrier will now guarantee ep_poll_callback or f_op->poll
1315 * (or both) will notice the readiness of an item.
1316 */
1317 smp_mb();
1318
1319 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 * Get current event bits. We can safely use the file* here because
1321 * its usage count has been increased by the caller of this function.
1322 */
Hans Verkuil626cf232012-03-23 15:02:27 -07001323 revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001326 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001327 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001329 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001330 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001331 if (!ep_is_linked(&epi->rdllink)) {
1332 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001333 __pm_stay_awake(epi->ws);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334
Davide Libenzic7ea7632007-05-15 01:40:47 -07001335 /* Notify waiting tasks that events are available */
1336 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e22007-08-30 16:10:22 -04001337 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001338 if (waitqueue_active(&ep->poll_wait))
1339 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 }
Tony Battersbye057e152009-03-31 15:24:15 -07001341 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343
1344 /* We have to call this outside the lock */
1345 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001346 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 return 0;
1349}
1350
Davide Libenzi296e2362009-03-31 15:24:11 -07001351static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1352 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001353{
1354 struct ep_send_events_data *esed = priv;
1355 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001356 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001357 struct epitem *epi;
1358 struct epoll_event __user *uevent;
Hans Verkuil626cf232012-03-23 15:02:27 -07001359 poll_table pt;
1360
1361 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -07001362
Davide Libenzi296e2362009-03-31 15:24:11 -07001363 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001364 * We can loop without lock because we are passed a task private list.
1365 * Items cannot vanish during the loop because ep_scan_ready_list() is
1366 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001367 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001368 for (eventcnt = 0, uevent = esed->events;
1369 !list_empty(head) && eventcnt < esed->maxevents;) {
1370 epi = list_first_entry(head, struct epitem, rdllink);
1371
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001372 /*
1373 * Activate ep->ws before deactivating epi->ws to prevent
1374 * triggering auto-suspend here (in case we reactive epi->ws
1375 * below).
1376 *
1377 * This could be rearranged to delay the deactivation of epi->ws
1378 * instead, but then epi->ws would temporarily be out of sync
1379 * with ep_is_linked().
1380 */
1381 if (epi->ws && epi->ws->active)
1382 __pm_stay_awake(ep->ws);
1383 __pm_relax(epi->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -07001384 list_del_init(&epi->rdllink);
1385
Hans Verkuil626cf232012-03-23 15:02:27 -07001386 pt._key = epi->event.events;
1387 revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
Davide Libenzi296e2362009-03-31 15:24:11 -07001388 epi->event.events;
Davide Libenzi5071f972009-03-31 15:24:10 -07001389
Davide Libenzi296e2362009-03-31 15:24:11 -07001390 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001391 * If the event mask intersect the caller-requested one,
1392 * deliver the event to userspace. Again, ep_scan_ready_list()
1393 * is holding "mtx", so no operations coming from userspace
1394 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001395 */
1396 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001397 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001398 __put_user(epi->event.data, &uevent->data)) {
1399 list_add(&epi->rdllink, head);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001400 __pm_stay_awake(epi->ws);
Davide Libenzi296e2362009-03-31 15:24:11 -07001401 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001402 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001403 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001404 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001405 if (epi->event.events & EPOLLONESHOT)
1406 epi->event.events &= EP_PRIVATE_BITS;
1407 else if (!(epi->event.events & EPOLLET)) {
1408 /*
1409 * If this file has been added with Level
1410 * Trigger mode, we need to insert back inside
1411 * the ready list, so that the next call to
1412 * epoll_wait() will check again the events
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001413 * availability. At this point, no one can insert
Davide Libenzi296e2362009-03-31 15:24:11 -07001414 * into ep->rdllist besides us. The epoll_ctl()
1415 * callers are locked out by
1416 * ep_scan_ready_list() holding "mtx" and the
1417 * poll callback will queue them in ep->ovflist.
1418 */
1419 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001420 __pm_stay_awake(epi->ws);
Davide Libenzi296e2362009-03-31 15:24:11 -07001421 }
1422 }
1423 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001424
1425 return eventcnt;
1426}
1427
Davide Libenzi296e2362009-03-31 15:24:11 -07001428static int ep_send_events(struct eventpoll *ep,
1429 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430{
Davide Libenzi5071f972009-03-31 15:24:10 -07001431 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001432
Davide Libenzi5071f972009-03-31 15:24:10 -07001433 esed.maxevents = maxevents;
1434 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001435
Nelson Elhaged8805e62011-10-31 17:13:14 -07001436 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437}
1438
Eric Dumazet0781b902011-02-01 15:52:35 -08001439static inline struct timespec ep_set_mstimeout(long ms)
1440{
1441 struct timespec now, ts = {
1442 .tv_sec = ms / MSEC_PER_SEC,
1443 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1444 };
1445
1446 ktime_get_ts(&now);
1447 return timespec_add_safe(now, ts);
1448}
1449
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001450/**
1451 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1452 * event buffer.
1453 *
1454 * @ep: Pointer to the eventpoll context.
1455 * @events: Pointer to the userspace buffer where the ready events should be
1456 * stored.
1457 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1458 * @timeout: Maximum timeout for the ready events fetch operation, in
1459 * milliseconds. If the @timeout is zero, the function will not block,
1460 * while if the @timeout is less than zero, the function will block
1461 * until at least one event has been retrieved (or an error
1462 * occurred).
1463 *
1464 * Returns: Returns the number of ready events which have been fetched, or an
1465 * error code, in case of error.
1466 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1468 int maxevents, long timeout)
1469{
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001470 int res = 0, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 unsigned long flags;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001472 long slack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001474 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001476 if (timeout > 0) {
Eric Dumazet0781b902011-02-01 15:52:35 -08001477 struct timespec end_time = ep_set_mstimeout(timeout);
1478
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001479 slack = select_estimate_accuracy(&end_time);
1480 to = &expires;
1481 *to = timespec_to_ktime(end_time);
1482 } else if (timeout == 0) {
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001483 /*
1484 * Avoid the unnecessary trip to the wait queue loop, if the
1485 * caller specified a non blocking operation.
1486 */
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001487 timed_out = 1;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001488 spin_lock_irqsave(&ep->lock, flags);
1489 goto check_events;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001492fetch_events:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001493 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001495 if (!ep_events_available(ep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 /*
1497 * We don't have any available event to return to the caller.
1498 * We need to sleep here, and we will be wake up by
1499 * ep_poll_callback() when events will become available.
1500 */
1501 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001502 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
1504 for (;;) {
1505 /*
1506 * We don't want to sleep if the ep_poll_callback() sends us
1507 * a wakeup in between. That's why we set the task state
1508 * to TASK_INTERRUPTIBLE before doing the checks.
1509 */
1510 set_current_state(TASK_INTERRUPTIBLE);
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001511 if (ep_events_available(ep) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 break;
1513 if (signal_pending(current)) {
1514 res = -EINTR;
1515 break;
1516 }
1517
Davide Libenzic7ea7632007-05-15 01:40:47 -07001518 spin_unlock_irqrestore(&ep->lock, flags);
Colin Cross3bd702c2013-05-06 23:50:16 +00001519 if (!freezable_schedule_hrtimeout_range(to, slack,
1520 HRTIMER_MODE_ABS))
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001521 timed_out = 1;
1522
Davide Libenzic7ea7632007-05-15 01:40:47 -07001523 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001525 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
1527 set_current_state(TASK_RUNNING);
1528 }
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001529check_events:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 /* Is it worth to try to dig for events ? */
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001531 eavail = ep_events_available(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
Davide Libenzic7ea7632007-05-15 01:40:47 -07001533 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534
1535 /*
1536 * Try to transfer events to user space. In case we get 0 events and
1537 * there's still timeout left over, we go trying again in search of
1538 * more luck.
1539 */
1540 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001541 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001542 goto fetch_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
1544 return res;
1545}
1546
Davide Libenzi22bacca2011-02-25 14:44:12 -08001547/**
1548 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1549 * API, to verify that adding an epoll file inside another
1550 * epoll structure, does not violate the constraints, in
1551 * terms of closed loops, or too deep chains (which can
1552 * result in excessive stack usage).
1553 *
1554 * @priv: Pointer to the epoll file to be currently checked.
1555 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1556 * data structure pointer.
1557 * @call_nests: Current dept of the @ep_call_nested() call stack.
1558 *
1559 * Returns: Returns zero if adding the epoll @file inside current epoll
1560 * structure @ep does not violate the constraints, or -1 otherwise.
1561 */
1562static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1563{
1564 int error = 0;
1565 struct file *file = priv;
1566 struct eventpoll *ep = file->private_data;
Jason Baron28d82dc2012-01-12 17:17:43 -08001567 struct eventpoll *ep_tovisit;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001568 struct rb_node *rbp;
1569 struct epitem *epi;
1570
Nelson Elhaged8805e62011-10-31 17:13:14 -07001571 mutex_lock_nested(&ep->mtx, call_nests + 1);
Jason Baron28d82dc2012-01-12 17:17:43 -08001572 ep->visited = 1;
1573 list_add(&ep->visited_list_link, &visited_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001574 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1575 epi = rb_entry(rbp, struct epitem, rbn);
1576 if (unlikely(is_file_epoll(epi->ffd.file))) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001577 ep_tovisit = epi->ffd.file->private_data;
1578 if (ep_tovisit->visited)
1579 continue;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001580 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Jason Baron28d82dc2012-01-12 17:17:43 -08001581 ep_loop_check_proc, epi->ffd.file,
1582 ep_tovisit, current);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001583 if (error != 0)
1584 break;
Jason Baron28d82dc2012-01-12 17:17:43 -08001585 } else {
1586 /*
1587 * If we've reached a file that is not associated with
1588 * an ep, then we need to check if the newly added
1589 * links are going to add too many wakeup paths. We do
1590 * this by adding it to the tfile_check_list, if it's
1591 * not already there, and calling reverse_path_check()
1592 * during ep_insert().
1593 */
1594 if (list_empty(&epi->ffd.file->f_tfile_llink))
1595 list_add(&epi->ffd.file->f_tfile_llink,
1596 &tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001597 }
1598 }
1599 mutex_unlock(&ep->mtx);
1600
1601 return error;
1602}
1603
1604/**
1605 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1606 * another epoll file (represented by @ep) does not create
1607 * closed loops or too deep chains.
1608 *
1609 * @ep: Pointer to the epoll private data structure.
1610 * @file: Pointer to the epoll file to be checked.
1611 *
1612 * Returns: Returns zero if adding the epoll @file inside current epoll
1613 * structure @ep does not violate the constraints, or -1 otherwise.
1614 */
1615static int ep_loop_check(struct eventpoll *ep, struct file *file)
1616{
Jason Baron28d82dc2012-01-12 17:17:43 -08001617 int ret;
1618 struct eventpoll *ep_cur, *ep_next;
1619
1620 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Davide Libenzi22bacca2011-02-25 14:44:12 -08001621 ep_loop_check_proc, file, ep, current);
Jason Baron28d82dc2012-01-12 17:17:43 -08001622 /* clear visited list */
1623 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1624 visited_list_link) {
1625 ep_cur->visited = 0;
1626 list_del(&ep_cur->visited_list_link);
1627 }
1628 return ret;
1629}
1630
1631static void clear_tfile_check_list(void)
1632{
1633 struct file *file;
1634
1635 /* first clear the tfile_check_list */
1636 while (!list_empty(&tfile_check_list)) {
1637 file = list_first_entry(&tfile_check_list, struct file,
1638 f_tfile_llink);
1639 list_del_init(&file->f_tfile_llink);
1640 }
1641 INIT_LIST_HEAD(&tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001642}
1643
Davide Libenzi7699acd2007-05-10 22:23:23 -07001644/*
Andrew Morton523723b2008-08-12 15:09:01 -07001645 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001646 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001647SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001648{
Jason Baron28d82dc2012-01-12 17:17:43 -08001649 int error, fd;
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001650 struct eventpoll *ep = NULL;
Jason Baron28d82dc2012-01-12 17:17:43 -08001651 struct file *file;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001652
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001653 /* Check the EPOLL_* constant for consistency. */
1654 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1655
Davide Libenzi296e2362009-03-31 15:24:11 -07001656 if (flags & ~EPOLL_CLOEXEC)
1657 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001658 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001659 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001660 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001661 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001662 if (error < 0)
1663 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001664 /*
1665 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001666 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001667 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001668 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1669 if (fd < 0) {
1670 error = fd;
1671 goto out_free_ep;
1672 }
1673 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001674 O_RDWR | (flags & O_CLOEXEC));
Jason Baron28d82dc2012-01-12 17:17:43 -08001675 if (IS_ERR(file)) {
1676 error = PTR_ERR(file);
1677 goto out_free_fd;
1678 }
1679 fd_install(fd, file);
1680 ep->file = file;
1681 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001682
Jason Baron28d82dc2012-01-12 17:17:43 -08001683out_free_fd:
1684 put_unused_fd(fd);
1685out_free_ep:
1686 ep_free(ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001687 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001688}
1689
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001690SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001691{
Davide Libenzibfe38912009-05-12 13:19:44 -07001692 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001693 return -EINVAL;
1694
1695 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001696}
1697
Davide Libenzi7699acd2007-05-10 22:23:23 -07001698/*
1699 * The following function implements the controller interface for
1700 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001701 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001702 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001703SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1704 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001705{
1706 int error;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001707 int did_lock_epmutex = 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001708 struct file *file, *tfile;
1709 struct eventpoll *ep;
1710 struct epitem *epi;
1711 struct epoll_event epds;
1712
Davide Libenzi7699acd2007-05-10 22:23:23 -07001713 error = -EFAULT;
1714 if (ep_op_has_event(op) &&
1715 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1716 goto error_return;
1717
1718 /* Get the "struct file *" for the eventpoll file */
1719 error = -EBADF;
1720 file = fget(epfd);
1721 if (!file)
1722 goto error_return;
1723
1724 /* Get the "struct file *" for the target file */
1725 tfile = fget(fd);
1726 if (!tfile)
1727 goto error_fput;
1728
1729 /* The target file descriptor must support poll */
1730 error = -EPERM;
1731 if (!tfile->f_op || !tfile->f_op->poll)
1732 goto error_tgt_fput;
1733
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001734 /* Check if EPOLLWAKEUP is allowed */
Michael Kerrisk854a5542012-07-17 21:37:27 +02001735 if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
Rafael J. Wysocki22b59672012-05-21 21:20:48 +02001736 epds.events &= ~EPOLLWAKEUP;
Arve Hjønnevåg0a907c62012-05-01 21:33:34 +02001737
Davide Libenzi7699acd2007-05-10 22:23:23 -07001738 /*
1739 * We have to check that the file structure underneath the file descriptor
1740 * the user passed to us _is_ an eventpoll file. And also we do not permit
1741 * adding an epoll file descriptor inside itself.
1742 */
1743 error = -EINVAL;
1744 if (file == tfile || !is_file_epoll(file))
1745 goto error_tgt_fput;
1746
1747 /*
1748 * At this point it is safe to assume that the "private_data" contains
1749 * our own data structure.
1750 */
1751 ep = file->private_data;
1752
Davide Libenzi22bacca2011-02-25 14:44:12 -08001753 /*
1754 * When we insert an epoll file descriptor, inside another epoll file
1755 * descriptor, there is the change of creating closed loops, which are
Jason Baron28d82dc2012-01-12 17:17:43 -08001756 * better be handled here, than in more critical paths. While we are
1757 * checking for loops we also determine the list of files reachable
1758 * and hang them on the tfile_check_list, so we can check that we
1759 * haven't created too many possible wakeup paths.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001760 *
Jason Baron28d82dc2012-01-12 17:17:43 -08001761 * We need to hold the epmutex across both ep_insert and ep_remove
1762 * b/c we want to make sure we are looking at a coherent view of
1763 * epoll network.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001764 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001765 if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
Davide Libenzi22bacca2011-02-25 14:44:12 -08001766 mutex_lock(&epmutex);
1767 did_lock_epmutex = 1;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001768 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001769 if (op == EPOLL_CTL_ADD) {
1770 if (is_file_epoll(tfile)) {
1771 error = -ELOOP;
Jason Baron13d51802012-04-25 16:01:47 -07001772 if (ep_loop_check(ep, tfile) != 0) {
1773 clear_tfile_check_list();
Jason Baron28d82dc2012-01-12 17:17:43 -08001774 goto error_tgt_fput;
Jason Baron13d51802012-04-25 16:01:47 -07001775 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001776 } else
1777 list_add(&tfile->f_tfile_llink, &tfile_check_list);
1778 }
Davide Libenzi22bacca2011-02-25 14:44:12 -08001779
Nelson Elhaged8805e62011-10-31 17:13:14 -07001780 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001781
Davide Libenzi67647d02007-05-15 01:40:52 -07001782 /*
1783 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1784 * above, we can be sure to be able to use the item looked up by
1785 * ep_find() till we release the mutex.
1786 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001787 epi = ep_find(ep, tfile, fd);
1788
1789 error = -EINVAL;
1790 switch (op) {
1791 case EPOLL_CTL_ADD:
1792 if (!epi) {
1793 epds.events |= POLLERR | POLLHUP;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001794 error = ep_insert(ep, &epds, tfile, fd);
1795 } else
1796 error = -EEXIST;
Jason Baron28d82dc2012-01-12 17:17:43 -08001797 clear_tfile_check_list();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001798 break;
1799 case EPOLL_CTL_DEL:
1800 if (epi)
1801 error = ep_remove(ep, epi);
1802 else
1803 error = -ENOENT;
1804 break;
1805 case EPOLL_CTL_MOD:
1806 if (epi) {
1807 epds.events |= POLLERR | POLLHUP;
1808 error = ep_modify(ep, epi, &epds);
1809 } else
1810 error = -ENOENT;
1811 break;
1812 }
Davide Libenzid47de162007-05-15 01:40:41 -07001813 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001814
1815error_tgt_fput:
Jason Baron28d82dc2012-01-12 17:17:43 -08001816 if (did_lock_epmutex)
Davide Libenzi22bacca2011-02-25 14:44:12 -08001817 mutex_unlock(&epmutex);
1818
Davide Libenzi7699acd2007-05-10 22:23:23 -07001819 fput(tfile);
1820error_fput:
1821 fput(file);
1822error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001823
1824 return error;
1825}
1826
1827/*
1828 * Implement the event wait interface for the eventpoll file. It is the kernel
1829 * part of the user space epoll_wait(2).
1830 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001831SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1832 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001833{
1834 int error;
1835 struct file *file;
1836 struct eventpoll *ep;
1837
Davide Libenzi7699acd2007-05-10 22:23:23 -07001838 /* The maximum number of event must be greater than zero */
1839 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1840 return -EINVAL;
1841
1842 /* Verify that the area passed by the user is writeable */
1843 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1844 error = -EFAULT;
1845 goto error_return;
1846 }
1847
1848 /* Get the "struct file *" for the eventpoll file */
1849 error = -EBADF;
1850 file = fget(epfd);
1851 if (!file)
1852 goto error_return;
1853
1854 /*
1855 * We have to check that the file structure underneath the fd
1856 * the user passed to us _is_ an eventpoll file.
1857 */
1858 error = -EINVAL;
1859 if (!is_file_epoll(file))
1860 goto error_fput;
1861
1862 /*
1863 * At this point it is safe to assume that the "private_data" contains
1864 * our own data structure.
1865 */
1866 ep = file->private_data;
1867
1868 /* Time to fish for events ... */
1869 error = ep_poll(ep, events, maxevents, timeout);
1870
1871error_fput:
1872 fput(file);
1873error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001874
1875 return error;
1876}
1877
Roland McGrathf3de2722008-04-30 00:53:09 -07001878#ifdef HAVE_SET_RESTORE_SIGMASK
Davide Libenzi7699acd2007-05-10 22:23:23 -07001879
1880/*
1881 * Implement the event wait interface for the eventpoll file. It is the kernel
1882 * part of the user space epoll_pwait(2).
1883 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001884SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1885 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1886 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001887{
1888 int error;
1889 sigset_t ksigmask, sigsaved;
1890
1891 /*
1892 * If the caller wants a certain signal mask to be set during the wait,
1893 * we apply it here.
1894 */
1895 if (sigmask) {
1896 if (sigsetsize != sizeof(sigset_t))
1897 return -EINVAL;
1898 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1899 return -EFAULT;
1900 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1901 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1902 }
1903
1904 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1905
1906 /*
1907 * If we changed the signal mask, we need to restore the original one.
1908 * In case we've got a signal while waiting, we do not restore the
1909 * signal mask yet, and we allow do_signal() to deliver the signal on
1910 * the way back to userspace, before the signal mask is restored.
1911 */
1912 if (sigmask) {
1913 if (error == -EINTR) {
1914 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001915 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001916 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001917 } else
1918 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1919 }
1920
1921 return error;
1922}
1923
Roland McGrathf3de2722008-04-30 00:53:09 -07001924#endif /* HAVE_SET_RESTORE_SIGMASK */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926static int __init eventpoll_init(void)
1927{
Davide Libenzi7ef99642008-12-01 13:13:55 -08001928 struct sysinfo si;
1929
1930 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08001931 /*
1932 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1933 */
1934 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08001935 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08001936 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937
Davide Libenzi22bacca2011-02-25 14:44:12 -08001938 /*
1939 * Initialize the structure used to perform epoll file descriptor
1940 * inclusion loops checks.
1941 */
1942 ep_nested_calls_init(&poll_loop_ncalls);
1943
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07001945 ep_nested_calls_init(&poll_safewake_ncalls);
1946
1947 /* Initialize the structure used to perform file's f_op->poll() calls */
1948 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949
1950 /* Allocates slab cache used to allocate "struct epitem" items */
1951 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001952 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
1954 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1955 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001956 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959}
Davide Libenzicea69242007-05-10 22:23:22 -07001960fs_initcall(eventpoll_init);