arch/x86/include/asm/spinlock.h

   1 #ifndef _ASM_X86_SPINLOCK_H
   2 #define _ASM_X86_SPINLOCK_H
   3
   4 #include <linux/atomic.h>
   5 #include <asm/page.h>
   6 #include <asm/processor.h>
   7 #include <linux/compiler.h>
   8 #include <asm/paravirt.h>
   9 /*
  10  * Your basic SMP spinlocks, allowing only a single CPU anywhere
  11  *
  12  * Simple spin lock operations.  There are two variants, one clears IRQ's
  13  * on the local processor, one does not.
  14  *
  15  * These are fair FIFO ticket locks, which are currently limited to 256
  16  * CPUs.
  17  *
  18  * (the type definitions are in asm/spinlock_types.h)
  19  */
  20
  21 #ifdef CONFIG_X86_32
  22 # define LOCK_PTR_REG "a"
  23 # define REG_PTR_MODE "k"
  24 #else
  25 # define LOCK_PTR_REG "D"
  26 # define REG_PTR_MODE "q"
  27 #endif
  28
  29 #if defined(CONFIG_X86_32) && \
  30         (defined(CONFIG_X86_OOSTORE) || defined(CONFIG_X86_PPRO_FENCE))
  31 /*
  32  * On PPro SMP or if we are using OOSTORE, we use a locked operation to unlock
  33  * (PPro errata 66, 92)
  34  */
  35 # define UNLOCK_LOCK_PREFIX LOCK_PREFIX
  36 #else
  37 # define UNLOCK_LOCK_PREFIX
  38 #endif
  39
  40 /*
  41  * Ticket locks are conceptually two parts, one indicating the current head of
  42  * the queue, and the other indicating the current tail. The lock is acquired
  43  * by atomically noting the tail and incrementing it by one (thus adding
  44  * ourself to the queue and noting our position), then waiting until the head
  45  * becomes equal to the the initial value of the tail.
  46  *
  47  * We use an xadd covering *both* parts of the lock, to increment the tail and
  48  * also load the position of the head, which takes care of memory ordering
  49  * issues and should be optimal for the uncontended case. Note the tail must be
  50  * in the high part, because a wide xadd increment of the low part would carry
  51  * up and contaminate the high part.
  52  *
  53  * With fewer than 2^8 possible CPUs, we can use x86's partial registers to
  54  * save some instructions and make the code more elegant. There really isn't
  55  * much between them in performance though, especially as locks are out of line.
  56  */
  57 #if (NR_CPUS < 256)
  58 #define TICKET_SHIFT 8
  59
  60 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
  61 {
  62         short inc = 0x0100;
  63
  64         asm volatile (
  65                 LOCK_PREFIX "xaddw %w0, %1\n"
  66                 "1:\t"
  67                 "cmpb %h0, %b0\n\t"
  68                 "je 2f\n\t"
  69                 "rep ; nop\n\t"
  70                 "movb %1, %b0\n\t"
  71                 /* don't need lfence here, because loads are in-order */
  72                 "jmp 1b\n"
  73                 "2:"
  74                 : "+Q" (inc), "+m" (lock->slock)
  75                 :
  76                 : "memory", "cc");
  77 }
  78
  79 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
  80 {
  81         int tmp, new;
  82
  83         asm volatile("movzwl %2, %0\n\t"
  84                      "cmpb %h0,%b0\n\t"
  85                      "leal 0x100(%" REG_PTR_MODE "0), %1\n\t"
  86                      "jne 1f\n\t"
  87                      LOCK_PREFIX "cmpxchgw %w1,%2\n\t"
  88                      "1:"
  89                      "sete %b1\n\t"
  90                      "movzbl %b1,%0\n\t"
  91                      : "=&a" (tmp), "=&q" (new), "+m" (lock->slock)
  92                      :
  93                      : "memory", "cc");
  94
  95         return tmp;
  96 }
  97
  98 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
  99 {
 100         asm volatile(UNLOCK_LOCK_PREFIX "incb %0"
 101                      : "+m" (lock->slock)
 102                      :
 103                      : "memory", "cc");
 104 }
 105 #else
 106 #define TICKET_SHIFT 16
 107
 108 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
 109 {
 110         int inc = 0x00010000;
 111         int tmp;
 112
 113         asm volatile(LOCK_PREFIX "xaddl %0, %1\n"
 114                      "movzwl %w0, %2\n\t"
 115                      "shrl $16, %0\n\t"
 116                      "1:\t"
 117                      "cmpl %0, %2\n\t"
 118                      "je 2f\n\t"
 119                      "rep ; nop\n\t"
 120                      "movzwl %1, %2\n\t"
 121                      /* don't need lfence here, because loads are in-order */
 122                      "jmp 1b\n"
 123                      "2:"
 124                      : "+r" (inc), "+m" (lock->slock), "=&r" (tmp)
 125                      :
 126                      : "memory", "cc");
 127 }
 128
 129 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
 130 {
 131         int tmp;
 132         int new;
 133
 134         asm volatile("movl %2,%0\n\t"
 135                      "movl %0,%1\n\t"
 136                      "roll $16, %0\n\t"
 137                      "cmpl %0,%1\n\t"
 138                      "leal 0x00010000(%" REG_PTR_MODE "0), %1\n\t"
 139                      "jne 1f\n\t"
 140                      LOCK_PREFIX "cmpxchgl %1,%2\n\t"
 141                      "1:"
 142                      "sete %b1\n\t"
 143                      "movzbl %b1,%0\n\t"
 144                      : "=&a" (tmp), "=&q" (new), "+m" (lock->slock)
 145                      :
 146                      : "memory", "cc");
 147
 148         return tmp;
 149 }
 150
 151 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
 152 {
 153         asm volatile(UNLOCK_LOCK_PREFIX "incw %0"
 154                      : "+m" (lock->slock)
 155                      :
 156                      : "memory", "cc");
 157 }
 158 #endif
 159
 160 static inline int __ticket_spin_is_locked(arch_spinlock_t *lock)
 161 {
 162         int tmp = ACCESS_ONCE(lock->slock);
 163
 164         return !!(((tmp >> TICKET_SHIFT) ^ tmp) & ((1 << TICKET_SHIFT) - 1));
 165 }
 166
 167 static inline int __ticket_spin_is_contended(arch_spinlock_t *lock)
 168 {
 169         int tmp = ACCESS_ONCE(lock->slock);
 170
 171         return (((tmp >> TICKET_SHIFT) - tmp) & ((1 << TICKET_SHIFT) - 1)) > 1;
 172 }
 173
 174 #ifndef CONFIG_PARAVIRT_SPINLOCKS
 175
 176 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
 177 {
 178         return __ticket_spin_is_locked(lock);
 179 }
 180
 181 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 182 {
 183         return __ticket_spin_is_contended(lock);
 184 }
 185 #define arch_spin_is_contended  arch_spin_is_contended
 186
 187 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
 188 {
 189         __ticket_spin_lock(lock);
 190 }
 191
 192 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
 193 {
 194         return __ticket_spin_trylock(lock);
 195 }
 196
 197 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
 198 {
 199         __ticket_spin_unlock(lock);
 200 }
 201
 202 static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
 203                                                   unsigned long flags)
 204 {
 205         arch_spin_lock(lock);
 206 }
 207
 208 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */
 209
 210 static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
 211 {
 212         while (arch_spin_is_locked(lock))
 213                 cpu_relax();
 214 }
 215
 216 /*
 217  * Read-write spinlocks, allowing multiple readers
 218  * but only one writer.
 219  *
 220  * NOTE! it is quite common to have readers in interrupts
 221  * but no interrupt writers. For those circumstances we
 222  * can "mix" irq-safe locks - any writer needs to get a
 223  * irq-safe write-lock, but readers can get non-irqsafe
 224  * read-locks.
 225  *
 226  * On x86, we implement read-write locks as a 32-bit counter
 227  * with the high bit (sign) being the "contended" bit.
 228  */
 229
 230 /**
 231  * read_can_lock - would read_trylock() succeed?
 232  * @lock: the rwlock in question.
 233  */
 234 static inline int arch_read_can_lock(arch_rwlock_t *lock)
 235 {
 236         return lock->lock > 0;
 237 }
 238
 239 /**
 240  * write_can_lock - would write_trylock() succeed?
 241  * @lock: the rwlock in question.
 242  */
 243 static inline int arch_write_can_lock(arch_rwlock_t *lock)
 244 {
 245         return lock->write == WRITE_LOCK_CMP;
 246 }
 247
 248 static inline void arch_read_lock(arch_rwlock_t *rw)
 249 {
 250         asm volatile(LOCK_PREFIX READ_LOCK_SIZE(dec) " (%0)\n\t"
 251                      "jns 1f\n"
 252                      "call __read_lock_failed\n\t"
 253                      "1:\n"
 254                      ::LOCK_PTR_REG (rw) : "memory");
 255 }
 256
 257 static inline void arch_write_lock(arch_rwlock_t *rw)
 258 {
 259         asm volatile(LOCK_PREFIX WRITE_LOCK_SUB(%1) "(%0)\n\t"
 260                      "jz 1f\n"
 261                      "call __write_lock_failed\n\t"
 262                      "1:\n"
 263                      ::LOCK_PTR_REG (&rw->write), "i" (RW_LOCK_BIAS)
 264                      : "memory");
 265 }
 266
 267 static inline int arch_read_trylock(arch_rwlock_t *lock)
 268 {
 269         READ_LOCK_ATOMIC(t) *count = (READ_LOCK_ATOMIC(t) *)lock;
 270
 271         if (READ_LOCK_ATOMIC(dec_return)(count) >= 0)
 272                 return 1;
 273         READ_LOCK_ATOMIC(inc)(count);
 274         return 0;
 275 }
 276
 277 static inline int arch_write_trylock(arch_rwlock_t *lock)
 278 {
 279         atomic_t *count = (atomic_t *)&lock->write;
 280
 281         if (atomic_sub_and_test(WRITE_LOCK_CMP, count))
 282                 return 1;
 283         atomic_add(WRITE_LOCK_CMP, count);
 284         return 0;
 285 }
 286
 287 static inline void arch_read_unlock(arch_rwlock_t *rw)
 288 {
 289         asm volatile(LOCK_PREFIX READ_LOCK_SIZE(inc) " %0"
 290                      :"+m" (rw->lock) : : "memory");
 291 }
 292
 293 static inline void arch_write_unlock(arch_rwlock_t *rw)
 294 {
 295         asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
 296                      : "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
 297 }
 298
 299 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
 300 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
 301
 302 #undef READ_LOCK_SIZE
 303 #undef READ_LOCK_ATOMIC
 304 #undef WRITE_LOCK_ADD
 305 #undef WRITE_LOCK_SUB
 306 #undef WRITE_LOCK_CMP
 307
 308 #define arch_spin_relax(lock)   cpu_relax()
 309 #define arch_read_relax(lock)   cpu_relax()
 310 #define arch_write_relax(lock)  cpu_relax()
 311
 312 /* The {read|write|spin}_lock() on x86 are full memory barriers. */
 313 static inline void smp_mb__after_lock(void) { }
 314 #define ARCH_HAS_SMP_MB_AFTER_LOCK
 315
 316 #endif /* _ASM_X86_SPINLOCK_H */