1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_X86_PERCPU_H
3 #define _ASM_X86_PERCPU_H
4
5 #ifdef CONFIG_X86_64
6 #define __percpu_seg gs
7 #else
8 #define __percpu_seg fs
9 #endif
10
11 #ifdef __ASSEMBLY__
12
13 #ifdef CONFIG_SMP
14 #define PER_CPU_VAR(var) %__percpu_seg:var
15 #else /* ! SMP */
16 #define PER_CPU_VAR(var) var
17 #endif /* SMP */
18
19 #ifdef CONFIG_X86_64_SMP
20 #define INIT_PER_CPU_VAR(var) init_per_cpu__##var
21 #else
22 #define INIT_PER_CPU_VAR(var) var
23 #endif
24
25 #else /* ...!ASSEMBLY */
26
27 #include <linux/kernel.h>
28 #include <linux/stringify.h>
29
30 #ifdef CONFIG_SMP
31 #define __percpu_prefix "%%"__stringify(__percpu_seg)":"
32 #define __my_cpu_offset this_cpu_read(this_cpu_off)
33
34 /*
35 * Compared to the generic __my_cpu_offset version, the following
36 * saves one instruction and avoids clobbering a temp register.
37 */
38 #define arch_raw_cpu_ptr(ptr) \
39 ({ \
40 unsigned long tcp_ptr__; \
41 asm ("add " __percpu_arg(1) ", %0" \
42 : "=r" (tcp_ptr__) \
43 : "m" (this_cpu_off), "0" (ptr)); \
44 (typeof(*(ptr)) __kernel __force *)tcp_ptr__; \
45 })
46 #else
47 #define __percpu_prefix ""
48 #endif
49
50 #define __percpu_arg(x) __percpu_prefix "%" #x
51
52 /*
53 * Initialized pointers to per-cpu variables needed for the boot
54 * processor need to use these macros to get the proper address
55 * offset from __per_cpu_load on SMP.
56 *
57 * There also must be an entry in vmlinux_64.lds.S
58 */
59 #define DECLARE_INIT_PER_CPU(var) \
60 extern typeof(var) init_per_cpu_var(var)
61
62 #ifdef CONFIG_X86_64_SMP
63 #define init_per_cpu_var(var) init_per_cpu__##var
64 #else
65 #define init_per_cpu_var(var) var
66 #endif
67
68 /* For arch-specific code, we can use direct single-insn ops (they
69 * don't give an lvalue though). */
70
71 #define __pcpu_type_1 u8
72 #define __pcpu_type_2 u16
73 #define __pcpu_type_4 u32
74 #define __pcpu_type_8 u64
75
76 #define __pcpu_cast_1(val) ((u8)(((unsigned long) val) & 0xff))
77 #define __pcpu_cast_2(val) ((u16)(((unsigned long) val) & 0xffff))
78 #define __pcpu_cast_4(val) ((u32)(((unsigned long) val) & 0xffffffff))
79 #define __pcpu_cast_8(val) ((u64)(val))
80
81 #define __pcpu_op1_1(op, dst) op "b " dst
82 #define __pcpu_op1_2(op, dst) op "w " dst
83 #define __pcpu_op1_4(op, dst) op "l " dst
84 #define __pcpu_op1_8(op, dst) op "q " dst
85
86 #define __pcpu_op2_1(op, src, dst) op "b " src ", " dst
87 #define __pcpu_op2_2(op, src, dst) op "w " src ", " dst
88 #define __pcpu_op2_4(op, src, dst) op "l " src ", " dst
89 #define __pcpu_op2_8(op, src, dst) op "q " src ", " dst
90
91 #define __pcpu_reg_1(mod, x) mod "q" (x)
92 #define __pcpu_reg_2(mod, x) mod "r" (x)
93 #define __pcpu_reg_4(mod, x) mod "r" (x)
94 #define __pcpu_reg_8(mod, x) mod "r" (x)
95
96 #define __pcpu_reg_imm_1(x) "qi" (x)
97 #define __pcpu_reg_imm_2(x) "ri" (x)
98 #define __pcpu_reg_imm_4(x) "ri" (x)
99 #define __pcpu_reg_imm_8(x) "re" (x)
100
101 #define percpu_to_op(size, qual, op, _var, _val) \
102 do { \
103 __pcpu_type_##size pto_val__ = __pcpu_cast_##size(_val); \
104 if (0) { \
105 typeof(_var) pto_tmp__; \
106 pto_tmp__ = (_val); \
107 (void)pto_tmp__; \
108 } \
109 asm qual(__pcpu_op2_##size(op, "%[val]", __percpu_arg([var])) \
110 : [var] "+m" (_var) \
111 : [val] __pcpu_reg_imm_##size(pto_val__)); \
112 } while (0)
113
114 #define percpu_unary_op(size, qual, op, _var) \
115 ({ \
116 asm qual (__pcpu_op1_##size(op, __percpu_arg([var])) \
117 : [var] "+m" (_var)); \
118 })
119
120 /*
121 * Generate a percpu add to memory instruction and optimize code
122 * if one is added or subtracted.
123 */
124 #define percpu_add_op(size, qual, var, val) \
125 do { \
126 const int pao_ID__ = (__builtin_constant_p(val) && \
127 ((val) == 1 || (val) == -1)) ? \
128 (int)(val) : 0; \
129 if (0) { \
130 typeof(var) pao_tmp__; \
131 pao_tmp__ = (val); \
132 (void)pao_tmp__; \
133 } \
134 if (pao_ID__ == 1) \
135 percpu_unary_op(size, qual, "inc", var); \
136 else if (pao_ID__ == -1) \
137 percpu_unary_op(size, qual, "dec", var); \
138 else \
139 percpu_to_op(size, qual, "add", var, val); \
140 } while (0)
141
142 #define percpu_from_op(size, qual, op, _var) \
143 ({ \
144 __pcpu_type_##size pfo_val__; \
145 asm qual (__pcpu_op2_##size(op, __percpu_arg([var]), "%[val]") \
146 : [val] __pcpu_reg_##size("=", pfo_val__) \
147 : [var] "m" (_var)); \
148 (typeof(_var))(unsigned long) pfo_val__; \
149 })
150
151 #define percpu_stable_op(size, op, _var) \
152 ({ \
153 __pcpu_type_##size pfo_val__; \
154 asm(__pcpu_op2_##size(op, __percpu_arg(P[var]), "%[val]") \
155 : [val] __pcpu_reg_##size("=", pfo_val__) \
156 : [var] "p" (&(_var))); \
157 (typeof(_var))(unsigned long) pfo_val__; \
158 })
159
160 /*
161 * Add return operation
162 */
163 #define percpu_add_return_op(size, qual, _var, _val) \
164 ({ \
165 __pcpu_type_##size paro_tmp__ = __pcpu_cast_##size(_val); \
166 asm qual (__pcpu_op2_##size("xadd", "%[tmp]", \
167 __percpu_arg([var])) \
168 : [tmp] __pcpu_reg_##size("+", paro_tmp__), \
169 [var] "+m" (_var) \
170 : : "memory"); \
171 (typeof(_var))(unsigned long) (paro_tmp__ + _val); \
172 })
173
174 /*
175 * xchg is implemented using cmpxchg without a lock prefix. xchg is
176 * expensive due to the implied lock prefix. The processor cannot prefetch
177 * cachelines if xchg is used.
178 */
179 #define percpu_xchg_op(size, qual, _var, _nval) \
180 ({ \
181 __pcpu_type_##size pxo_old__; \
182 __pcpu_type_##size pxo_new__ = __pcpu_cast_##size(_nval); \
183 asm qual (__pcpu_op2_##size("mov", __percpu_arg([var]), \
184 "%[oval]") \
185 "\n1:\t" \
186 __pcpu_op2_##size("cmpxchg", "%[nval]", \
187 __percpu_arg([var])) \
188 "\n\tjnz 1b" \
189 : [oval] "=&a" (pxo_old__), \
190 [var] "+m" (_var) \
191 : [nval] __pcpu_reg_##size(, pxo_new__) \
192 : "memory"); \
193 (typeof(_var))(unsigned long) pxo_old__; \
194 })
195
196 /*
197 * cmpxchg has no such implied lock semantics as a result it is much
198 * more efficient for cpu local operations.
199 */
200 #define percpu_cmpxchg_op(size, qual, _var, _oval, _nval) \
201 ({ \
202 __pcpu_type_##size pco_old__ = __pcpu_cast_##size(_oval); \
203 __pcpu_type_##size pco_new__ = __pcpu_cast_##size(_nval); \
204 asm qual (__pcpu_op2_##size("cmpxchg", "%[nval]", \
205 __percpu_arg([var])) \
206 : [oval] "+a" (pco_old__), \
207 [var] "+m" (_var) \
208 : [nval] __pcpu_reg_##size(, pco_new__) \
209 : "memory"); \
210 (typeof(_var))(unsigned long) pco_old__; \
211 })
212
213 /*
214 * this_cpu_read() makes gcc load the percpu variable every time it is
215 * accessed while this_cpu_read_stable() allows the value to be cached.
216 * this_cpu_read_stable() is more efficient and can be used if its value
217 * is guaranteed to be valid across cpus. The current users include
218 * get_current() and get_thread_info() both of which are actually
219 * per-thread variables implemented as per-cpu variables and thus
220 * stable for the duration of the respective task.
221 */
222 #define this_cpu_read_stable_1(pcp) percpu_stable_op(1, "mov", pcp)
223 #define this_cpu_read_stable_2(pcp) percpu_stable_op(2, "mov", pcp)
224 #define this_cpu_read_stable_4(pcp) percpu_stable_op(4, "mov", pcp)
225 #define this_cpu_read_stable_8(pcp) percpu_stable_op(8, "mov", pcp)
226 #define this_cpu_read_stable(pcp) __pcpu_size_call_return(this_cpu_read_stable_, pcp)
227
228 #define raw_cpu_read_1(pcp) percpu_from_op(1, , "mov", pcp)
229 #define raw_cpu_read_2(pcp) percpu_from_op(2, , "mov", pcp)
230 #define raw_cpu_read_4(pcp) percpu_from_op(4, , "mov", pcp)
231
232 #define raw_cpu_write_1(pcp, val) percpu_to_op(1, , "mov", (pcp), val)
233 #define raw_cpu_write_2(pcp, val) percpu_to_op(2, , "mov", (pcp), val)
234 #define raw_cpu_write_4(pcp, val) percpu_to_op(4, , "mov", (pcp), val)
235 #define raw_cpu_add_1(pcp, val) percpu_add_op(1, , (pcp), val)
236 #define raw_cpu_add_2(pcp, val) percpu_add_op(2, , (pcp), val)
237 #define raw_cpu_add_4(pcp, val) percpu_add_op(4, , (pcp), val)
238 #define raw_cpu_and_1(pcp, val) percpu_to_op(1, , "and", (pcp), val)
239 #define raw_cpu_and_2(pcp, val) percpu_to_op(2, , "and", (pcp), val)
240 #define raw_cpu_and_4(pcp, val) percpu_to_op(4, , "and", (pcp), val)
241 #define raw_cpu_or_1(pcp, val) percpu_to_op(1, , "or", (pcp), val)
242 #define raw_cpu_or_2(pcp, val) percpu_to_op(2, , "or", (pcp), val)
243 #define raw_cpu_or_4(pcp, val) percpu_to_op(4, , "or", (pcp), val)
244
245 /*
246 * raw_cpu_xchg() can use a load-store since it is not required to be
247 * IRQ-safe.
248 */
249 #define raw_percpu_xchg_op(var, nval) \
250 ({ \
251 typeof(var) pxo_ret__ = raw_cpu_read(var); \
252 raw_cpu_write(var, (nval)); \
253 pxo_ret__; \
254 })
255
256 #define raw_cpu_xchg_1(pcp, val) raw_percpu_xchg_op(pcp, val)
257 #define raw_cpu_xchg_2(pcp, val) raw_percpu_xchg_op(pcp, val)
258 #define raw_cpu_xchg_4(pcp, val) raw_percpu_xchg_op(pcp, val)
259
260 #define this_cpu_read_1(pcp) percpu_from_op(1, volatile, "mov", pcp)
261 #define this_cpu_read_2(pcp) percpu_from_op(2, volatile, "mov", pcp)
262 #define this_cpu_read_4(pcp) percpu_from_op(4, volatile, "mov", pcp)
263 #define this_cpu_write_1(pcp, val) percpu_to_op(1, volatile, "mov", (pcp), val)
264 #define this_cpu_write_2(pcp, val) percpu_to_op(2, volatile, "mov", (pcp), val)
265 #define this_cpu_write_4(pcp, val) percpu_to_op(4, volatile, "mov", (pcp), val)
266 #define this_cpu_add_1(pcp, val) percpu_add_op(1, volatile, (pcp), val)
267 #define this_cpu_add_2(pcp, val) percpu_add_op(2, volatile, (pcp), val)
268 #define this_cpu_add_4(pcp, val) percpu_add_op(4, volatile, (pcp), val)
269 #define this_cpu_and_1(pcp, val) percpu_to_op(1, volatile, "and", (pcp), val)
270 #define this_cpu_and_2(pcp, val) percpu_to_op(2, volatile, "and", (pcp), val)
271 #define this_cpu_and_4(pcp, val) percpu_to_op(4, volatile, "and", (pcp), val)
272 #define this_cpu_or_1(pcp, val) percpu_to_op(1, volatile, "or", (pcp), val)
273 #define this_cpu_or_2(pcp, val) percpu_to_op(2, volatile, "or", (pcp), val)
274 #define this_cpu_or_4(pcp, val) percpu_to_op(4, volatile, "or", (pcp), val)
275 #define this_cpu_xchg_1(pcp, nval) percpu_xchg_op(1, volatile, pcp, nval)
276 #define this_cpu_xchg_2(pcp, nval) percpu_xchg_op(2, volatile, pcp, nval)
277 #define this_cpu_xchg_4(pcp, nval) percpu_xchg_op(4, volatile, pcp, nval)
278
279 #define raw_cpu_add_return_1(pcp, val) percpu_add_return_op(1, , pcp, val)
280 #define raw_cpu_add_return_2(pcp, val) percpu_add_return_op(2, , pcp, val)
281 #define raw_cpu_add_return_4(pcp, val) percpu_add_return_op(4, , pcp, val)
282 #define raw_cpu_cmpxchg_1(pcp, oval, nval) percpu_cmpxchg_op(1, , pcp, oval, nval)
283 #define raw_cpu_cmpxchg_2(pcp, oval, nval) percpu_cmpxchg_op(2, , pcp, oval, nval)
284 #define raw_cpu_cmpxchg_4(pcp, oval, nval) percpu_cmpxchg_op(4, , pcp, oval, nval)
285
286 #define this_cpu_add_return_1(pcp, val) percpu_add_return_op(1, volatile, pcp, val)
287 #define this_cpu_add_return_2(pcp, val) percpu_add_return_op(2, volatile, pcp, val)
288 #define this_cpu_add_return_4(pcp, val) percpu_add_return_op(4, volatile, pcp, val)
289 #define this_cpu_cmpxchg_1(pcp, oval, nval) percpu_cmpxchg_op(1, volatile, pcp, oval, nval)
290 #define this_cpu_cmpxchg_2(pcp, oval, nval) percpu_cmpxchg_op(2, volatile, pcp, oval, nval)
291 #define this_cpu_cmpxchg_4(pcp, oval, nval) percpu_cmpxchg_op(4, volatile, pcp, oval, nval)
292
293 #ifdef CONFIG_X86_CMPXCHG64
294 #define percpu_cmpxchg8b_double(pcp1, pcp2, o1, o2, n1, n2) \
295 ({ \
296 bool __ret; \
297 typeof(pcp1) __o1 = (o1), __n1 = (n1); \
298 typeof(pcp2) __o2 = (o2), __n2 = (n2); \
299 asm volatile("cmpxchg8b "__percpu_arg(1) \
300 CC_SET(z) \
301 : CC_OUT(z) (__ret), "+m" (pcp1), "+m" (pcp2), "+a" (__o1), "+d" (__o2) \
302 : "b" (__n1), "c" (__n2)); \
303 __ret; \
304 })
305
306 #define raw_cpu_cmpxchg_double_4 percpu_cmpxchg8b_double
307 #define this_cpu_cmpxchg_double_4 percpu_cmpxchg8b_double
308 #endif /* CONFIG_X86_CMPXCHG64 */
309
310 /*
311 * Per cpu atomic 64 bit operations are only available under 64 bit.
312 * 32 bit must fall back to generic operations.
313 */
314 #ifdef CONFIG_X86_64
315 #define raw_cpu_read_8(pcp) percpu_from_op(8, , "mov", pcp)
316 #define raw_cpu_write_8(pcp, val) percpu_to_op(8, , "mov", (pcp), val)
317 #define raw_cpu_add_8(pcp, val) percpu_add_op(8, , (pcp), val)
318 #define raw_cpu_and_8(pcp, val) percpu_to_op(8, , "and", (pcp), val)
319 #define raw_cpu_or_8(pcp, val) percpu_to_op(8, , "or", (pcp), val)
320 #define raw_cpu_add_return_8(pcp, val) percpu_add_return_op(8, , pcp, val)
321 #define raw_cpu_xchg_8(pcp, nval) raw_percpu_xchg_op(pcp, nval)
322 #define raw_cpu_cmpxchg_8(pcp, oval, nval) percpu_cmpxchg_op(8, , pcp, oval, nval)
323
324 #define this_cpu_read_8(pcp) percpu_from_op(8, volatile, "mov", pcp)
325 #define this_cpu_write_8(pcp, val) percpu_to_op(8, volatile, "mov", (pcp), val)
326 #define this_cpu_add_8(pcp, val) percpu_add_op(8, volatile, (pcp), val)
327 #define this_cpu_and_8(pcp, val) percpu_to_op(8, volatile, "and", (pcp), val)
328 #define this_cpu_or_8(pcp, val) percpu_to_op(8, volatile, "or", (pcp), val)
329 #define this_cpu_add_return_8(pcp, val) percpu_add_return_op(8, volatile, pcp, val)
330 #define this_cpu_xchg_8(pcp, nval) percpu_xchg_op(8, volatile, pcp, nval)
331 #define this_cpu_cmpxchg_8(pcp, oval, nval) percpu_cmpxchg_op(8, volatile, pcp, oval, nval)
332
333 /*
334 * Pretty complex macro to generate cmpxchg16 instruction. The instruction
335 * is not supported on early AMD64 processors so we must be able to emulate
336 * it in software. The address used in the cmpxchg16 instruction must be
337 * aligned to a 16 byte boundary.
338 */
339 #define percpu_cmpxchg16b_double(pcp1, pcp2, o1, o2, n1, n2) \
340 ({ \
341 bool __ret; \
342 typeof(pcp1) __o1 = (o1), __n1 = (n1); \
343 typeof(pcp2) __o2 = (o2), __n2 = (n2); \
344 alternative_io("leaq %P1,%%rsi\n\tcall this_cpu_cmpxchg16b_emu\n\t", \
345 "cmpxchg16b " __percpu_arg(1) "\n\tsetz %0\n\t", \
346 X86_FEATURE_CX16, \
347 ASM_OUTPUT2("=a" (__ret), "+m" (pcp1), \
348 "+m" (pcp2), "+d" (__o2)), \
349 "b" (__n1), "c" (__n2), "a" (__o1) : "rsi"); \
350 __ret; \
351 })
352
353 #define raw_cpu_cmpxchg_double_8 percpu_cmpxchg16b_double
354 #define this_cpu_cmpxchg_double_8 percpu_cmpxchg16b_double
355
356 #endif
357
x86_this_cpu_constant_test_bit(unsigned int nr,const unsigned long __percpu * addr)358 static __always_inline bool x86_this_cpu_constant_test_bit(unsigned int nr,
359 const unsigned long __percpu *addr)
360 {
361 unsigned long __percpu *a =
362 (unsigned long __percpu *)addr + nr / BITS_PER_LONG;
363
364 #ifdef CONFIG_X86_64
365 return ((1UL << (nr % BITS_PER_LONG)) & raw_cpu_read_8(*a)) != 0;
366 #else
367 return ((1UL << (nr % BITS_PER_LONG)) & raw_cpu_read_4(*a)) != 0;
368 #endif
369 }
370
x86_this_cpu_variable_test_bit(int nr,const unsigned long __percpu * addr)371 static inline bool x86_this_cpu_variable_test_bit(int nr,
372 const unsigned long __percpu *addr)
373 {
374 bool oldbit;
375
376 asm volatile("btl "__percpu_arg(2)",%1"
377 CC_SET(c)
378 : CC_OUT(c) (oldbit)
379 : "m" (*(unsigned long __percpu *)addr), "Ir" (nr));
380
381 return oldbit;
382 }
383
384 #define x86_this_cpu_test_bit(nr, addr) \
385 (__builtin_constant_p((nr)) \
386 ? x86_this_cpu_constant_test_bit((nr), (addr)) \
387 : x86_this_cpu_variable_test_bit((nr), (addr)))
388
389
390 #include <asm-generic/percpu.h>
391
392 /* We can use this directly for local CPU (faster). */
393 DECLARE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off);
394
395 #endif /* !__ASSEMBLY__ */
396
397 #ifdef CONFIG_SMP
398
399 /*
400 * Define the "EARLY_PER_CPU" macros. These are used for some per_cpu
401 * variables that are initialized and accessed before there are per_cpu
402 * areas allocated.
403 */
404
405 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue) \
406 DEFINE_PER_CPU(_type, _name) = _initvalue; \
407 __typeof__(_type) _name##_early_map[NR_CPUS] __initdata = \
408 { [0 ... NR_CPUS-1] = _initvalue }; \
409 __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
410
411 #define DEFINE_EARLY_PER_CPU_READ_MOSTLY(_type, _name, _initvalue) \
412 DEFINE_PER_CPU_READ_MOSTLY(_type, _name) = _initvalue; \
413 __typeof__(_type) _name##_early_map[NR_CPUS] __initdata = \
414 { [0 ... NR_CPUS-1] = _initvalue }; \
415 __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
416
417 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name) \
418 EXPORT_PER_CPU_SYMBOL(_name)
419
420 #define DECLARE_EARLY_PER_CPU(_type, _name) \
421 DECLARE_PER_CPU(_type, _name); \
422 extern __typeof__(_type) *_name##_early_ptr; \
423 extern __typeof__(_type) _name##_early_map[]
424
425 #define DECLARE_EARLY_PER_CPU_READ_MOSTLY(_type, _name) \
426 DECLARE_PER_CPU_READ_MOSTLY(_type, _name); \
427 extern __typeof__(_type) *_name##_early_ptr; \
428 extern __typeof__(_type) _name##_early_map[]
429
430 #define early_per_cpu_ptr(_name) (_name##_early_ptr)
431 #define early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
432 #define early_per_cpu(_name, _cpu) \
433 *(early_per_cpu_ptr(_name) ? \
434 &early_per_cpu_ptr(_name)[_cpu] : \
435 &per_cpu(_name, _cpu))
436
437 #else /* !CONFIG_SMP */
438 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue) \
439 DEFINE_PER_CPU(_type, _name) = _initvalue
440
441 #define DEFINE_EARLY_PER_CPU_READ_MOSTLY(_type, _name, _initvalue) \
442 DEFINE_PER_CPU_READ_MOSTLY(_type, _name) = _initvalue
443
444 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name) \
445 EXPORT_PER_CPU_SYMBOL(_name)
446
447 #define DECLARE_EARLY_PER_CPU(_type, _name) \
448 DECLARE_PER_CPU(_type, _name)
449
450 #define DECLARE_EARLY_PER_CPU_READ_MOSTLY(_type, _name) \
451 DECLARE_PER_CPU_READ_MOSTLY(_type, _name)
452
453 #define early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
454 #define early_per_cpu_ptr(_name) NULL
455 /* no early_per_cpu_map() */
456
457 #endif /* !CONFIG_SMP */
458
459 #endif /* _ASM_X86_PERCPU_H */
460