xref: /openbmc/linux/arch/x86/include/asm/percpu.h (revision 0da85d1e)
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_prefix		"%%"__stringify(__percpu_seg)":"
49 #define __my_cpu_offset		this_cpu_read(this_cpu_off)
50 
51 /*
52  * Compared to the generic __my_cpu_offset version, the following
53  * saves one instruction and avoids clobbering a temp register.
54  */
55 #define arch_raw_cpu_ptr(ptr)				\
56 ({							\
57 	unsigned long tcp_ptr__;			\
58 	asm volatile("add " __percpu_arg(1) ", %0"	\
59 		     : "=r" (tcp_ptr__)			\
60 		     : "m" (this_cpu_off), "0" (ptr));	\
61 	(typeof(*(ptr)) __kernel __force *)tcp_ptr__;	\
62 })
63 #else
64 #define __percpu_prefix		""
65 #endif
66 
67 #define __percpu_arg(x)		__percpu_prefix "%" #x
68 
69 /*
70  * Initialized pointers to per-cpu variables needed for the boot
71  * processor need to use these macros to get the proper address
72  * offset from __per_cpu_load on SMP.
73  *
74  * There also must be an entry in vmlinux_64.lds.S
75  */
76 #define DECLARE_INIT_PER_CPU(var) \
77        extern typeof(var) init_per_cpu_var(var)
78 
79 #ifdef CONFIG_X86_64_SMP
80 #define init_per_cpu_var(var)  init_per_cpu__##var
81 #else
82 #define init_per_cpu_var(var)  var
83 #endif
84 
85 /* For arch-specific code, we can use direct single-insn ops (they
86  * don't give an lvalue though). */
87 extern void __bad_percpu_size(void);
88 
89 #define percpu_to_op(op, var, val)			\
90 do {							\
91 	typedef typeof(var) pto_T__;			\
92 	if (0) {					\
93 		pto_T__ pto_tmp__;			\
94 		pto_tmp__ = (val);			\
95 		(void)pto_tmp__;			\
96 	}						\
97 	switch (sizeof(var)) {				\
98 	case 1:						\
99 		asm(op "b %1,"__percpu_arg(0)		\
100 		    : "+m" (var)			\
101 		    : "qi" ((pto_T__)(val)));		\
102 		break;					\
103 	case 2:						\
104 		asm(op "w %1,"__percpu_arg(0)		\
105 		    : "+m" (var)			\
106 		    : "ri" ((pto_T__)(val)));		\
107 		break;					\
108 	case 4:						\
109 		asm(op "l %1,"__percpu_arg(0)		\
110 		    : "+m" (var)			\
111 		    : "ri" ((pto_T__)(val)));		\
112 		break;					\
113 	case 8:						\
114 		asm(op "q %1,"__percpu_arg(0)		\
115 		    : "+m" (var)			\
116 		    : "re" ((pto_T__)(val)));		\
117 		break;					\
118 	default: __bad_percpu_size();			\
119 	}						\
120 } while (0)
121 
122 /*
123  * Generate a percpu add to memory instruction and optimize code
124  * if one is added or subtracted.
125  */
126 #define percpu_add_op(var, val)						\
127 do {									\
128 	typedef typeof(var) pao_T__;					\
129 	const int pao_ID__ = (__builtin_constant_p(val) &&		\
130 			      ((val) == 1 || (val) == -1)) ?		\
131 				(int)(val) : 0;				\
132 	if (0) {							\
133 		pao_T__ pao_tmp__;					\
134 		pao_tmp__ = (val);					\
135 		(void)pao_tmp__;					\
136 	}								\
137 	switch (sizeof(var)) {						\
138 	case 1:								\
139 		if (pao_ID__ == 1)					\
140 			asm("incb "__percpu_arg(0) : "+m" (var));	\
141 		else if (pao_ID__ == -1)				\
142 			asm("decb "__percpu_arg(0) : "+m" (var));	\
143 		else							\
144 			asm("addb %1, "__percpu_arg(0)			\
145 			    : "+m" (var)				\
146 			    : "qi" ((pao_T__)(val)));			\
147 		break;							\
148 	case 2:								\
149 		if (pao_ID__ == 1)					\
150 			asm("incw "__percpu_arg(0) : "+m" (var));	\
151 		else if (pao_ID__ == -1)				\
152 			asm("decw "__percpu_arg(0) : "+m" (var));	\
153 		else							\
154 			asm("addw %1, "__percpu_arg(0)			\
155 			    : "+m" (var)				\
156 			    : "ri" ((pao_T__)(val)));			\
157 		break;							\
158 	case 4:								\
159 		if (pao_ID__ == 1)					\
160 			asm("incl "__percpu_arg(0) : "+m" (var));	\
161 		else if (pao_ID__ == -1)				\
162 			asm("decl "__percpu_arg(0) : "+m" (var));	\
163 		else							\
164 			asm("addl %1, "__percpu_arg(0)			\
165 			    : "+m" (var)				\
166 			    : "ri" ((pao_T__)(val)));			\
167 		break;							\
168 	case 8:								\
169 		if (pao_ID__ == 1)					\
170 			asm("incq "__percpu_arg(0) : "+m" (var));	\
171 		else if (pao_ID__ == -1)				\
172 			asm("decq "__percpu_arg(0) : "+m" (var));	\
173 		else							\
174 			asm("addq %1, "__percpu_arg(0)			\
175 			    : "+m" (var)				\
176 			    : "re" ((pao_T__)(val)));			\
177 		break;							\
178 	default: __bad_percpu_size();					\
179 	}								\
180 } while (0)
181 
182 #define percpu_from_op(op, var)				\
183 ({							\
184 	typeof(var) pfo_ret__;				\
185 	switch (sizeof(var)) {				\
186 	case 1:						\
187 		asm(op "b "__percpu_arg(1)",%0"		\
188 		    : "=q" (pfo_ret__)			\
189 		    : "m" (var));			\
190 		break;					\
191 	case 2:						\
192 		asm(op "w "__percpu_arg(1)",%0"		\
193 		    : "=r" (pfo_ret__)			\
194 		    : "m" (var));			\
195 		break;					\
196 	case 4:						\
197 		asm(op "l "__percpu_arg(1)",%0"		\
198 		    : "=r" (pfo_ret__)			\
199 		    : "m" (var));			\
200 		break;					\
201 	case 8:						\
202 		asm(op "q "__percpu_arg(1)",%0"		\
203 		    : "=r" (pfo_ret__)			\
204 		    : "m" (var));			\
205 		break;					\
206 	default: __bad_percpu_size();			\
207 	}						\
208 	pfo_ret__;					\
209 })
210 
211 #define percpu_stable_op(op, var)			\
212 ({							\
213 	typeof(var) pfo_ret__;				\
214 	switch (sizeof(var)) {				\
215 	case 1:						\
216 		asm(op "b "__percpu_arg(P1)",%0"	\
217 		    : "=q" (pfo_ret__)			\
218 		    : "p" (&(var)));			\
219 		break;					\
220 	case 2:						\
221 		asm(op "w "__percpu_arg(P1)",%0"	\
222 		    : "=r" (pfo_ret__)			\
223 		    : "p" (&(var)));			\
224 		break;					\
225 	case 4:						\
226 		asm(op "l "__percpu_arg(P1)",%0"	\
227 		    : "=r" (pfo_ret__)			\
228 		    : "p" (&(var)));			\
229 		break;					\
230 	case 8:						\
231 		asm(op "q "__percpu_arg(P1)",%0"	\
232 		    : "=r" (pfo_ret__)			\
233 		    : "p" (&(var)));			\
234 		break;					\
235 	default: __bad_percpu_size();			\
236 	}						\
237 	pfo_ret__;					\
238 })
239 
240 #define percpu_unary_op(op, var)			\
241 ({							\
242 	switch (sizeof(var)) {				\
243 	case 1:						\
244 		asm(op "b "__percpu_arg(0)		\
245 		    : "+m" (var));			\
246 		break;					\
247 	case 2:						\
248 		asm(op "w "__percpu_arg(0)		\
249 		    : "+m" (var));			\
250 		break;					\
251 	case 4:						\
252 		asm(op "l "__percpu_arg(0)		\
253 		    : "+m" (var));			\
254 		break;					\
255 	case 8:						\
256 		asm(op "q "__percpu_arg(0)		\
257 		    : "+m" (var));			\
258 		break;					\
259 	default: __bad_percpu_size();			\
260 	}						\
261 })
262 
263 /*
264  * Add return operation
265  */
266 #define percpu_add_return_op(var, val)					\
267 ({									\
268 	typeof(var) paro_ret__ = val;					\
269 	switch (sizeof(var)) {						\
270 	case 1:								\
271 		asm("xaddb %0, "__percpu_arg(1)				\
272 			    : "+q" (paro_ret__), "+m" (var)		\
273 			    : : "memory");				\
274 		break;							\
275 	case 2:								\
276 		asm("xaddw %0, "__percpu_arg(1)				\
277 			    : "+r" (paro_ret__), "+m" (var)		\
278 			    : : "memory");				\
279 		break;							\
280 	case 4:								\
281 		asm("xaddl %0, "__percpu_arg(1)				\
282 			    : "+r" (paro_ret__), "+m" (var)		\
283 			    : : "memory");				\
284 		break;							\
285 	case 8:								\
286 		asm("xaddq %0, "__percpu_arg(1)				\
287 			    : "+re" (paro_ret__), "+m" (var)		\
288 			    : : "memory");				\
289 		break;							\
290 	default: __bad_percpu_size();					\
291 	}								\
292 	paro_ret__ += val;						\
293 	paro_ret__;							\
294 })
295 
296 /*
297  * xchg is implemented using cmpxchg without a lock prefix. xchg is
298  * expensive due to the implied lock prefix.  The processor cannot prefetch
299  * cachelines if xchg is used.
300  */
301 #define percpu_xchg_op(var, nval)					\
302 ({									\
303 	typeof(var) pxo_ret__;						\
304 	typeof(var) pxo_new__ = (nval);					\
305 	switch (sizeof(var)) {						\
306 	case 1:								\
307 		asm("\n\tmov "__percpu_arg(1)",%%al"			\
308 		    "\n1:\tcmpxchgb %2, "__percpu_arg(1)		\
309 		    "\n\tjnz 1b"					\
310 			    : "=&a" (pxo_ret__), "+m" (var)		\
311 			    : "q" (pxo_new__)				\
312 			    : "memory");				\
313 		break;							\
314 	case 2:								\
315 		asm("\n\tmov "__percpu_arg(1)",%%ax"			\
316 		    "\n1:\tcmpxchgw %2, "__percpu_arg(1)		\
317 		    "\n\tjnz 1b"					\
318 			    : "=&a" (pxo_ret__), "+m" (var)		\
319 			    : "r" (pxo_new__)				\
320 			    : "memory");				\
321 		break;							\
322 	case 4:								\
323 		asm("\n\tmov "__percpu_arg(1)",%%eax"			\
324 		    "\n1:\tcmpxchgl %2, "__percpu_arg(1)		\
325 		    "\n\tjnz 1b"					\
326 			    : "=&a" (pxo_ret__), "+m" (var)		\
327 			    : "r" (pxo_new__)				\
328 			    : "memory");				\
329 		break;							\
330 	case 8:								\
331 		asm("\n\tmov "__percpu_arg(1)",%%rax"			\
332 		    "\n1:\tcmpxchgq %2, "__percpu_arg(1)		\
333 		    "\n\tjnz 1b"					\
334 			    : "=&a" (pxo_ret__), "+m" (var)		\
335 			    : "r" (pxo_new__)				\
336 			    : "memory");				\
337 		break;							\
338 	default: __bad_percpu_size();					\
339 	}								\
340 	pxo_ret__;							\
341 })
342 
343 /*
344  * cmpxchg has no such implied lock semantics as a result it is much
345  * more efficient for cpu local operations.
346  */
347 #define percpu_cmpxchg_op(var, oval, nval)				\
348 ({									\
349 	typeof(var) pco_ret__;						\
350 	typeof(var) pco_old__ = (oval);					\
351 	typeof(var) pco_new__ = (nval);					\
352 	switch (sizeof(var)) {						\
353 	case 1:								\
354 		asm("cmpxchgb %2, "__percpu_arg(1)			\
355 			    : "=a" (pco_ret__), "+m" (var)		\
356 			    : "q" (pco_new__), "0" (pco_old__)		\
357 			    : "memory");				\
358 		break;							\
359 	case 2:								\
360 		asm("cmpxchgw %2, "__percpu_arg(1)			\
361 			    : "=a" (pco_ret__), "+m" (var)		\
362 			    : "r" (pco_new__), "0" (pco_old__)		\
363 			    : "memory");				\
364 		break;							\
365 	case 4:								\
366 		asm("cmpxchgl %2, "__percpu_arg(1)			\
367 			    : "=a" (pco_ret__), "+m" (var)		\
368 			    : "r" (pco_new__), "0" (pco_old__)		\
369 			    : "memory");				\
370 		break;							\
371 	case 8:								\
372 		asm("cmpxchgq %2, "__percpu_arg(1)			\
373 			    : "=a" (pco_ret__), "+m" (var)		\
374 			    : "r" (pco_new__), "0" (pco_old__)		\
375 			    : "memory");				\
376 		break;							\
377 	default: __bad_percpu_size();					\
378 	}								\
379 	pco_ret__;							\
380 })
381 
382 /*
383  * this_cpu_read() makes gcc load the percpu variable every time it is
384  * accessed while this_cpu_read_stable() allows the value to be cached.
385  * this_cpu_read_stable() is more efficient and can be used if its value
386  * is guaranteed to be valid across cpus.  The current users include
387  * get_current() and get_thread_info() both of which are actually
388  * per-thread variables implemented as per-cpu variables and thus
389  * stable for the duration of the respective task.
390  */
391 #define this_cpu_read_stable(var)	percpu_stable_op("mov", var)
392 
393 #define raw_cpu_read_1(pcp)		percpu_from_op("mov", pcp)
394 #define raw_cpu_read_2(pcp)		percpu_from_op("mov", pcp)
395 #define raw_cpu_read_4(pcp)		percpu_from_op("mov", pcp)
396 
397 #define raw_cpu_write_1(pcp, val)	percpu_to_op("mov", (pcp), val)
398 #define raw_cpu_write_2(pcp, val)	percpu_to_op("mov", (pcp), val)
399 #define raw_cpu_write_4(pcp, val)	percpu_to_op("mov", (pcp), val)
400 #define raw_cpu_add_1(pcp, val)		percpu_add_op((pcp), val)
401 #define raw_cpu_add_2(pcp, val)		percpu_add_op((pcp), val)
402 #define raw_cpu_add_4(pcp, val)		percpu_add_op((pcp), val)
403 #define raw_cpu_and_1(pcp, val)		percpu_to_op("and", (pcp), val)
404 #define raw_cpu_and_2(pcp, val)		percpu_to_op("and", (pcp), val)
405 #define raw_cpu_and_4(pcp, val)		percpu_to_op("and", (pcp), val)
406 #define raw_cpu_or_1(pcp, val)		percpu_to_op("or", (pcp), val)
407 #define raw_cpu_or_2(pcp, val)		percpu_to_op("or", (pcp), val)
408 #define raw_cpu_or_4(pcp, val)		percpu_to_op("or", (pcp), val)
409 #define raw_cpu_xchg_1(pcp, val)	percpu_xchg_op(pcp, val)
410 #define raw_cpu_xchg_2(pcp, val)	percpu_xchg_op(pcp, val)
411 #define raw_cpu_xchg_4(pcp, val)	percpu_xchg_op(pcp, val)
412 
413 #define this_cpu_read_1(pcp)		percpu_from_op("mov", pcp)
414 #define this_cpu_read_2(pcp)		percpu_from_op("mov", pcp)
415 #define this_cpu_read_4(pcp)		percpu_from_op("mov", pcp)
416 #define this_cpu_write_1(pcp, val)	percpu_to_op("mov", (pcp), val)
417 #define this_cpu_write_2(pcp, val)	percpu_to_op("mov", (pcp), val)
418 #define this_cpu_write_4(pcp, val)	percpu_to_op("mov", (pcp), val)
419 #define this_cpu_add_1(pcp, val)	percpu_add_op((pcp), val)
420 #define this_cpu_add_2(pcp, val)	percpu_add_op((pcp), val)
421 #define this_cpu_add_4(pcp, val)	percpu_add_op((pcp), val)
422 #define this_cpu_and_1(pcp, val)	percpu_to_op("and", (pcp), val)
423 #define this_cpu_and_2(pcp, val)	percpu_to_op("and", (pcp), val)
424 #define this_cpu_and_4(pcp, val)	percpu_to_op("and", (pcp), val)
425 #define this_cpu_or_1(pcp, val)		percpu_to_op("or", (pcp), val)
426 #define this_cpu_or_2(pcp, val)		percpu_to_op("or", (pcp), val)
427 #define this_cpu_or_4(pcp, val)		percpu_to_op("or", (pcp), val)
428 #define this_cpu_xchg_1(pcp, nval)	percpu_xchg_op(pcp, nval)
429 #define this_cpu_xchg_2(pcp, nval)	percpu_xchg_op(pcp, nval)
430 #define this_cpu_xchg_4(pcp, nval)	percpu_xchg_op(pcp, nval)
431 
432 #define raw_cpu_add_return_1(pcp, val)		percpu_add_return_op(pcp, val)
433 #define raw_cpu_add_return_2(pcp, val)		percpu_add_return_op(pcp, val)
434 #define raw_cpu_add_return_4(pcp, val)		percpu_add_return_op(pcp, val)
435 #define raw_cpu_cmpxchg_1(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
436 #define raw_cpu_cmpxchg_2(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
437 #define raw_cpu_cmpxchg_4(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
438 
439 #define this_cpu_add_return_1(pcp, val)		percpu_add_return_op(pcp, val)
440 #define this_cpu_add_return_2(pcp, val)		percpu_add_return_op(pcp, val)
441 #define this_cpu_add_return_4(pcp, val)		percpu_add_return_op(pcp, val)
442 #define this_cpu_cmpxchg_1(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
443 #define this_cpu_cmpxchg_2(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
444 #define this_cpu_cmpxchg_4(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
445 
446 #ifdef CONFIG_X86_CMPXCHG64
447 #define percpu_cmpxchg8b_double(pcp1, pcp2, o1, o2, n1, n2)		\
448 ({									\
449 	bool __ret;							\
450 	typeof(pcp1) __o1 = (o1), __n1 = (n1);				\
451 	typeof(pcp2) __o2 = (o2), __n2 = (n2);				\
452 	asm volatile("cmpxchg8b "__percpu_arg(1)"\n\tsetz %0\n\t"	\
453 		    : "=a" (__ret), "+m" (pcp1), "+m" (pcp2), "+d" (__o2) \
454 		    :  "b" (__n1), "c" (__n2), "a" (__o1));		\
455 	__ret;								\
456 })
457 
458 #define raw_cpu_cmpxchg_double_4	percpu_cmpxchg8b_double
459 #define this_cpu_cmpxchg_double_4	percpu_cmpxchg8b_double
460 #endif /* CONFIG_X86_CMPXCHG64 */
461 
462 /*
463  * Per cpu atomic 64 bit operations are only available under 64 bit.
464  * 32 bit must fall back to generic operations.
465  */
466 #ifdef CONFIG_X86_64
467 #define raw_cpu_read_8(pcp)			percpu_from_op("mov", pcp)
468 #define raw_cpu_write_8(pcp, val)		percpu_to_op("mov", (pcp), val)
469 #define raw_cpu_add_8(pcp, val)			percpu_add_op((pcp), val)
470 #define raw_cpu_and_8(pcp, val)			percpu_to_op("and", (pcp), val)
471 #define raw_cpu_or_8(pcp, val)			percpu_to_op("or", (pcp), val)
472 #define raw_cpu_add_return_8(pcp, val)		percpu_add_return_op(pcp, val)
473 #define raw_cpu_xchg_8(pcp, nval)		percpu_xchg_op(pcp, nval)
474 #define raw_cpu_cmpxchg_8(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
475 
476 #define this_cpu_read_8(pcp)			percpu_from_op("mov", pcp)
477 #define this_cpu_write_8(pcp, val)		percpu_to_op("mov", (pcp), val)
478 #define this_cpu_add_8(pcp, val)		percpu_add_op((pcp), val)
479 #define this_cpu_and_8(pcp, val)		percpu_to_op("and", (pcp), val)
480 #define this_cpu_or_8(pcp, val)			percpu_to_op("or", (pcp), val)
481 #define this_cpu_add_return_8(pcp, val)		percpu_add_return_op(pcp, val)
482 #define this_cpu_xchg_8(pcp, nval)		percpu_xchg_op(pcp, nval)
483 #define this_cpu_cmpxchg_8(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
484 
485 /*
486  * Pretty complex macro to generate cmpxchg16 instruction.  The instruction
487  * is not supported on early AMD64 processors so we must be able to emulate
488  * it in software.  The address used in the cmpxchg16 instruction must be
489  * aligned to a 16 byte boundary.
490  */
491 #define percpu_cmpxchg16b_double(pcp1, pcp2, o1, o2, n1, n2)		\
492 ({									\
493 	bool __ret;							\
494 	typeof(pcp1) __o1 = (o1), __n1 = (n1);				\
495 	typeof(pcp2) __o2 = (o2), __n2 = (n2);				\
496 	alternative_io("leaq %P1,%%rsi\n\tcall this_cpu_cmpxchg16b_emu\n\t", \
497 		       "cmpxchg16b " __percpu_arg(1) "\n\tsetz %0\n\t",	\
498 		       X86_FEATURE_CX16,				\
499 		       ASM_OUTPUT2("=a" (__ret), "+m" (pcp1),		\
500 				   "+m" (pcp2), "+d" (__o2)),		\
501 		       "b" (__n1), "c" (__n2), "a" (__o1) : "rsi");	\
502 	__ret;								\
503 })
504 
505 #define raw_cpu_cmpxchg_double_8	percpu_cmpxchg16b_double
506 #define this_cpu_cmpxchg_double_8	percpu_cmpxchg16b_double
507 
508 #endif
509 
510 /* This is not atomic against other CPUs -- CPU preemption needs to be off */
511 #define x86_test_and_clear_bit_percpu(bit, var)				\
512 ({									\
513 	int old__;							\
514 	asm volatile("btr %2,"__percpu_arg(1)"\n\tsbbl %0,%0"		\
515 		     : "=r" (old__), "+m" (var)				\
516 		     : "dIr" (bit));					\
517 	old__;								\
518 })
519 
520 static __always_inline int x86_this_cpu_constant_test_bit(unsigned int nr,
521                         const unsigned long __percpu *addr)
522 {
523 	unsigned long __percpu *a = (unsigned long *)addr + nr / BITS_PER_LONG;
524 
525 #ifdef CONFIG_X86_64
526 	return ((1UL << (nr % BITS_PER_LONG)) & raw_cpu_read_8(*a)) != 0;
527 #else
528 	return ((1UL << (nr % BITS_PER_LONG)) & raw_cpu_read_4(*a)) != 0;
529 #endif
530 }
531 
532 static inline int x86_this_cpu_variable_test_bit(int nr,
533                         const unsigned long __percpu *addr)
534 {
535 	int oldbit;
536 
537 	asm volatile("bt "__percpu_arg(2)",%1\n\t"
538 			"sbb %0,%0"
539 			: "=r" (oldbit)
540 			: "m" (*(unsigned long *)addr), "Ir" (nr));
541 
542 	return oldbit;
543 }
544 
545 #define x86_this_cpu_test_bit(nr, addr)			\
546 	(__builtin_constant_p((nr))			\
547 	 ? x86_this_cpu_constant_test_bit((nr), (addr))	\
548 	 : x86_this_cpu_variable_test_bit((nr), (addr)))
549 
550 
551 #include <asm-generic/percpu.h>
552 
553 /* We can use this directly for local CPU (faster). */
554 DECLARE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off);
555 
556 #endif /* !__ASSEMBLY__ */
557 
558 #ifdef CONFIG_SMP
559 
560 /*
561  * Define the "EARLY_PER_CPU" macros.  These are used for some per_cpu
562  * variables that are initialized and accessed before there are per_cpu
563  * areas allocated.
564  */
565 
566 #define	DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)			\
567 	DEFINE_PER_CPU(_type, _name) = _initvalue;			\
568 	__typeof__(_type) _name##_early_map[NR_CPUS] __initdata =	\
569 				{ [0 ... NR_CPUS-1] = _initvalue };	\
570 	__typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
571 
572 #define DEFINE_EARLY_PER_CPU_READ_MOSTLY(_type, _name, _initvalue)	\
573 	DEFINE_PER_CPU_READ_MOSTLY(_type, _name) = _initvalue;		\
574 	__typeof__(_type) _name##_early_map[NR_CPUS] __initdata =	\
575 				{ [0 ... NR_CPUS-1] = _initvalue };	\
576 	__typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
577 
578 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)			\
579 	EXPORT_PER_CPU_SYMBOL(_name)
580 
581 #define DECLARE_EARLY_PER_CPU(_type, _name)			\
582 	DECLARE_PER_CPU(_type, _name);				\
583 	extern __typeof__(_type) *_name##_early_ptr;		\
584 	extern __typeof__(_type)  _name##_early_map[]
585 
586 #define DECLARE_EARLY_PER_CPU_READ_MOSTLY(_type, _name)		\
587 	DECLARE_PER_CPU_READ_MOSTLY(_type, _name);		\
588 	extern __typeof__(_type) *_name##_early_ptr;		\
589 	extern __typeof__(_type)  _name##_early_map[]
590 
591 #define	early_per_cpu_ptr(_name) (_name##_early_ptr)
592 #define	early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
593 #define	early_per_cpu(_name, _cpu) 				\
594 	*(early_per_cpu_ptr(_name) ?				\
595 		&early_per_cpu_ptr(_name)[_cpu] :		\
596 		&per_cpu(_name, _cpu))
597 
598 #else	/* !CONFIG_SMP */
599 #define	DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)		\
600 	DEFINE_PER_CPU(_type, _name) = _initvalue
601 
602 #define DEFINE_EARLY_PER_CPU_READ_MOSTLY(_type, _name, _initvalue)	\
603 	DEFINE_PER_CPU_READ_MOSTLY(_type, _name) = _initvalue
604 
605 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)			\
606 	EXPORT_PER_CPU_SYMBOL(_name)
607 
608 #define DECLARE_EARLY_PER_CPU(_type, _name)			\
609 	DECLARE_PER_CPU(_type, _name)
610 
611 #define DECLARE_EARLY_PER_CPU_READ_MOSTLY(_type, _name)		\
612 	DECLARE_PER_CPU_READ_MOSTLY(_type, _name)
613 
614 #define	early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
615 #define	early_per_cpu_ptr(_name) NULL
616 /* no early_per_cpu_map() */
617 
618 #endif	/* !CONFIG_SMP */
619 
620 #endif /* _ASM_X86_PERCPU_H */
621