GNU Linux-libre 5.4.274-gnu1
[releases.git] / arch / xtensa / include / asm / cmpxchg.h
1 /*
2  * Atomic xchg and cmpxchg operations.
3  *
4  * This file is subject to the terms and conditions of the GNU General Public
5  * License.  See the file "COPYING" in the main directory of this archive
6  * for more details.
7  *
8  * Copyright (C) 2001 - 2005 Tensilica Inc.
9  */
10
11 #ifndef _XTENSA_CMPXCHG_H
12 #define _XTENSA_CMPXCHG_H
13
14 #ifndef __ASSEMBLY__
15
16 #include <linux/bits.h>
17 #include <linux/stringify.h>
18
19 /*
20  * cmpxchg
21  */
22
23 static inline unsigned long
24 __cmpxchg_u32(volatile int *p, int old, int new)
25 {
26 #if XCHAL_HAVE_EXCLUSIVE
27         unsigned long tmp, result;
28
29         __asm__ __volatile__(
30                         "1:     l32ex   %0, %3\n"
31                         "       bne     %0, %4, 2f\n"
32                         "       mov     %1, %2\n"
33                         "       s32ex   %1, %3\n"
34                         "       getex   %1\n"
35                         "       beqz    %1, 1b\n"
36                         "2:\n"
37                         : "=&a" (result), "=&a" (tmp)
38                         : "a" (new), "a" (p), "a" (old)
39                         : "memory"
40                         );
41
42         return result;
43 #elif XCHAL_HAVE_S32C1I
44         __asm__ __volatile__(
45                         "       wsr     %2, scompare1\n"
46                         "       s32c1i  %0, %1, 0\n"
47                         : "+a" (new)
48                         : "a" (p), "a" (old)
49                         : "memory"
50                         );
51
52         return new;
53 #else
54         __asm__ __volatile__(
55                         "       rsil    a15, "__stringify(TOPLEVEL)"\n"
56                         "       l32i    %0, %1, 0\n"
57                         "       bne     %0, %2, 1f\n"
58                         "       s32i    %3, %1, 0\n"
59                         "1:\n"
60                         "       wsr     a15, ps\n"
61                         "       rsync\n"
62                         : "=&a" (old)
63                         : "a" (p), "a" (old), "r" (new)
64                         : "a15", "memory");
65         return old;
66 #endif
67 }
68 /* This function doesn't exist, so you'll get a linker error
69  * if something tries to do an invalid cmpxchg(). */
70
71 extern void __cmpxchg_called_with_bad_pointer(void);
72
73 static __inline__ unsigned long
74 __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
75 {
76         switch (size) {
77         case 4:  return __cmpxchg_u32(ptr, old, new);
78         default: __cmpxchg_called_with_bad_pointer();
79                  return old;
80         }
81 }
82
83 #define cmpxchg(ptr,o,n)                                                      \
84         ({ __typeof__(*(ptr)) _o_ = (o);                                      \
85            __typeof__(*(ptr)) _n_ = (n);                                      \
86            (__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_,          \
87                                         (unsigned long)_n_, sizeof (*(ptr))); \
88         })
89
90 #include <asm-generic/cmpxchg-local.h>
91
92 static inline unsigned long __cmpxchg_local(volatile void *ptr,
93                                       unsigned long old,
94                                       unsigned long new, int size)
95 {
96         switch (size) {
97         case 4:
98                 return __cmpxchg_u32(ptr, old, new);
99         default:
100                 return __cmpxchg_local_generic(ptr, old, new, size);
101         }
102
103         return old;
104 }
105
106 /*
107  * cmpxchg_local and cmpxchg64_local are atomic wrt current CPU. Always make
108  * them available.
109  */
110 #define cmpxchg_local(ptr, o, n)                                               \
111         ((__typeof__(*(ptr)))__cmpxchg_local_generic((ptr), (unsigned long)(o),\
112                         (unsigned long)(n), sizeof(*(ptr))))
113 #define cmpxchg64_local(ptr, o, n) __cmpxchg64_local_generic((ptr), (o), (n))
114 #define cmpxchg64(ptr, o, n)    cmpxchg64_local((ptr), (o), (n))
115
116 /*
117  * xchg_u32
118  *
119  * Note that a15 is used here because the register allocation
120  * done by the compiler is not guaranteed and a window overflow
121  * may not occur between the rsil and wsr instructions. By using
122  * a15 in the rsil, the machine is guaranteed to be in a state
123  * where no register reference will cause an overflow.
124  */
125
126 static inline unsigned long xchg_u32(volatile int * m, unsigned long val)
127 {
128 #if XCHAL_HAVE_EXCLUSIVE
129         unsigned long tmp, result;
130
131         __asm__ __volatile__(
132                         "1:     l32ex   %0, %3\n"
133                         "       mov     %1, %2\n"
134                         "       s32ex   %1, %3\n"
135                         "       getex   %1\n"
136                         "       beqz    %1, 1b\n"
137                         : "=&a" (result), "=&a" (tmp)
138                         : "a" (val), "a" (m)
139                         : "memory"
140                         );
141
142         return result;
143 #elif XCHAL_HAVE_S32C1I
144         unsigned long tmp, result;
145         __asm__ __volatile__(
146                         "1:     l32i    %1, %2, 0\n"
147                         "       mov     %0, %3\n"
148                         "       wsr     %1, scompare1\n"
149                         "       s32c1i  %0, %2, 0\n"
150                         "       bne     %0, %1, 1b\n"
151                         : "=&a" (result), "=&a" (tmp)
152                         : "a" (m), "a" (val)
153                         : "memory"
154                         );
155         return result;
156 #else
157         unsigned long tmp;
158         __asm__ __volatile__(
159                         "       rsil    a15, "__stringify(TOPLEVEL)"\n"
160                         "       l32i    %0, %1, 0\n"
161                         "       s32i    %2, %1, 0\n"
162                         "       wsr     a15, ps\n"
163                         "       rsync\n"
164                         : "=&a" (tmp)
165                         : "a" (m), "a" (val)
166                         : "a15", "memory");
167         return tmp;
168 #endif
169 }
170
171 #define xchg(ptr,x) \
172         ((__typeof__(*(ptr)))__xchg((unsigned long)(x),(ptr),sizeof(*(ptr))))
173
174 static inline u32 xchg_small(volatile void *ptr, u32 x, int size)
175 {
176         int off = (unsigned long)ptr % sizeof(u32);
177         volatile u32 *p = ptr - off;
178 #ifdef __BIG_ENDIAN
179         int bitoff = (sizeof(u32) - size - off) * BITS_PER_BYTE;
180 #else
181         int bitoff = off * BITS_PER_BYTE;
182 #endif
183         u32 bitmask = ((0x1 << size * BITS_PER_BYTE) - 1) << bitoff;
184         u32 oldv, newv;
185         u32 ret;
186
187         do {
188                 oldv = READ_ONCE(*p);
189                 ret = (oldv & bitmask) >> bitoff;
190                 newv = (oldv & ~bitmask) | (x << bitoff);
191         } while (__cmpxchg_u32(p, oldv, newv) != oldv);
192
193         return ret;
194 }
195
196 /*
197  * This only works if the compiler isn't horribly bad at optimizing.
198  * gcc-2.5.8 reportedly can't handle this, but I define that one to
199  * be dead anyway.
200  */
201
202 extern void __xchg_called_with_bad_pointer(void);
203
204 static __inline__ unsigned long
205 __xchg(unsigned long x, volatile void * ptr, int size)
206 {
207         switch (size) {
208         case 1:
209                 return xchg_small(ptr, x, 1);
210         case 2:
211                 return xchg_small(ptr, x, 2);
212         case 4:
213                 return xchg_u32(ptr, x);
214         default:
215                 __xchg_called_with_bad_pointer();
216                 return x;
217         }
218 }
219
220 #endif /* __ASSEMBLY__ */
221
222 #endif /* _XTENSA_CMPXCHG_H */