xref: /openbmc/linux/arch/x86/include/asm/cmpxchg_32.h (revision b6dcefde)
1 #ifndef _ASM_X86_CMPXCHG_32_H
2 #define _ASM_X86_CMPXCHG_32_H
3 
4 #include <linux/bitops.h> /* for LOCK_PREFIX */
5 
6 /*
7  * Note: if you use set64_bit(), __cmpxchg64(), or their variants, you
8  *       you need to test for the feature in boot_cpu_data.
9  */
10 
11 extern void __xchg_wrong_size(void);
12 
13 /*
14  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
15  * Note 2: xchg has side effect, so that attribute volatile is necessary,
16  *	  but generally the primitive is invalid, *ptr is output argument. --ANK
17  */
18 
19 struct __xchg_dummy {
20 	unsigned long a[100];
21 };
22 #define __xg(x) ((struct __xchg_dummy *)(x))
23 
24 #define __xchg(x, ptr, size)						\
25 ({									\
26 	__typeof(*(ptr)) __x = (x);					\
27 	switch (size) {							\
28 	case 1:								\
29 		asm volatile("xchgb %b0,%1"				\
30 			     : "=q" (__x)				\
31 			     : "m" (*__xg(ptr)), "0" (__x)		\
32 			     : "memory");				\
33 		break;							\
34 	case 2:								\
35 		asm volatile("xchgw %w0,%1"				\
36 			     : "=r" (__x)				\
37 			     : "m" (*__xg(ptr)), "0" (__x)		\
38 			     : "memory");				\
39 		break;							\
40 	case 4:								\
41 		asm volatile("xchgl %0,%1"				\
42 			     : "=r" (__x)				\
43 			     : "m" (*__xg(ptr)), "0" (__x)		\
44 			     : "memory");				\
45 		break;							\
46 	default:							\
47 		__xchg_wrong_size();					\
48 	}								\
49 	__x;								\
50 })
51 
52 #define xchg(ptr, v)							\
53 	__xchg((v), (ptr), sizeof(*ptr))
54 
55 /*
56  * The semantics of XCHGCMP8B are a bit strange, this is why
57  * there is a loop and the loading of %%eax and %%edx has to
58  * be inside. This inlines well in most cases, the cached
59  * cost is around ~38 cycles. (in the future we might want
60  * to do an SIMD/3DNOW!/MMX/FPU 64-bit store here, but that
61  * might have an implicit FPU-save as a cost, so it's not
62  * clear which path to go.)
63  *
64  * cmpxchg8b must be used with the lock prefix here to allow
65  * the instruction to be executed atomically, see page 3-102
66  * of the instruction set reference 24319102.pdf. We need
67  * the reader side to see the coherent 64bit value.
68  */
69 static inline void __set_64bit(unsigned long long *ptr,
70 			       unsigned int low, unsigned int high)
71 {
72 	asm volatile("\n1:\t"
73 		     "movl (%0), %%eax\n\t"
74 		     "movl 4(%0), %%edx\n\t"
75 		     LOCK_PREFIX "cmpxchg8b (%0)\n\t"
76 		     "jnz 1b"
77 		     : /* no outputs */
78 		     : "D"(ptr),
79 		       "b"(low),
80 		       "c"(high)
81 		     : "ax", "dx", "memory");
82 }
83 
84 static inline void __set_64bit_constant(unsigned long long *ptr,
85 					unsigned long long value)
86 {
87 	__set_64bit(ptr, (unsigned int)value, (unsigned int)(value >> 32));
88 }
89 
90 #define ll_low(x)	*(((unsigned int *)&(x)) + 0)
91 #define ll_high(x)	*(((unsigned int *)&(x)) + 1)
92 
93 static inline void __set_64bit_var(unsigned long long *ptr,
94 				   unsigned long long value)
95 {
96 	__set_64bit(ptr, ll_low(value), ll_high(value));
97 }
98 
99 #define set_64bit(ptr, value)			\
100 	(__builtin_constant_p((value))		\
101 	 ? __set_64bit_constant((ptr), (value))	\
102 	 : __set_64bit_var((ptr), (value)))
103 
104 #define _set_64bit(ptr, value)						\
105 	(__builtin_constant_p(value)					\
106 	 ? __set_64bit(ptr, (unsigned int)(value),			\
107 		       (unsigned int)((value) >> 32))			\
108 	 : __set_64bit(ptr, ll_low((value)), ll_high((value))))
109 
110 extern void __cmpxchg_wrong_size(void);
111 
112 /*
113  * Atomic compare and exchange.  Compare OLD with MEM, if identical,
114  * store NEW in MEM.  Return the initial value in MEM.  Success is
115  * indicated by comparing RETURN with OLD.
116  */
117 #define __raw_cmpxchg(ptr, old, new, size, lock)			\
118 ({									\
119 	__typeof__(*(ptr)) __ret;					\
120 	__typeof__(*(ptr)) __old = (old);				\
121 	__typeof__(*(ptr)) __new = (new);				\
122 	switch (size) {							\
123 	case 1:								\
124 		asm volatile(lock "cmpxchgb %b1,%2"			\
125 			     : "=a"(__ret)				\
126 			     : "q"(__new), "m"(*__xg(ptr)), "0"(__old)	\
127 			     : "memory");				\
128 		break;							\
129 	case 2:								\
130 		asm volatile(lock "cmpxchgw %w1,%2"			\
131 			     : "=a"(__ret)				\
132 			     : "r"(__new), "m"(*__xg(ptr)), "0"(__old)	\
133 			     : "memory");				\
134 		break;							\
135 	case 4:								\
136 		asm volatile(lock "cmpxchgl %1,%2"			\
137 			     : "=a"(__ret)				\
138 			     : "r"(__new), "m"(*__xg(ptr)), "0"(__old)	\
139 			     : "memory");				\
140 		break;							\
141 	default:							\
142 		__cmpxchg_wrong_size();					\
143 	}								\
144 	__ret;								\
145 })
146 
147 #define __cmpxchg(ptr, old, new, size)					\
148 	__raw_cmpxchg((ptr), (old), (new), (size), LOCK_PREFIX)
149 
150 #define __sync_cmpxchg(ptr, old, new, size)				\
151 	__raw_cmpxchg((ptr), (old), (new), (size), "lock; ")
152 
153 #define __cmpxchg_local(ptr, old, new, size)				\
154 	__raw_cmpxchg((ptr), (old), (new), (size), "")
155 
156 #ifdef CONFIG_X86_CMPXCHG
157 #define __HAVE_ARCH_CMPXCHG 1
158 
159 #define cmpxchg(ptr, old, new)						\
160 	__cmpxchg((ptr), (old), (new), sizeof(*ptr))
161 
162 #define sync_cmpxchg(ptr, old, new)					\
163 	__sync_cmpxchg((ptr), (old), (new), sizeof(*ptr))
164 
165 #define cmpxchg_local(ptr, old, new)					\
166 	__cmpxchg_local((ptr), (old), (new), sizeof(*ptr))
167 #endif
168 
169 #ifdef CONFIG_X86_CMPXCHG64
170 #define cmpxchg64(ptr, o, n)						\
171 	((__typeof__(*(ptr)))__cmpxchg64((ptr), (unsigned long long)(o), \
172 					 (unsigned long long)(n)))
173 #define cmpxchg64_local(ptr, o, n)					\
174 	((__typeof__(*(ptr)))__cmpxchg64_local((ptr), (unsigned long long)(o), \
175 					       (unsigned long long)(n)))
176 #endif
177 
178 static inline unsigned long long __cmpxchg64(volatile void *ptr,
179 					     unsigned long long old,
180 					     unsigned long long new)
181 {
182 	unsigned long long prev;
183 	asm volatile(LOCK_PREFIX "cmpxchg8b %3"
184 		     : "=A"(prev)
185 		     : "b"((unsigned long)new),
186 		       "c"((unsigned long)(new >> 32)),
187 		       "m"(*__xg(ptr)),
188 		       "0"(old)
189 		     : "memory");
190 	return prev;
191 }
192 
193 static inline unsigned long long __cmpxchg64_local(volatile void *ptr,
194 						   unsigned long long old,
195 						   unsigned long long new)
196 {
197 	unsigned long long prev;
198 	asm volatile("cmpxchg8b %3"
199 		     : "=A"(prev)
200 		     : "b"((unsigned long)new),
201 		       "c"((unsigned long)(new >> 32)),
202 		       "m"(*__xg(ptr)),
203 		       "0"(old)
204 		     : "memory");
205 	return prev;
206 }
207 
208 #ifndef CONFIG_X86_CMPXCHG
209 /*
210  * Building a kernel capable running on 80386. It may be necessary to
211  * simulate the cmpxchg on the 80386 CPU. For that purpose we define
212  * a function for each of the sizes we support.
213  */
214 
215 extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
216 extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
217 extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
218 
219 static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
220 					unsigned long new, int size)
221 {
222 	switch (size) {
223 	case 1:
224 		return cmpxchg_386_u8(ptr, old, new);
225 	case 2:
226 		return cmpxchg_386_u16(ptr, old, new);
227 	case 4:
228 		return cmpxchg_386_u32(ptr, old, new);
229 	}
230 	return old;
231 }
232 
233 #define cmpxchg(ptr, o, n)						\
234 ({									\
235 	__typeof__(*(ptr)) __ret;					\
236 	if (likely(boot_cpu_data.x86 > 3))				\
237 		__ret = (__typeof__(*(ptr)))__cmpxchg((ptr),		\
238 				(unsigned long)(o), (unsigned long)(n),	\
239 				sizeof(*(ptr)));			\
240 	else								\
241 		__ret = (__typeof__(*(ptr)))cmpxchg_386((ptr),		\
242 				(unsigned long)(o), (unsigned long)(n),	\
243 				sizeof(*(ptr)));			\
244 	__ret;								\
245 })
246 #define cmpxchg_local(ptr, o, n)					\
247 ({									\
248 	__typeof__(*(ptr)) __ret;					\
249 	if (likely(boot_cpu_data.x86 > 3))				\
250 		__ret = (__typeof__(*(ptr)))__cmpxchg_local((ptr),	\
251 				(unsigned long)(o), (unsigned long)(n),	\
252 				sizeof(*(ptr)));			\
253 	else								\
254 		__ret = (__typeof__(*(ptr)))cmpxchg_386((ptr),		\
255 				(unsigned long)(o), (unsigned long)(n),	\
256 				sizeof(*(ptr)));			\
257 	__ret;								\
258 })
259 #endif
260 
261 #ifndef CONFIG_X86_CMPXCHG64
262 /*
263  * Building a kernel capable running on 80386 and 80486. It may be necessary
264  * to simulate the cmpxchg8b on the 80386 and 80486 CPU.
265  */
266 
267 extern unsigned long long cmpxchg_486_u64(volatile void *, u64, u64);
268 
269 #define cmpxchg64(ptr, o, n)					\
270 ({								\
271 	__typeof__(*(ptr)) __ret;				\
272 	__typeof__(*(ptr)) __old = (o);				\
273 	__typeof__(*(ptr)) __new = (n);				\
274 	alternative_io("call cmpxchg8b_emu",			\
275 			"lock; cmpxchg8b (%%esi)" ,		\
276 		       X86_FEATURE_CX8,				\
277 		       "=A" (__ret),				\
278 		       "S" ((ptr)), "0" (__old),		\
279 		       "b" ((unsigned int)__new),		\
280 		       "c" ((unsigned int)(__new>>32))		\
281 		       : "memory");				\
282 	__ret; })
283 
284 
285 
286 #define cmpxchg64_local(ptr, o, n)					\
287 ({									\
288 	__typeof__(*(ptr)) __ret;					\
289 	if (likely(boot_cpu_data.x86 > 4))				\
290 		__ret = (__typeof__(*(ptr)))__cmpxchg64_local((ptr),	\
291 				(unsigned long long)(o),		\
292 				(unsigned long long)(n));		\
293 	else								\
294 		__ret = (__typeof__(*(ptr)))cmpxchg_486_u64((ptr),	\
295 				(unsigned long long)(o),		\
296 				(unsigned long long)(n));		\
297 	__ret;								\
298 })
299 
300 #endif
301 
302 #endif /* _ASM_X86_CMPXCHG_32_H */
303