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		percpu_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 __this_cpu_ptr(ptr)				\
56 ({							\
57 	unsigned long tcp_ptr__;			\
58 	__verify_pcpu_ptr(ptr);				\
59 	asm volatile("add " __percpu_arg(1) ", %0"	\
60 		     : "=r" (tcp_ptr__)			\
61 		     : "m" (this_cpu_off), "0" (ptr));	\
62 	(typeof(*(ptr)) __kernel __force *)tcp_ptr__;	\
63 })
64 #else
65 #define __percpu_prefix		""
66 #endif
67 
68 #define __percpu_arg(x)		__percpu_prefix "%P" #x
69 
70 /*
71  * Initialized pointers to per-cpu variables needed for the boot
72  * processor need to use these macros to get the proper address
73  * offset from __per_cpu_load on SMP.
74  *
75  * There also must be an entry in vmlinux_64.lds.S
76  */
77 #define DECLARE_INIT_PER_CPU(var) \
78        extern typeof(var) init_per_cpu_var(var)
79 
80 #ifdef CONFIG_X86_64_SMP
81 #define init_per_cpu_var(var)  init_per_cpu__##var
82 #else
83 #define init_per_cpu_var(var)  var
84 #endif
85 
86 /* For arch-specific code, we can use direct single-insn ops (they
87  * don't give an lvalue though). */
88 extern void __bad_percpu_size(void);
89 
90 #define percpu_to_op(op, var, val)			\
91 do {							\
92 	typedef typeof(var) pto_T__;			\
93 	if (0) {					\
94 		pto_T__ pto_tmp__;			\
95 		pto_tmp__ = (val);			\
96 		(void)pto_tmp__;			\
97 	}						\
98 	switch (sizeof(var)) {				\
99 	case 1:						\
100 		asm(op "b %1,"__percpu_arg(0)		\
101 		    : "+m" (var)			\
102 		    : "qi" ((pto_T__)(val)));		\
103 		break;					\
104 	case 2:						\
105 		asm(op "w %1,"__percpu_arg(0)		\
106 		    : "+m" (var)			\
107 		    : "ri" ((pto_T__)(val)));		\
108 		break;					\
109 	case 4:						\
110 		asm(op "l %1,"__percpu_arg(0)		\
111 		    : "+m" (var)			\
112 		    : "ri" ((pto_T__)(val)));		\
113 		break;					\
114 	case 8:						\
115 		asm(op "q %1,"__percpu_arg(0)		\
116 		    : "+m" (var)			\
117 		    : "re" ((pto_T__)(val)));		\
118 		break;					\
119 	default: __bad_percpu_size();			\
120 	}						\
121 } while (0)
122 
123 /*
124  * Generate a percpu add to memory instruction and optimize code
125  * if one is added or subtracted.
126  */
127 #define percpu_add_op(var, val)						\
128 do {									\
129 	typedef typeof(var) pao_T__;					\
130 	const int pao_ID__ = (__builtin_constant_p(val) &&		\
131 			      ((val) == 1 || (val) == -1)) ? (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, constraint)		\
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 		    : constraint);			\
190 		break;					\
191 	case 2:						\
192 		asm(op "w "__percpu_arg(1)",%0"		\
193 		    : "=r" (pfo_ret__)			\
194 		    : constraint);			\
195 		break;					\
196 	case 4:						\
197 		asm(op "l "__percpu_arg(1)",%0"		\
198 		    : "=r" (pfo_ret__)			\
199 		    : constraint);			\
200 		break;					\
201 	case 8:						\
202 		asm(op "q "__percpu_arg(1)",%0"		\
203 		    : "=r" (pfo_ret__)			\
204 		    : constraint);			\
205 		break;					\
206 	default: __bad_percpu_size();			\
207 	}						\
208 	pfo_ret__;					\
209 })
210 
211 #define percpu_unary_op(op, var)			\
212 ({							\
213 	switch (sizeof(var)) {				\
214 	case 1:						\
215 		asm(op "b "__percpu_arg(0)		\
216 		    : "+m" (var));			\
217 		break;					\
218 	case 2:						\
219 		asm(op "w "__percpu_arg(0)		\
220 		    : "+m" (var));			\
221 		break;					\
222 	case 4:						\
223 		asm(op "l "__percpu_arg(0)		\
224 		    : "+m" (var));			\
225 		break;					\
226 	case 8:						\
227 		asm(op "q "__percpu_arg(0)		\
228 		    : "+m" (var));			\
229 		break;					\
230 	default: __bad_percpu_size();			\
231 	}						\
232 })
233 
234 /*
235  * Add return operation
236  */
237 #define percpu_add_return_op(var, val)					\
238 ({									\
239 	typeof(var) paro_ret__ = val;					\
240 	switch (sizeof(var)) {						\
241 	case 1:								\
242 		asm("xaddb %0, "__percpu_arg(1)				\
243 			    : "+q" (paro_ret__), "+m" (var)		\
244 			    : : "memory");				\
245 		break;							\
246 	case 2:								\
247 		asm("xaddw %0, "__percpu_arg(1)				\
248 			    : "+r" (paro_ret__), "+m" (var)		\
249 			    : : "memory");				\
250 		break;							\
251 	case 4:								\
252 		asm("xaddl %0, "__percpu_arg(1)				\
253 			    : "+r" (paro_ret__), "+m" (var)		\
254 			    : : "memory");				\
255 		break;							\
256 	case 8:								\
257 		asm("xaddq %0, "__percpu_arg(1)				\
258 			    : "+re" (paro_ret__), "+m" (var)		\
259 			    : : "memory");				\
260 		break;							\
261 	default: __bad_percpu_size();					\
262 	}								\
263 	paro_ret__ += val;						\
264 	paro_ret__;							\
265 })
266 
267 /*
268  * xchg is implemented using cmpxchg without a lock prefix. xchg is
269  * expensive due to the implied lock prefix.  The processor cannot prefetch
270  * cachelines if xchg is used.
271  */
272 #define percpu_xchg_op(var, nval)					\
273 ({									\
274 	typeof(var) pxo_ret__;						\
275 	typeof(var) pxo_new__ = (nval);					\
276 	switch (sizeof(var)) {						\
277 	case 1:								\
278 		asm("\n\tmov "__percpu_arg(1)",%%al"			\
279 		    "\n1:\tcmpxchgb %2, "__percpu_arg(1)		\
280 		    "\n\tjnz 1b"					\
281 			    : "=&a" (pxo_ret__), "+m" (var)		\
282 			    : "q" (pxo_new__)				\
283 			    : "memory");				\
284 		break;							\
285 	case 2:								\
286 		asm("\n\tmov "__percpu_arg(1)",%%ax"			\
287 		    "\n1:\tcmpxchgw %2, "__percpu_arg(1)		\
288 		    "\n\tjnz 1b"					\
289 			    : "=&a" (pxo_ret__), "+m" (var)		\
290 			    : "r" (pxo_new__)				\
291 			    : "memory");				\
292 		break;							\
293 	case 4:								\
294 		asm("\n\tmov "__percpu_arg(1)",%%eax"			\
295 		    "\n1:\tcmpxchgl %2, "__percpu_arg(1)		\
296 		    "\n\tjnz 1b"					\
297 			    : "=&a" (pxo_ret__), "+m" (var)		\
298 			    : "r" (pxo_new__)				\
299 			    : "memory");				\
300 		break;							\
301 	case 8:								\
302 		asm("\n\tmov "__percpu_arg(1)",%%rax"			\
303 		    "\n1:\tcmpxchgq %2, "__percpu_arg(1)		\
304 		    "\n\tjnz 1b"					\
305 			    : "=&a" (pxo_ret__), "+m" (var)		\
306 			    : "r" (pxo_new__)				\
307 			    : "memory");				\
308 		break;							\
309 	default: __bad_percpu_size();					\
310 	}								\
311 	pxo_ret__;							\
312 })
313 
314 /*
315  * cmpxchg has no such implied lock semantics as a result it is much
316  * more efficient for cpu local operations.
317  */
318 #define percpu_cmpxchg_op(var, oval, nval)				\
319 ({									\
320 	typeof(var) pco_ret__;						\
321 	typeof(var) pco_old__ = (oval);					\
322 	typeof(var) pco_new__ = (nval);					\
323 	switch (sizeof(var)) {						\
324 	case 1:								\
325 		asm("cmpxchgb %2, "__percpu_arg(1)			\
326 			    : "=a" (pco_ret__), "+m" (var)		\
327 			    : "q" (pco_new__), "0" (pco_old__)		\
328 			    : "memory");				\
329 		break;							\
330 	case 2:								\
331 		asm("cmpxchgw %2, "__percpu_arg(1)			\
332 			    : "=a" (pco_ret__), "+m" (var)		\
333 			    : "r" (pco_new__), "0" (pco_old__)		\
334 			    : "memory");				\
335 		break;							\
336 	case 4:								\
337 		asm("cmpxchgl %2, "__percpu_arg(1)			\
338 			    : "=a" (pco_ret__), "+m" (var)		\
339 			    : "r" (pco_new__), "0" (pco_old__)		\
340 			    : "memory");				\
341 		break;							\
342 	case 8:								\
343 		asm("cmpxchgq %2, "__percpu_arg(1)			\
344 			    : "=a" (pco_ret__), "+m" (var)		\
345 			    : "r" (pco_new__), "0" (pco_old__)		\
346 			    : "memory");				\
347 		break;							\
348 	default: __bad_percpu_size();					\
349 	}								\
350 	pco_ret__;							\
351 })
352 
353 /*
354  * percpu_read() makes gcc load the percpu variable every time it is
355  * accessed while percpu_read_stable() allows the value to be cached.
356  * percpu_read_stable() is more efficient and can be used if its value
357  * is guaranteed to be valid across cpus.  The current users include
358  * get_current() and get_thread_info() both of which are actually
359  * per-thread variables implemented as per-cpu variables and thus
360  * stable for the duration of the respective task.
361  */
362 #define percpu_read(var)		percpu_from_op("mov", var, "m" (var))
363 #define percpu_read_stable(var)		percpu_from_op("mov", var, "p" (&(var)))
364 #define percpu_write(var, val)		percpu_to_op("mov", var, val)
365 #define percpu_add(var, val)		percpu_add_op(var, val)
366 #define percpu_sub(var, val)		percpu_add_op(var, -(val))
367 #define percpu_and(var, val)		percpu_to_op("and", var, val)
368 #define percpu_or(var, val)		percpu_to_op("or", var, val)
369 #define percpu_xor(var, val)		percpu_to_op("xor", var, val)
370 #define percpu_inc(var)		percpu_unary_op("inc", var)
371 
372 #define __this_cpu_read_1(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
373 #define __this_cpu_read_2(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
374 #define __this_cpu_read_4(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
375 
376 #define __this_cpu_write_1(pcp, val)	percpu_to_op("mov", (pcp), val)
377 #define __this_cpu_write_2(pcp, val)	percpu_to_op("mov", (pcp), val)
378 #define __this_cpu_write_4(pcp, val)	percpu_to_op("mov", (pcp), val)
379 #define __this_cpu_add_1(pcp, val)	percpu_add_op((pcp), val)
380 #define __this_cpu_add_2(pcp, val)	percpu_add_op((pcp), val)
381 #define __this_cpu_add_4(pcp, val)	percpu_add_op((pcp), val)
382 #define __this_cpu_and_1(pcp, val)	percpu_to_op("and", (pcp), val)
383 #define __this_cpu_and_2(pcp, val)	percpu_to_op("and", (pcp), val)
384 #define __this_cpu_and_4(pcp, val)	percpu_to_op("and", (pcp), val)
385 #define __this_cpu_or_1(pcp, val)	percpu_to_op("or", (pcp), val)
386 #define __this_cpu_or_2(pcp, val)	percpu_to_op("or", (pcp), val)
387 #define __this_cpu_or_4(pcp, val)	percpu_to_op("or", (pcp), val)
388 #define __this_cpu_xor_1(pcp, val)	percpu_to_op("xor", (pcp), val)
389 #define __this_cpu_xor_2(pcp, val)	percpu_to_op("xor", (pcp), val)
390 #define __this_cpu_xor_4(pcp, val)	percpu_to_op("xor", (pcp), val)
391 /*
392  * Generic fallback operations for __this_cpu_xchg_[1-4] are okay and much
393  * faster than an xchg with forced lock semantics.
394  */
395 #define __this_cpu_xchg_8(pcp, nval)	percpu_xchg_op(pcp, nval)
396 #define __this_cpu_cmpxchg_8(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
397 
398 #define this_cpu_read_1(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
399 #define this_cpu_read_2(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
400 #define this_cpu_read_4(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
401 #define this_cpu_write_1(pcp, val)	percpu_to_op("mov", (pcp), val)
402 #define this_cpu_write_2(pcp, val)	percpu_to_op("mov", (pcp), val)
403 #define this_cpu_write_4(pcp, val)	percpu_to_op("mov", (pcp), val)
404 #define this_cpu_add_1(pcp, val)	percpu_add_op((pcp), val)
405 #define this_cpu_add_2(pcp, val)	percpu_add_op((pcp), val)
406 #define this_cpu_add_4(pcp, val)	percpu_add_op((pcp), val)
407 #define this_cpu_and_1(pcp, val)	percpu_to_op("and", (pcp), val)
408 #define this_cpu_and_2(pcp, val)	percpu_to_op("and", (pcp), val)
409 #define this_cpu_and_4(pcp, val)	percpu_to_op("and", (pcp), val)
410 #define this_cpu_or_1(pcp, val)		percpu_to_op("or", (pcp), val)
411 #define this_cpu_or_2(pcp, val)		percpu_to_op("or", (pcp), val)
412 #define this_cpu_or_4(pcp, val)		percpu_to_op("or", (pcp), val)
413 #define this_cpu_xor_1(pcp, val)	percpu_to_op("xor", (pcp), val)
414 #define this_cpu_xor_2(pcp, val)	percpu_to_op("xor", (pcp), val)
415 #define this_cpu_xor_4(pcp, val)	percpu_to_op("xor", (pcp), val)
416 #define this_cpu_xchg_1(pcp, nval)	percpu_xchg_op(pcp, nval)
417 #define this_cpu_xchg_2(pcp, nval)	percpu_xchg_op(pcp, nval)
418 #define this_cpu_xchg_4(pcp, nval)	percpu_xchg_op(pcp, nval)
419 
420 #define irqsafe_cpu_add_1(pcp, val)	percpu_add_op((pcp), val)
421 #define irqsafe_cpu_add_2(pcp, val)	percpu_add_op((pcp), val)
422 #define irqsafe_cpu_add_4(pcp, val)	percpu_add_op((pcp), val)
423 #define irqsafe_cpu_and_1(pcp, val)	percpu_to_op("and", (pcp), val)
424 #define irqsafe_cpu_and_2(pcp, val)	percpu_to_op("and", (pcp), val)
425 #define irqsafe_cpu_and_4(pcp, val)	percpu_to_op("and", (pcp), val)
426 #define irqsafe_cpu_or_1(pcp, val)	percpu_to_op("or", (pcp), val)
427 #define irqsafe_cpu_or_2(pcp, val)	percpu_to_op("or", (pcp), val)
428 #define irqsafe_cpu_or_4(pcp, val)	percpu_to_op("or", (pcp), val)
429 #define irqsafe_cpu_xor_1(pcp, val)	percpu_to_op("xor", (pcp), val)
430 #define irqsafe_cpu_xor_2(pcp, val)	percpu_to_op("xor", (pcp), val)
431 #define irqsafe_cpu_xor_4(pcp, val)	percpu_to_op("xor", (pcp), val)
432 #define irqsafe_cpu_xchg_1(pcp, nval)	percpu_xchg_op(pcp, nval)
433 #define irqsafe_cpu_xchg_2(pcp, nval)	percpu_xchg_op(pcp, nval)
434 #define irqsafe_cpu_xchg_4(pcp, nval)	percpu_xchg_op(pcp, nval)
435 
436 #ifndef CONFIG_M386
437 #define __this_cpu_add_return_1(pcp, val) percpu_add_return_op(pcp, val)
438 #define __this_cpu_add_return_2(pcp, val) percpu_add_return_op(pcp, val)
439 #define __this_cpu_add_return_4(pcp, val) percpu_add_return_op(pcp, val)
440 #define __this_cpu_cmpxchg_1(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
441 #define __this_cpu_cmpxchg_2(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
442 #define __this_cpu_cmpxchg_4(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
443 
444 #define this_cpu_add_return_1(pcp, val)	percpu_add_return_op(pcp, val)
445 #define this_cpu_add_return_2(pcp, val)	percpu_add_return_op(pcp, val)
446 #define this_cpu_add_return_4(pcp, val)	percpu_add_return_op(pcp, val)
447 #define this_cpu_cmpxchg_1(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
448 #define this_cpu_cmpxchg_2(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
449 #define this_cpu_cmpxchg_4(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
450 
451 #define irqsafe_cpu_cmpxchg_1(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
452 #define irqsafe_cpu_cmpxchg_2(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
453 #define irqsafe_cpu_cmpxchg_4(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
454 #endif /* !CONFIG_M386 */
455 
456 #ifdef CONFIG_X86_CMPXCHG64
457 #define percpu_cmpxchg8b_double(pcp1, o1, o2, n1, n2)			\
458 ({									\
459 	char __ret;							\
460 	typeof(o1) __o1 = o1;						\
461 	typeof(o1) __n1 = n1;						\
462 	typeof(o2) __o2 = o2;						\
463 	typeof(o2) __n2 = n2;						\
464 	typeof(o2) __dummy = n2;					\
465 	asm volatile("cmpxchg8b "__percpu_arg(1)"\n\tsetz %0\n\t"	\
466 		    : "=a"(__ret), "=m" (pcp1), "=d"(__dummy)		\
467 		    :  "b"(__n1), "c"(__n2), "a"(__o1), "d"(__o2));	\
468 	__ret;								\
469 })
470 
471 #define __this_cpu_cmpxchg_double_4(pcp1, pcp2, o1, o2, n1, n2)		percpu_cmpxchg8b_double(pcp1, o1, o2, n1, n2)
472 #define this_cpu_cmpxchg_double_4(pcp1, pcp2, o1, o2, n1, n2)		percpu_cmpxchg8b_double(pcp1, o1, o2, n1, n2)
473 #define irqsafe_cpu_cmpxchg_double_4(pcp1, pcp2, o1, o2, n1, n2)	percpu_cmpxchg8b_double(pcp1, o1, o2, n1, n2)
474 #endif /* CONFIG_X86_CMPXCHG64 */
475 
476 /*
477  * Per cpu atomic 64 bit operations are only available under 64 bit.
478  * 32 bit must fall back to generic operations.
479  */
480 #ifdef CONFIG_X86_64
481 #define __this_cpu_read_8(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
482 #define __this_cpu_write_8(pcp, val)	percpu_to_op("mov", (pcp), val)
483 #define __this_cpu_add_8(pcp, val)	percpu_add_op((pcp), val)
484 #define __this_cpu_and_8(pcp, val)	percpu_to_op("and", (pcp), val)
485 #define __this_cpu_or_8(pcp, val)	percpu_to_op("or", (pcp), val)
486 #define __this_cpu_xor_8(pcp, val)	percpu_to_op("xor", (pcp), val)
487 #define __this_cpu_add_return_8(pcp, val) percpu_add_return_op(pcp, val)
488 
489 #define this_cpu_read_8(pcp)		percpu_from_op("mov", (pcp), "m"(pcp))
490 #define this_cpu_write_8(pcp, val)	percpu_to_op("mov", (pcp), val)
491 #define this_cpu_add_8(pcp, val)	percpu_add_op((pcp), val)
492 #define this_cpu_and_8(pcp, val)	percpu_to_op("and", (pcp), val)
493 #define this_cpu_or_8(pcp, val)		percpu_to_op("or", (pcp), val)
494 #define this_cpu_xor_8(pcp, val)	percpu_to_op("xor", (pcp), val)
495 #define this_cpu_add_return_8(pcp, val)	percpu_add_return_op(pcp, val)
496 #define this_cpu_xchg_8(pcp, nval)	percpu_xchg_op(pcp, nval)
497 #define this_cpu_cmpxchg_8(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
498 
499 #define irqsafe_cpu_add_8(pcp, val)	percpu_add_op((pcp), val)
500 #define irqsafe_cpu_and_8(pcp, val)	percpu_to_op("and", (pcp), val)
501 #define irqsafe_cpu_or_8(pcp, val)	percpu_to_op("or", (pcp), val)
502 #define irqsafe_cpu_xor_8(pcp, val)	percpu_to_op("xor", (pcp), val)
503 #define irqsafe_cpu_xchg_8(pcp, nval)	percpu_xchg_op(pcp, nval)
504 #define irqsafe_cpu_cmpxchg_8(pcp, oval, nval)	percpu_cmpxchg_op(pcp, oval, nval)
505 
506 /*
507  * Pretty complex macro to generate cmpxchg16 instruction.  The instruction
508  * is not supported on early AMD64 processors so we must be able to emulate
509  * it in software.  The address used in the cmpxchg16 instruction must be
510  * aligned to a 16 byte boundary.
511  */
512 #define percpu_cmpxchg16b_double(pcp1, o1, o2, n1, n2)			\
513 ({									\
514 	char __ret;							\
515 	typeof(o1) __o1 = o1;						\
516 	typeof(o1) __n1 = n1;						\
517 	typeof(o2) __o2 = o2;						\
518 	typeof(o2) __n2 = n2;						\
519 	typeof(o2) __dummy;						\
520 	alternative_io("call this_cpu_cmpxchg16b_emu\n\t" P6_NOP4,	\
521 		       "cmpxchg16b " __percpu_prefix "(%%rsi)\n\tsetz %0\n\t",	\
522 		       X86_FEATURE_CX16,				\
523 		       ASM_OUTPUT2("=a"(__ret), "=d"(__dummy)),		\
524 		       "S" (&pcp1), "b"(__n1), "c"(__n2),		\
525 		       "a"(__o1), "d"(__o2) : "memory");		\
526 	__ret;								\
527 })
528 
529 #define __this_cpu_cmpxchg_double_8(pcp1, pcp2, o1, o2, n1, n2)		percpu_cmpxchg16b_double(pcp1, o1, o2, n1, n2)
530 #define this_cpu_cmpxchg_double_8(pcp1, pcp2, o1, o2, n1, n2)		percpu_cmpxchg16b_double(pcp1, o1, o2, n1, n2)
531 #define irqsafe_cpu_cmpxchg_double_8(pcp1, pcp2, o1, o2, n1, n2)	percpu_cmpxchg16b_double(pcp1, o1, o2, n1, n2)
532 
533 #endif
534 
535 /* This is not atomic against other CPUs -- CPU preemption needs to be off */
536 #define x86_test_and_clear_bit_percpu(bit, var)				\
537 ({									\
538 	int old__;							\
539 	asm volatile("btr %2,"__percpu_arg(1)"\n\tsbbl %0,%0"		\
540 		     : "=r" (old__), "+m" (var)				\
541 		     : "dIr" (bit));					\
542 	old__;								\
543 })
544 
545 #include <asm-generic/percpu.h>
546 
547 /* We can use this directly for local CPU (faster). */
548 DECLARE_PER_CPU(unsigned long, this_cpu_off);
549 
550 #endif /* !__ASSEMBLY__ */
551 
552 #ifdef CONFIG_SMP
553 
554 /*
555  * Define the "EARLY_PER_CPU" macros.  These are used for some per_cpu
556  * variables that are initialized and accessed before there are per_cpu
557  * areas allocated.
558  */
559 
560 #define	DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)			\
561 	DEFINE_PER_CPU(_type, _name) = _initvalue;			\
562 	__typeof__(_type) _name##_early_map[NR_CPUS] __initdata =	\
563 				{ [0 ... NR_CPUS-1] = _initvalue };	\
564 	__typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
565 
566 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)			\
567 	EXPORT_PER_CPU_SYMBOL(_name)
568 
569 #define DECLARE_EARLY_PER_CPU(_type, _name)			\
570 	DECLARE_PER_CPU(_type, _name);				\
571 	extern __typeof__(_type) *_name##_early_ptr;		\
572 	extern __typeof__(_type)  _name##_early_map[]
573 
574 #define	early_per_cpu_ptr(_name) (_name##_early_ptr)
575 #define	early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
576 #define	early_per_cpu(_name, _cpu) 				\
577 	*(early_per_cpu_ptr(_name) ?				\
578 		&early_per_cpu_ptr(_name)[_cpu] :		\
579 		&per_cpu(_name, _cpu))
580 
581 #else	/* !CONFIG_SMP */
582 #define	DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)		\
583 	DEFINE_PER_CPU(_type, _name) = _initvalue
584 
585 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)			\
586 	EXPORT_PER_CPU_SYMBOL(_name)
587 
588 #define DECLARE_EARLY_PER_CPU(_type, _name)			\
589 	DECLARE_PER_CPU(_type, _name)
590 
591 #define	early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
592 #define	early_per_cpu_ptr(_name) NULL
593 /* no early_per_cpu_map() */
594 
595 #endif	/* !CONFIG_SMP */
596 
597 #endif /* _ASM_X86_PERCPU_H */
598