Merge branch 'fix/hda' into for-linus
[sfrench/cifs-2.6.git] / arch / x86 / include / asm / percpu.h
1 #ifndef _ASM_X86_PERCPU_H
2 #define _ASM_X86_PERCPU_H
3
4 #ifdef CONFIG_X86_64
5 #define __percpu_seg            gs
6 #define __percpu_mov_op         movq
7 #else
8 #define __percpu_seg            fs
9 #define __percpu_mov_op         movl
10 #endif
11
12 #ifdef __ASSEMBLY__
13
14 /*
15  * PER_CPU finds an address of a per-cpu variable.
16  *
17  * Args:
18  *    var - variable name
19  *    reg - 32bit register
20  *
21  * The resulting address is stored in the "reg" argument.
22  *
23  * Example:
24  *    PER_CPU(cpu_gdt_descr, %ebx)
25  */
26 #ifdef CONFIG_SMP
27 #define PER_CPU(var, reg)                                               \
28         __percpu_mov_op %__percpu_seg:per_cpu__this_cpu_off, reg;       \
29         lea per_cpu__##var(reg), reg
30 #define PER_CPU_VAR(var)        %__percpu_seg:per_cpu__##var
31 #else /* ! SMP */
32 #define PER_CPU(var, reg)                                               \
33         __percpu_mov_op $per_cpu__##var, reg
34 #define PER_CPU_VAR(var)        per_cpu__##var
35 #endif  /* SMP */
36
37 #ifdef CONFIG_X86_64_SMP
38 #define INIT_PER_CPU_VAR(var)  init_per_cpu__##var
39 #else
40 #define INIT_PER_CPU_VAR(var)  per_cpu__##var
41 #endif
42
43 #else /* ...!ASSEMBLY */
44
45 #include <linux/kernel.h>
46 #include <linux/stringify.h>
47
48 #ifdef CONFIG_SMP
49 #define __percpu_arg(x)         "%%"__stringify(__percpu_seg)":%P" #x
50 #define __my_cpu_offset         percpu_read(this_cpu_off)
51 #else
52 #define __percpu_arg(x)         "%P" #x
53 #endif
54
55 /*
56  * Initialized pointers to per-cpu variables needed for the boot
57  * processor need to use these macros to get the proper address
58  * offset from __per_cpu_load on SMP.
59  *
60  * There also must be an entry in vmlinux_64.lds.S
61  */
62 #define DECLARE_INIT_PER_CPU(var) \
63        extern typeof(per_cpu_var(var)) init_per_cpu_var(var)
64
65 #ifdef CONFIG_X86_64_SMP
66 #define init_per_cpu_var(var)  init_per_cpu__##var
67 #else
68 #define init_per_cpu_var(var)  per_cpu_var(var)
69 #endif
70
71 /* For arch-specific code, we can use direct single-insn ops (they
72  * don't give an lvalue though). */
73 extern void __bad_percpu_size(void);
74
75 #define percpu_to_op(op, var, val)                      \
76 do {                                                    \
77         typedef typeof(var) pto_T__;                    \
78         if (0) {                                        \
79                 pto_T__ pto_tmp__;                      \
80                 pto_tmp__ = (val);                      \
81         }                                               \
82         switch (sizeof(var)) {                          \
83         case 1:                                         \
84                 asm(op "b %1,"__percpu_arg(0)           \
85                     : "+m" (var)                        \
86                     : "qi" ((pto_T__)(val)));           \
87                 break;                                  \
88         case 2:                                         \
89                 asm(op "w %1,"__percpu_arg(0)           \
90                     : "+m" (var)                        \
91                     : "ri" ((pto_T__)(val)));           \
92                 break;                                  \
93         case 4:                                         \
94                 asm(op "l %1,"__percpu_arg(0)           \
95                     : "+m" (var)                        \
96                     : "ri" ((pto_T__)(val)));           \
97                 break;                                  \
98         case 8:                                         \
99                 asm(op "q %1,"__percpu_arg(0)           \
100                     : "+m" (var)                        \
101                     : "re" ((pto_T__)(val)));           \
102                 break;                                  \
103         default: __bad_percpu_size();                   \
104         }                                               \
105 } while (0)
106
107 #define percpu_from_op(op, var, constraint)             \
108 ({                                                      \
109         typeof(var) pfo_ret__;                          \
110         switch (sizeof(var)) {                          \
111         case 1:                                         \
112                 asm(op "b "__percpu_arg(1)",%0"         \
113                     : "=q" (pfo_ret__)                  \
114                     : constraint);                      \
115                 break;                                  \
116         case 2:                                         \
117                 asm(op "w "__percpu_arg(1)",%0"         \
118                     : "=r" (pfo_ret__)                  \
119                     : constraint);                      \
120                 break;                                  \
121         case 4:                                         \
122                 asm(op "l "__percpu_arg(1)",%0"         \
123                     : "=r" (pfo_ret__)                  \
124                     : constraint);                      \
125                 break;                                  \
126         case 8:                                         \
127                 asm(op "q "__percpu_arg(1)",%0"         \
128                     : "=r" (pfo_ret__)                  \
129                     : constraint);                      \
130                 break;                                  \
131         default: __bad_percpu_size();                   \
132         }                                               \
133         pfo_ret__;                                      \
134 })
135
136 /*
137  * percpu_read() makes gcc load the percpu variable every time it is
138  * accessed while percpu_read_stable() allows the value to be cached.
139  * percpu_read_stable() is more efficient and can be used if its value
140  * is guaranteed to be valid across cpus.  The current users include
141  * get_current() and get_thread_info() both of which are actually
142  * per-thread variables implemented as per-cpu variables and thus
143  * stable for the duration of the respective task.
144  */
145 #define percpu_read(var)        percpu_from_op("mov", per_cpu__##var,   \
146                                                "m" (per_cpu__##var))
147 #define percpu_read_stable(var) percpu_from_op("mov", per_cpu__##var,   \
148                                                "p" (&per_cpu__##var))
149 #define percpu_write(var, val)  percpu_to_op("mov", per_cpu__##var, val)
150 #define percpu_add(var, val)    percpu_to_op("add", per_cpu__##var, val)
151 #define percpu_sub(var, val)    percpu_to_op("sub", per_cpu__##var, val)
152 #define percpu_and(var, val)    percpu_to_op("and", per_cpu__##var, val)
153 #define percpu_or(var, val)     percpu_to_op("or", per_cpu__##var, val)
154 #define percpu_xor(var, val)    percpu_to_op("xor", per_cpu__##var, val)
155
156 #define __this_cpu_read_1(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
157 #define __this_cpu_read_2(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
158 #define __this_cpu_read_4(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
159
160 #define __this_cpu_write_1(pcp, val)    percpu_to_op("mov", (pcp), val)
161 #define __this_cpu_write_2(pcp, val)    percpu_to_op("mov", (pcp), val)
162 #define __this_cpu_write_4(pcp, val)    percpu_to_op("mov", (pcp), val)
163 #define __this_cpu_add_1(pcp, val)      percpu_to_op("add", (pcp), val)
164 #define __this_cpu_add_2(pcp, val)      percpu_to_op("add", (pcp), val)
165 #define __this_cpu_add_4(pcp, val)      percpu_to_op("add", (pcp), val)
166 #define __this_cpu_and_1(pcp, val)      percpu_to_op("and", (pcp), val)
167 #define __this_cpu_and_2(pcp, val)      percpu_to_op("and", (pcp), val)
168 #define __this_cpu_and_4(pcp, val)      percpu_to_op("and", (pcp), val)
169 #define __this_cpu_or_1(pcp, val)       percpu_to_op("or", (pcp), val)
170 #define __this_cpu_or_2(pcp, val)       percpu_to_op("or", (pcp), val)
171 #define __this_cpu_or_4(pcp, val)       percpu_to_op("or", (pcp), val)
172 #define __this_cpu_xor_1(pcp, val)      percpu_to_op("xor", (pcp), val)
173 #define __this_cpu_xor_2(pcp, val)      percpu_to_op("xor", (pcp), val)
174 #define __this_cpu_xor_4(pcp, val)      percpu_to_op("xor", (pcp), val)
175
176 #define this_cpu_read_1(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
177 #define this_cpu_read_2(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
178 #define this_cpu_read_4(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
179 #define this_cpu_write_1(pcp, val)      percpu_to_op("mov", (pcp), val)
180 #define this_cpu_write_2(pcp, val)      percpu_to_op("mov", (pcp), val)
181 #define this_cpu_write_4(pcp, val)      percpu_to_op("mov", (pcp), val)
182 #define this_cpu_add_1(pcp, val)        percpu_to_op("add", (pcp), val)
183 #define this_cpu_add_2(pcp, val)        percpu_to_op("add", (pcp), val)
184 #define this_cpu_add_4(pcp, val)        percpu_to_op("add", (pcp), val)
185 #define this_cpu_and_1(pcp, val)        percpu_to_op("and", (pcp), val)
186 #define this_cpu_and_2(pcp, val)        percpu_to_op("and", (pcp), val)
187 #define this_cpu_and_4(pcp, val)        percpu_to_op("and", (pcp), val)
188 #define this_cpu_or_1(pcp, val)         percpu_to_op("or", (pcp), val)
189 #define this_cpu_or_2(pcp, val)         percpu_to_op("or", (pcp), val)
190 #define this_cpu_or_4(pcp, val)         percpu_to_op("or", (pcp), val)
191 #define this_cpu_xor_1(pcp, val)        percpu_to_op("xor", (pcp), val)
192 #define this_cpu_xor_2(pcp, val)        percpu_to_op("xor", (pcp), val)
193 #define this_cpu_xor_4(pcp, val)        percpu_to_op("xor", (pcp), val)
194
195 #define irqsafe_cpu_add_1(pcp, val)     percpu_to_op("add", (pcp), val)
196 #define irqsafe_cpu_add_2(pcp, val)     percpu_to_op("add", (pcp), val)
197 #define irqsafe_cpu_add_4(pcp, val)     percpu_to_op("add", (pcp), val)
198 #define irqsafe_cpu_and_1(pcp, val)     percpu_to_op("and", (pcp), val)
199 #define irqsafe_cpu_and_2(pcp, val)     percpu_to_op("and", (pcp), val)
200 #define irqsafe_cpu_and_4(pcp, val)     percpu_to_op("and", (pcp), val)
201 #define irqsafe_cpu_or_1(pcp, val)      percpu_to_op("or", (pcp), val)
202 #define irqsafe_cpu_or_2(pcp, val)      percpu_to_op("or", (pcp), val)
203 #define irqsafe_cpu_or_4(pcp, val)      percpu_to_op("or", (pcp), val)
204 #define irqsafe_cpu_xor_1(pcp, val)     percpu_to_op("xor", (pcp), val)
205 #define irqsafe_cpu_xor_2(pcp, val)     percpu_to_op("xor", (pcp), val)
206 #define irqsafe_cpu_xor_4(pcp, val)     percpu_to_op("xor", (pcp), val)
207
208 /*
209  * Per cpu atomic 64 bit operations are only available under 64 bit.
210  * 32 bit must fall back to generic operations.
211  */
212 #ifdef CONFIG_X86_64
213 #define __this_cpu_read_8(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
214 #define __this_cpu_write_8(pcp, val)    percpu_to_op("mov", (pcp), val)
215 #define __this_cpu_add_8(pcp, val)      percpu_to_op("add", (pcp), val)
216 #define __this_cpu_and_8(pcp, val)      percpu_to_op("and", (pcp), val)
217 #define __this_cpu_or_8(pcp, val)       percpu_to_op("or", (pcp), val)
218 #define __this_cpu_xor_8(pcp, val)      percpu_to_op("xor", (pcp), val)
219
220 #define this_cpu_read_8(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
221 #define this_cpu_write_8(pcp, val)      percpu_to_op("mov", (pcp), val)
222 #define this_cpu_add_8(pcp, val)        percpu_to_op("add", (pcp), val)
223 #define this_cpu_and_8(pcp, val)        percpu_to_op("and", (pcp), val)
224 #define this_cpu_or_8(pcp, val)         percpu_to_op("or", (pcp), val)
225 #define this_cpu_xor_8(pcp, val)        percpu_to_op("xor", (pcp), val)
226
227 #define irqsafe_cpu_add_8(pcp, val)     percpu_to_op("add", (pcp), val)
228 #define irqsafe_cpu_and_8(pcp, val)     percpu_to_op("and", (pcp), val)
229 #define irqsafe_cpu_or_8(pcp, val)      percpu_to_op("or", (pcp), val)
230 #define irqsafe_cpu_xor_8(pcp, val)     percpu_to_op("xor", (pcp), val)
231
232 #endif
233
234 /* This is not atomic against other CPUs -- CPU preemption needs to be off */
235 #define x86_test_and_clear_bit_percpu(bit, var)                         \
236 ({                                                                      \
237         int old__;                                                      \
238         asm volatile("btr %2,"__percpu_arg(1)"\n\tsbbl %0,%0"           \
239                      : "=r" (old__), "+m" (per_cpu__##var)              \
240                      : "dIr" (bit));                                    \
241         old__;                                                          \
242 })
243
244 #include <asm-generic/percpu.h>
245
246 /* We can use this directly for local CPU (faster). */
247 DECLARE_PER_CPU(unsigned long, this_cpu_off);
248
249 #endif /* !__ASSEMBLY__ */
250
251 #ifdef CONFIG_SMP
252
253 /*
254  * Define the "EARLY_PER_CPU" macros.  These are used for some per_cpu
255  * variables that are initialized and accessed before there are per_cpu
256  * areas allocated.
257  */
258
259 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)                  \
260         DEFINE_PER_CPU(_type, _name) = _initvalue;                      \
261         __typeof__(_type) _name##_early_map[NR_CPUS] __initdata =       \
262                                 { [0 ... NR_CPUS-1] = _initvalue };     \
263         __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
264
265 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)                      \
266         EXPORT_PER_CPU_SYMBOL(_name)
267
268 #define DECLARE_EARLY_PER_CPU(_type, _name)                     \
269         DECLARE_PER_CPU(_type, _name);                          \
270         extern __typeof__(_type) *_name##_early_ptr;            \
271         extern __typeof__(_type)  _name##_early_map[]
272
273 #define early_per_cpu_ptr(_name) (_name##_early_ptr)
274 #define early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
275 #define early_per_cpu(_name, _cpu)                              \
276         *(early_per_cpu_ptr(_name) ?                            \
277                 &early_per_cpu_ptr(_name)[_cpu] :               \
278                 &per_cpu(_name, _cpu))
279
280 #else   /* !CONFIG_SMP */
281 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)          \
282         DEFINE_PER_CPU(_type, _name) = _initvalue
283
284 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)                      \
285         EXPORT_PER_CPU_SYMBOL(_name)
286
287 #define DECLARE_EARLY_PER_CPU(_type, _name)                     \
288         DECLARE_PER_CPU(_type, _name)
289
290 #define early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
291 #define early_per_cpu_ptr(_name) NULL
292 /* no early_per_cpu_map() */
293
294 #endif  /* !CONFIG_SMP */
295
296 #endif /* _ASM_X86_PERCPU_H */