1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef ASM_X86_CMPXCHG_H
3#define ASM_X86_CMPXCHG_H
4
5#include <linux/compiler.h>
6#include <asm/cpufeatures.h>
7#include <asm/alternative.h> /* Provides LOCK_PREFIX */
8
9/*
10 * Non-existent functions to indicate usage errors at link time
11 * (or compile-time if the compiler implements __compiletime_error().
12 */
13extern void __xchg_wrong_size(void)
14 __compiletime_error("Bad argument size for xchg");
15extern void __cmpxchg_wrong_size(void)
16 __compiletime_error("Bad argument size for cmpxchg");
17extern void __xadd_wrong_size(void)
18 __compiletime_error("Bad argument size for xadd");
19extern void __add_wrong_size(void)
20 __compiletime_error("Bad argument size for add");
21
22/*
23 * Constants for operation sizes. On 32-bit, the 64-bit size it set to
24 * -1 because sizeof will never return -1, thereby making those switch
25 * case statements guaranteeed dead code which the compiler will
26 * eliminate, and allowing the "missing symbol in the default case" to
27 * indicate a usage error.
28 */
29#define __X86_CASE_B 1
30#define __X86_CASE_W 2
31#define __X86_CASE_L 4
32#ifdef CONFIG_64BIT
33#define __X86_CASE_Q 8
34#else
35#define __X86_CASE_Q -1 /* sizeof will never return -1 */
36#endif
37
38/*
39 * An exchange-type operation, which takes a value and a pointer, and
40 * returns the old value.
41 */
42#define __xchg_op(ptr, arg, op, lock) \
43 ({ \
44 __typeof__ (*(ptr)) __ret = (arg); \
45 switch (sizeof(*(ptr))) { \
46 case __X86_CASE_B: \
47 asm volatile (lock #op "b %b0, %1\n" \
48 : "+q" (__ret), "+m" (*(ptr)) \
49 : : "memory", "cc"); \
50 break; \
51 case __X86_CASE_W: \
52 asm volatile (lock #op "w %w0, %1\n" \
53 : "+r" (__ret), "+m" (*(ptr)) \
54 : : "memory", "cc"); \
55 break; \
56 case __X86_CASE_L: \
57 asm volatile (lock #op "l %0, %1\n" \
58 : "+r" (__ret), "+m" (*(ptr)) \
59 : : "memory", "cc"); \
60 break; \
61 case __X86_CASE_Q: \
62 asm volatile (lock #op "q %q0, %1\n" \
63 : "+r" (__ret), "+m" (*(ptr)) \
64 : : "memory", "cc"); \
65 break; \
66 default: \
67 __ ## op ## _wrong_size(); \
68 } \
69 __ret; \
70 })
71
72/*
73 * Note: no "lock" prefix even on SMP: xchg always implies lock anyway.
74 * Since this is generally used to protect other memory information, we
75 * use "asm volatile" and "memory" clobbers to prevent gcc from moving
76 * information around.
77 */
78#define arch_xchg(ptr, v) __xchg_op((ptr), (v), xchg, "")
79
80/*
81 * Atomic compare and exchange. Compare OLD with MEM, if identical,
82 * store NEW in MEM. Return the initial value in MEM. Success is
83 * indicated by comparing RETURN with OLD.
84 */
85#define __raw_cmpxchg(ptr, old, new, size, lock) \
86({ \
87 __typeof__(*(ptr)) __ret; \
88 __typeof__(*(ptr)) __old = (old); \
89 __typeof__(*(ptr)) __new = (new); \
90 switch (size) { \
91 case __X86_CASE_B: \
92 { \
93 volatile u8 *__ptr = (volatile u8 *)(ptr); \
94 asm volatile(lock "cmpxchgb %2,%1" \
95 : "=a" (__ret), "+m" (*__ptr) \
96 : "q" (__new), "0" (__old) \
97 : "memory"); \
98 break; \
99 } \
100 case __X86_CASE_W: \
101 { \
102 volatile u16 *__ptr = (volatile u16 *)(ptr); \
103 asm volatile(lock "cmpxchgw %2,%1" \
104 : "=a" (__ret), "+m" (*__ptr) \
105 : "r" (__new), "0" (__old) \
106 : "memory"); \
107 break; \
108 } \
109 case __X86_CASE_L: \
110 { \
111 volatile u32 *__ptr = (volatile u32 *)(ptr); \
112 asm volatile(lock "cmpxchgl %2,%1" \
113 : "=a" (__ret), "+m" (*__ptr) \
114 : "r" (__new), "0" (__old) \
115 : "memory"); \
116 break; \
117 } \
118 case __X86_CASE_Q: \
119 { \
120 volatile u64 *__ptr = (volatile u64 *)(ptr); \
121 asm volatile(lock "cmpxchgq %2,%1" \
122 : "=a" (__ret), "+m" (*__ptr) \
123 : "r" (__new), "0" (__old) \
124 : "memory"); \
125 break; \
126 } \
127 default: \
128 __cmpxchg_wrong_size(); \
129 } \
130 __ret; \
131})
132
133#define __cmpxchg(ptr, old, new, size) \
134 __raw_cmpxchg((ptr), (old), (new), (size), LOCK_PREFIX)
135
136#define __sync_cmpxchg(ptr, old, new, size) \
137 __raw_cmpxchg((ptr), (old), (new), (size), "lock; ")
138
139#define __cmpxchg_local(ptr, old, new, size) \
140 __raw_cmpxchg((ptr), (old), (new), (size), "")
141
142#ifdef CONFIG_X86_32
143# include <asm/cmpxchg_32.h>
144#else
145# include <asm/cmpxchg_64.h>
146#endif
147
148#define arch_cmpxchg(ptr, old, new) \
149 __cmpxchg(ptr, old, new, sizeof(*(ptr)))
150
151#define arch_sync_cmpxchg(ptr, old, new) \
152 __sync_cmpxchg(ptr, old, new, sizeof(*(ptr)))
153
154#define arch_cmpxchg_local(ptr, old, new) \
155 __cmpxchg_local(ptr, old, new, sizeof(*(ptr)))
156
157
158#define __raw_try_cmpxchg(_ptr, _pold, _new, size, lock) \
159({ \
160 bool success; \
161 __typeof__(_ptr) _old = (__typeof__(_ptr))(_pold); \
162 __typeof__(*(_ptr)) __old = *_old; \
163 __typeof__(*(_ptr)) __new = (_new); \
164 switch (size) { \
165 case __X86_CASE_B: \
166 { \
167 volatile u8 *__ptr = (volatile u8 *)(_ptr); \
168 asm volatile(lock "cmpxchgb %[new], %[ptr]" \
169 CC_SET(z) \
170 : CC_OUT(z) (success), \
171 [ptr] "+m" (*__ptr), \
172 [old] "+a" (__old) \
173 : [new] "q" (__new) \
174 : "memory"); \
175 break; \
176 } \
177 case __X86_CASE_W: \
178 { \
179 volatile u16 *__ptr = (volatile u16 *)(_ptr); \
180 asm volatile(lock "cmpxchgw %[new], %[ptr]" \
181 CC_SET(z) \
182 : CC_OUT(z) (success), \
183 [ptr] "+m" (*__ptr), \
184 [old] "+a" (__old) \
185 : [new] "r" (__new) \
186 : "memory"); \
187 break; \
188 } \
189 case __X86_CASE_L: \
190 { \
191 volatile u32 *__ptr = (volatile u32 *)(_ptr); \
192 asm volatile(lock "cmpxchgl %[new], %[ptr]" \
193 CC_SET(z) \
194 : CC_OUT(z) (success), \
195 [ptr] "+m" (*__ptr), \
196 [old] "+a" (__old) \
197 : [new] "r" (__new) \
198 : "memory"); \
199 break; \
200 } \
201 case __X86_CASE_Q: \
202 { \
203 volatile u64 *__ptr = (volatile u64 *)(_ptr); \
204 asm volatile(lock "cmpxchgq %[new], %[ptr]" \
205 CC_SET(z) \
206 : CC_OUT(z) (success), \
207 [ptr] "+m" (*__ptr), \
208 [old] "+a" (__old) \
209 : [new] "r" (__new) \
210 : "memory"); \
211 break; \
212 } \
213 default: \
214 __cmpxchg_wrong_size(); \
215 } \
216 if (unlikely(!success)) \
217 *_old = __old; \
218 likely(success); \
219})
220
221#define __try_cmpxchg(ptr, pold, new, size) \
222 __raw_try_cmpxchg((ptr), (pold), (new), (size), LOCK_PREFIX)
223
224#define try_cmpxchg(ptr, pold, new) \
225 __try_cmpxchg((ptr), (pold), (new), sizeof(*(ptr)))
226
227/*
228 * xadd() adds "inc" to "*ptr" and atomically returns the previous
229 * value of "*ptr".
230 *
231 * xadd() is locked when multiple CPUs are online
232 */
233#define __xadd(ptr, inc, lock) __xchg_op((ptr), (inc), xadd, lock)
234#define xadd(ptr, inc) __xadd((ptr), (inc), LOCK_PREFIX)
235
236#define __cmpxchg_double(pfx, p1, p2, o1, o2, n1, n2) \
237({ \
238 bool __ret; \
239 __typeof__(*(p1)) __old1 = (o1), __new1 = (n1); \
240 __typeof__(*(p2)) __old2 = (o2), __new2 = (n2); \
241 BUILD_BUG_ON(sizeof(*(p1)) != sizeof(long)); \
242 BUILD_BUG_ON(sizeof(*(p2)) != sizeof(long)); \
243 VM_BUG_ON((unsigned long)(p1) % (2 * sizeof(long))); \
244 VM_BUG_ON((unsigned long)((p1) + 1) != (unsigned long)(p2)); \
245 asm volatile(pfx "cmpxchg%c5b %1" \
246 CC_SET(e) \
247 : CC_OUT(e) (__ret), \
248 "+m" (*(p1)), "+m" (*(p2)), \
249 "+a" (__old1), "+d" (__old2) \
250 : "i" (2 * sizeof(long)), \
251 "b" (__new1), "c" (__new2)); \
252 __ret; \
253})
254
255#define arch_cmpxchg_double(p1, p2, o1, o2, n1, n2) \
256 __cmpxchg_double(LOCK_PREFIX, p1, p2, o1, o2, n1, n2)
257
258#define arch_cmpxchg_double_local(p1, p2, o1, o2, n1, n2) \
259 __cmpxchg_double(, p1, p2, o1, o2, n1, n2)
260
261#endif /* ASM_X86_CMPXCHG_H */
262