include/asm-x86/spinlock.h

   1 #ifndef _X86_SPINLOCK_H_
   2 #define _X86_SPINLOCK_H_
   3
   4 #include <asm/atomic.h>
   5 #include <asm/rwlock.h>
   6 #include <asm/page.h>
   7 #include <asm/processor.h>
   8 #include <linux/compiler.h>
   9
  10 /*
  11  * Your basic SMP spinlocks, allowing only a single CPU anywhere
  12  *
  13  * Simple spin lock operations.  There are two variants, one clears IRQ's
  14  * on the local processor, one does not.
  15  *
  16  * These are fair FIFO ticket locks, which are currently limited to 256
  17  * CPUs.
  18  *
  19  * (the type definitions are in asm/spinlock_types.h)
  20  */
  21
  22 #ifdef CONFIG_X86_32
  23 typedef char _slock_t;
  24 # define LOCK_INS_DEC "decb"
  25 # define LOCK_INS_XCH "xchgb"
  26 # define LOCK_INS_MOV "movb"
  27 # define LOCK_INS_CMP "cmpb"
  28 # define LOCK_PTR_REG "a"
  29 #else
  30 typedef int _slock_t;
  31 # define LOCK_INS_DEC "decl"
  32 # define LOCK_INS_XCH "xchgl"
  33 # define LOCK_INS_MOV "movl"
  34 # define LOCK_INS_CMP "cmpl"
  35 # define LOCK_PTR_REG "D"
  36 #endif
  37
  38 #if defined(CONFIG_X86_32) && \
  39         (defined(CONFIG_X86_OOSTORE) || defined(CONFIG_X86_PPRO_FENCE))
  40 /*
  41  * On PPro SMP or if we are using OOSTORE, we use a locked operation to unlock
  42  * (PPro errata 66, 92)
  43  */
  44 # define UNLOCK_LOCK_PREFIX LOCK_PREFIX
  45 #else
  46 # define UNLOCK_LOCK_PREFIX
  47 #endif
  48
  49 /*
  50  * Ticket locks are conceptually two parts, one indicating the current head of
  51  * the queue, and the other indicating the current tail. The lock is acquired
  52  * by atomically noting the tail and incrementing it by one (thus adding
  53  * ourself to the queue and noting our position), then waiting until the head
  54  * becomes equal to the the initial value of the tail.
  55  *
  56  * We use an xadd covering *both* parts of the lock, to increment the tail and
  57  * also load the position of the head, which takes care of memory ordering
  58  * issues and should be optimal for the uncontended case. Note the tail must be
  59  * in the high part, because a wide xadd increment of the low part would carry
  60  * up and contaminate the high part.
  61  *
  62  * With fewer than 2^8 possible CPUs, we can use x86's partial registers to
  63  * save some instructions and make the code more elegant. There really isn't
  64  * much between them in performance though, especially as locks are out of line.
  65  */
  66 #if (NR_CPUS < 256)
  67 static inline int __raw_spin_is_locked(raw_spinlock_t *lock)
  68 {
  69         int tmp = *(volatile signed int *)(&(lock)->slock);
  70
  71         return (((tmp >> 8) & 0xff) != (tmp & 0xff));
  72 }
  73
  74 static inline int __raw_spin_is_contended(raw_spinlock_t *lock)
  75 {
  76         int tmp = *(volatile signed int *)(&(lock)->slock);
  77
  78         return (((tmp >> 8) & 0xff) - (tmp & 0xff)) > 1;
  79 }
  80
  81 static __always_inline void __raw_spin_lock(raw_spinlock_t *lock)
  82 {
  83         short inc = 0x0100;
  84
  85         asm volatile (
  86                 LOCK_PREFIX "xaddw %w0, %1\n"
  87                 "1:\t"
  88                 "cmpb %h0, %b0\n\t"
  89                 "je 2f\n\t"
  90                 "rep ; nop\n\t"
  91                 "movb %1, %b0\n\t"
  92                 /* don't need lfence here, because loads are in-order */
  93                 "jmp 1b\n"
  94                 "2:"
  95                 : "+Q" (inc), "+m" (lock->slock)
  96                 :
  97                 : "memory", "cc");
  98 }
  99
 100 #define __raw_spin_lock_flags(lock, flags) __raw_spin_lock(lock)
 101
 102 static __always_inline int __raw_spin_trylock(raw_spinlock_t *lock)
 103 {
 104         int tmp;
 105         short new;
 106
 107         asm volatile("movw %2,%w0\n\t"
 108                      "cmpb %h0,%b0\n\t"
 109                      "jne 1f\n\t"
 110                      "movw %w0,%w1\n\t"
 111                      "incb %h1\n\t"
 112                      "lock ; cmpxchgw %w1,%2\n\t"
 113                      "1:"
 114                      "sete %b1\n\t"
 115                      "movzbl %b1,%0\n\t"
 116                      : "=&a" (tmp), "=Q" (new), "+m" (lock->slock)
 117                      :
 118                      : "memory", "cc");
 119
 120         return tmp;
 121 }
 122
 123 static __always_inline void __raw_spin_unlock(raw_spinlock_t *lock)
 124 {
 125         asm volatile(UNLOCK_LOCK_PREFIX "incb %0"
 126                      : "+m" (lock->slock)
 127                      :
 128                      : "memory", "cc");
 129 }
 130 #else
 131 static inline int __raw_spin_is_locked(raw_spinlock_t *lock)
 132 {
 133         int tmp = *(volatile signed int *)(&(lock)->slock);
 134
 135         return (((tmp >> 16) & 0xffff) != (tmp & 0xffff));
 136 }
 137
 138 static inline int __raw_spin_is_contended(raw_spinlock_t *lock)
 139 {
 140         int tmp = *(volatile signed int *)(&(lock)->slock);
 141
 142         return (((tmp >> 16) & 0xffff) - (tmp & 0xffff)) > 1;
 143 }
 144
 145 static __always_inline void __raw_spin_lock(raw_spinlock_t *lock)
 146 {
 147         int inc = 0x00010000;
 148         int tmp;
 149
 150         asm volatile("lock ; xaddl %0, %1\n"
 151                      "movzwl %w0, %2\n\t"
 152                      "shrl $16, %0\n\t"
 153                      "1:\t"
 154                      "cmpl %0, %2\n\t"
 155                      "je 2f\n\t"
 156                      "rep ; nop\n\t"
 157                      "movzwl %1, %2\n\t"
 158                      /* don't need lfence here, because loads are in-order */
 159                      "jmp 1b\n"
 160                      "2:"
 161                      : "+Q" (inc), "+m" (lock->slock), "=r" (tmp)
 162                      :
 163                      : "memory", "cc");
 164 }
 165
 166 #define __raw_spin_lock_flags(lock, flags) __raw_spin_lock(lock)
 167
 168 static __always_inline int __raw_spin_trylock(raw_spinlock_t *lock)
 169 {
 170         int tmp;
 171         int new;
 172
 173         asm volatile("movl %2,%0\n\t"
 174                      "movl %0,%1\n\t"
 175                      "roll $16, %0\n\t"
 176                      "cmpl %0,%1\n\t"
 177                      "jne 1f\n\t"
 178                      "addl $0x00010000, %1\n\t"
 179                      "lock ; cmpxchgl %1,%2\n\t"
 180                      "1:"
 181                      "sete %b1\n\t"
 182                      "movzbl %b1,%0\n\t"
 183                      : "=&a" (tmp), "=r" (new), "+m" (lock->slock)
 184                      :
 185                      : "memory", "cc");
 186
 187         return tmp;
 188 }
 189
 190 static __always_inline void __raw_spin_unlock(raw_spinlock_t *lock)
 191 {
 192         asm volatile(UNLOCK_LOCK_PREFIX "incw %0"
 193                      : "+m" (lock->slock)
 194                      :
 195                      : "memory", "cc");
 196 }
 197 #endif
 198
 199 static inline void __raw_spin_unlock_wait(raw_spinlock_t *lock)
 200 {
 201         while (__raw_spin_is_locked(lock))
 202                 cpu_relax();
 203 }
 204
 205 /*
 206  * Read-write spinlocks, allowing multiple readers
 207  * but only one writer.
 208  *
 209  * NOTE! it is quite common to have readers in interrupts
 210  * but no interrupt writers. For those circumstances we
 211  * can "mix" irq-safe locks - any writer needs to get a
 212  * irq-safe write-lock, but readers can get non-irqsafe
 213  * read-locks.
 214  *
 215  * On x86, we implement read-write locks as a 32-bit counter
 216  * with the high bit (sign) being the "contended" bit.
 217  */
 218
 219 /**
 220  * read_can_lock - would read_trylock() succeed?
 221  * @lock: the rwlock in question.
 222  */
 223 static inline int __raw_read_can_lock(raw_rwlock_t *lock)
 224 {
 225         return (int)(lock)->lock > 0;
 226 }
 227
 228 /**
 229  * write_can_lock - would write_trylock() succeed?
 230  * @lock: the rwlock in question.
 231  */
 232 static inline int __raw_write_can_lock(raw_rwlock_t *lock)
 233 {
 234         return (lock)->lock == RW_LOCK_BIAS;
 235 }
 236
 237 static inline void __raw_read_lock(raw_rwlock_t *rw)
 238 {
 239         asm volatile(LOCK_PREFIX " subl $1,(%0)\n\t"
 240                      "jns 1f\n"
 241                      "call __read_lock_failed\n\t"
 242                      "1:\n"
 243                      ::LOCK_PTR_REG (rw) : "memory");
 244 }
 245
 246 static inline void __raw_write_lock(raw_rwlock_t *rw)
 247 {
 248         asm volatile(LOCK_PREFIX " subl %1,(%0)\n\t"
 249                      "jz 1f\n"
 250                      "call __write_lock_failed\n\t"
 251                      "1:\n"
 252                      ::LOCK_PTR_REG (rw), "i" (RW_LOCK_BIAS) : "memory");
 253 }
 254
 255 static inline int __raw_read_trylock(raw_rwlock_t *lock)
 256 {
 257         atomic_t *count = (atomic_t *)lock;
 258
 259         atomic_dec(count);
 260         if (atomic_read(count) >= 0)
 261                 return 1;
 262         atomic_inc(count);
 263         return 0;
 264 }
 265
 266 static inline int __raw_write_trylock(raw_rwlock_t *lock)
 267 {
 268         atomic_t *count = (atomic_t *)lock;
 269
 270         if (atomic_sub_and_test(RW_LOCK_BIAS, count))
 271                 return 1;
 272         atomic_add(RW_LOCK_BIAS, count);
 273         return 0;
 274 }
 275
 276 static inline void __raw_read_unlock(raw_rwlock_t *rw)
 277 {
 278         asm volatile(LOCK_PREFIX "incl %0" :"+m" (rw->lock) : : "memory");
 279 }
 280
 281 static inline void __raw_write_unlock(raw_rwlock_t *rw)
 282 {
 283         asm volatile(LOCK_PREFIX "addl %1, %0"
 284                      : "+m" (rw->lock) : "i" (RW_LOCK_BIAS) : "memory");
 285 }
 286
 287 #define _raw_spin_relax(lock)   cpu_relax()
 288 #define _raw_read_relax(lock)   cpu_relax()
 289 #define _raw_write_relax(lock)  cpu_relax()
 290
 291 #endif