Linux-2.6.12-rc2
[linux-flexiantxendom0-natty.git] / arch / x86_64 / kernel / semaphore.c
1 /*
2  * x86_64 semaphore implementation.
3  *
4  * (C) Copyright 1999 Linus Torvalds
5  *
6  * Portions Copyright 1999 Red Hat, Inc.
7  *
8  *      This program is free software; you can redistribute it and/or
9  *      modify it under the terms of the GNU General Public License
10  *      as published by the Free Software Foundation; either version
11  *      2 of the License, or (at your option) any later version.
12  *
13  * rw semaphores implemented November 1999 by Benjamin LaHaise <bcrl@kvack.org>
14  */
15 #include <linux/config.h>
16 #include <linux/sched.h>
17 #include <linux/init.h>
18 #include <asm/errno.h>
19
20 #include <asm/semaphore.h>
21
22 /*
23  * Semaphores are implemented using a two-way counter:
24  * The "count" variable is decremented for each process
25  * that tries to acquire the semaphore, while the "sleeping"
26  * variable is a count of such acquires.
27  *
28  * Notably, the inline "up()" and "down()" functions can
29  * efficiently test if they need to do any extra work (up
30  * needs to do something only if count was negative before
31  * the increment operation.
32  *
33  * "sleeping" and the contention routine ordering is protected
34  * by the spinlock in the semaphore's waitqueue head.
35  *
36  * Note that these functions are only called when there is
37  * contention on the lock, and as such all this is the
38  * "non-critical" part of the whole semaphore business. The
39  * critical part is the inline stuff in <asm/semaphore.h>
40  * where we want to avoid any extra jumps and calls.
41  */
42
43 /*
44  * Logic:
45  *  - only on a boundary condition do we need to care. When we go
46  *    from a negative count to a non-negative, we wake people up.
47  *  - when we go from a non-negative count to a negative do we
48  *    (a) synchronize with the "sleeper" count and (b) make sure
49  *    that we're on the wakeup list before we synchronize so that
50  *    we cannot lose wakeup events.
51  */
52
53 void __up(struct semaphore *sem)
54 {
55         wake_up(&sem->wait);
56 }
57
58 void __sched __down(struct semaphore * sem)
59 {
60         struct task_struct *tsk = current;
61         DECLARE_WAITQUEUE(wait, tsk);
62         unsigned long flags;
63
64         tsk->state = TASK_UNINTERRUPTIBLE;
65         spin_lock_irqsave(&sem->wait.lock, flags);
66         add_wait_queue_exclusive_locked(&sem->wait, &wait);
67
68         sem->sleepers++;
69         for (;;) {
70                 int sleepers = sem->sleepers;
71
72                 /*
73                  * Add "everybody else" into it. They aren't
74                  * playing, because we own the spinlock in
75                  * the wait_queue_head.
76                  */
77                 if (!atomic_add_negative(sleepers - 1, &sem->count)) {
78                         sem->sleepers = 0;
79                         break;
80                 }
81                 sem->sleepers = 1;      /* us - see -1 above */
82                 spin_unlock_irqrestore(&sem->wait.lock, flags);
83
84                 schedule();
85
86                 spin_lock_irqsave(&sem->wait.lock, flags);
87                 tsk->state = TASK_UNINTERRUPTIBLE;
88         }
89         remove_wait_queue_locked(&sem->wait, &wait);
90         wake_up_locked(&sem->wait);
91         spin_unlock_irqrestore(&sem->wait.lock, flags);
92         tsk->state = TASK_RUNNING;
93 }
94
95 int __sched __down_interruptible(struct semaphore * sem)
96 {
97         int retval = 0;
98         struct task_struct *tsk = current;
99         DECLARE_WAITQUEUE(wait, tsk);
100         unsigned long flags;
101
102         tsk->state = TASK_INTERRUPTIBLE;
103         spin_lock_irqsave(&sem->wait.lock, flags);
104         add_wait_queue_exclusive_locked(&sem->wait, &wait);
105
106         sem->sleepers++;
107         for (;;) {
108                 int sleepers = sem->sleepers;
109
110                 /*
111                  * With signals pending, this turns into
112                  * the trylock failure case - we won't be
113                  * sleeping, and we* can't get the lock as
114                  * it has contention. Just correct the count
115                  * and exit.
116                  */
117                 if (signal_pending(current)) {
118                         retval = -EINTR;
119                         sem->sleepers = 0;
120                         atomic_add(sleepers, &sem->count);
121                         break;
122                 }
123
124                 /*
125                  * Add "everybody else" into it. They aren't
126                  * playing, because we own the spinlock in
127                  * wait_queue_head. The "-1" is because we're
128                  * still hoping to get the semaphore.
129                  */
130                 if (!atomic_add_negative(sleepers - 1, &sem->count)) {
131                         sem->sleepers = 0;
132                         break;
133                 }
134                 sem->sleepers = 1;      /* us - see -1 above */
135                 spin_unlock_irqrestore(&sem->wait.lock, flags);
136
137                 schedule();
138
139                 spin_lock_irqsave(&sem->wait.lock, flags);
140                 tsk->state = TASK_INTERRUPTIBLE;
141         }
142         remove_wait_queue_locked(&sem->wait, &wait);
143         wake_up_locked(&sem->wait);
144         spin_unlock_irqrestore(&sem->wait.lock, flags);
145
146         tsk->state = TASK_RUNNING;
147         return retval;
148 }
149
150 /*
151  * Trylock failed - make sure we correct for
152  * having decremented the count.
153  *
154  * We could have done the trylock with a
155  * single "cmpxchg" without failure cases,
156  * but then it wouldn't work on a 386.
157  */
158 int __down_trylock(struct semaphore * sem)
159 {
160         int sleepers;
161         unsigned long flags;
162
163         spin_lock_irqsave(&sem->wait.lock, flags);
164         sleepers = sem->sleepers + 1;
165         sem->sleepers = 0;
166
167         /*
168          * Add "everybody else" and us into it. They aren't
169          * playing, because we own the spinlock in the
170          * wait_queue_head.
171          */
172         if (!atomic_add_negative(sleepers, &sem->count)) {
173                 wake_up_locked(&sem->wait);
174         }
175
176         spin_unlock_irqrestore(&sem->wait.lock, flags);
177         return 1;
178 }
179
180