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 <[email protected]>
   7  * Copyright (C) 1999 Walt Drummond <[email protected]>
   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
  19 #define arch_spin_lock_init(x)                  ((x)->lock = 0)
  20
  21 /*
  22  * Ticket locks are conceptually two parts, one indicating the current head of
  23  * the queue, and the other indicating the current tail. The lock is acquired
  24  * by atomically noting the tail and incrementing it by one (thus adding
  25  * ourself to the queue and noting our position), then waiting until the head
  26  * becomes equal to the the initial value of the tail.
  27  * The pad bits in the middle are used to prevent the next_ticket number
  28  * overflowing into the now_serving number.
  29  *
  30  *   31             17  16    15  14                    0
  31  *  +----------------------------------------------------+
  32  *  |  now_serving     | padding |   next_ticket         |
  33  *  +----------------------------------------------------+
  34  */
  35
  36 #define TICKET_SHIFT    17
  37 #define TICKET_BITS     15
  38 #define TICKET_MASK     ((1 << TICKET_BITS) - 1)
  39
  40 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
  41 {
  42         int     *p = (int *)&lock->lock, ticket, serve;
  43
  44         ticket = ia64_fetchadd(1, p, acq);
  45
  46         if (!(((ticket >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
  47                 return;
  48
  49         ia64_invala();
  50
  51         for (;;) {
  52                 asm volatile ("ld4.c.nc %0=[%1]" : "=r"(serve) : "r"(p) : "memory");
  53
  54                 if (!(((serve >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
  55                         return;
  56                 cpu_relax();
  57         }
  58 }
  59
  60 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
  61 {
  62         int tmp = ACCESS_ONCE(lock->lock);
  63
  64         if (!(((tmp >> TICKET_SHIFT) ^ tmp) & TICKET_MASK))
  65                 return ia64_cmpxchg(acq, &lock->lock, tmp, tmp + 1, sizeof (tmp)) == tmp;
  66         return 0;
  67 }
  68
  69 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
  70 {
  71         unsigned short  *p = (unsigned short *)&lock->lock + 1, tmp;
  72
  73         asm volatile ("ld2.bias %0=[%1]" : "=r"(tmp) : "r"(p));
  74         ACCESS_ONCE(*p) = (tmp + 2) & ~1;
  75 }
  76
  77 static __always_inline void __ticket_spin_unlock_wait(arch_spinlock_t *lock)
  78 {
  79         int     *p = (int *)&lock->lock, ticket;
  80
  81         ia64_invala();
  82
  83         for (;;) {
  84                 asm volatile ("ld4.c.nc %0=[%1]" : "=r"(ticket) : "r"(p) : "memory");
  85                 if (!(((ticket >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
  86                         return;
  87                 cpu_relax();
  88         }
  89 }
  90
  91 static inline int __ticket_spin_is_locked(arch_spinlock_t *lock)
  92 {
  93         long tmp = ACCESS_ONCE(lock->lock);
  94
  95         return !!(((tmp >> TICKET_SHIFT) ^ tmp) & TICKET_MASK);
  96 }
  97
  98 static inline int __ticket_spin_is_contended(arch_spinlock_t *lock)
  99 {
 100         long tmp = ACCESS_ONCE(lock->lock);
 101
 102         return ((tmp - (tmp >> TICKET_SHIFT)) & TICKET_MASK) > 1;
 103 }
 104
 105 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
 106 {
 107         return __ticket_spin_is_locked(lock);
 108 }
 109
 110 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 111 {
 112         return __ticket_spin_is_contended(lock);
 113 }
 114 #define arch_spin_is_contended  arch_spin_is_contended
 115
 116 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
 117 {
 118         __ticket_spin_lock(lock);
 119 }
 120
 121 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
 122 {
 123         return __ticket_spin_trylock(lock);
 124 }
 125
 126 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
 127 {
 128         __ticket_spin_unlock(lock);
 129 }
 130
 131 static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
 132                                                   unsigned long flags)
 133 {
 134         arch_spin_lock(lock);
 135 }
 136
 137 static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
 138 {
 139         __ticket_spin_unlock_wait(lock);
 140 }
 141
 142 #define arch_read_can_lock(rw)          (*(volatile int *)(rw) >= 0)
 143 #define arch_write_can_lock(rw) (*(volatile int *)(rw) == 0)
 144
 145 #ifdef ASM_SUPPORTED
 146
 147 static __always_inline void
 148 arch_read_lock_flags(arch_rwlock_t *lock, unsigned long flags)
 149 {
 150         __asm__ __volatile__ (
 151                 "tbit.nz p6, p0 = %1,%2\n"
 152                 "br.few 3f\n"
 153                 "1:\n"
 154                 "fetchadd4.rel r2 = [%0], -1;;\n"
 155                 "(p6) ssm psr.i\n"
 156                 "2:\n"
 157                 "hint @pause\n"
 158                 "ld4 r2 = [%0];;\n"
 159                 "cmp4.lt p7,p0 = r2, r0\n"
 160                 "(p7) br.cond.spnt.few 2b\n"
 161                 "(p6) rsm psr.i\n"
 162                 ";;\n"
 163                 "3:\n"
 164                 "fetchadd4.acq r2 = [%0], 1;;\n"
 165                 "cmp4.lt p7,p0 = r2, r0\n"
 166                 "(p7) br.cond.spnt.few 1b\n"
 167                 : : "r"(lock), "r"(flags), "i"(IA64_PSR_I_BIT)
 168                 : "p6", "p7", "r2", "memory");
 169 }
 170
 171 #define arch_read_lock(lock) arch_read_lock_flags(lock, 0)
 172
 173 #else /* !ASM_SUPPORTED */
 174
 175 #define arch_read_lock_flags(rw, flags) arch_read_lock(rw)
 176
 177 #define arch_read_lock(rw)                                                              \
 178 do {                                                                                    \
 179         arch_rwlock_t *__read_lock_ptr = (rw);                                          \
 180                                                                                         \
 181         while (unlikely(ia64_fetchadd(1, (int *) __read_lock_ptr, acq) < 0)) {          \
 182                 ia64_fetchadd(-1, (int *) __read_lock_ptr, rel);                        \
 183                 while (*(volatile int *)__read_lock_ptr < 0)                            \
 184                         cpu_relax();                                                    \
 185         }                                                                               \
 186 } while (0)
 187
 188 #endif /* !ASM_SUPPORTED */
 189
 190 #define arch_read_unlock(rw)                                    \
 191 do {                                                            \
 192         arch_rwlock_t *__read_lock_ptr = (rw);                  \
 193         ia64_fetchadd(-1, (int *) __read_lock_ptr, rel);        \
 194 } while (0)
 195
 196 #ifdef ASM_SUPPORTED
 197
 198 static __always_inline void
 199 arch_write_lock_flags(arch_rwlock_t *lock, unsigned long flags)
 200 {
 201         __asm__ __volatile__ (
 202                 "tbit.nz p6, p0 = %1, %2\n"
 203                 "mov ar.ccv = r0\n"
 204                 "dep r29 = -1, r0, 31, 1\n"
 205                 "br.few 3f;;\n"
 206                 "1:\n"
 207                 "(p6) ssm psr.i\n"
 208                 "2:\n"
 209                 "hint @pause\n"
 210                 "ld4 r2 = [%0];;\n"
 211                 "cmp4.eq p0,p7 = r0, r2\n"
 212                 "(p7) br.cond.spnt.few 2b\n"
 213                 "(p6) rsm psr.i\n"
 214                 ";;\n"
 215                 "3:\n"
 216                 "cmpxchg4.acq r2 = [%0], r29, ar.ccv;;\n"
 217                 "cmp4.eq p0,p7 = r0, r2\n"
 218                 "(p7) br.cond.spnt.few 1b;;\n"
 219                 : : "r"(lock), "r"(flags), "i"(IA64_PSR_I_BIT)
 220                 : "ar.ccv", "p6", "p7", "r2", "r29", "memory");
 221 }
 222
 223 #define arch_write_lock(rw) arch_write_lock_flags(rw, 0)
 224
 225 #define arch_write_trylock(rw)                                                  \
 226 ({                                                                              \
 227         register long result;                                                   \
 228                                                                                 \
 229         __asm__ __volatile__ (                                                  \
 230                 "mov ar.ccv = r0\n"                                             \
 231                 "dep r29 = -1, r0, 31, 1;;\n"                                   \
 232                 "cmpxchg4.acq %0 = [%1], r29, ar.ccv\n"                         \
 233                 : "=r"(result) : "r"(rw) : "ar.ccv", "r29", "memory");          \
 234         (result == 0);                                                          \
 235 })
 236
 237 static inline void arch_write_unlock(arch_rwlock_t *x)
 238 {
 239         u8 *y = (u8 *)x;
 240         barrier();
 241         asm volatile ("st1.rel.nta [%0] = r0\n\t" :: "r"(y+3) : "memory" );
 242 }
 243
 244 #else /* !ASM_SUPPORTED */
 245
 246 #define arch_write_lock_flags(l, flags) arch_write_lock(l)
 247
 248 #define arch_write_lock(l)                                                              \
 249 ({                                                                                      \
 250         __u64 ia64_val, ia64_set_val = ia64_dep_mi(-1, 0, 31, 1);                       \
 251         __u32 *ia64_write_lock_ptr = (__u32 *) (l);                                     \
 252         do {                                                                            \
 253                 while (*ia64_write_lock_ptr)                                            \
 254                         ia64_barrier();                                                 \
 255                 ia64_val = ia64_cmpxchg4_acq(ia64_write_lock_ptr, ia64_set_val, 0);     \
 256         } while (ia64_val);                                                             \
 257 })
 258
 259 #define arch_write_trylock(rw)                                          \
 260 ({                                                                      \
 261         __u64 ia64_val;                                                 \
 262         __u64 ia64_set_val = ia64_dep_mi(-1, 0, 31,1);                  \
 263         ia64_val = ia64_cmpxchg4_acq((__u32 *)(rw), ia64_set_val, 0);   \
 264         (ia64_val == 0);                                                \
 265 })
 266
 267 static inline void arch_write_unlock(arch_rwlock_t *x)
 268 {
 269         barrier();
 270         x->write_lock = 0;
 271 }
 272
 273 #endif /* !ASM_SUPPORTED */
 274
 275 static inline int arch_read_trylock(arch_rwlock_t *x)
 276 {
 277         union {
 278                 arch_rwlock_t lock;
 279                 __u32 word;
 280         } old, new;
 281         old.lock = new.lock = *x;
 282         old.lock.write_lock = new.lock.write_lock = 0;
 283         ++new.lock.read_counter;
 284         return (u32)ia64_cmpxchg4_acq((__u32 *)(x), new.word, old.word) == old.word;
 285 }
 286
 287 #define arch_spin_relax(lock)   cpu_relax()
 288 #define arch_read_relax(lock)   cpu_relax()
 289 #define arch_write_relax(lock)  cpu_relax()
 290
 291 #endif /*  _ASM_IA64_SPINLOCK_H */