xref: /linux/arch/sparc/include/asm/cmpxchg_64.h (revision 55ec81f7517fad09135f65552cea0a3ee84fff30)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* 64-bit atomic xchg() and cmpxchg() definitions.
3  *
4  * Copyright (C) 1996, 1997, 2000 David S. Miller (davem@redhat.com)
5  */
6 
7 #ifndef __ARCH_SPARC64_CMPXCHG__
8 #define __ARCH_SPARC64_CMPXCHG__
9 
10 static inline unsigned long
11 __cmpxchg_u32(volatile int *m, int old, int new)
12 {
13 	__asm__ __volatile__("cas [%2], %3, %0"
14 			     : "=&r" (new)
15 			     : "0" (new), "r" (m), "r" (old)
16 			     : "memory");
17 
18 	return new;
19 }
20 
21 static inline unsigned long xchg32(__volatile__ unsigned int *m, unsigned int val)
22 {
23 	unsigned long tmp1, tmp2;
24 
25 	__asm__ __volatile__(
26 "	mov		%0, %1\n"
27 "1:	lduw		[%4], %2\n"
28 "	cas		[%4], %2, %0\n"
29 "	cmp		%2, %0\n"
30 "	bne,a,pn	%%icc, 1b\n"
31 "	 mov		%1, %0\n"
32 	: "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
33 	: "0" (val), "r" (m)
34 	: "cc", "memory");
35 	return val;
36 }
37 
38 static inline unsigned long xchg64(__volatile__ unsigned long *m, unsigned long val)
39 {
40 	unsigned long tmp1, tmp2;
41 
42 	__asm__ __volatile__(
43 "	mov		%0, %1\n"
44 "1:	ldx		[%4], %2\n"
45 "	casx		[%4], %2, %0\n"
46 "	cmp		%2, %0\n"
47 "	bne,a,pn	%%xcc, 1b\n"
48 "	 mov		%1, %0\n"
49 	: "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
50 	: "0" (val), "r" (m)
51 	: "cc", "memory");
52 	return val;
53 }
54 
55 #define arch_xchg(ptr,x)							\
56 ({	__typeof__(*(ptr)) __ret;					\
57 	__ret = (__typeof__(*(ptr)))					\
58 		__arch_xchg((unsigned long)(x), (ptr), sizeof(*(ptr)));	\
59 	__ret;								\
60 })
61 
62 void __xchg_called_with_bad_pointer(void);
63 
64 /*
65  * Use 4 byte cas instruction to achieve 2 byte xchg. Main logic
66  * here is to get the bit shift of the byte we are interested in.
67  * The XOR is handy for reversing the bits for big-endian byte order.
68  */
69 static inline unsigned long
70 xchg16(__volatile__ unsigned short *m, unsigned short val)
71 {
72 	unsigned long maddr = (unsigned long)m;
73 	int bit_shift = (((unsigned long)m & 2) ^ 2) << 3;
74 	unsigned int mask = 0xffff << bit_shift;
75 	unsigned int *ptr = (unsigned int  *) (maddr & ~2);
76 	unsigned int old32, new32, load32;
77 
78 	/* Read the old value */
79 	load32 = *ptr;
80 
81 	do {
82 		old32 = load32;
83 		new32 = (load32 & (~mask)) | val << bit_shift;
84 		load32 = __cmpxchg_u32(ptr, old32, new32);
85 	} while (load32 != old32);
86 
87 	return (load32 & mask) >> bit_shift;
88 }
89 
90 static __always_inline unsigned long
91 __arch_xchg(unsigned long x, __volatile__ void * ptr, int size)
92 {
93 	switch (size) {
94 	case 2:
95 		return xchg16(ptr, x);
96 	case 4:
97 		return xchg32(ptr, x);
98 	case 8:
99 		return xchg64(ptr, x);
100 	}
101 	__xchg_called_with_bad_pointer();
102 	return x;
103 }
104 
105 /*
106  * Atomic compare and exchange.  Compare OLD with MEM, if identical,
107  * store NEW in MEM.  Return the initial value in MEM.  Success is
108  * indicated by comparing RETURN with OLD.
109  */
110 
111 #include <asm-generic/cmpxchg-local.h>
112 
113 
114 static inline unsigned long
115 __cmpxchg_u64(volatile long *m, unsigned long old, unsigned long new)
116 {
117 	__asm__ __volatile__("casx [%2], %3, %0"
118 			     : "=&r" (new)
119 			     : "0" (new), "r" (m), "r" (old)
120 			     : "memory");
121 
122 	return new;
123 }
124 
125 /*
126  * Use 4 byte cas instruction to achieve 1 byte cmpxchg. Main logic
127  * here is to get the bit shift of the byte we are interested in.
128  * The XOR is handy for reversing the bits for big-endian byte order
129  */
130 static inline unsigned long
131 __cmpxchg_u8(volatile unsigned char *m, unsigned char old, unsigned char new)
132 {
133 	unsigned long maddr = (unsigned long)m;
134 	int bit_shift = (((unsigned long)m & 3) ^ 3) << 3;
135 	unsigned int mask = 0xff << bit_shift;
136 	unsigned int *ptr = (unsigned int *) (maddr & ~3);
137 	unsigned int old32, new32, load;
138 	unsigned int load32 = *ptr;
139 
140 	do {
141 		new32 = (load32 & ~mask) | (new << bit_shift);
142 		old32 = (load32 & ~mask) | (old << bit_shift);
143 		load32 = __cmpxchg_u32(ptr, old32, new32);
144 		if (load32 == old32)
145 			return old;
146 		load = (load32 & mask) >> bit_shift;
147 	} while (load == old);
148 
149 	return load;
150 }
151 
152 /* This function doesn't exist, so you'll get a linker error
153    if something tries to do an invalid cmpxchg().  */
154 void __cmpxchg_called_with_bad_pointer(void);
155 
156 static inline unsigned long
157 __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
158 {
159 	switch (size) {
160 		case 1:
161 			return __cmpxchg_u8(ptr, old, new);
162 		case 4:
163 			return __cmpxchg_u32(ptr, old, new);
164 		case 8:
165 			return __cmpxchg_u64(ptr, old, new);
166 	}
167 	__cmpxchg_called_with_bad_pointer();
168 	return old;
169 }
170 
171 #define arch_cmpxchg(ptr,o,n)						 \
172   ({									 \
173      __typeof__(*(ptr)) _o_ = (o);					 \
174      __typeof__(*(ptr)) _n_ = (n);					 \
175      (__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_,		 \
176 				    (unsigned long)_n_, sizeof(*(ptr))); \
177   })
178 
179 /*
180  * cmpxchg_local and cmpxchg64_local are atomic wrt current CPU. Always make
181  * them available.
182  */
183 
184 static inline unsigned long __cmpxchg_local(volatile void *ptr,
185 				      unsigned long old,
186 				      unsigned long new, int size)
187 {
188 	switch (size) {
189 	case 4:
190 	case 8:	return __cmpxchg(ptr, old, new, size);
191 	default:
192 		return __generic_cmpxchg_local(ptr, old, new, size);
193 	}
194 
195 	return old;
196 }
197 
198 #define arch_cmpxchg_local(ptr, o, n)				  	\
199 	((__typeof__(*(ptr)))__cmpxchg_local((ptr), (unsigned long)(o),	\
200 			(unsigned long)(n), sizeof(*(ptr))))
201 #define arch_cmpxchg64_local(ptr, o, n)					\
202   ({									\
203 	BUILD_BUG_ON(sizeof(*(ptr)) != 8);				\
204 	arch_cmpxchg_local((ptr), (o), (n));					\
205   })
206 #define arch_cmpxchg64(ptr, o, n)	arch_cmpxchg64_local((ptr), (o), (n))
207 
208 #endif /* __ARCH_SPARC64_CMPXCHG__ */
209