arch/i386/kernel/semaphore.c

   1 /*
   2  * i386 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/err.h>
  18 #include <linux/init.h>
  19 #include <asm/semaphore.h>
  20
  21 /*
  22  * Semaphores are implemented using a two-way counter:
  23  * The "count" variable is decremented for each process
  24  * that tries to acquire the semaphore, while the "sleeping"
  25  * variable is a count of such acquires.
  26  *
  27  * Notably, the inline "up()" and "down()" functions can
  28  * efficiently test if they need to do any extra work (up
  29  * needs to do something only if count was negative before
  30  * the increment operation.
  31  *
  32  * "sleeping" and the contention routine ordering is protected
  33  * by the spinlock in the semaphore's waitqueue head.
  34  *
  35  * Note that these functions are only called when there is
  36  * contention on the lock, and as such all this is the
  37  * "non-critical" part of the whole semaphore business. The
  38  * critical part is the inline stuff in <asm/semaphore.h>
  39  * where we want to avoid any extra jumps and calls.
  40  */
  41
  42 /*
  43  * Logic:
  44  *  - only on a boundary condition do we need to care. When we go
  45  *    from a negative count to a non-negative, we wake people up.
  46  *  - when we go from a non-negative count to a negative do we
  47  *    (a) synchronize with the "sleeper" count and (b) make sure
  48  *    that we're on the wakeup list before we synchronize so that
  49  *    we cannot lose wakeup events.
  50  */
  51
  52 static fastcall void __attribute_used__  __up(struct semaphore *sem)
  53 {
  54         wake_up(&sem->wait);
  55 }
  56
  57 static fastcall void __attribute_used__ __sched __down(struct semaphore * sem)
  58 {
  59         struct task_struct *tsk = current;
  60         DECLARE_WAITQUEUE(wait, tsk);
  61         unsigned long flags;
  62
  63         tsk->state = TASK_UNINTERRUPTIBLE;
  64         spin_lock_irqsave(&sem->wait.lock, flags);
  65         add_wait_queue_exclusive_locked(&sem->wait, &wait);
  66
  67         sem->sleepers++;
  68         for (;;) {
  69                 int sleepers = sem->sleepers;
  70
  71                 /*
  72                  * Add "everybody else" into it. They aren't
  73                  * playing, because we own the spinlock in
  74                  * the wait_queue_head.
  75                  */
  76                 if (!atomic_add_negative(sleepers - 1, &sem->count)) {
  77                         sem->sleepers = 0;
  78                         break;
  79                 }
  80                 sem->sleepers = 1;      /* us - see -1 above */
  81                 spin_unlock_irqrestore(&sem->wait.lock, flags);
  82
  83                 schedule();
  84
  85                 spin_lock_irqsave(&sem->wait.lock, flags);
  86                 tsk->state = TASK_UNINTERRUPTIBLE;
  87         }
  88         remove_wait_queue_locked(&sem->wait, &wait);
  89         wake_up_locked(&sem->wait);
  90         spin_unlock_irqrestore(&sem->wait.lock, flags);
  91         tsk->state = TASK_RUNNING;
  92 }
  93
  94 static fastcall int __attribute_used__ __sched __down_interruptible(struct semaphore * sem)
  95 {
  96         int retval = 0;
  97         struct task_struct *tsk = current;
  98         DECLARE_WAITQUEUE(wait, tsk);
  99         unsigned long flags;
 100
 101         tsk->state = TASK_INTERRUPTIBLE;
 102         spin_lock_irqsave(&sem->wait.lock, flags);
 103         add_wait_queue_exclusive_locked(&sem->wait, &wait);
 104
 105         sem->sleepers++;
 106         for (;;) {
 107                 int sleepers = sem->sleepers;
 108
 109                 /*
 110                  * With signals pending, this turns into
 111                  * the trylock failure case - we won't be
 112                  * sleeping, and we* can't get the lock as
 113                  * it has contention. Just correct the count
 114                  * and exit.
 115                  */
 116                 if (signal_pending(current)) {
 117                         retval = -EINTR;
 118                         sem->sleepers = 0;
 119                         atomic_add(sleepers, &sem->count);
 120                         break;
 121                 }
 122
 123                 /*
 124                  * Add "everybody else" into it. They aren't
 125                  * playing, because we own the spinlock in
 126                  * wait_queue_head. The "-1" is because we're
 127                  * still hoping to get the semaphore.
 128                  */
 129                 if (!atomic_add_negative(sleepers - 1, &sem->count)) {
 130                         sem->sleepers = 0;
 131                         break;
 132                 }
 133                 sem->sleepers = 1;      /* us - see -1 above */
 134                 spin_unlock_irqrestore(&sem->wait.lock, flags);
 135
 136                 schedule();
 137
 138                 spin_lock_irqsave(&sem->wait.lock, flags);
 139                 tsk->state = TASK_INTERRUPTIBLE;
 140         }
 141         remove_wait_queue_locked(&sem->wait, &wait);
 142         wake_up_locked(&sem->wait);
 143         spin_unlock_irqrestore(&sem->wait.lock, flags);
 144
 145         tsk->state = TASK_RUNNING;
 146         return retval;
 147 }
 148
 149 /*
 150  * Trylock failed - make sure we correct for
 151  * having decremented the count.
 152  *
 153  * We could have done the trylock with a
 154  * single "cmpxchg" without failure cases,
 155  * but then it wouldn't work on a 386.
 156  */
 157 static fastcall int __attribute_used__ __down_trylock(struct semaphore * sem)
 158 {
 159         int sleepers;
 160         unsigned long flags;
 161
 162         spin_lock_irqsave(&sem->wait.lock, flags);
 163         sleepers = sem->sleepers + 1;
 164         sem->sleepers = 0;
 165
 166         /*
 167          * Add "everybody else" and us into it. They aren't
 168          * playing, because we own the spinlock in the
 169          * wait_queue_head.
 170          */
 171         if (!atomic_add_negative(sleepers, &sem->count)) {
 172                 wake_up_locked(&sem->wait);
 173         }
 174
 175         spin_unlock_irqrestore(&sem->wait.lock, flags);
 176         return 1;
 177 }
 178
 179
 180 /*
 181  * The semaphore operations have a special calling sequence that
 182  * allow us to do a simpler in-line version of them. These routines
 183  * need to convert that sequence back into the C sequence when
 184  * there is contention on the semaphore.
 185  *
 186  * %eax contains the semaphore pointer on entry. Save the C-clobbered
 187  * registers (%eax, %edx and %ecx) except %eax whish is either a return
 188  * value or just clobbered..
 189  */
 190 asm(
 191 ".section .sched.text\n"
 192 ".align 4\n"
 193 ".globl __down_failed\n"
 194 "__down_failed:\n\t"
 195 #if defined(CONFIG_FRAME_POINTER)
 196         "pushl %ebp\n\t"
 197         "movl  %esp,%ebp\n\t"
 198 #endif
 199         "pushl %edx\n\t"
 200         "pushl %ecx\n\t"
 201         "call __down\n\t"
 202         "popl %ecx\n\t"
 203         "popl %edx\n\t"
 204 #if defined(CONFIG_FRAME_POINTER)
 205         "movl %ebp,%esp\n\t"
 206         "popl %ebp\n\t"
 207 #endif
 208         "ret"
 209 );
 210
 211 asm(
 212 ".section .sched.text\n"
 213 ".align 4\n"
 214 ".globl __down_failed_interruptible\n"
 215 "__down_failed_interruptible:\n\t"
 216 #if defined(CONFIG_FRAME_POINTER)
 217         "pushl %ebp\n\t"
 218         "movl  %esp,%ebp\n\t"
 219 #endif
 220         "pushl %edx\n\t"
 221         "pushl %ecx\n\t"
 222         "call __down_interruptible\n\t"
 223         "popl %ecx\n\t"
 224         "popl %edx\n\t"
 225 #if defined(CONFIG_FRAME_POINTER)
 226         "movl %ebp,%esp\n\t"
 227         "popl %ebp\n\t"
 228 #endif
 229         "ret"
 230 );
 231
 232 asm(
 233 ".section .sched.text\n"
 234 ".align 4\n"
 235 ".globl __down_failed_trylock\n"
 236 "__down_failed_trylock:\n\t"
 237 #if defined(CONFIG_FRAME_POINTER)
 238         "pushl %ebp\n\t"
 239         "movl  %esp,%ebp\n\t"
 240 #endif
 241         "pushl %edx\n\t"
 242         "pushl %ecx\n\t"
 243         "call __down_trylock\n\t"
 244         "popl %ecx\n\t"
 245         "popl %edx\n\t"
 246 #if defined(CONFIG_FRAME_POINTER)
 247         "movl %ebp,%esp\n\t"
 248         "popl %ebp\n\t"
 249 #endif
 250         "ret"
 251 );
 252
 253 asm(
 254 ".section .sched.text\n"
 255 ".align 4\n"
 256 ".globl __up_wakeup\n"
 257 "__up_wakeup:\n\t"
 258         "pushl %edx\n\t"
 259         "pushl %ecx\n\t"
 260         "call __up\n\t"
 261         "popl %ecx\n\t"
 262         "popl %edx\n\t"
 263         "ret"
 264 );
 265
 266 /*
 267  * rw spinlock fallbacks
 268  */
 269 #if defined(CONFIG_SMP)
 270 asm(
 271 ".section .sched.text\n"
 272 ".align 4\n"
 273 ".globl __write_lock_failed\n"
 274 "__write_lock_failed:\n\t"
 275         LOCK "addl      $" RW_LOCK_BIAS_STR ",(%eax)\n"
 276 "1:     rep; nop\n\t"
 277         "cmpl   $" RW_LOCK_BIAS_STR ",(%eax)\n\t"
 278         "jne    1b\n\t"
 279         LOCK "subl      $" RW_LOCK_BIAS_STR ",(%eax)\n\t"
 280         "jnz    __write_lock_failed\n\t"
 281         "ret"
 282 );
 283
 284 asm(
 285 ".section .sched.text\n"
 286 ".align 4\n"
 287 ".globl __read_lock_failed\n"
 288 "__read_lock_failed:\n\t"
 289         LOCK "incl      (%eax)\n"
 290 "1:     rep; nop\n\t"
 291         "cmpl   $1,(%eax)\n\t"
 292         "js     1b\n\t"
 293         LOCK "decl      (%eax)\n\t"
 294         "js     __read_lock_failed\n\t"
 295         "ret"
 296 );
 297 #endif