1 #ifndef __ASM_CMPXCHG_H
2 #define __ASM_CMPXCHG_H
4 #include <linux/bitops.h> /* for LOCK_PREFIX */
7 * Note: if you use set64_bit(), __cmpxchg64(), or their variants, you
8 * you need to test for the feature in boot_cpu_data.
11 #define xchg(ptr,v) ((__typeof__(*(ptr)))__xchg((unsigned long)(v),(ptr),sizeof(*(ptr))))
13 struct __xchg_dummy { unsigned long a[100]; };
14 #define __xg(x) ((struct __xchg_dummy *)(x))
17 * The semantics of XCHGCMP8B are a bit strange, this is why
18 * there is a loop and the loading of %%eax and %%edx has to
19 * be inside. This inlines well in most cases, the cached
20 * cost is around ~38 cycles. (in the future we might want
21 * to do an SIMD/3DNOW!/MMX/FPU 64-bit store here, but that
22 * might have an implicit FPU-save as a cost, so it's not
23 * clear which path to go.)
25 * cmpxchg8b must be used with the lock prefix here to allow
26 * the instruction to be executed atomically, see page 3-102
27 * of the instruction set reference 24319102.pdf. We need
28 * the reader side to see the coherent 64bit value.
30 static inline void __set_64bit (unsigned long long * ptr,
31 unsigned int low, unsigned int high)
33 __asm__ __volatile__ (
35 "movl (%0), %%eax\n\t"
36 "movl 4(%0), %%edx\n\t"
37 LOCK_PREFIX "cmpxchg8b (%0)\n\t"
43 : "ax","dx","memory");
46 static inline void __set_64bit_constant (unsigned long long *ptr,
47 unsigned long long value)
49 __set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
51 #define ll_low(x) *(((unsigned int*)&(x))+0)
52 #define ll_high(x) *(((unsigned int*)&(x))+1)
54 static inline void __set_64bit_var (unsigned long long *ptr,
55 unsigned long long value)
57 __set_64bit(ptr,ll_low(value), ll_high(value));
60 #define set_64bit(ptr,value) \
61 (__builtin_constant_p(value) ? \
62 __set_64bit_constant(ptr, value) : \
63 __set_64bit_var(ptr, value) )
65 #define _set_64bit(ptr,value) \
66 (__builtin_constant_p(value) ? \
67 __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
68 __set_64bit(ptr, ll_low(value), ll_high(value)) )
71 * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
72 * Note 2: xchg has side effect, so that attribute volatile is necessary,
73 * but generally the primitive is invalid, *ptr is output argument. --ANK
75 static inline unsigned long __xchg(unsigned long x, volatile void * ptr, int size)
79 __asm__ __volatile__("xchgb %b0,%1"
81 :"m" (*__xg(ptr)), "0" (x)
85 __asm__ __volatile__("xchgw %w0,%1"
87 :"m" (*__xg(ptr)), "0" (x)
91 __asm__ __volatile__("xchgl %0,%1"
93 :"m" (*__xg(ptr)), "0" (x)
101 * Atomic compare and exchange. Compare OLD with MEM, if identical,
102 * store NEW in MEM. Return the initial value in MEM. Success is
103 * indicated by comparing RETURN with OLD.
106 #ifdef CONFIG_X86_CMPXCHG
107 #define __HAVE_ARCH_CMPXCHG 1
108 #define cmpxchg(ptr, o, n) \
109 ((__typeof__(*(ptr)))__cmpxchg((ptr), (unsigned long)(o), \
110 (unsigned long)(n), sizeof(*(ptr))))
111 #define sync_cmpxchg(ptr, o, n) \
112 ((__typeof__(*(ptr)))__sync_cmpxchg((ptr), (unsigned long)(o), \
113 (unsigned long)(n), sizeof(*(ptr))))
114 #define cmpxchg_local(ptr, o, n) \
115 ((__typeof__(*(ptr)))__cmpxchg_local((ptr), (unsigned long)(o), \
116 (unsigned long)(n), sizeof(*(ptr))))
119 #ifdef CONFIG_X86_CMPXCHG64
120 #define cmpxchg64(ptr, o, n) \
121 ((__typeof__(*(ptr)))__cmpxchg64((ptr), (unsigned long long)(o), \
122 (unsigned long long)(n)))
123 #define cmpxchg64_local(ptr, o, n) \
124 ((__typeof__(*(ptr)))__cmpxchg64_local((ptr), (unsigned long long)(o),\
125 (unsigned long long)(n)))
128 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
129 unsigned long new, int size)
134 __asm__ __volatile__(LOCK_PREFIX "cmpxchgb %b1,%2"
136 : "q"(new), "m"(*__xg(ptr)), "0"(old)
140 __asm__ __volatile__(LOCK_PREFIX "cmpxchgw %w1,%2"
142 : "r"(new), "m"(*__xg(ptr)), "0"(old)
146 __asm__ __volatile__(LOCK_PREFIX "cmpxchgl %1,%2"
148 : "r"(new), "m"(*__xg(ptr)), "0"(old)
156 * Always use locked operations when touching memory shared with a
157 * hypervisor, since the system may be SMP even if the guest kernel
160 static inline unsigned long __sync_cmpxchg(volatile void *ptr,
162 unsigned long new, int size)
167 __asm__ __volatile__("lock; cmpxchgb %b1,%2"
169 : "q"(new), "m"(*__xg(ptr)), "0"(old)
173 __asm__ __volatile__("lock; cmpxchgw %w1,%2"
175 : "r"(new), "m"(*__xg(ptr)), "0"(old)
179 __asm__ __volatile__("lock; cmpxchgl %1,%2"
181 : "r"(new), "m"(*__xg(ptr)), "0"(old)
188 static inline unsigned long __cmpxchg_local(volatile void *ptr,
189 unsigned long old, unsigned long new, int size)
194 __asm__ __volatile__("cmpxchgb %b1,%2"
196 : "q"(new), "m"(*__xg(ptr)), "0"(old)
200 __asm__ __volatile__("cmpxchgw %w1,%2"
202 : "r"(new), "m"(*__xg(ptr)), "0"(old)
206 __asm__ __volatile__("cmpxchgl %1,%2"
208 : "r"(new), "m"(*__xg(ptr)), "0"(old)
215 static inline unsigned long long __cmpxchg64(volatile void *ptr,
216 unsigned long long old, unsigned long long new)
218 unsigned long long prev;
219 __asm__ __volatile__(LOCK_PREFIX "cmpxchg8b %3"
221 : "b"((unsigned long)new),
222 "c"((unsigned long)(new >> 32)),
229 static inline unsigned long long __cmpxchg64_local(volatile void *ptr,
230 unsigned long long old, unsigned long long new)
232 unsigned long long prev;
233 __asm__ __volatile__("cmpxchg8b %3"
235 : "b"((unsigned long)new),
236 "c"((unsigned long)(new >> 32)),
243 #ifndef CONFIG_X86_CMPXCHG
245 * Building a kernel capable running on 80386. It may be necessary to
246 * simulate the cmpxchg on the 80386 CPU. For that purpose we define
247 * a function for each of the sizes we support.
250 extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
251 extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
252 extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
254 static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
255 unsigned long new, int size)
259 return cmpxchg_386_u8(ptr, old, new);
261 return cmpxchg_386_u16(ptr, old, new);
263 return cmpxchg_386_u32(ptr, old, new);
268 #define cmpxchg(ptr, o, n) \
270 __typeof__(*(ptr)) __ret; \
271 if (likely(boot_cpu_data.x86 > 3)) \
272 __ret = __cmpxchg((ptr), (unsigned long)(o), \
273 (unsigned long)(n), sizeof(*(ptr))); \
275 __ret = cmpxchg_386((ptr), (unsigned long)(o), \
276 (unsigned long)(n), sizeof(*(ptr))); \
279 #define cmpxchg_local(ptr, o, n) \
281 __typeof__(*(ptr)) __ret; \
282 if (likely(boot_cpu_data.x86 > 3)) \
283 __ret = __cmpxchg_local((ptr), (unsigned long)(o), \
284 (unsigned long)(n), sizeof(*(ptr))); \
286 __ret = cmpxchg_386((ptr), (unsigned long)(o), \
287 (unsigned long)(n), sizeof(*(ptr))); \
292 #ifndef CONFIG_X86_CMPXCHG64
294 * Building a kernel capable running on 80386 and 80486. It may be necessary
295 * to simulate the cmpxchg8b on the 80386 and 80486 CPU.
298 extern unsigned long long cmpxchg_486_u64(volatile void *, u64, u64);
300 #define cmpxchg64(ptr, o, n) \
302 __typeof__(*(ptr)) __ret; \
303 if (likely(boot_cpu_data.x86 > 4)) \
304 __ret = __cmpxchg64((ptr), (unsigned long long)(o), \
305 (unsigned long long)(n)); \
307 __ret = cmpxchg_486_u64((ptr), (unsigned long long)(o), \
308 (unsigned long long)(n)); \
311 #define cmpxchg64_local(ptr, o, n) \
313 __typeof__(*(ptr)) __ret; \
314 if (likely(boot_cpu_data.x86 > 4)) \
315 __ret = __cmpxchg64_local((ptr), (unsigned long long)(o), \
316 (unsigned long long)(n)); \
318 __ret = cmpxchg_486_u64((ptr), (unsigned long long)(o), \
319 (unsigned long long)(n)); \