blob: 9fec1836057a9298ad9f121afcea03c64388eb88 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Davide Libenzi5071f972009-03-31 15:24:10 -07002 * fs/eventpoll.c (Efficient event retrieval implementation)
3 * Copyright (C) 2001,...,2009 Davide Libenzi
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * Davide Libenzi <davidel@xmailserver.org>
11 *
12 */
13
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/init.h>
15#include <linux/kernel.h>
16#include <linux/sched.h>
17#include <linux/fs.h>
18#include <linux/file.h>
19#include <linux/signal.h>
20#include <linux/errno.h>
21#include <linux/mm.h>
22#include <linux/slab.h>
23#include <linux/poll.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/string.h>
25#include <linux/list.h>
26#include <linux/hash.h>
27#include <linux/spinlock.h>
28#include <linux/syscalls.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/rbtree.h>
30#include <linux/wait.h>
31#include <linux/eventpoll.h>
32#include <linux/mount.h>
33#include <linux/bitops.h>
Arjan van de Ven144efe32006-03-23 03:00:32 -080034#include <linux/mutex.h>
Davide Libenzida66f7c2007-05-10 22:23:21 -070035#include <linux/anon_inodes.h>
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020036#include <linux/device.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include <asm/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <asm/io.h>
39#include <asm/mman.h>
Arun Sharma600634972011-07-26 16:09:06 -070040#include <linux/atomic.h>
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -080041#include <linux/proc_fs.h>
42#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*
45 * LOCKING:
46 * There are three level of locking required by epoll :
47 *
Arjan van de Ven144efe32006-03-23 03:00:32 -080048 * 1) epmutex (mutex)
Davide Libenzic7ea7632007-05-15 01:40:47 -070049 * 2) ep->mtx (mutex)
50 * 3) ep->lock (spinlock)
Linus Torvalds1da177e2005-04-16 15:20:36 -070051 *
52 * The acquire order is the one listed above, from 1 to 3.
53 * We need a spinlock (ep->lock) because we manipulate objects
54 * from inside the poll callback, that might be triggered from
55 * a wake_up() that in turn might be called from IRQ context.
56 * So we can't sleep inside the poll callback and hence we need
57 * a spinlock. During the event transfer loop (from kernel to
58 * user space) we could end up sleeping due a copy_to_user(), so
59 * we need a lock that will allow us to sleep. This lock is a
Davide Libenzid47de162007-05-15 01:40:41 -070060 * mutex (ep->mtx). It is acquired during the event transfer loop,
61 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
62 * Then we also need a global mutex to serialize eventpoll_release_file()
63 * and ep_free().
64 * This mutex is acquired by ep_free() during the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 * cleanup path and it is also acquired by eventpoll_release_file()
66 * if a file has been pushed inside an epoll set and it is then
Daniel Balutabf6a41d2011-01-30 23:42:29 +020067 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
Davide Libenzi22bacca2011-02-25 14:44:12 -080068 * It is also acquired when inserting an epoll fd onto another epoll
69 * fd. We do this so that we walk the epoll tree and ensure that this
70 * insertion does not create a cycle of epoll file descriptors, which
71 * could lead to deadlock. We need a global mutex to prevent two
72 * simultaneous inserts (A into B and B into A) from racing and
73 * constructing a cycle without either insert observing that it is
74 * going to.
Nelson Elhaged8805e62011-10-31 17:13:14 -070075 * It is necessary to acquire multiple "ep->mtx"es at once in the
76 * case when one epoll fd is added to another. In this case, we
77 * always acquire the locks in the order of nesting (i.e. after
78 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
79 * before e2->mtx). Since we disallow cycles of epoll file
80 * descriptors, this ensures that the mutexes are well-ordered. In
81 * order to communicate this nesting to lockdep, when walking a tree
82 * of epoll file descriptors, we use the current recursion depth as
83 * the lockdep subkey.
Davide Libenzid47de162007-05-15 01:40:41 -070084 * It is possible to drop the "ep->mtx" and to use the global
85 * mutex "epmutex" (together with "ep->lock") to have it working,
86 * but having "ep->mtx" will make the interface more scalable.
Arjan van de Ven144efe32006-03-23 03:00:32 -080087 * Events that require holding "epmutex" are very rare, while for
Davide Libenzid47de162007-05-15 01:40:41 -070088 * normal operations the epoll private "ep->mtx" will guarantee
89 * a better scalability.
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 */
91
Linus Torvalds1da177e2005-04-16 15:20:36 -070092/* Epoll private bits inside the event mask */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +020093#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
Davide Libenzi5071f972009-03-31 15:24:10 -070095/* Maximum number of nesting allowed inside epoll sets */
96#define EP_MAX_NESTS 4
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Davide Libenzib6119672006-10-11 01:21:44 -070098#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
99
Davide Libenzid47de162007-05-15 01:40:41 -0700100#define EP_UNACTIVE_PTR ((void *) -1L)
101
Davide Libenzi7ef99642008-12-01 13:13:55 -0800102#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
103
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104struct epoll_filefd {
105 struct file *file;
106 int fd;
107};
108
109/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700110 * Structure used to track possible nested calls, for too deep recursions
111 * and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700113struct nested_call_node {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 struct list_head llink;
Davide Libenzi5071f972009-03-31 15:24:10 -0700115 void *cookie;
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700116 void *ctx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117};
118
119/*
Davide Libenzi5071f972009-03-31 15:24:10 -0700120 * This structure is used as collector for nested calls, to check for
121 * maximum recursion dept and loop cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700123struct nested_calls {
124 struct list_head tasks_call_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 spinlock_t lock;
126};
127
128/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 * Each file descriptor added to the eventpoll interface will
Davide Libenzi6192bd52007-05-08 00:25:41 -0700130 * have an entry of this type linked to the "rbr" RB tree.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 */
132struct epitem {
Davide Libenzi67647d02007-05-15 01:40:52 -0700133 /* RB tree node used to link this structure to the eventpoll RB tree */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 struct rb_node rbn;
135
136 /* List header used to link this structure to the eventpoll ready list */
137 struct list_head rdllink;
138
Davide Libenzic7ea7632007-05-15 01:40:47 -0700139 /*
140 * Works together "struct eventpoll"->ovflist in keeping the
141 * single linked chain of items.
142 */
143 struct epitem *next;
144
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 /* The file descriptor information this item refers to */
146 struct epoll_filefd ffd;
147
148 /* Number of active wait queue attached to poll operations */
149 int nwait;
150
151 /* List containing poll wait queues */
152 struct list_head pwqlist;
153
154 /* The "container" of this item */
155 struct eventpoll *ep;
156
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 /* List header used to link this item to the "struct file" items list */
158 struct list_head fllink;
Davide Libenzid47de162007-05-15 01:40:41 -0700159
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200160 /* wakeup_source used when EPOLLWAKEUP is set */
161 struct wakeup_source *ws;
162
Davide Libenzic7ea7632007-05-15 01:40:47 -0700163 /* The structure that describe the interested events and the source fd */
164 struct epoll_event event;
Davide Libenzid47de162007-05-15 01:40:41 -0700165};
166
167/*
168 * This structure is stored inside the "private_data" member of the file
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200169 * structure and represents the main data structure for the eventpoll
Davide Libenzid47de162007-05-15 01:40:41 -0700170 * interface.
171 */
172struct eventpoll {
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200173 /* Protect the access to this structure */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700174 spinlock_t lock;
Davide Libenzid47de162007-05-15 01:40:41 -0700175
176 /*
177 * This mutex is used to ensure that files are not removed
178 * while epoll is using them. This is held during the event
179 * collection loop, the file cleanup path, the epoll file exit
180 * code and the ctl operations.
181 */
182 struct mutex mtx;
183
184 /* Wait queue used by sys_epoll_wait() */
185 wait_queue_head_t wq;
186
187 /* Wait queue used by file->poll() */
188 wait_queue_head_t poll_wait;
189
190 /* List of ready file descriptors */
191 struct list_head rdllist;
192
Davide Libenzi67647d02007-05-15 01:40:52 -0700193 /* RB tree root used to store monitored fd structs */
Davide Libenzid47de162007-05-15 01:40:41 -0700194 struct rb_root rbr;
195
196 /*
197 * This is a single linked list that chains all the "struct epitem" that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300198 * happened while transferring ready events to userspace w/out
Davide Libenzid47de162007-05-15 01:40:41 -0700199 * holding ->lock.
200 */
201 struct epitem *ovflist;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800202
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200203 /* wakeup_source used when ep_scan_ready_list is running */
204 struct wakeup_source *ws;
205
Davide Libenzi7ef99642008-12-01 13:13:55 -0800206 /* The user that created the eventpoll descriptor */
207 struct user_struct *user;
Jason Baron28d82dc2012-01-12 17:17:43 -0800208
209 struct file *file;
210
211 /* used to optimize loop detection check */
212 int visited;
213 struct list_head visited_list_link;
Davide Libenzid47de162007-05-15 01:40:41 -0700214};
215
216/* Wait structure used by the poll hooks */
217struct eppoll_entry {
218 /* List header used to link this structure to the "struct epitem" */
219 struct list_head llink;
220
221 /* The "base" pointer is set to the container "struct epitem" */
Tony Battersby4f0989d2009-03-31 15:24:16 -0700222 struct epitem *base;
Davide Libenzid47de162007-05-15 01:40:41 -0700223
224 /*
225 * Wait queue item that will be linked to the target file wait
226 * queue head.
227 */
228 wait_queue_t wait;
229
230 /* The wait queue head that linked the "wait" wait queue item */
231 wait_queue_head_t *whead;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232};
233
234/* Wrapper struct used by poll queueing */
235struct ep_pqueue {
236 poll_table pt;
237 struct epitem *epi;
238};
239
Davide Libenzi5071f972009-03-31 15:24:10 -0700240/* Used by the ep_send_events() function as callback private data */
241struct ep_send_events_data {
242 int maxevents;
243 struct epoll_event __user *events;
244};
245
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246/*
Davide Libenzi7ef99642008-12-01 13:13:55 -0800247 * Configuration options available inside /proc/sys/fs/epoll/
248 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800249/* Maximum number of epoll watched descriptors, per user */
Robin Holt52bd19f72011-01-12 17:00:01 -0800250static long max_user_watches __read_mostly;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800251
252/*
Davide Libenzid47de162007-05-15 01:40:41 -0700253 * This mutex is used to serialize ep_free() and eventpoll_release_file().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 */
Davide Libenzi7ef99642008-12-01 13:13:55 -0800255static DEFINE_MUTEX(epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256
Davide Libenzi22bacca2011-02-25 14:44:12 -0800257/* Used to check for epoll file descriptor inclusion loops */
258static struct nested_calls poll_loop_ncalls;
259
Davide Libenzi5071f972009-03-31 15:24:10 -0700260/* Used for safe wake up implementation */
261static struct nested_calls poll_safewake_ncalls;
262
263/* Used to call file's f_op->poll() under the nested calls boundaries */
264static struct nested_calls poll_readywalk_ncalls;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
266/* Slab cache used to allocate "struct epitem" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800267static struct kmem_cache *epi_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
269/* Slab cache used to allocate "struct eppoll_entry" */
Christoph Lametere18b8902006-12-06 20:33:20 -0800270static struct kmem_cache *pwq_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271
Jason Baron28d82dc2012-01-12 17:17:43 -0800272/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
273static LIST_HEAD(visited_list);
274
275/*
276 * List of files with newly added links, where we may need to limit the number
277 * of emanating paths. Protected by the epmutex.
278 */
279static LIST_HEAD(tfile_check_list);
280
Davide Libenzi7ef99642008-12-01 13:13:55 -0800281#ifdef CONFIG_SYSCTL
282
283#include <linux/sysctl.h>
284
Robin Holt52bd19f72011-01-12 17:00:01 -0800285static long zero;
286static long long_max = LONG_MAX;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800287
288ctl_table epoll_table[] = {
289 {
Davide Libenzi7ef99642008-12-01 13:13:55 -0800290 .procname = "max_user_watches",
291 .data = &max_user_watches,
Robin Holt52bd19f72011-01-12 17:00:01 -0800292 .maxlen = sizeof(max_user_watches),
Davide Libenzi7ef99642008-12-01 13:13:55 -0800293 .mode = 0644,
Robin Holt52bd19f72011-01-12 17:00:01 -0800294 .proc_handler = proc_doulongvec_minmax,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800295 .extra1 = &zero,
Robin Holt52bd19f72011-01-12 17:00:01 -0800296 .extra2 = &long_max,
Davide Libenzi7ef99642008-12-01 13:13:55 -0800297 },
Eric W. Biedermanab092032009-11-05 14:25:10 -0800298 { }
Davide Libenzi7ef99642008-12-01 13:13:55 -0800299};
300#endif /* CONFIG_SYSCTL */
301
Jason Baron28d82dc2012-01-12 17:17:43 -0800302static const struct file_operations eventpoll_fops;
303
304static inline int is_file_epoll(struct file *f)
305{
306 return f->f_op == &eventpoll_fops;
307}
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700308
Davide Libenzi67647d02007-05-15 01:40:52 -0700309/* Setup the structure that is used as key for the RB tree */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700310static inline void ep_set_ffd(struct epoll_filefd *ffd,
311 struct file *file, int fd)
312{
313 ffd->file = file;
314 ffd->fd = fd;
315}
316
Davide Libenzi67647d02007-05-15 01:40:52 -0700317/* Compare RB tree keys */
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700318static inline int ep_cmp_ffd(struct epoll_filefd *p1,
319 struct epoll_filefd *p2)
320{
321 return (p1->file > p2->file ? +1:
322 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
323}
324
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700325/* Tells us if the item is currently linked */
326static inline int ep_is_linked(struct list_head *p)
327{
328 return !list_empty(p);
329}
330
Oleg Nesterov971316f2012-02-24 20:07:29 +0100331static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
332{
333 return container_of(p, struct eppoll_entry, wait);
334}
335
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700336/* Get the "struct epitem" from a wait queue pointer */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700337static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700338{
339 return container_of(p, struct eppoll_entry, wait)->base;
340}
341
342/* Get the "struct epitem" from an epoll queue wrapper */
Davide Libenzicdac75e2008-04-29 00:58:34 -0700343static inline struct epitem *ep_item_from_epqueue(poll_table *p)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700344{
345 return container_of(p, struct ep_pqueue, pt)->epi;
346}
347
348/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
Davide Libenzi6192bd52007-05-08 00:25:41 -0700349static inline int ep_op_has_event(int op)
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700350{
Andrew Mortona80a6b82012-11-08 15:53:35 -0800351 return op != EPOLL_CTL_DEL;
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700352}
353
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354/* Initialize the poll safe wake up structure */
Davide Libenzi5071f972009-03-31 15:24:10 -0700355static void ep_nested_calls_init(struct nested_calls *ncalls)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356{
Davide Libenzi5071f972009-03-31 15:24:10 -0700357 INIT_LIST_HEAD(&ncalls->tasks_call_list);
358 spin_lock_init(&ncalls->lock);
359}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360
Davide Libenzi5071f972009-03-31 15:24:10 -0700361/**
Davide Libenzi3fb0e582011-03-22 16:34:46 -0700362 * ep_events_available - Checks if ready events might be available.
363 *
364 * @ep: Pointer to the eventpoll context.
365 *
366 * Returns: Returns a value different than zero if ready events are available,
367 * or zero otherwise.
368 */
369static inline int ep_events_available(struct eventpoll *ep)
370{
371 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
372}
373
374/**
Davide Libenzi5071f972009-03-31 15:24:10 -0700375 * ep_call_nested - Perform a bound (possibly) nested call, by checking
376 * that the recursion limit is not exceeded, and that
377 * the same nested call (by the meaning of same cookie) is
378 * no re-entered.
379 *
380 * @ncalls: Pointer to the nested_calls structure to be used for this call.
381 * @max_nests: Maximum number of allowed nesting calls.
382 * @nproc: Nested call core function pointer.
383 * @priv: Opaque data to be passed to the @nproc callback.
384 * @cookie: Cookie to be used to identify this nested call.
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700385 * @ctx: This instance context.
Davide Libenzi5071f972009-03-31 15:24:10 -0700386 *
387 * Returns: Returns the code returned by the @nproc callback, or -1 if
388 * the maximum recursion limit has been exceeded.
389 */
390static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
391 int (*nproc)(void *, void *, int), void *priv,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700392 void *cookie, void *ctx)
Davide Libenzi5071f972009-03-31 15:24:10 -0700393{
394 int error, call_nests = 0;
395 unsigned long flags;
Davide Libenzi5071f972009-03-31 15:24:10 -0700396 struct list_head *lsthead = &ncalls->tasks_call_list;
397 struct nested_call_node *tncur;
398 struct nested_call_node tnode;
399
400 spin_lock_irqsave(&ncalls->lock, flags);
401
402 /*
403 * Try to see if the current task is already inside this wakeup call.
404 * We use a list here, since the population inside this set is always
405 * very much limited.
406 */
407 list_for_each_entry(tncur, lsthead, llink) {
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700408 if (tncur->ctx == ctx &&
Davide Libenzi5071f972009-03-31 15:24:10 -0700409 (tncur->cookie == cookie || ++call_nests > max_nests)) {
410 /*
411 * Ops ... loop detected or maximum nest level reached.
412 * We abort this wake by breaking the cycle itself.
413 */
Tony Battersbyabff55c2009-03-31 15:24:13 -0700414 error = -1;
415 goto out_unlock;
Davide Libenzi5071f972009-03-31 15:24:10 -0700416 }
417 }
418
419 /* Add the current task and cookie to the list */
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700420 tnode.ctx = ctx;
Davide Libenzi5071f972009-03-31 15:24:10 -0700421 tnode.cookie = cookie;
422 list_add(&tnode.llink, lsthead);
423
424 spin_unlock_irqrestore(&ncalls->lock, flags);
425
426 /* Call the nested function */
427 error = (*nproc)(priv, cookie, call_nests);
428
429 /* Remove the current task from the list */
430 spin_lock_irqsave(&ncalls->lock, flags);
431 list_del(&tnode.llink);
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700432out_unlock:
Davide Libenzi5071f972009-03-31 15:24:10 -0700433 spin_unlock_irqrestore(&ncalls->lock, flags);
434
435 return error;
436}
437
Steven Rostedt02edc6f2012-03-23 15:02:27 -0700438/*
439 * As described in commit 0ccf831cb lockdep: annotate epoll
440 * the use of wait queues used by epoll is done in a very controlled
441 * manner. Wake ups can nest inside each other, but are never done
442 * with the same locking. For example:
443 *
444 * dfd = socket(...);
445 * efd1 = epoll_create();
446 * efd2 = epoll_create();
447 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
448 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
449 *
450 * When a packet arrives to the device underneath "dfd", the net code will
451 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
452 * callback wakeup entry on that queue, and the wake_up() performed by the
453 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
454 * (efd1) notices that it may have some event ready, so it needs to wake up
455 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
456 * that ends up in another wake_up(), after having checked about the
457 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
458 * avoid stack blasting.
459 *
460 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
461 * this special case of epoll.
462 */
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700463#ifdef CONFIG_DEBUG_LOCK_ALLOC
464static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
465 unsigned long events, int subclass)
466{
467 unsigned long flags;
468
469 spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
470 wake_up_locked_poll(wqueue, events);
471 spin_unlock_irqrestore(&wqueue->lock, flags);
472}
473#else
474static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
475 unsigned long events, int subclass)
476{
477 wake_up_poll(wqueue, events);
478}
479#endif
480
Davide Libenzi5071f972009-03-31 15:24:10 -0700481static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
482{
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700483 ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
484 1 + call_nests);
Davide Libenzi5071f972009-03-31 15:24:10 -0700485 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486}
487
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488/*
489 * Perform a safe wake up of the poll wait list. The problem is that
490 * with the new callback'd wake up system, it is possible that the
491 * poll callback is reentered from inside the call to wake_up() done
492 * on the poll wait queue head. The rule is that we cannot reenter the
Davide Libenzi5071f972009-03-31 15:24:10 -0700493 * wake up code from the same task more than EP_MAX_NESTS times,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 * and we cannot reenter the same wait queue head at all. This will
495 * enable to have a hierarchy of epoll file descriptor of no more than
Davide Libenzi5071f972009-03-31 15:24:10 -0700496 * EP_MAX_NESTS deep.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 */
Davide Libenzi5071f972009-03-31 15:24:10 -0700498static void ep_poll_safewake(wait_queue_head_t *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499{
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700500 int this_cpu = get_cpu();
501
Davide Libenzi5071f972009-03-31 15:24:10 -0700502 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700503 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
504
505 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506}
507
Oleg Nesterov971316f2012-02-24 20:07:29 +0100508static void ep_remove_wait_queue(struct eppoll_entry *pwq)
509{
510 wait_queue_head_t *whead;
511
512 rcu_read_lock();
513 /* If it is cleared by POLLFREE, it should be rcu-safe */
514 whead = rcu_dereference(pwq->whead);
515 if (whead)
516 remove_wait_queue(whead, &pwq->wait);
517 rcu_read_unlock();
518}
519
Davide Libenzi7699acd2007-05-10 22:23:23 -0700520/*
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700521 * This function unregisters poll callbacks from the associated file
522 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
523 * ep_free).
Davide Libenzi7699acd2007-05-10 22:23:23 -0700524 */
525static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
526{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700527 struct list_head *lsthead = &epi->pwqlist;
528 struct eppoll_entry *pwq;
529
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700530 while (!list_empty(lsthead)) {
531 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700532
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700533 list_del(&pwq->llink);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100534 ep_remove_wait_queue(pwq);
Tony Battersbyd1bc90d2009-03-31 15:24:15 -0700535 kmem_cache_free(pwq_cache, pwq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700536 }
537}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538
Davide Libenzi5071f972009-03-31 15:24:10 -0700539/**
540 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
541 * the scan code, to call f_op->poll(). Also allows for
542 * O(NumReady) performance.
543 *
544 * @ep: Pointer to the epoll private data structure.
545 * @sproc: Pointer to the scan callback.
546 * @priv: Private opaque data passed to the @sproc callback.
Nelson Elhaged8805e62011-10-31 17:13:14 -0700547 * @depth: The current depth of recursive f_op->poll calls.
Davide Libenzi5071f972009-03-31 15:24:10 -0700548 *
549 * Returns: The same integer error code returned by the @sproc callback.
550 */
551static int ep_scan_ready_list(struct eventpoll *ep,
552 int (*sproc)(struct eventpoll *,
553 struct list_head *, void *),
Nelson Elhaged8805e62011-10-31 17:13:14 -0700554 void *priv,
555 int depth)
Davide Libenzi5071f972009-03-31 15:24:10 -0700556{
557 int error, pwake = 0;
558 unsigned long flags;
559 struct epitem *epi, *nepi;
Davide Libenzi296e2362009-03-31 15:24:11 -0700560 LIST_HEAD(txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700561
562 /*
563 * We need to lock this because we could be hit by
Tony Battersbye057e152009-03-31 15:24:15 -0700564 * eventpoll_release_file() and epoll_ctl().
Davide Libenzi5071f972009-03-31 15:24:10 -0700565 */
Nelson Elhaged8805e62011-10-31 17:13:14 -0700566 mutex_lock_nested(&ep->mtx, depth);
Davide Libenzi5071f972009-03-31 15:24:10 -0700567
568 /*
569 * Steal the ready list, and re-init the original one to the
570 * empty list. Also, set ep->ovflist to NULL so that events
571 * happening while looping w/out locks, are not lost. We cannot
572 * have the poll callback to queue directly on ep->rdllist,
573 * because we want the "sproc" callback to be able to do it
574 * in a lockless way.
575 */
576 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi296e2362009-03-31 15:24:11 -0700577 list_splice_init(&ep->rdllist, &txlist);
Davide Libenzi5071f972009-03-31 15:24:10 -0700578 ep->ovflist = NULL;
579 spin_unlock_irqrestore(&ep->lock, flags);
580
581 /*
582 * Now call the callback function.
583 */
584 error = (*sproc)(ep, &txlist, priv);
585
586 spin_lock_irqsave(&ep->lock, flags);
587 /*
588 * During the time we spent inside the "sproc" callback, some
589 * other events might have been queued by the poll callback.
590 * We re-insert them inside the main ready-list here.
591 */
592 for (nepi = ep->ovflist; (epi = nepi) != NULL;
593 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
594 /*
595 * We need to check if the item is already in the list.
596 * During the "sproc" callback execution time, items are
597 * queued into ->ovflist but the "txlist" might already
598 * contain them, and the list_splice() below takes care of them.
599 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200600 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700601 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200602 __pm_stay_awake(epi->ws);
603 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700604 }
605 /*
606 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
607 * releasing the lock, events will be queued in the normal way inside
608 * ep->rdllist.
609 */
610 ep->ovflist = EP_UNACTIVE_PTR;
611
612 /*
613 * Quickly re-inject items left on "txlist".
614 */
615 list_splice(&txlist, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200616 __pm_relax(ep->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700617
618 if (!list_empty(&ep->rdllist)) {
619 /*
Davide Libenzi296e2362009-03-31 15:24:11 -0700620 * Wake up (if active) both the eventpoll wait list and
621 * the ->poll() wait list (delayed after we release the lock).
Davide Libenzi5071f972009-03-31 15:24:10 -0700622 */
623 if (waitqueue_active(&ep->wq))
624 wake_up_locked(&ep->wq);
625 if (waitqueue_active(&ep->poll_wait))
626 pwake++;
627 }
628 spin_unlock_irqrestore(&ep->lock, flags);
629
630 mutex_unlock(&ep->mtx);
631
632 /* We have to call this outside the lock */
633 if (pwake)
634 ep_poll_safewake(&ep->poll_wait);
635
636 return error;
637}
638
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700640 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
Davide Libenzic7ea7632007-05-15 01:40:47 -0700641 * all the associated resources. Must be called with "mtx" held.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700642 */
643static int ep_remove(struct eventpoll *ep, struct epitem *epi)
644{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700645 unsigned long flags;
646 struct file *file = epi->ffd.file;
647
648 /*
649 * Removes poll wait queue hooks. We _have_ to do this without holding
650 * the "ep->lock" otherwise a deadlock might occur. This because of the
651 * sequence of the lock acquisition. Here we do "ep->lock" then the wait
652 * queue head lock when unregistering the wait queue. The wakeup callback
653 * will run by holding the wait queue head lock and will call our callback
654 * that will try to get "ep->lock".
655 */
656 ep_unregister_pollwait(ep, epi);
657
658 /* Remove the current item from the list of epoll hooks */
Jonathan Corbet68499912009-02-06 13:52:43 -0700659 spin_lock(&file->f_lock);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700660 if (ep_is_linked(&epi->fllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -0700661 list_del_init(&epi->fllink);
Jonathan Corbet68499912009-02-06 13:52:43 -0700662 spin_unlock(&file->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
Davide Libenzicdac75e2008-04-29 00:58:34 -0700664 rb_erase(&epi->rbn, &ep->rbr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665
Davide Libenzic7ea7632007-05-15 01:40:47 -0700666 spin_lock_irqsave(&ep->lock, flags);
667 if (ep_is_linked(&epi->rdllink))
668 list_del_init(&epi->rdllink);
669 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200671 wakeup_source_unregister(epi->ws);
672
Davide Libenzi7699acd2007-05-10 22:23:23 -0700673 /* At this point it is safe to free the eventpoll item */
Davide Libenzic7ea7632007-05-15 01:40:47 -0700674 kmem_cache_free(epi_cache, epi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675
Robin Holt52bd19f72011-01-12 17:00:01 -0800676 atomic_long_dec(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800677
Davide Libenzic7ea7632007-05-15 01:40:47 -0700678 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679}
680
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681static void ep_free(struct eventpoll *ep)
682{
683 struct rb_node *rbp;
684 struct epitem *epi;
685
686 /* We need to release all tasks waiting for these file */
687 if (waitqueue_active(&ep->poll_wait))
Davide Libenzi5071f972009-03-31 15:24:10 -0700688 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
690 /*
691 * We need to lock this because we could be hit by
692 * eventpoll_release_file() while we're freeing the "struct eventpoll".
Davide Libenzid47de162007-05-15 01:40:41 -0700693 * We do not need to hold "ep->mtx" here because the epoll file
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 * is on the way to be removed and no one has references to it
695 * anymore. The only hit might come from eventpoll_release_file() but
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300696 * holding "epmutex" is sufficient here.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 */
Arjan van de Ven144efe32006-03-23 03:00:32 -0800698 mutex_lock(&epmutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
700 /*
701 * Walks through the whole tree by unregistering poll callbacks.
702 */
703 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
704 epi = rb_entry(rbp, struct epitem, rbn);
705
706 ep_unregister_pollwait(ep, epi);
707 }
708
709 /*
Davide Libenzi6192bd52007-05-08 00:25:41 -0700710 * Walks through the whole tree by freeing each "struct epitem". At this
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 * point we are sure no poll callbacks will be lingering around, and also by
Davide Libenzid47de162007-05-15 01:40:41 -0700712 * holding "epmutex" we can be sure that no file cleanup code will hit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 * us during this operation. So we can avoid the lock on "ep->lock".
714 */
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700715 while ((rbp = rb_first(&ep->rbr)) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 epi = rb_entry(rbp, struct epitem, rbn);
717 ep_remove(ep, epi);
718 }
719
Arjan van de Ven144efe32006-03-23 03:00:32 -0800720 mutex_unlock(&epmutex);
Davide Libenzid47de162007-05-15 01:40:41 -0700721 mutex_destroy(&ep->mtx);
Davide Libenzi7ef99642008-12-01 13:13:55 -0800722 free_uid(ep->user);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200723 wakeup_source_unregister(ep->ws);
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700724 kfree(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725}
726
Davide Libenzi7699acd2007-05-10 22:23:23 -0700727static int ep_eventpoll_release(struct inode *inode, struct file *file)
728{
729 struct eventpoll *ep = file->private_data;
730
Davide Libenzif0ee9aa2007-05-15 01:40:57 -0700731 if (ep)
Davide Libenzi7699acd2007-05-10 22:23:23 -0700732 ep_free(ep);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700733
Davide Libenzi7699acd2007-05-10 22:23:23 -0700734 return 0;
735}
736
Davide Libenzi296e2362009-03-31 15:24:11 -0700737static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
738 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -0700739{
740 struct epitem *epi, *tmp;
Hans Verkuil626cf232012-03-23 15:02:27 -0700741 poll_table pt;
Davide Libenzi5071f972009-03-31 15:24:10 -0700742
Hans Verkuil626cf232012-03-23 15:02:27 -0700743 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -0700744 list_for_each_entry_safe(epi, tmp, head, rdllink) {
Hans Verkuil626cf232012-03-23 15:02:27 -0700745 pt._key = epi->event.events;
746 if (epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
Davide Libenzi5071f972009-03-31 15:24:10 -0700747 epi->event.events)
748 return POLLIN | POLLRDNORM;
Davide Libenzi296e2362009-03-31 15:24:11 -0700749 else {
Davide Libenzi5071f972009-03-31 15:24:10 -0700750 /*
751 * Item has been dropped into the ready list by the poll
752 * callback, but it's not actually ready, as far as
753 * caller requested events goes. We can remove it here.
754 */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200755 __pm_relax(epi->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -0700756 list_del_init(&epi->rdllink);
Davide Libenzi296e2362009-03-31 15:24:11 -0700757 }
Davide Libenzi5071f972009-03-31 15:24:10 -0700758 }
759
760 return 0;
761}
762
763static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
764{
Nelson Elhaged8805e62011-10-31 17:13:14 -0700765 return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
Davide Libenzi5071f972009-03-31 15:24:10 -0700766}
767
Davide Libenzi7699acd2007-05-10 22:23:23 -0700768static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
769{
Davide Libenzi5071f972009-03-31 15:24:10 -0700770 int pollflags;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700771 struct eventpoll *ep = file->private_data;
772
773 /* Insert inside our poll wait queue */
774 poll_wait(file, &ep->poll_wait, wait);
775
Davide Libenzi5071f972009-03-31 15:24:10 -0700776 /*
777 * Proceed to find out if wanted events are really available inside
778 * the ready list. This need to be done under ep_call_nested()
779 * supervision, since the call to f_op->poll() done on listed files
780 * could re-enter here.
781 */
782 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
Davide Libenzi3fe4a972009-06-17 16:25:58 -0700783 ep_poll_readyevents_proc, ep, ep, current);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700784
Davide Libenzi296e2362009-03-31 15:24:11 -0700785 return pollflags != -1 ? pollflags : 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700786}
787
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800788#ifdef CONFIG_PROC_FS
789static int ep_show_fdinfo(struct seq_file *m, struct file *f)
790{
791 struct eventpoll *ep = f->private_data;
792 struct rb_node *rbp;
793 int ret = 0;
794
795 mutex_lock(&ep->mtx);
796 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
797 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
798
799 ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
800 epi->ffd.fd, epi->event.events,
801 (long long)epi->event.data);
802 if (ret)
803 break;
804 }
805 mutex_unlock(&ep->mtx);
806
807 return ret;
808}
809#endif
810
Davide Libenzi7699acd2007-05-10 22:23:23 -0700811/* File callbacks that implement the eventpoll file behaviour */
812static const struct file_operations eventpoll_fops = {
Cyrill Gorcunov138d22b2012-12-17 16:05:02 -0800813#ifdef CONFIG_PROC_FS
814 .show_fdinfo = ep_show_fdinfo,
815#endif
Davide Libenzi7699acd2007-05-10 22:23:23 -0700816 .release = ep_eventpoll_release,
Arnd Bergmann6038f372010-08-15 18:52:59 +0200817 .poll = ep_eventpoll_poll,
818 .llseek = noop_llseek,
Davide Libenzi7699acd2007-05-10 22:23:23 -0700819};
820
Davide Libenzi7699acd2007-05-10 22:23:23 -0700821/*
822 * This is called from eventpoll_release() to unlink files from the eventpoll
823 * interface. We need to have this facility to cleanup correctly files that are
824 * closed without being removed from the eventpoll interface.
825 */
826void eventpoll_release_file(struct file *file)
827{
828 struct list_head *lsthead = &file->f_ep_links;
829 struct eventpoll *ep;
830 struct epitem *epi;
831
832 /*
Jonathan Corbet68499912009-02-06 13:52:43 -0700833 * We don't want to get "file->f_lock" because it is not
Davide Libenzi7699acd2007-05-10 22:23:23 -0700834 * necessary. It is not necessary because we're in the "struct file"
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300835 * cleanup path, and this means that no one is using this file anymore.
Davide Libenzi5071f972009-03-31 15:24:10 -0700836 * So, for example, epoll_ctl() cannot hit here since if we reach this
Davide Libenzi67647d02007-05-15 01:40:52 -0700837 * point, the file counter already went to zero and fget() would fail.
Davide Libenzid47de162007-05-15 01:40:41 -0700838 * The only hit might come from ep_free() but by holding the mutex
Davide Libenzi7699acd2007-05-10 22:23:23 -0700839 * will correctly serialize the operation. We do need to acquire
Davide Libenzid47de162007-05-15 01:40:41 -0700840 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
Davide Libenzi7699acd2007-05-10 22:23:23 -0700841 * from anywhere but ep_free().
Jonathan Corbet68499912009-02-06 13:52:43 -0700842 *
843 * Besides, ep_remove() acquires the lock, so we can't hold it here.
Davide Libenzi7699acd2007-05-10 22:23:23 -0700844 */
845 mutex_lock(&epmutex);
846
847 while (!list_empty(lsthead)) {
848 epi = list_first_entry(lsthead, struct epitem, fllink);
849
850 ep = epi->ep;
851 list_del_init(&epi->fllink);
Nelson Elhaged8805e62011-10-31 17:13:14 -0700852 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700853 ep_remove(ep, epi);
Davide Libenzid47de162007-05-15 01:40:41 -0700854 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700855 }
856
857 mutex_unlock(&epmutex);
858}
859
860static int ep_alloc(struct eventpoll **pep)
861{
Davide Libenzi7ef99642008-12-01 13:13:55 -0800862 int error;
863 struct user_struct *user;
864 struct eventpoll *ep;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700865
Davide Libenzi7ef99642008-12-01 13:13:55 -0800866 user = get_current_user();
Davide Libenzi7ef99642008-12-01 13:13:55 -0800867 error = -ENOMEM;
868 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
869 if (unlikely(!ep))
870 goto free_uid;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700871
Davide Libenzic7ea7632007-05-15 01:40:47 -0700872 spin_lock_init(&ep->lock);
Davide Libenzid47de162007-05-15 01:40:41 -0700873 mutex_init(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700874 init_waitqueue_head(&ep->wq);
875 init_waitqueue_head(&ep->poll_wait);
876 INIT_LIST_HEAD(&ep->rdllist);
877 ep->rbr = RB_ROOT;
Davide Libenzid47de162007-05-15 01:40:41 -0700878 ep->ovflist = EP_UNACTIVE_PTR;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800879 ep->user = user;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700880
881 *pep = ep;
882
Davide Libenzi7699acd2007-05-10 22:23:23 -0700883 return 0;
Davide Libenzi7ef99642008-12-01 13:13:55 -0800884
885free_uid:
886 free_uid(user);
887 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -0700888}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
890/*
Davide Libenzic7ea7632007-05-15 01:40:47 -0700891 * Search the file inside the eventpoll tree. The RB tree operations
892 * are protected by the "mtx" mutex, and ep_find() must be called with
893 * "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 */
895static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
896{
897 int kcmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 struct rb_node *rbp;
899 struct epitem *epi, *epir = NULL;
900 struct epoll_filefd ffd;
901
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700902 ep_set_ffd(&ffd, file, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 for (rbp = ep->rbr.rb_node; rbp; ) {
904 epi = rb_entry(rbp, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -0700905 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 if (kcmp > 0)
907 rbp = rbp->rb_right;
908 else if (kcmp < 0)
909 rbp = rbp->rb_left;
910 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 epir = epi;
912 break;
913 }
914 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 return epir;
917}
918
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919/*
Davide Libenzi7699acd2007-05-10 22:23:23 -0700920 * This is the callback that is passed to the wait queue wakeup
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200921 * mechanism. It is called by the stored file descriptors when they
Davide Libenzi7699acd2007-05-10 22:23:23 -0700922 * have events to report.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 */
Davide Libenzi7699acd2007-05-10 22:23:23 -0700924static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925{
Davide Libenzi7699acd2007-05-10 22:23:23 -0700926 int pwake = 0;
927 unsigned long flags;
928 struct epitem *epi = ep_item_from_wait(wait);
929 struct eventpoll *ep = epi->ep;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930
Oleg Nesterov971316f2012-02-24 20:07:29 +0100931 if ((unsigned long)key & POLLFREE) {
932 ep_pwq_from_wait(wait)->whead = NULL;
933 /*
934 * whead = NULL above can race with ep_remove_wait_queue()
935 * which can do another remove_wait_queue() after us, so we
936 * can't use __remove_wait_queue(). whead->lock is held by
937 * the caller.
938 */
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100939 list_del_init(&wait->task_list);
Oleg Nesterov971316f2012-02-24 20:07:29 +0100940 }
Oleg Nesterovd80e7312012-02-24 20:07:11 +0100941
Davide Libenzic7ea7632007-05-15 01:40:47 -0700942 spin_lock_irqsave(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700943
944 /*
945 * If the event mask does not contain any poll(2) event, we consider the
946 * descriptor to be disabled. This condition is likely the effect of the
947 * EPOLLONESHOT bit that disables the descriptor when an event is received,
948 * until the next EPOLL_CTL_MOD will be issued.
949 */
950 if (!(epi->event.events & ~EP_PRIVATE_BITS))
Davide Libenzid47de162007-05-15 01:40:41 -0700951 goto out_unlock;
952
953 /*
Davide Libenzi2dfa4ee2009-03-31 15:24:22 -0700954 * Check the events coming with the callback. At this stage, not
955 * every device reports the events in the "key" parameter of the
956 * callback. We need to be able to handle both cases here, hence the
957 * test for "key" != NULL before the event match test.
958 */
959 if (key && !((unsigned long) key & epi->event.events))
960 goto out_unlock;
961
962 /*
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200963 * If we are transferring events to userspace, we can hold no locks
Davide Libenzid47de162007-05-15 01:40:41 -0700964 * (because we're accessing user memory, and because of linux f_op->poll()
Daniel Balutabf6a41d2011-01-30 23:42:29 +0200965 * semantics). All the events that happen during that period of time are
Davide Libenzid47de162007-05-15 01:40:41 -0700966 * chained in ep->ovflist and requeued later on.
967 */
968 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
969 if (epi->next == EP_UNACTIVE_PTR) {
970 epi->next = ep->ovflist;
971 ep->ovflist = epi;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200972 if (epi->ws) {
973 /*
974 * Activate ep->ws since epi->ws may get
975 * deactivated at any time.
976 */
977 __pm_stay_awake(ep->ws);
978 }
979
Davide Libenzid47de162007-05-15 01:40:41 -0700980 }
981 goto out_unlock;
982 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700983
984 /* If this file is already in the ready list we exit soon */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200985 if (!ep_is_linked(&epi->rdllink)) {
Davide Libenzi5071f972009-03-31 15:24:10 -0700986 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +0200987 __pm_stay_awake(epi->ws);
988 }
Davide Libenzi7699acd2007-05-10 22:23:23 -0700989
Davide Libenzi7699acd2007-05-10 22:23:23 -0700990 /*
991 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
992 * wait list.
993 */
994 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -0400995 wake_up_locked(&ep->wq);
Davide Libenzi7699acd2007-05-10 22:23:23 -0700996 if (waitqueue_active(&ep->poll_wait))
997 pwake++;
998
Davide Libenzid47de162007-05-15 01:40:41 -0700999out_unlock:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001000 spin_unlock_irqrestore(&ep->lock, flags);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001001
1002 /* We have to call this outside the lock */
1003 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001004 ep_poll_safewake(&ep->poll_wait);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001005
1006 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007}
1008
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009/*
1010 * This is the callback that is used to add our wait queue to the
1011 * target file wakeup lists.
1012 */
1013static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1014 poll_table *pt)
1015{
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001016 struct epitem *epi = ep_item_from_epqueue(pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 struct eppoll_entry *pwq;
1018
Christoph Lametere94b1762006-12-06 20:33:17 -08001019 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1021 pwq->whead = whead;
1022 pwq->base = epi;
1023 add_wait_queue(whead, &pwq->wait);
1024 list_add_tail(&pwq->llink, &epi->pwqlist);
1025 epi->nwait++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001026 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 /* We have to signal that an error occurred */
1028 epi->nwait = -1;
Davide Libenzi296e2362009-03-31 15:24:11 -07001029 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030}
1031
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1033{
1034 int kcmp;
1035 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1036 struct epitem *epic;
1037
1038 while (*p) {
1039 parent = *p;
1040 epic = rb_entry(parent, struct epitem, rbn);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001041 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 if (kcmp > 0)
1043 p = &parent->rb_right;
1044 else
1045 p = &parent->rb_left;
1046 }
1047 rb_link_node(&epi->rbn, parent, p);
1048 rb_insert_color(&epi->rbn, &ep->rbr);
1049}
1050
Andrew Mortona80a6b82012-11-08 15:53:35 -08001051
1052
Jason Baron28d82dc2012-01-12 17:17:43 -08001053#define PATH_ARR_SIZE 5
1054/*
1055 * These are the number paths of length 1 to 5, that we are allowing to emanate
1056 * from a single file of interest. For example, we allow 1000 paths of length
1057 * 1, to emanate from each file of interest. This essentially represents the
1058 * potential wakeup paths, which need to be limited in order to avoid massive
1059 * uncontrolled wakeup storms. The common use case should be a single ep which
1060 * is connected to n file sources. In this case each file source has 1 path
1061 * of length 1. Thus, the numbers below should be more than sufficient. These
1062 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1063 * and delete can't add additional paths. Protected by the epmutex.
1064 */
1065static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1066static int path_count[PATH_ARR_SIZE];
1067
1068static int path_count_inc(int nests)
1069{
Jason Baron93dc6102012-03-16 16:34:03 -04001070 /* Allow an arbitrary number of depth 1 paths */
1071 if (nests == 0)
1072 return 0;
1073
Jason Baron28d82dc2012-01-12 17:17:43 -08001074 if (++path_count[nests] > path_limits[nests])
1075 return -1;
1076 return 0;
1077}
1078
1079static void path_count_init(void)
1080{
1081 int i;
1082
1083 for (i = 0; i < PATH_ARR_SIZE; i++)
1084 path_count[i] = 0;
1085}
1086
1087static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1088{
1089 int error = 0;
1090 struct file *file = priv;
1091 struct file *child_file;
1092 struct epitem *epi;
1093
1094 list_for_each_entry(epi, &file->f_ep_links, fllink) {
1095 child_file = epi->ep->file;
1096 if (is_file_epoll(child_file)) {
1097 if (list_empty(&child_file->f_ep_links)) {
1098 if (path_count_inc(call_nests)) {
1099 error = -1;
1100 break;
1101 }
1102 } else {
1103 error = ep_call_nested(&poll_loop_ncalls,
1104 EP_MAX_NESTS,
1105 reverse_path_check_proc,
1106 child_file, child_file,
1107 current);
1108 }
1109 if (error != 0)
1110 break;
1111 } else {
1112 printk(KERN_ERR "reverse_path_check_proc: "
1113 "file is not an ep!\n");
1114 }
1115 }
1116 return error;
1117}
1118
1119/**
1120 * reverse_path_check - The tfile_check_list is list of file *, which have
1121 * links that are proposed to be newly added. We need to
1122 * make sure that those added links don't add too many
1123 * paths such that we will spend all our time waking up
1124 * eventpoll objects.
1125 *
1126 * Returns: Returns zero if the proposed links don't create too many paths,
1127 * -1 otherwise.
1128 */
1129static int reverse_path_check(void)
1130{
Jason Baron28d82dc2012-01-12 17:17:43 -08001131 int error = 0;
1132 struct file *current_file;
1133
1134 /* let's call this for all tfiles */
1135 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001136 path_count_init();
1137 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1138 reverse_path_check_proc, current_file,
1139 current_file, current);
1140 if (error)
1141 break;
1142 }
1143 return error;
1144}
1145
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001146static int ep_create_wakeup_source(struct epitem *epi)
1147{
1148 const char *name;
1149
1150 if (!epi->ep->ws) {
1151 epi->ep->ws = wakeup_source_register("eventpoll");
1152 if (!epi->ep->ws)
1153 return -ENOMEM;
1154 }
1155
1156 name = epi->ffd.file->f_path.dentry->d_name.name;
1157 epi->ws = wakeup_source_register(name);
1158 if (!epi->ws)
1159 return -ENOMEM;
1160
1161 return 0;
1162}
1163
1164static void ep_destroy_wakeup_source(struct epitem *epi)
1165{
1166 wakeup_source_unregister(epi->ws);
1167 epi->ws = NULL;
1168}
1169
Davide Libenzic7ea7632007-05-15 01:40:47 -07001170/*
1171 * Must be called with "mtx" held.
1172 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1174 struct file *tfile, int fd)
1175{
1176 int error, revents, pwake = 0;
1177 unsigned long flags;
Robin Holt52bd19f72011-01-12 17:00:01 -08001178 long user_watches;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 struct epitem *epi;
1180 struct ep_pqueue epq;
1181
Robin Holt52bd19f72011-01-12 17:00:01 -08001182 user_watches = atomic_long_read(&ep->user->epoll_watches);
1183 if (unlikely(user_watches >= max_user_watches))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001184 return -ENOSPC;
Christoph Lametere94b1762006-12-06 20:33:17 -08001185 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
Davide Libenzi7ef99642008-12-01 13:13:55 -08001186 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187
1188 /* Item initialization follow here ... */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 INIT_LIST_HEAD(&epi->rdllink);
1190 INIT_LIST_HEAD(&epi->fllink);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 INIT_LIST_HEAD(&epi->pwqlist);
1192 epi->ep = ep;
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001193 ep_set_ffd(&epi->ffd, tfile, fd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 epi->event = *event;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 epi->nwait = 0;
Davide Libenzid47de162007-05-15 01:40:41 -07001196 epi->next = EP_UNACTIVE_PTR;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001197 if (epi->event.events & EPOLLWAKEUP) {
1198 error = ep_create_wakeup_source(epi);
1199 if (error)
1200 goto error_create_wakeup_source;
1201 } else {
1202 epi->ws = NULL;
1203 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204
1205 /* Initialize the poll table using the queue callback */
1206 epq.epi = epi;
1207 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
Hans Verkuil626cf232012-03-23 15:02:27 -07001208 epq.pt._key = event->events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209
1210 /*
1211 * Attach the item to the poll hooks and get current event bits.
1212 * We can safely use the file* here because its usage count has
Davide Libenzic7ea7632007-05-15 01:40:47 -07001213 * been increased by the caller of this function. Note that after
1214 * this operation completes, the poll callback can start hitting
1215 * the new item.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 */
1217 revents = tfile->f_op->poll(tfile, &epq.pt);
1218
1219 /*
1220 * We have to check if something went wrong during the poll wait queue
1221 * install process. Namely an allocation for a wait queue failed due
1222 * high memory pressure.
1223 */
Davide Libenzi7ef99642008-12-01 13:13:55 -08001224 error = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 if (epi->nwait < 0)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001226 goto error_unregister;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
1228 /* Add the current item to the list of active epoll hook for this file */
Jonathan Corbet68499912009-02-06 13:52:43 -07001229 spin_lock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 list_add_tail(&epi->fllink, &tfile->f_ep_links);
Jonathan Corbet68499912009-02-06 13:52:43 -07001231 spin_unlock(&tfile->f_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232
Davide Libenzic7ea7632007-05-15 01:40:47 -07001233 /*
1234 * Add the current item to the RB tree. All RB tree operations are
1235 * protected by "mtx", and ep_insert() is called with "mtx" held.
1236 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 ep_rbtree_insert(ep, epi);
1238
Jason Baron28d82dc2012-01-12 17:17:43 -08001239 /* now check if we've created too many backpaths */
1240 error = -EINVAL;
1241 if (reverse_path_check())
1242 goto error_remove_epi;
1243
Davide Libenzic7ea7632007-05-15 01:40:47 -07001244 /* We have to drop the new item inside our item list to keep track of it */
1245 spin_lock_irqsave(&ep->lock, flags);
1246
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 /* If the file is already "ready" we drop it inside the ready list */
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001248 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001250 __pm_stay_awake(epi->ws);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251
1252 /* Notify waiting tasks that events are available */
1253 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -04001254 wake_up_locked(&ep->wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 if (waitqueue_active(&ep->poll_wait))
1256 pwake++;
1257 }
1258
Davide Libenzic7ea7632007-05-15 01:40:47 -07001259 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
Robin Holt52bd19f72011-01-12 17:00:01 -08001261 atomic_long_inc(&ep->user->epoll_watches);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001262
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 /* We have to call this outside the lock */
1264 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001265 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 return 0;
1268
Jason Baron28d82dc2012-01-12 17:17:43 -08001269error_remove_epi:
1270 spin_lock(&tfile->f_lock);
1271 if (ep_is_linked(&epi->fllink))
1272 list_del_init(&epi->fllink);
1273 spin_unlock(&tfile->f_lock);
1274
1275 rb_erase(&epi->rbn, &ep->rbr);
1276
Davide Libenzi7699acd2007-05-10 22:23:23 -07001277error_unregister:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 ep_unregister_pollwait(ep, epi);
1279
1280 /*
1281 * We need to do this because an event could have been arrived on some
Davide Libenzi67647d02007-05-15 01:40:52 -07001282 * allocated wait queue. Note that we don't care about the ep->ovflist
1283 * list, since that is used/cleaned only inside a section bound by "mtx".
1284 * And ep_insert() is called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001286 spin_lock_irqsave(&ep->lock, flags);
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001287 if (ep_is_linked(&epi->rdllink))
Davide Libenzi6192bd52007-05-08 00:25:41 -07001288 list_del_init(&epi->rdllink);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001289 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001291 wakeup_source_unregister(epi->ws);
1292
1293error_create_wakeup_source:
Pekka Enbergb030a4d2005-06-23 00:10:03 -07001294 kmem_cache_free(epi_cache, epi);
Davide Libenzi7ef99642008-12-01 13:13:55 -08001295
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 return error;
1297}
1298
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299/*
1300 * Modify the interest event mask by dropping an event if the new mask
Davide Libenzic7ea7632007-05-15 01:40:47 -07001301 * has a match in the current file status. Must be called with "mtx" held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 */
1303static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1304{
1305 int pwake = 0;
1306 unsigned int revents;
Hans Verkuil626cf232012-03-23 15:02:27 -07001307 poll_table pt;
1308
1309 init_poll_funcptr(&pt, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310
1311 /*
Tony Battersbye057e152009-03-31 15:24:15 -07001312 * Set the new event interest mask before calling f_op->poll();
1313 * otherwise we might miss an event that happens between the
1314 * f_op->poll() call and the new event set registering.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 */
Eric Wong128dd172013-01-01 21:20:27 +00001316 epi->event.events = event->events; /* need barrier below */
Hans Verkuil626cf232012-03-23 15:02:27 -07001317 pt._key = event->events;
Tony Battersbye057e152009-03-31 15:24:15 -07001318 epi->event.data = event->data; /* protected by mtx */
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001319 if (epi->event.events & EPOLLWAKEUP) {
1320 if (!epi->ws)
1321 ep_create_wakeup_source(epi);
1322 } else if (epi->ws) {
1323 ep_destroy_wakeup_source(epi);
1324 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325
1326 /*
Eric Wong128dd172013-01-01 21:20:27 +00001327 * The following barrier has two effects:
1328 *
1329 * 1) Flush epi changes above to other CPUs. This ensures
1330 * we do not miss events from ep_poll_callback if an
1331 * event occurs immediately after we call f_op->poll().
1332 * We need this because we did not take ep->lock while
1333 * changing epi above (but ep_poll_callback does take
1334 * ep->lock).
1335 *
1336 * 2) We also need to ensure we do not miss _past_ events
1337 * when calling f_op->poll(). This barrier also
1338 * pairs with the barrier in wq_has_sleeper (see
1339 * comments for wq_has_sleeper).
1340 *
1341 * This barrier will now guarantee ep_poll_callback or f_op->poll
1342 * (or both) will notice the readiness of an item.
1343 */
1344 smp_mb();
1345
1346 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 * Get current event bits. We can safely use the file* here because
1348 * its usage count has been increased by the caller of this function.
1349 */
Hans Verkuil626cf232012-03-23 15:02:27 -07001350 revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 /*
Davide Libenzic7ea7632007-05-15 01:40:47 -07001353 * If the item is "hot" and it is not registered inside the ready
Davide Libenzi67647d02007-05-15 01:40:52 -07001354 * list, push it inside.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 */
Davide Libenzic7ea7632007-05-15 01:40:47 -07001356 if (revents & event->events) {
Tony Battersbye057e152009-03-31 15:24:15 -07001357 spin_lock_irq(&ep->lock);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001358 if (!ep_is_linked(&epi->rdllink)) {
1359 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001360 __pm_stay_awake(epi->ws);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
Davide Libenzic7ea7632007-05-15 01:40:47 -07001362 /* Notify waiting tasks that events are available */
1363 if (waitqueue_active(&ep->wq))
Matthew Wilcox4a6e9e2c2007-08-30 16:10:22 -04001364 wake_up_locked(&ep->wq);
Davide Libenzic7ea7632007-05-15 01:40:47 -07001365 if (waitqueue_active(&ep->poll_wait))
1366 pwake++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 }
Tony Battersbye057e152009-03-31 15:24:15 -07001368 spin_unlock_irq(&ep->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370
1371 /* We have to call this outside the lock */
1372 if (pwake)
Davide Libenzi5071f972009-03-31 15:24:10 -07001373 ep_poll_safewake(&ep->poll_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374
1375 return 0;
1376}
1377
Davide Libenzi296e2362009-03-31 15:24:11 -07001378static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1379 void *priv)
Davide Libenzi5071f972009-03-31 15:24:10 -07001380{
1381 struct ep_send_events_data *esed = priv;
1382 int eventcnt;
Davide Libenzi296e2362009-03-31 15:24:11 -07001383 unsigned int revents;
Davide Libenzi5071f972009-03-31 15:24:10 -07001384 struct epitem *epi;
1385 struct epoll_event __user *uevent;
Hans Verkuil626cf232012-03-23 15:02:27 -07001386 poll_table pt;
1387
1388 init_poll_funcptr(&pt, NULL);
Davide Libenzi5071f972009-03-31 15:24:10 -07001389
Davide Libenzi296e2362009-03-31 15:24:11 -07001390 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001391 * We can loop without lock because we are passed a task private list.
1392 * Items cannot vanish during the loop because ep_scan_ready_list() is
1393 * holding "mtx" during this call.
Davide Libenzi296e2362009-03-31 15:24:11 -07001394 */
Davide Libenzi5071f972009-03-31 15:24:10 -07001395 for (eventcnt = 0, uevent = esed->events;
1396 !list_empty(head) && eventcnt < esed->maxevents;) {
1397 epi = list_first_entry(head, struct epitem, rdllink);
1398
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001399 /*
1400 * Activate ep->ws before deactivating epi->ws to prevent
1401 * triggering auto-suspend here (in case we reactive epi->ws
1402 * below).
1403 *
1404 * This could be rearranged to delay the deactivation of epi->ws
1405 * instead, but then epi->ws would temporarily be out of sync
1406 * with ep_is_linked().
1407 */
1408 if (epi->ws && epi->ws->active)
1409 __pm_stay_awake(ep->ws);
1410 __pm_relax(epi->ws);
Davide Libenzi5071f972009-03-31 15:24:10 -07001411 list_del_init(&epi->rdllink);
1412
Hans Verkuil626cf232012-03-23 15:02:27 -07001413 pt._key = epi->event.events;
1414 revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
Davide Libenzi296e2362009-03-31 15:24:11 -07001415 epi->event.events;
Davide Libenzi5071f972009-03-31 15:24:10 -07001416
Davide Libenzi296e2362009-03-31 15:24:11 -07001417 /*
Davide Libenzi5071f972009-03-31 15:24:10 -07001418 * If the event mask intersect the caller-requested one,
1419 * deliver the event to userspace. Again, ep_scan_ready_list()
1420 * is holding "mtx", so no operations coming from userspace
1421 * can change the item.
Davide Libenzi296e2362009-03-31 15:24:11 -07001422 */
1423 if (revents) {
Davide Libenzi5071f972009-03-31 15:24:10 -07001424 if (__put_user(revents, &uevent->events) ||
Tony Battersbyd0305882009-03-31 15:24:14 -07001425 __put_user(epi->event.data, &uevent->data)) {
1426 list_add(&epi->rdllink, head);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001427 __pm_stay_awake(epi->ws);
Davide Libenzi296e2362009-03-31 15:24:11 -07001428 return eventcnt ? eventcnt : -EFAULT;
Tony Battersbyd0305882009-03-31 15:24:14 -07001429 }
Davide Libenzi296e2362009-03-31 15:24:11 -07001430 eventcnt++;
Davide Libenzi5071f972009-03-31 15:24:10 -07001431 uevent++;
Davide Libenzi296e2362009-03-31 15:24:11 -07001432 if (epi->event.events & EPOLLONESHOT)
1433 epi->event.events &= EP_PRIVATE_BITS;
1434 else if (!(epi->event.events & EPOLLET)) {
1435 /*
1436 * If this file has been added with Level
1437 * Trigger mode, we need to insert back inside
1438 * the ready list, so that the next call to
1439 * epoll_wait() will check again the events
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001440 * availability. At this point, no one can insert
Davide Libenzi296e2362009-03-31 15:24:11 -07001441 * into ep->rdllist besides us. The epoll_ctl()
1442 * callers are locked out by
1443 * ep_scan_ready_list() holding "mtx" and the
1444 * poll callback will queue them in ep->ovflist.
1445 */
1446 list_add_tail(&epi->rdllink, &ep->rdllist);
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001447 __pm_stay_awake(epi->ws);
Davide Libenzi296e2362009-03-31 15:24:11 -07001448 }
1449 }
1450 }
Davide Libenzi5071f972009-03-31 15:24:10 -07001451
1452 return eventcnt;
1453}
1454
Davide Libenzi296e2362009-03-31 15:24:11 -07001455static int ep_send_events(struct eventpoll *ep,
1456 struct epoll_event __user *events, int maxevents)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457{
Davide Libenzi5071f972009-03-31 15:24:10 -07001458 struct ep_send_events_data esed;
Davide Libenzi6192bd52007-05-08 00:25:41 -07001459
Davide Libenzi5071f972009-03-31 15:24:10 -07001460 esed.maxevents = maxevents;
1461 esed.events = events;
Davide Libenzid47de162007-05-15 01:40:41 -07001462
Nelson Elhaged8805e62011-10-31 17:13:14 -07001463 return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464}
1465
Eric Dumazet0781b902011-02-01 15:52:35 -08001466static inline struct timespec ep_set_mstimeout(long ms)
1467{
1468 struct timespec now, ts = {
1469 .tv_sec = ms / MSEC_PER_SEC,
1470 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1471 };
1472
1473 ktime_get_ts(&now);
1474 return timespec_add_safe(now, ts);
1475}
1476
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001477/**
1478 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1479 * event buffer.
1480 *
1481 * @ep: Pointer to the eventpoll context.
1482 * @events: Pointer to the userspace buffer where the ready events should be
1483 * stored.
1484 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1485 * @timeout: Maximum timeout for the ready events fetch operation, in
1486 * milliseconds. If the @timeout is zero, the function will not block,
1487 * while if the @timeout is less than zero, the function will block
1488 * until at least one event has been retrieved (or an error
1489 * occurred).
1490 *
1491 * Returns: Returns the number of ready events which have been fetched, or an
1492 * error code, in case of error.
1493 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1495 int maxevents, long timeout)
1496{
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001497 int res = 0, eavail, timed_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 unsigned long flags;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001499 long slack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 wait_queue_t wait;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001501 ktime_t expires, *to = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001503 if (timeout > 0) {
Eric Dumazet0781b902011-02-01 15:52:35 -08001504 struct timespec end_time = ep_set_mstimeout(timeout);
1505
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001506 slack = select_estimate_accuracy(&end_time);
1507 to = &expires;
1508 *to = timespec_to_ktime(end_time);
1509 } else if (timeout == 0) {
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001510 /*
1511 * Avoid the unnecessary trip to the wait queue loop, if the
1512 * caller specified a non blocking operation.
1513 */
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001514 timed_out = 1;
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001515 spin_lock_irqsave(&ep->lock, flags);
1516 goto check_events;
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001517 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001519fetch_events:
Davide Libenzic7ea7632007-05-15 01:40:47 -07001520 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001522 if (!ep_events_available(ep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 /*
1524 * We don't have any available event to return to the caller.
1525 * We need to sleep here, and we will be wake up by
1526 * ep_poll_callback() when events will become available.
1527 */
1528 init_waitqueue_entry(&wait, current);
Changli Gaoa93d2f12010-05-07 14:33:26 +08001529 __add_wait_queue_exclusive(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530
1531 for (;;) {
1532 /*
1533 * We don't want to sleep if the ep_poll_callback() sends us
1534 * a wakeup in between. That's why we set the task state
1535 * to TASK_INTERRUPTIBLE before doing the checks.
1536 */
1537 set_current_state(TASK_INTERRUPTIBLE);
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001538 if (ep_events_available(ep) || timed_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 break;
1540 if (signal_pending(current)) {
1541 res = -EINTR;
1542 break;
1543 }
1544
Davide Libenzic7ea7632007-05-15 01:40:47 -07001545 spin_unlock_irqrestore(&ep->lock, flags);
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001546 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1547 timed_out = 1;
1548
Davide Libenzic7ea7632007-05-15 01:40:47 -07001549 spin_lock_irqsave(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 }
Davide Libenzi3419b232006-06-25 05:48:14 -07001551 __remove_wait_queue(&ep->wq, &wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552
1553 set_current_state(TASK_RUNNING);
1554 }
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001555check_events:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 /* Is it worth to try to dig for events ? */
Davide Libenzi3fb0e582011-03-22 16:34:46 -07001557 eavail = ep_events_available(ep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558
Davide Libenzic7ea7632007-05-15 01:40:47 -07001559 spin_unlock_irqrestore(&ep->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560
1561 /*
1562 * Try to transfer events to user space. In case we get 0 events and
1563 * there's still timeout left over, we go trying again in search of
1564 * more luck.
1565 */
1566 if (!res && eavail &&
Shawn Bohrer95aac7b2010-10-27 15:34:54 -07001567 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
Shawn Bohrerf4d93ad2011-03-22 16:34:47 -07001568 goto fetch_events;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
1570 return res;
1571}
1572
Davide Libenzi22bacca2011-02-25 14:44:12 -08001573/**
1574 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1575 * API, to verify that adding an epoll file inside another
1576 * epoll structure, does not violate the constraints, in
1577 * terms of closed loops, or too deep chains (which can
1578 * result in excessive stack usage).
1579 *
1580 * @priv: Pointer to the epoll file to be currently checked.
1581 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1582 * data structure pointer.
1583 * @call_nests: Current dept of the @ep_call_nested() call stack.
1584 *
1585 * Returns: Returns zero if adding the epoll @file inside current epoll
1586 * structure @ep does not violate the constraints, or -1 otherwise.
1587 */
1588static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1589{
1590 int error = 0;
1591 struct file *file = priv;
1592 struct eventpoll *ep = file->private_data;
Jason Baron28d82dc2012-01-12 17:17:43 -08001593 struct eventpoll *ep_tovisit;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001594 struct rb_node *rbp;
1595 struct epitem *epi;
1596
Nelson Elhaged8805e62011-10-31 17:13:14 -07001597 mutex_lock_nested(&ep->mtx, call_nests + 1);
Jason Baron28d82dc2012-01-12 17:17:43 -08001598 ep->visited = 1;
1599 list_add(&ep->visited_list_link, &visited_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001600 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1601 epi = rb_entry(rbp, struct epitem, rbn);
1602 if (unlikely(is_file_epoll(epi->ffd.file))) {
Jason Baron28d82dc2012-01-12 17:17:43 -08001603 ep_tovisit = epi->ffd.file->private_data;
1604 if (ep_tovisit->visited)
1605 continue;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001606 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Jason Baron28d82dc2012-01-12 17:17:43 -08001607 ep_loop_check_proc, epi->ffd.file,
1608 ep_tovisit, current);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001609 if (error != 0)
1610 break;
Jason Baron28d82dc2012-01-12 17:17:43 -08001611 } else {
1612 /*
1613 * If we've reached a file that is not associated with
1614 * an ep, then we need to check if the newly added
1615 * links are going to add too many wakeup paths. We do
1616 * this by adding it to the tfile_check_list, if it's
1617 * not already there, and calling reverse_path_check()
1618 * during ep_insert().
1619 */
1620 if (list_empty(&epi->ffd.file->f_tfile_llink))
1621 list_add(&epi->ffd.file->f_tfile_llink,
1622 &tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001623 }
1624 }
1625 mutex_unlock(&ep->mtx);
1626
1627 return error;
1628}
1629
1630/**
1631 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1632 * another epoll file (represented by @ep) does not create
1633 * closed loops or too deep chains.
1634 *
1635 * @ep: Pointer to the epoll private data structure.
1636 * @file: Pointer to the epoll file to be checked.
1637 *
1638 * Returns: Returns zero if adding the epoll @file inside current epoll
1639 * structure @ep does not violate the constraints, or -1 otherwise.
1640 */
1641static int ep_loop_check(struct eventpoll *ep, struct file *file)
1642{
Jason Baron28d82dc2012-01-12 17:17:43 -08001643 int ret;
1644 struct eventpoll *ep_cur, *ep_next;
1645
1646 ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
Davide Libenzi22bacca2011-02-25 14:44:12 -08001647 ep_loop_check_proc, file, ep, current);
Jason Baron28d82dc2012-01-12 17:17:43 -08001648 /* clear visited list */
1649 list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1650 visited_list_link) {
1651 ep_cur->visited = 0;
1652 list_del(&ep_cur->visited_list_link);
1653 }
1654 return ret;
1655}
1656
1657static void clear_tfile_check_list(void)
1658{
1659 struct file *file;
1660
1661 /* first clear the tfile_check_list */
1662 while (!list_empty(&tfile_check_list)) {
1663 file = list_first_entry(&tfile_check_list, struct file,
1664 f_tfile_llink);
1665 list_del_init(&file->f_tfile_llink);
1666 }
1667 INIT_LIST_HEAD(&tfile_check_list);
Davide Libenzi22bacca2011-02-25 14:44:12 -08001668}
1669
Davide Libenzi7699acd2007-05-10 22:23:23 -07001670/*
Andrew Morton523723b2008-08-12 15:09:01 -07001671 * Open an eventpoll file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001672 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001673SYSCALL_DEFINE1(epoll_create1, int, flags)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001674{
Jason Baron28d82dc2012-01-12 17:17:43 -08001675 int error, fd;
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001676 struct eventpoll *ep = NULL;
Jason Baron28d82dc2012-01-12 17:17:43 -08001677 struct file *file;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001678
Ulrich Dreppere38b36f2008-07-23 21:29:42 -07001679 /* Check the EPOLL_* constant for consistency. */
1680 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1681
Davide Libenzi296e2362009-03-31 15:24:11 -07001682 if (flags & ~EPOLL_CLOEXEC)
1683 return -EINVAL;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001684 /*
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001685 * Create the internal data structure ("struct eventpoll").
Davide Libenzi7699acd2007-05-10 22:23:23 -07001686 */
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001687 error = ep_alloc(&ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001688 if (error < 0)
1689 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001690 /*
1691 * Creates all the items needed to setup an eventpoll file. That is,
Al Viro2030a422008-02-23 06:46:49 -05001692 * a file structure and a free file descriptor.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001693 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001694 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1695 if (fd < 0) {
1696 error = fd;
1697 goto out_free_ep;
1698 }
1699 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
Roland Dreier628ff7c2009-12-18 09:41:24 -08001700 O_RDWR | (flags & O_CLOEXEC));
Jason Baron28d82dc2012-01-12 17:17:43 -08001701 if (IS_ERR(file)) {
1702 error = PTR_ERR(file);
1703 goto out_free_fd;
1704 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001705 ep->file = file;
Al Viro98022742012-08-17 22:42:36 -04001706 fd_install(fd, file);
Jason Baron28d82dc2012-01-12 17:17:43 -08001707 return fd;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001708
Jason Baron28d82dc2012-01-12 17:17:43 -08001709out_free_fd:
1710 put_unused_fd(fd);
1711out_free_ep:
1712 ep_free(ep);
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001713 return error;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001714}
1715
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001716SYSCALL_DEFINE1(epoll_create, int, size)
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001717{
Davide Libenzibfe38912009-05-12 13:19:44 -07001718 if (size <= 0)
Ulrich Drepper9fe5ad92008-07-23 21:29:43 -07001719 return -EINVAL;
1720
1721 return sys_epoll_create1(0);
Ulrich Dreppera0998b52008-07-23 21:29:27 -07001722}
1723
Davide Libenzi7699acd2007-05-10 22:23:23 -07001724/*
1725 * The following function implements the controller interface for
1726 * the eventpoll file that enables the insertion/removal/change of
Davide Libenzi67647d02007-05-15 01:40:52 -07001727 * file descriptors inside the interest set.
Davide Libenzi7699acd2007-05-10 22:23:23 -07001728 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001729SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1730 struct epoll_event __user *, event)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001731{
1732 int error;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001733 int did_lock_epmutex = 0;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001734 struct file *file, *tfile;
1735 struct eventpoll *ep;
1736 struct epitem *epi;
1737 struct epoll_event epds;
1738
Davide Libenzi7699acd2007-05-10 22:23:23 -07001739 error = -EFAULT;
1740 if (ep_op_has_event(op) &&
1741 copy_from_user(&epds, event, sizeof(struct epoll_event)))
1742 goto error_return;
1743
1744 /* Get the "struct file *" for the eventpoll file */
1745 error = -EBADF;
1746 file = fget(epfd);
1747 if (!file)
1748 goto error_return;
1749
1750 /* Get the "struct file *" for the target file */
1751 tfile = fget(fd);
1752 if (!tfile)
1753 goto error_fput;
1754
1755 /* The target file descriptor must support poll */
1756 error = -EPERM;
1757 if (!tfile->f_op || !tfile->f_op->poll)
1758 goto error_tgt_fput;
1759
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001760 /* Check if EPOLLWAKEUP is allowed */
Michael Kerriskd9914cf2012-07-17 21:37:27 +02001761 if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
Rafael J. Wysockia8159412012-05-21 21:20:48 +02001762 epds.events &= ~EPOLLWAKEUP;
Arve Hjønnevåg4d7e30d2012-05-01 21:33:34 +02001763
Davide Libenzi7699acd2007-05-10 22:23:23 -07001764 /*
1765 * We have to check that the file structure underneath the file descriptor
1766 * the user passed to us _is_ an eventpoll file. And also we do not permit
1767 * adding an epoll file descriptor inside itself.
1768 */
1769 error = -EINVAL;
1770 if (file == tfile || !is_file_epoll(file))
1771 goto error_tgt_fput;
1772
1773 /*
1774 * At this point it is safe to assume that the "private_data" contains
1775 * our own data structure.
1776 */
1777 ep = file->private_data;
1778
Davide Libenzi22bacca2011-02-25 14:44:12 -08001779 /*
1780 * When we insert an epoll file descriptor, inside another epoll file
1781 * descriptor, there is the change of creating closed loops, which are
Jason Baron28d82dc2012-01-12 17:17:43 -08001782 * better be handled here, than in more critical paths. While we are
1783 * checking for loops we also determine the list of files reachable
1784 * and hang them on the tfile_check_list, so we can check that we
1785 * haven't created too many possible wakeup paths.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001786 *
Jason Baron28d82dc2012-01-12 17:17:43 -08001787 * We need to hold the epmutex across both ep_insert and ep_remove
1788 * b/c we want to make sure we are looking at a coherent view of
1789 * epoll network.
Davide Libenzi22bacca2011-02-25 14:44:12 -08001790 */
Jason Baron28d82dc2012-01-12 17:17:43 -08001791 if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
Davide Libenzi22bacca2011-02-25 14:44:12 -08001792 mutex_lock(&epmutex);
1793 did_lock_epmutex = 1;
Davide Libenzi22bacca2011-02-25 14:44:12 -08001794 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001795 if (op == EPOLL_CTL_ADD) {
1796 if (is_file_epoll(tfile)) {
1797 error = -ELOOP;
Jason Baron13d51802012-04-25 16:01:47 -07001798 if (ep_loop_check(ep, tfile) != 0) {
1799 clear_tfile_check_list();
Jason Baron28d82dc2012-01-12 17:17:43 -08001800 goto error_tgt_fput;
Jason Baron13d51802012-04-25 16:01:47 -07001801 }
Jason Baron28d82dc2012-01-12 17:17:43 -08001802 } else
1803 list_add(&tfile->f_tfile_llink, &tfile_check_list);
1804 }
Davide Libenzi22bacca2011-02-25 14:44:12 -08001805
Nelson Elhaged8805e62011-10-31 17:13:14 -07001806 mutex_lock_nested(&ep->mtx, 0);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001807
Davide Libenzi67647d02007-05-15 01:40:52 -07001808 /*
1809 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1810 * above, we can be sure to be able to use the item looked up by
1811 * ep_find() till we release the mutex.
1812 */
Davide Libenzi7699acd2007-05-10 22:23:23 -07001813 epi = ep_find(ep, tfile, fd);
1814
1815 error = -EINVAL;
1816 switch (op) {
1817 case EPOLL_CTL_ADD:
1818 if (!epi) {
1819 epds.events |= POLLERR | POLLHUP;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001820 error = ep_insert(ep, &epds, tfile, fd);
1821 } else
1822 error = -EEXIST;
Jason Baron28d82dc2012-01-12 17:17:43 -08001823 clear_tfile_check_list();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001824 break;
1825 case EPOLL_CTL_DEL:
1826 if (epi)
1827 error = ep_remove(ep, epi);
1828 else
1829 error = -ENOENT;
1830 break;
1831 case EPOLL_CTL_MOD:
1832 if (epi) {
1833 epds.events |= POLLERR | POLLHUP;
1834 error = ep_modify(ep, epi, &epds);
1835 } else
1836 error = -ENOENT;
1837 break;
1838 }
Davide Libenzid47de162007-05-15 01:40:41 -07001839 mutex_unlock(&ep->mtx);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001840
1841error_tgt_fput:
Jason Baron28d82dc2012-01-12 17:17:43 -08001842 if (did_lock_epmutex)
Davide Libenzi22bacca2011-02-25 14:44:12 -08001843 mutex_unlock(&epmutex);
1844
Davide Libenzi7699acd2007-05-10 22:23:23 -07001845 fput(tfile);
1846error_fput:
1847 fput(file);
1848error_return:
Davide Libenzi7699acd2007-05-10 22:23:23 -07001849
1850 return error;
1851}
1852
1853/*
1854 * Implement the event wait interface for the eventpoll file. It is the kernel
1855 * part of the user space epoll_wait(2).
1856 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001857SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1858 int, maxevents, int, timeout)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001859{
Al Viro2903ff02012-08-28 12:52:22 -04001860 int error;
1861 struct fd f;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001862 struct eventpoll *ep;
1863
Davide Libenzi7699acd2007-05-10 22:23:23 -07001864 /* The maximum number of event must be greater than zero */
1865 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1866 return -EINVAL;
1867
1868 /* Verify that the area passed by the user is writeable */
Al Viro2903ff02012-08-28 12:52:22 -04001869 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1870 return -EFAULT;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001871
1872 /* Get the "struct file *" for the eventpoll file */
Al Viro2903ff02012-08-28 12:52:22 -04001873 f = fdget(epfd);
1874 if (!f.file)
1875 return -EBADF;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001876
1877 /*
1878 * We have to check that the file structure underneath the fd
1879 * the user passed to us _is_ an eventpoll file.
1880 */
1881 error = -EINVAL;
Al Viro2903ff02012-08-28 12:52:22 -04001882 if (!is_file_epoll(f.file))
Davide Libenzi7699acd2007-05-10 22:23:23 -07001883 goto error_fput;
1884
1885 /*
1886 * At this point it is safe to assume that the "private_data" contains
1887 * our own data structure.
1888 */
Al Viro2903ff02012-08-28 12:52:22 -04001889 ep = f.file->private_data;
Davide Libenzi7699acd2007-05-10 22:23:23 -07001890
1891 /* Time to fish for events ... */
1892 error = ep_poll(ep, events, maxevents, timeout);
1893
1894error_fput:
Al Viro2903ff02012-08-28 12:52:22 -04001895 fdput(f);
Davide Libenzi7699acd2007-05-10 22:23:23 -07001896 return error;
1897}
1898
Davide Libenzi7699acd2007-05-10 22:23:23 -07001899/*
1900 * Implement the event wait interface for the eventpoll file. It is the kernel
1901 * part of the user space epoll_pwait(2).
1902 */
Heiko Carstens5a8a82b2009-01-14 14:14:25 +01001903SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1904 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1905 size_t, sigsetsize)
Davide Libenzi7699acd2007-05-10 22:23:23 -07001906{
1907 int error;
1908 sigset_t ksigmask, sigsaved;
1909
1910 /*
1911 * If the caller wants a certain signal mask to be set during the wait,
1912 * we apply it here.
1913 */
1914 if (sigmask) {
1915 if (sigsetsize != sizeof(sigset_t))
1916 return -EINVAL;
1917 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1918 return -EFAULT;
1919 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1920 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1921 }
1922
1923 error = sys_epoll_wait(epfd, events, maxevents, timeout);
1924
1925 /*
1926 * If we changed the signal mask, we need to restore the original one.
1927 * In case we've got a signal while waiting, we do not restore the
1928 * signal mask yet, and we allow do_signal() to deliver the signal on
1929 * the way back to userspace, before the signal mask is restored.
1930 */
1931 if (sigmask) {
1932 if (error == -EINTR) {
1933 memcpy(&current->saved_sigmask, &sigsaved,
Davide Libenzic7ea7632007-05-15 01:40:47 -07001934 sizeof(sigsaved));
Roland McGrath4e4c22c2008-04-30 00:53:06 -07001935 set_restore_sigmask();
Davide Libenzi7699acd2007-05-10 22:23:23 -07001936 } else
1937 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1938 }
1939
1940 return error;
1941}
1942
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943static int __init eventpoll_init(void)
1944{
Davide Libenzi7ef99642008-12-01 13:13:55 -08001945 struct sysinfo si;
1946
1947 si_meminfo(&si);
Davide Libenzi9df04e12009-01-29 14:25:26 -08001948 /*
1949 * Allows top 4% of lomem to be allocated for epoll watches (per user).
1950 */
1951 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
Davide Libenzi7ef99642008-12-01 13:13:55 -08001952 EP_ITEM_COST;
Robin Holt52bd19f72011-01-12 17:00:01 -08001953 BUG_ON(max_user_watches < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
Davide Libenzi22bacca2011-02-25 14:44:12 -08001955 /*
1956 * Initialize the structure used to perform epoll file descriptor
1957 * inclusion loops checks.
1958 */
1959 ep_nested_calls_init(&poll_loop_ncalls);
1960
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 /* Initialize the structure used to perform safe poll wait head wake ups */
Davide Libenzi5071f972009-03-31 15:24:10 -07001962 ep_nested_calls_init(&poll_safewake_ncalls);
1963
1964 /* Initialize the structure used to perform file's f_op->poll() calls */
1965 ep_nested_calls_init(&poll_readywalk_ncalls);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966
1967 /* Allocates slab cache used to allocate "struct epitem" items */
1968 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001969 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970
1971 /* Allocates slab cache used to allocate "struct eppoll_entry" */
1972 pwq_cache = kmem_cache_create("eventpoll_pwq",
Davide Libenzibb57c3e2009-03-31 15:24:12 -07001973 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976}
Davide Libenzicea69242007-05-10 22:23:22 -07001977fs_initcall(eventpoll_init);