Merge branch 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6.git] / kernel / wait.c
1 /*
2  * Generic waiting primitives.
3  *
4  * (C) 2004 William Irwin, Oracle
5  */
6 #include <linux/init.h>
7 #include <linux/module.h>
8 #include <linux/sched.h>
9 #include <linux/mm.h>
10 #include <linux/wait.h>
11 #include <linux/hash.h>
12
13 void init_waitqueue_head(wait_queue_head_t *q)
14 {
15         spin_lock_init(&q->lock);
16         INIT_LIST_HEAD(&q->task_list);
17 }
18
19 EXPORT_SYMBOL(init_waitqueue_head);
20
21 void add_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
22 {
23         unsigned long flags;
24
25         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
26         spin_lock_irqsave(&q->lock, flags);
27         __add_wait_queue(q, wait);
28         spin_unlock_irqrestore(&q->lock, flags);
29 }
30 EXPORT_SYMBOL(add_wait_queue);
31
32 void add_wait_queue_exclusive(wait_queue_head_t *q, wait_queue_t *wait)
33 {
34         unsigned long flags;
35
36         wait->flags |= WQ_FLAG_EXCLUSIVE;
37         spin_lock_irqsave(&q->lock, flags);
38         __add_wait_queue_tail(q, wait);
39         spin_unlock_irqrestore(&q->lock, flags);
40 }
41 EXPORT_SYMBOL(add_wait_queue_exclusive);
42
43 void remove_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
44 {
45         unsigned long flags;
46
47         spin_lock_irqsave(&q->lock, flags);
48         __remove_wait_queue(q, wait);
49         spin_unlock_irqrestore(&q->lock, flags);
50 }
51 EXPORT_SYMBOL(remove_wait_queue);
52
53
54 /*
55  * Note: we use "set_current_state()" _after_ the wait-queue add,
56  * because we need a memory barrier there on SMP, so that any
57  * wake-function that tests for the wait-queue being active
58  * will be guaranteed to see waitqueue addition _or_ subsequent
59  * tests in this thread will see the wakeup having taken place.
60  *
61  * The spin_unlock() itself is semi-permeable and only protects
62  * one way (it only protects stuff inside the critical region and
63  * stops them from bleeding out - it would still allow subsequent
64  * loads to move into the critical region).
65  */
66 void
67 prepare_to_wait(wait_queue_head_t *q, wait_queue_t *wait, int state)
68 {
69         unsigned long flags;
70
71         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
72         spin_lock_irqsave(&q->lock, flags);
73         if (list_empty(&wait->task_list))
74                 __add_wait_queue(q, wait);
75         set_current_state(state);
76         spin_unlock_irqrestore(&q->lock, flags);
77 }
78 EXPORT_SYMBOL(prepare_to_wait);
79
80 void
81 prepare_to_wait_exclusive(wait_queue_head_t *q, wait_queue_t *wait, int state)
82 {
83         unsigned long flags;
84
85         wait->flags |= WQ_FLAG_EXCLUSIVE;
86         spin_lock_irqsave(&q->lock, flags);
87         if (list_empty(&wait->task_list))
88                 __add_wait_queue_tail(q, wait);
89         set_current_state(state);
90         spin_unlock_irqrestore(&q->lock, flags);
91 }
92 EXPORT_SYMBOL(prepare_to_wait_exclusive);
93
94 void finish_wait(wait_queue_head_t *q, wait_queue_t *wait)
95 {
96         unsigned long flags;
97
98         __set_current_state(TASK_RUNNING);
99         /*
100          * We can check for list emptiness outside the lock
101          * IFF:
102          *  - we use the "careful" check that verifies both
103          *    the next and prev pointers, so that there cannot
104          *    be any half-pending updates in progress on other
105          *    CPU's that we haven't seen yet (and that might
106          *    still change the stack area.
107          * and
108          *  - all other users take the lock (ie we can only
109          *    have _one_ other CPU that looks at or modifies
110          *    the list).
111          */
112         if (!list_empty_careful(&wait->task_list)) {
113                 spin_lock_irqsave(&q->lock, flags);
114                 list_del_init(&wait->task_list);
115                 spin_unlock_irqrestore(&q->lock, flags);
116         }
117 }
118 EXPORT_SYMBOL(finish_wait);
119
120 int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
121 {
122         int ret = default_wake_function(wait, mode, sync, key);
123
124         if (ret)
125                 list_del_init(&wait->task_list);
126         return ret;
127 }
128 EXPORT_SYMBOL(autoremove_wake_function);
129
130 int wake_bit_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
131 {
132         struct wait_bit_key *key = arg;
133         struct wait_bit_queue *wait_bit
134                 = container_of(wait, struct wait_bit_queue, wait);
135
136         if (wait_bit->key.flags != key->flags ||
137                         wait_bit->key.bit_nr != key->bit_nr ||
138                         test_bit(key->bit_nr, key->flags))
139                 return 0;
140         else
141                 return autoremove_wake_function(wait, mode, sync, key);
142 }
143 EXPORT_SYMBOL(wake_bit_function);
144
145 /*
146  * To allow interruptible waiting and asynchronous (i.e. nonblocking)
147  * waiting, the actions of __wait_on_bit() and __wait_on_bit_lock() are
148  * permitted return codes. Nonzero return codes halt waiting and return.
149  */
150 int __sched
151 __wait_on_bit(wait_queue_head_t *wq, struct wait_bit_queue *q,
152                         int (*action)(void *), unsigned mode)
153 {
154         int ret = 0;
155
156         do {
157                 prepare_to_wait(wq, &q->wait, mode);
158                 if (test_bit(q->key.bit_nr, q->key.flags))
159                         ret = (*action)(q->key.flags);
160         } while (test_bit(q->key.bit_nr, q->key.flags) && !ret);
161         finish_wait(wq, &q->wait);
162         return ret;
163 }
164 EXPORT_SYMBOL(__wait_on_bit);
165
166 int __sched out_of_line_wait_on_bit(void *word, int bit,
167                                         int (*action)(void *), unsigned mode)
168 {
169         wait_queue_head_t *wq = bit_waitqueue(word, bit);
170         DEFINE_WAIT_BIT(wait, word, bit);
171
172         return __wait_on_bit(wq, &wait, action, mode);
173 }
174 EXPORT_SYMBOL(out_of_line_wait_on_bit);
175
176 int __sched
177 __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q,
178                         int (*action)(void *), unsigned mode)
179 {
180         int ret = 0;
181
182         do {
183                 prepare_to_wait_exclusive(wq, &q->wait, mode);
184                 if (test_bit(q->key.bit_nr, q->key.flags)) {
185                         if ((ret = (*action)(q->key.flags)))
186                                 break;
187                 }
188         } while (test_and_set_bit(q->key.bit_nr, q->key.flags));
189         finish_wait(wq, &q->wait);
190         return ret;
191 }
192 EXPORT_SYMBOL(__wait_on_bit_lock);
193
194 int __sched out_of_line_wait_on_bit_lock(void *word, int bit,
195                                         int (*action)(void *), unsigned mode)
196 {
197         wait_queue_head_t *wq = bit_waitqueue(word, bit);
198         DEFINE_WAIT_BIT(wait, word, bit);
199
200         return __wait_on_bit_lock(wq, &wait, action, mode);
201 }
202 EXPORT_SYMBOL(out_of_line_wait_on_bit_lock);
203
204 void __wake_up_bit(wait_queue_head_t *wq, void *word, int bit)
205 {
206         struct wait_bit_key key = __WAIT_BIT_KEY_INITIALIZER(word, bit);
207         if (waitqueue_active(wq))
208                 __wake_up(wq, TASK_NORMAL, 1, &key);
209 }
210 EXPORT_SYMBOL(__wake_up_bit);
211
212 /**
213  * wake_up_bit - wake up a waiter on a bit
214  * @word: the word being waited on, a kernel virtual address
215  * @bit: the bit of the word being waited on
216  *
217  * There is a standard hashed waitqueue table for generic use. This
218  * is the part of the hashtable's accessor API that wakes up waiters
219  * on a bit. For instance, if one were to have waiters on a bitflag,
220  * one would call wake_up_bit() after clearing the bit.
221  *
222  * In order for this to function properly, as it uses waitqueue_active()
223  * internally, some kind of memory barrier must be done prior to calling
224  * this. Typically, this will be smp_mb__after_clear_bit(), but in some
225  * cases where bitflags are manipulated non-atomically under a lock, one
226  * may need to use a less regular barrier, such fs/inode.c's smp_mb(),
227  * because spin_unlock() does not guarantee a memory barrier.
228  */
229 void wake_up_bit(void *word, int bit)
230 {
231         __wake_up_bit(bit_waitqueue(word, bit), word, bit);
232 }
233 EXPORT_SYMBOL(wake_up_bit);
234
235 wait_queue_head_t *bit_waitqueue(void *word, int bit)
236 {
237         const int shift = BITS_PER_LONG == 32 ? 5 : 6;
238         const struct zone *zone = page_zone(virt_to_page(word));
239         unsigned long val = (unsigned long)word << shift | bit;
240
241         return &zone->wait_table[hash_long(val, zone->wait_table_bits)];
242 }
243 EXPORT_SYMBOL(bit_waitqueue);