Merge branch 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/rzhang/linux
[linux-2.6-block.git] / kernel / locking / rwsem-spinlock.c
1 /* rwsem-spinlock.c: R/W semaphores: contention handling functions for
2  * generic spinlock implementation
3  *
4  * Copyright (c) 2001   David Howells (dhowells@redhat.com).
5  * - Derived partially from idea by Andrea Arcangeli <andrea@suse.de>
6  * - Derived also from comments by Linus
7  */
8 #include <linux/rwsem.h>
9 #include <linux/sched.h>
10 #include <linux/export.h>
11
12 enum rwsem_waiter_type {
13         RWSEM_WAITING_FOR_WRITE,
14         RWSEM_WAITING_FOR_READ
15 };
16
17 struct rwsem_waiter {
18         struct list_head list;
19         struct task_struct *task;
20         enum rwsem_waiter_type type;
21 };
22
23 int rwsem_is_locked(struct rw_semaphore *sem)
24 {
25         int ret = 1;
26         unsigned long flags;
27
28         if (raw_spin_trylock_irqsave(&sem->wait_lock, flags)) {
29                 ret = (sem->count != 0);
30                 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
31         }
32         return ret;
33 }
34 EXPORT_SYMBOL(rwsem_is_locked);
35
36 /*
37  * initialise the semaphore
38  */
39 void __init_rwsem(struct rw_semaphore *sem, const char *name,
40                   struct lock_class_key *key)
41 {
42 #ifdef CONFIG_DEBUG_LOCK_ALLOC
43         /*
44          * Make sure we are not reinitializing a held semaphore:
45          */
46         debug_check_no_locks_freed((void *)sem, sizeof(*sem));
47         lockdep_init_map(&sem->dep_map, name, key, 0);
48 #endif
49         sem->count = 0;
50         raw_spin_lock_init(&sem->wait_lock);
51         INIT_LIST_HEAD(&sem->wait_list);
52 }
53 EXPORT_SYMBOL(__init_rwsem);
54
55 /*
56  * handle the lock release when processes blocked on it that can now run
57  * - if we come here, then:
58  *   - the 'active count' _reached_ zero
59  *   - the 'waiting count' is non-zero
60  * - the spinlock must be held by the caller
61  * - woken process blocks are discarded from the list after having task zeroed
62  * - writers are only woken if wakewrite is non-zero
63  */
64 static inline struct rw_semaphore *
65 __rwsem_do_wake(struct rw_semaphore *sem, int wakewrite)
66 {
67         struct rwsem_waiter *waiter;
68         struct task_struct *tsk;
69         int woken;
70
71         waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
72
73         if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
74                 if (wakewrite)
75                         /* Wake up a writer. Note that we do not grant it the
76                          * lock - it will have to acquire it when it runs. */
77                         wake_up_process(waiter->task);
78                 goto out;
79         }
80
81         /* grant an infinite number of read locks to the front of the queue */
82         woken = 0;
83         do {
84                 struct list_head *next = waiter->list.next;
85
86                 list_del(&waiter->list);
87                 tsk = waiter->task;
88                 /*
89                  * Make sure we do not wakeup the next reader before
90                  * setting the nil condition to grant the next reader;
91                  * otherwise we could miss the wakeup on the other
92                  * side and end up sleeping again. See the pairing
93                  * in rwsem_down_read_failed().
94                  */
95                 smp_mb();
96                 waiter->task = NULL;
97                 wake_up_process(tsk);
98                 put_task_struct(tsk);
99                 woken++;
100                 if (next == &sem->wait_list)
101                         break;
102                 waiter = list_entry(next, struct rwsem_waiter, list);
103         } while (waiter->type != RWSEM_WAITING_FOR_WRITE);
104
105         sem->count += woken;
106
107  out:
108         return sem;
109 }
110
111 /*
112  * wake a single writer
113  */
114 static inline struct rw_semaphore *
115 __rwsem_wake_one_writer(struct rw_semaphore *sem)
116 {
117         struct rwsem_waiter *waiter;
118
119         waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
120         wake_up_process(waiter->task);
121
122         return sem;
123 }
124
125 /*
126  * get a read lock on the semaphore
127  */
128 void __sched __down_read(struct rw_semaphore *sem)
129 {
130         struct rwsem_waiter waiter;
131         unsigned long flags;
132
133         raw_spin_lock_irqsave(&sem->wait_lock, flags);
134
135         if (sem->count >= 0 && list_empty(&sem->wait_list)) {
136                 /* granted */
137                 sem->count++;
138                 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
139                 goto out;
140         }
141
142         set_current_state(TASK_UNINTERRUPTIBLE);
143
144         /* set up my own style of waitqueue */
145         waiter.task = current;
146         waiter.type = RWSEM_WAITING_FOR_READ;
147         get_task_struct(current);
148
149         list_add_tail(&waiter.list, &sem->wait_list);
150
151         /* we don't need to touch the semaphore struct anymore */
152         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
153
154         /* wait to be given the lock */
155         for (;;) {
156                 if (!waiter.task)
157                         break;
158                 schedule();
159                 set_current_state(TASK_UNINTERRUPTIBLE);
160         }
161
162         __set_current_state(TASK_RUNNING);
163  out:
164         ;
165 }
166
167 /*
168  * trylock for reading -- returns 1 if successful, 0 if contention
169  */
170 int __down_read_trylock(struct rw_semaphore *sem)
171 {
172         unsigned long flags;
173         int ret = 0;
174
175
176         raw_spin_lock_irqsave(&sem->wait_lock, flags);
177
178         if (sem->count >= 0 && list_empty(&sem->wait_list)) {
179                 /* granted */
180                 sem->count++;
181                 ret = 1;
182         }
183
184         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
185
186         return ret;
187 }
188
189 /*
190  * get a write lock on the semaphore
191  */
192 int __sched __down_write_common(struct rw_semaphore *sem, int state)
193 {
194         struct rwsem_waiter waiter;
195         unsigned long flags;
196         int ret = 0;
197
198         raw_spin_lock_irqsave(&sem->wait_lock, flags);
199
200         /* set up my own style of waitqueue */
201         waiter.task = current;
202         waiter.type = RWSEM_WAITING_FOR_WRITE;
203         list_add_tail(&waiter.list, &sem->wait_list);
204
205         /* wait for someone to release the lock */
206         for (;;) {
207                 /*
208                  * That is the key to support write lock stealing: allows the
209                  * task already on CPU to get the lock soon rather than put
210                  * itself into sleep and waiting for system woke it or someone
211                  * else in the head of the wait list up.
212                  */
213                 if (sem->count == 0)
214                         break;
215                 if (signal_pending_state(state, current)) {
216                         ret = -EINTR;
217                         goto out;
218                 }
219                 set_current_state(state);
220                 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
221                 schedule();
222                 raw_spin_lock_irqsave(&sem->wait_lock, flags);
223         }
224         /* got the lock */
225         sem->count = -1;
226 out:
227         list_del(&waiter.list);
228
229         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
230
231         return ret;
232 }
233
234 void __sched __down_write(struct rw_semaphore *sem)
235 {
236         __down_write_common(sem, TASK_UNINTERRUPTIBLE);
237 }
238
239 int __sched __down_write_killable(struct rw_semaphore *sem)
240 {
241         return __down_write_common(sem, TASK_KILLABLE);
242 }
243
244 /*
245  * trylock for writing -- returns 1 if successful, 0 if contention
246  */
247 int __down_write_trylock(struct rw_semaphore *sem)
248 {
249         unsigned long flags;
250         int ret = 0;
251
252         raw_spin_lock_irqsave(&sem->wait_lock, flags);
253
254         if (sem->count == 0) {
255                 /* got the lock */
256                 sem->count = -1;
257                 ret = 1;
258         }
259
260         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
261
262         return ret;
263 }
264
265 /*
266  * release a read lock on the semaphore
267  */
268 void __up_read(struct rw_semaphore *sem)
269 {
270         unsigned long flags;
271
272         raw_spin_lock_irqsave(&sem->wait_lock, flags);
273
274         if (--sem->count == 0 && !list_empty(&sem->wait_list))
275                 sem = __rwsem_wake_one_writer(sem);
276
277         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
278 }
279
280 /*
281  * release a write lock on the semaphore
282  */
283 void __up_write(struct rw_semaphore *sem)
284 {
285         unsigned long flags;
286
287         raw_spin_lock_irqsave(&sem->wait_lock, flags);
288
289         sem->count = 0;
290         if (!list_empty(&sem->wait_list))
291                 sem = __rwsem_do_wake(sem, 1);
292
293         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
294 }
295
296 /*
297  * downgrade a write lock into a read lock
298  * - just wake up any readers at the front of the queue
299  */
300 void __downgrade_write(struct rw_semaphore *sem)
301 {
302         unsigned long flags;
303
304         raw_spin_lock_irqsave(&sem->wait_lock, flags);
305
306         sem->count = 1;
307         if (!list_empty(&sem->wait_list))
308                 sem = __rwsem_do_wake(sem, 0);
309
310         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
311 }
312