x86-32, mm: Add an initial page table for core bootstrapping
[pandora-kernel.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:this_cpu_off, reg;                \
29         lea var(reg), reg
30 #define PER_CPU_VAR(var)        %__percpu_seg:var
31 #else /* ! SMP */
32 #define PER_CPU(var, reg)       __percpu_mov_op $var, reg
33 #define PER_CPU_VAR(var)        var
34 #endif  /* SMP */
35
36 #ifdef CONFIG_X86_64_SMP
37 #define INIT_PER_CPU_VAR(var)  init_per_cpu__##var
38 #else
39 #define INIT_PER_CPU_VAR(var)  var
40 #endif
41
42 #else /* ...!ASSEMBLY */
43
44 #include <linux/kernel.h>
45 #include <linux/stringify.h>
46
47 #ifdef CONFIG_SMP
48 #define __percpu_arg(x)         "%%"__stringify(__percpu_seg)":%P" #x
49 #define __my_cpu_offset         percpu_read(this_cpu_off)
50 #else
51 #define __percpu_arg(x)         "%P" #x
52 #endif
53
54 /*
55  * Initialized pointers to per-cpu variables needed for the boot
56  * processor need to use these macros to get the proper address
57  * offset from __per_cpu_load on SMP.
58  *
59  * There also must be an entry in vmlinux_64.lds.S
60  */
61 #define DECLARE_INIT_PER_CPU(var) \
62        extern typeof(var) init_per_cpu_var(var)
63
64 #ifdef CONFIG_X86_64_SMP
65 #define init_per_cpu_var(var)  init_per_cpu__##var
66 #else
67 #define init_per_cpu_var(var)  var
68 #endif
69
70 /* For arch-specific code, we can use direct single-insn ops (they
71  * don't give an lvalue though). */
72 extern void __bad_percpu_size(void);
73
74 #define percpu_to_op(op, var, val)                      \
75 do {                                                    \
76         typedef typeof(var) pto_T__;                    \
77         if (0) {                                        \
78                 pto_T__ pto_tmp__;                      \
79                 pto_tmp__ = (val);                      \
80                 (void)pto_tmp__;                        \
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 /*
108  * Generate a percpu add to memory instruction and optimize code
109  * if one is added or subtracted.
110  */
111 #define percpu_add_op(var, val)                                         \
112 do {                                                                    \
113         typedef typeof(var) pao_T__;                                    \
114         const int pao_ID__ = (__builtin_constant_p(val) &&              \
115                               ((val) == 1 || (val) == -1)) ? (val) : 0; \
116         if (0) {                                                        \
117                 pao_T__ pao_tmp__;                                      \
118                 pao_tmp__ = (val);                                      \
119                 (void)pao_tmp__;                                        \
120         }                                                               \
121         switch (sizeof(var)) {                                          \
122         case 1:                                                         \
123                 if (pao_ID__ == 1)                                      \
124                         asm("incb "__percpu_arg(0) : "+m" (var));       \
125                 else if (pao_ID__ == -1)                                \
126                         asm("decb "__percpu_arg(0) : "+m" (var));       \
127                 else                                                    \
128                         asm("addb %1, "__percpu_arg(0)                  \
129                             : "+m" (var)                                \
130                             : "qi" ((pao_T__)(val)));                   \
131                 break;                                                  \
132         case 2:                                                         \
133                 if (pao_ID__ == 1)                                      \
134                         asm("incw "__percpu_arg(0) : "+m" (var));       \
135                 else if (pao_ID__ == -1)                                \
136                         asm("decw "__percpu_arg(0) : "+m" (var));       \
137                 else                                                    \
138                         asm("addw %1, "__percpu_arg(0)                  \
139                             : "+m" (var)                                \
140                             : "ri" ((pao_T__)(val)));                   \
141                 break;                                                  \
142         case 4:                                                         \
143                 if (pao_ID__ == 1)                                      \
144                         asm("incl "__percpu_arg(0) : "+m" (var));       \
145                 else if (pao_ID__ == -1)                                \
146                         asm("decl "__percpu_arg(0) : "+m" (var));       \
147                 else                                                    \
148                         asm("addl %1, "__percpu_arg(0)                  \
149                             : "+m" (var)                                \
150                             : "ri" ((pao_T__)(val)));                   \
151                 break;                                                  \
152         case 8:                                                         \
153                 if (pao_ID__ == 1)                                      \
154                         asm("incq "__percpu_arg(0) : "+m" (var));       \
155                 else if (pao_ID__ == -1)                                \
156                         asm("decq "__percpu_arg(0) : "+m" (var));       \
157                 else                                                    \
158                         asm("addq %1, "__percpu_arg(0)                  \
159                             : "+m" (var)                                \
160                             : "re" ((pao_T__)(val)));                   \
161                 break;                                                  \
162         default: __bad_percpu_size();                                   \
163         }                                                               \
164 } while (0)
165
166 #define percpu_from_op(op, var, constraint)             \
167 ({                                                      \
168         typeof(var) pfo_ret__;                          \
169         switch (sizeof(var)) {                          \
170         case 1:                                         \
171                 asm(op "b "__percpu_arg(1)",%0"         \
172                     : "=q" (pfo_ret__)                  \
173                     : constraint);                      \
174                 break;                                  \
175         case 2:                                         \
176                 asm(op "w "__percpu_arg(1)",%0"         \
177                     : "=r" (pfo_ret__)                  \
178                     : constraint);                      \
179                 break;                                  \
180         case 4:                                         \
181                 asm(op "l "__percpu_arg(1)",%0"         \
182                     : "=r" (pfo_ret__)                  \
183                     : constraint);                      \
184                 break;                                  \
185         case 8:                                         \
186                 asm(op "q "__percpu_arg(1)",%0"         \
187                     : "=r" (pfo_ret__)                  \
188                     : constraint);                      \
189                 break;                                  \
190         default: __bad_percpu_size();                   \
191         }                                               \
192         pfo_ret__;                                      \
193 })
194
195 #define percpu_unary_op(op, var)                        \
196 ({                                                      \
197         switch (sizeof(var)) {                          \
198         case 1:                                         \
199                 asm(op "b "__percpu_arg(0)              \
200                     : "+m" (var));                      \
201                 break;                                  \
202         case 2:                                         \
203                 asm(op "w "__percpu_arg(0)              \
204                     : "+m" (var));                      \
205                 break;                                  \
206         case 4:                                         \
207                 asm(op "l "__percpu_arg(0)              \
208                     : "+m" (var));                      \
209                 break;                                  \
210         case 8:                                         \
211                 asm(op "q "__percpu_arg(0)              \
212                     : "+m" (var));                      \
213                 break;                                  \
214         default: __bad_percpu_size();                   \
215         }                                               \
216 })
217
218 /*
219  * percpu_read() makes gcc load the percpu variable every time it is
220  * accessed while percpu_read_stable() allows the value to be cached.
221  * percpu_read_stable() is more efficient and can be used if its value
222  * is guaranteed to be valid across cpus.  The current users include
223  * get_current() and get_thread_info() both of which are actually
224  * per-thread variables implemented as per-cpu variables and thus
225  * stable for the duration of the respective task.
226  */
227 #define percpu_read(var)                percpu_from_op("mov", var, "m" (var))
228 #define percpu_read_stable(var)         percpu_from_op("mov", var, "p" (&(var)))
229 #define percpu_write(var, val)          percpu_to_op("mov", var, val)
230 #define percpu_add(var, val)            percpu_add_op(var, val)
231 #define percpu_sub(var, val)            percpu_add_op(var, -(val))
232 #define percpu_and(var, val)            percpu_to_op("and", var, val)
233 #define percpu_or(var, val)             percpu_to_op("or", var, val)
234 #define percpu_xor(var, val)            percpu_to_op("xor", var, val)
235 #define percpu_inc(var)         percpu_unary_op("inc", var)
236
237 #define __this_cpu_read_1(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
238 #define __this_cpu_read_2(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
239 #define __this_cpu_read_4(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
240
241 #define __this_cpu_write_1(pcp, val)    percpu_to_op("mov", (pcp), val)
242 #define __this_cpu_write_2(pcp, val)    percpu_to_op("mov", (pcp), val)
243 #define __this_cpu_write_4(pcp, val)    percpu_to_op("mov", (pcp), val)
244 #define __this_cpu_add_1(pcp, val)      percpu_add_op((pcp), val)
245 #define __this_cpu_add_2(pcp, val)      percpu_add_op((pcp), val)
246 #define __this_cpu_add_4(pcp, val)      percpu_add_op((pcp), val)
247 #define __this_cpu_and_1(pcp, val)      percpu_to_op("and", (pcp), val)
248 #define __this_cpu_and_2(pcp, val)      percpu_to_op("and", (pcp), val)
249 #define __this_cpu_and_4(pcp, val)      percpu_to_op("and", (pcp), val)
250 #define __this_cpu_or_1(pcp, val)       percpu_to_op("or", (pcp), val)
251 #define __this_cpu_or_2(pcp, val)       percpu_to_op("or", (pcp), val)
252 #define __this_cpu_or_4(pcp, val)       percpu_to_op("or", (pcp), val)
253 #define __this_cpu_xor_1(pcp, val)      percpu_to_op("xor", (pcp), val)
254 #define __this_cpu_xor_2(pcp, val)      percpu_to_op("xor", (pcp), val)
255 #define __this_cpu_xor_4(pcp, val)      percpu_to_op("xor", (pcp), val)
256
257 #define this_cpu_read_1(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
258 #define this_cpu_read_2(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
259 #define this_cpu_read_4(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
260 #define this_cpu_write_1(pcp, val)      percpu_to_op("mov", (pcp), val)
261 #define this_cpu_write_2(pcp, val)      percpu_to_op("mov", (pcp), val)
262 #define this_cpu_write_4(pcp, val)      percpu_to_op("mov", (pcp), val)
263 #define this_cpu_add_1(pcp, val)        percpu_add_op((pcp), val)
264 #define this_cpu_add_2(pcp, val)        percpu_add_op((pcp), val)
265 #define this_cpu_add_4(pcp, val)        percpu_add_op((pcp), val)
266 #define this_cpu_and_1(pcp, val)        percpu_to_op("and", (pcp), val)
267 #define this_cpu_and_2(pcp, val)        percpu_to_op("and", (pcp), val)
268 #define this_cpu_and_4(pcp, val)        percpu_to_op("and", (pcp), val)
269 #define this_cpu_or_1(pcp, val)         percpu_to_op("or", (pcp), val)
270 #define this_cpu_or_2(pcp, val)         percpu_to_op("or", (pcp), val)
271 #define this_cpu_or_4(pcp, val)         percpu_to_op("or", (pcp), val)
272 #define this_cpu_xor_1(pcp, val)        percpu_to_op("xor", (pcp), val)
273 #define this_cpu_xor_2(pcp, val)        percpu_to_op("xor", (pcp), val)
274 #define this_cpu_xor_4(pcp, val)        percpu_to_op("xor", (pcp), val)
275
276 #define irqsafe_cpu_add_1(pcp, val)     percpu_add_op((pcp), val)
277 #define irqsafe_cpu_add_2(pcp, val)     percpu_add_op((pcp), val)
278 #define irqsafe_cpu_add_4(pcp, val)     percpu_add_op((pcp), val)
279 #define irqsafe_cpu_and_1(pcp, val)     percpu_to_op("and", (pcp), val)
280 #define irqsafe_cpu_and_2(pcp, val)     percpu_to_op("and", (pcp), val)
281 #define irqsafe_cpu_and_4(pcp, val)     percpu_to_op("and", (pcp), val)
282 #define irqsafe_cpu_or_1(pcp, val)      percpu_to_op("or", (pcp), val)
283 #define irqsafe_cpu_or_2(pcp, val)      percpu_to_op("or", (pcp), val)
284 #define irqsafe_cpu_or_4(pcp, val)      percpu_to_op("or", (pcp), val)
285 #define irqsafe_cpu_xor_1(pcp, val)     percpu_to_op("xor", (pcp), val)
286 #define irqsafe_cpu_xor_2(pcp, val)     percpu_to_op("xor", (pcp), val)
287 #define irqsafe_cpu_xor_4(pcp, val)     percpu_to_op("xor", (pcp), val)
288
289 /*
290  * Per cpu atomic 64 bit operations are only available under 64 bit.
291  * 32 bit must fall back to generic operations.
292  */
293 #ifdef CONFIG_X86_64
294 #define __this_cpu_read_8(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
295 #define __this_cpu_write_8(pcp, val)    percpu_to_op("mov", (pcp), val)
296 #define __this_cpu_add_8(pcp, val)      percpu_add_op((pcp), val)
297 #define __this_cpu_and_8(pcp, val)      percpu_to_op("and", (pcp), val)
298 #define __this_cpu_or_8(pcp, val)       percpu_to_op("or", (pcp), val)
299 #define __this_cpu_xor_8(pcp, val)      percpu_to_op("xor", (pcp), val)
300
301 #define this_cpu_read_8(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
302 #define this_cpu_write_8(pcp, val)      percpu_to_op("mov", (pcp), val)
303 #define this_cpu_add_8(pcp, val)        percpu_add_op((pcp), val)
304 #define this_cpu_and_8(pcp, val)        percpu_to_op("and", (pcp), val)
305 #define this_cpu_or_8(pcp, val)         percpu_to_op("or", (pcp), val)
306 #define this_cpu_xor_8(pcp, val)        percpu_to_op("xor", (pcp), val)
307
308 #define irqsafe_cpu_add_8(pcp, val)     percpu_add_op((pcp), val)
309 #define irqsafe_cpu_and_8(pcp, val)     percpu_to_op("and", (pcp), val)
310 #define irqsafe_cpu_or_8(pcp, val)      percpu_to_op("or", (pcp), val)
311 #define irqsafe_cpu_xor_8(pcp, val)     percpu_to_op("xor", (pcp), val)
312
313 #endif
314
315 /* This is not atomic against other CPUs -- CPU preemption needs to be off */
316 #define x86_test_and_clear_bit_percpu(bit, var)                         \
317 ({                                                                      \
318         int old__;                                                      \
319         asm volatile("btr %2,"__percpu_arg(1)"\n\tsbbl %0,%0"           \
320                      : "=r" (old__), "+m" (var)                         \
321                      : "dIr" (bit));                                    \
322         old__;                                                          \
323 })
324
325 #include <asm-generic/percpu.h>
326
327 /* We can use this directly for local CPU (faster). */
328 DECLARE_PER_CPU(unsigned long, this_cpu_off);
329
330 #endif /* !__ASSEMBLY__ */
331
332 #ifdef CONFIG_SMP
333
334 /*
335  * Define the "EARLY_PER_CPU" macros.  These are used for some per_cpu
336  * variables that are initialized and accessed before there are per_cpu
337  * areas allocated.
338  */
339
340 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)                  \
341         DEFINE_PER_CPU(_type, _name) = _initvalue;                      \
342         __typeof__(_type) _name##_early_map[NR_CPUS] __initdata =       \
343                                 { [0 ... NR_CPUS-1] = _initvalue };     \
344         __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
345
346 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)                      \
347         EXPORT_PER_CPU_SYMBOL(_name)
348
349 #define DECLARE_EARLY_PER_CPU(_type, _name)                     \
350         DECLARE_PER_CPU(_type, _name);                          \
351         extern __typeof__(_type) *_name##_early_ptr;            \
352         extern __typeof__(_type)  _name##_early_map[]
353
354 #define early_per_cpu_ptr(_name) (_name##_early_ptr)
355 #define early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
356 #define early_per_cpu(_name, _cpu)                              \
357         *(early_per_cpu_ptr(_name) ?                            \
358                 &early_per_cpu_ptr(_name)[_cpu] :               \
359                 &per_cpu(_name, _cpu))
360
361 #else   /* !CONFIG_SMP */
362 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)          \
363         DEFINE_PER_CPU(_type, _name) = _initvalue
364
365 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)                      \
366         EXPORT_PER_CPU_SYMBOL(_name)
367
368 #define DECLARE_EARLY_PER_CPU(_type, _name)                     \
369         DECLARE_PER_CPU(_type, _name)
370
371 #define early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
372 #define early_per_cpu_ptr(_name) NULL
373 /* no early_per_cpu_map() */
374
375 #endif  /* !CONFIG_SMP */
376
377 #endif /* _ASM_X86_PERCPU_H */