arch/ia64/include/asm/spinlock.h

   1 #ifndef _ASM_IA64_SPINLOCK_H
   2 #define _ASM_IA64_SPINLOCK_H
   3
   4 /*
   5  * Copyright (C) 1998-2003 Hewlett-Packard Co
   6  *      David Mosberger-Tang <davidm@hpl.hp.com>
   7  * Copyright (C) 1999 Walt Drummond <drummond@valinux.com>
   8  *
   9  * This file is used for SMP configurations only.
  10  */
  11
  12 #include <linux/compiler.h>
  13 #include <linux/kernel.h>
  14 #include <linux/bitops.h>
  15
  16 #include <linux/atomic.h>
  17 #include <asm/intrinsics.h>
  18 #include <asm/barrier.h>
  19 #include <asm/processor.h>
  20
  21 #define arch_spin_lock_init(x)                  ((x)->lock = 0)
  22
  23 /*
  24  * Ticket locks are conceptually two parts, one indicating the current head of
  25  * the queue, and the other indicating the current tail. The lock is acquired
  26  * by atomically noting the tail and incrementing it by one (thus adding
  27  * ourself to the queue and noting our position), then waiting until the head
  28  * becomes equal to the the initial value of the tail.
  29  * The pad bits in the middle are used to prevent the next_ticket number
  30  * overflowing into the now_serving number.
  31  *
  32  *   31             17  16    15  14                    0
  33  *  +----------------------------------------------------+
  34  *  |  now_serving     | padding |   next_ticket         |
  35  *  +----------------------------------------------------+
  36  */
  37
  38 #define TICKET_SHIFT    17
  39 #define TICKET_BITS     15
  40 #define TICKET_MASK     ((1 << TICKET_BITS) - 1)
  41
  42 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
  43 {
  44         int     *p = (int *)&lock->lock, ticket, serve;
  45
  46         ticket = ia64_fetchadd(1, p, acq);
  47
  48         if (!(((ticket >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
  49                 return;
  50
  51         ia64_invala();
  52
  53         for (;;) {
  54                 asm volatile ("ld4.c.nc %0=[%1]" : "=r"(serve) : "r"(p) : "memory");
  55
  56                 if (!(((serve >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
  57                         return;
  58                 cpu_relax();
  59         }
  60 }
  61
  62 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
  63 {
  64         int tmp = ACCESS_ONCE(lock->lock);
  65
  66         if (!(((tmp >> TICKET_SHIFT) ^ tmp) & TICKET_MASK))
  67                 return ia64_cmpxchg(acq, &lock->lock, tmp, tmp + 1, sizeof (tmp)) == tmp;
  68         return 0;
  69 }
  70
  71 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
  72 {
  73         unsigned short  *p = (unsigned short *)&lock->lock + 1, tmp;
  74
  75         asm volatile ("ld2.bias %0=[%1]" : "=r"(tmp) : "r"(p));
  76         ACCESS_ONCE(*p) = (tmp + 2) & ~1;
  77 }
  78
  79 static inline int __ticket_spin_is_locked(arch_spinlock_t *lock)
  80 {
  81         long tmp = ACCESS_ONCE(lock->lock);
  82
  83         return !!(((tmp >> TICKET_SHIFT) ^ tmp) & TICKET_MASK);
  84 }
  85
  86 static inline int __ticket_spin_is_contended(arch_spinlock_t *lock)
  87 {
  88         long tmp = ACCESS_ONCE(lock->lock);
  89
  90         return ((tmp - (tmp >> TICKET_SHIFT)) & TICKET_MASK) > 1;
  91 }
  92
  93 static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
  94 {
  95         return !(((lock.lock >> TICKET_SHIFT) ^ lock.lock) & TICKET_MASK);
  96 }
  97
  98 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
  99 {
 100         return __ticket_spin_is_locked(lock);
 101 }
 102
 103 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 104 {
 105         return __ticket_spin_is_contended(lock);
 106 }
 107 #define arch_spin_is_contended  arch_spin_is_contended
 108
 109 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
 110 {
 111         __ticket_spin_lock(lock);
 112 }
 113
 114 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
 115 {
 116         return __ticket_spin_trylock(lock);
 117 }
 118
 119 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
 120 {
 121         __ticket_spin_unlock(lock);
 122 }
 123
 124 static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
 125                                                   unsigned long flags)
 126 {
 127         arch_spin_lock(lock);
 128 }
 129
 130 #define arch_read_can_lock(rw)          (*(volatile int *)(rw) >= 0)
 131 #define arch_write_can_lock(rw) (*(volatile int *)(rw) == 0)
 132
 133 #ifdef ASM_SUPPORTED
 134
 135 static __always_inline void
 136 arch_read_lock_flags(arch_rwlock_t *lock, unsigned long flags)
 137 {
 138         __asm__ __volatile__ (
 139                 "tbit.nz p6, p0 = %1,%2\n"
 140                 "br.few 3f\n"
 141                 "1:\n"
 142                 "fetchadd4.rel r2 = [%0], -1;;\n"
 143                 "(p6) ssm psr.i\n"
 144                 "2:\n"
 145                 "hint @pause\n"
 146                 "ld4 r2 = [%0];;\n"
 147                 "cmp4.lt p7,p0 = r2, r0\n"
 148                 "(p7) br.cond.spnt.few 2b\n"
 149                 "(p6) rsm psr.i\n"
 150                 ";;\n"
 151                 "3:\n"
 152                 "fetchadd4.acq r2 = [%0], 1;;\n"
 153                 "cmp4.lt p7,p0 = r2, r0\n"
 154                 "(p7) br.cond.spnt.few 1b\n"
 155                 : : "r"(lock), "r"(flags), "i"(IA64_PSR_I_BIT)
 156                 : "p6", "p7", "r2", "memory");
 157 }
 158
 159 #define arch_read_lock(lock) arch_read_lock_flags(lock, 0)
 160
 161 #else /* !ASM_SUPPORTED */
 162
 163 #define arch_read_lock_flags(rw, flags) arch_read_lock(rw)
 164
 165 #define arch_read_lock(rw)                                                              \
 166 do {                                                                                    \
 167         arch_rwlock_t *__read_lock_ptr = (rw);                                          \
 168                                                                                         \
 169         while (unlikely(ia64_fetchadd(1, (int *) __read_lock_ptr, acq) < 0)) {          \
 170                 ia64_fetchadd(-1, (int *) __read_lock_ptr, rel);                        \
 171                 while (*(volatile int *)__read_lock_ptr < 0)                            \
 172                         cpu_relax();                                                    \
 173         }                                                                               \
 174 } while (0)
 175
 176 #endif /* !ASM_SUPPORTED */
 177
 178 #define arch_read_unlock(rw)                                    \
 179 do {                                                            \
 180         arch_rwlock_t *__read_lock_ptr = (rw);                  \
 181         ia64_fetchadd(-1, (int *) __read_lock_ptr, rel);        \
 182 } while (0)
 183
 184 #ifdef ASM_SUPPORTED
 185
 186 static __always_inline void
 187 arch_write_lock_flags(arch_rwlock_t *lock, unsigned long flags)
 188 {
 189         __asm__ __volatile__ (
 190                 "tbit.nz p6, p0 = %1, %2\n"
 191                 "mov ar.ccv = r0\n"
 192                 "dep r29 = -1, r0, 31, 1\n"
 193                 "br.few 3f;;\n"
 194                 "1:\n"
 195                 "(p6) ssm psr.i\n"
 196                 "2:\n"
 197                 "hint @pause\n"
 198                 "ld4 r2 = [%0];;\n"
 199                 "cmp4.eq p0,p7 = r0, r2\n"
 200                 "(p7) br.cond.spnt.few 2b\n"
 201                 "(p6) rsm psr.i\n"
 202                 ";;\n"
 203                 "3:\n"
 204                 "cmpxchg4.acq r2 = [%0], r29, ar.ccv;;\n"
 205                 "cmp4.eq p0,p7 = r0, r2\n"
 206                 "(p7) br.cond.spnt.few 1b;;\n"
 207                 : : "r"(lock), "r"(flags), "i"(IA64_PSR_I_BIT)
 208                 : "ar.ccv", "p6", "p7", "r2", "r29", "memory");
 209 }
 210
 211 #define arch_write_lock(rw) arch_write_lock_flags(rw, 0)
 212
 213 #define arch_write_trylock(rw)                                                  \
 214 ({                                                                              \
 215         register long result;                                                   \
 216                                                                                 \
 217         __asm__ __volatile__ (                                                  \
 218                 "mov ar.ccv = r0\n"                                             \
 219                 "dep r29 = -1, r0, 31, 1;;\n"                                   \
 220                 "cmpxchg4.acq %0 = [%1], r29, ar.ccv\n"                         \
 221                 : "=r"(result) : "r"(rw) : "ar.ccv", "r29", "memory");          \
 222         (result == 0);                                                          \
 223 })
 224
 225 static inline void arch_write_unlock(arch_rwlock_t *x)
 226 {
 227         u8 *y = (u8 *)x;
 228         barrier();
 229         asm volatile ("st1.rel.nta [%0] = r0\n\t" :: "r"(y+3) : "memory" );
 230 }
 231
 232 #else /* !ASM_SUPPORTED */
 233
 234 #define arch_write_lock_flags(l, flags) arch_write_lock(l)
 235
 236 #define arch_write_lock(l)                                                              \
 237 ({                                                                                      \
 238         __u64 ia64_val, ia64_set_val = ia64_dep_mi(-1, 0, 31, 1);                       \
 239         __u32 *ia64_write_lock_ptr = (__u32 *) (l);                                     \
 240         do {                                                                            \
 241                 while (*ia64_write_lock_ptr)                                            \
 242                         ia64_barrier();                                                 \
 243                 ia64_val = ia64_cmpxchg4_acq(ia64_write_lock_ptr, ia64_set_val, 0);     \
 244         } while (ia64_val);                                                             \
 245 })
 246
 247 #define arch_write_trylock(rw)                                          \
 248 ({                                                                      \
 249         __u64 ia64_val;                                                 \
 250         __u64 ia64_set_val = ia64_dep_mi(-1, 0, 31,1);                  \
 251         ia64_val = ia64_cmpxchg4_acq((__u32 *)(rw), ia64_set_val, 0);   \
 252         (ia64_val == 0);                                                \
 253 })
 254
 255 static inline void arch_write_unlock(arch_rwlock_t *x)
 256 {
 257         barrier();
 258         x->write_lock = 0;
 259 }
 260
 261 #endif /* !ASM_SUPPORTED */
 262
 263 static inline int arch_read_trylock(arch_rwlock_t *x)
 264 {
 265         union {
 266                 arch_rwlock_t lock;
 267                 __u32 word;
 268         } old, new;
 269         old.lock = new.lock = *x;
 270         old.lock.write_lock = new.lock.write_lock = 0;
 271         ++new.lock.read_counter;
 272         return (u32)ia64_cmpxchg4_acq((__u32 *)(x), new.word, old.word) == old.word;
 273 }
 274
 275 #define arch_spin_relax(lock)   cpu_relax()
 276 #define arch_read_relax(lock)   cpu_relax()
 277 #define arch_write_relax(lock)  cpu_relax()
 278
 279 #endif /*  _ASM_IA64_SPINLOCK_H */