1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_X86_PERCPU_H
3 #define _ASM_X86_PERCPU_H
6 #define __percpu_seg gs
7 #define __percpu_mov_op movq
9 #define __percpu_seg fs
10 #define __percpu_mov_op movl
16 * PER_CPU finds an address of a per-cpu variable.
20 * reg - 32bit register
22 * The resulting address is stored in the "reg" argument.
25 * PER_CPU(cpu_gdt_descr, %ebx)
28 #define PER_CPU(var, reg) \
29 __percpu_mov_op %__percpu_seg:this_cpu_off, reg; \
31 #define PER_CPU_VAR(var) %__percpu_seg:var
33 #define PER_CPU(var, reg) __percpu_mov_op $var, reg
34 #define PER_CPU_VAR(var) var
37 #ifdef CONFIG_X86_64_SMP
38 #define INIT_PER_CPU_VAR(var) init_per_cpu__##var
40 #define INIT_PER_CPU_VAR(var) var
43 #else /* ...!ASSEMBLY */
45 #include <linux/kernel.h>
46 #include <linux/stringify.h>
49 #define __percpu_prefix "%%"__stringify(__percpu_seg)":"
50 #define __my_cpu_offset this_cpu_read(this_cpu_off)
53 * Compared to the generic __my_cpu_offset version, the following
54 * saves one instruction and avoids clobbering a temp register.
56 #define arch_raw_cpu_ptr(ptr) \
58 unsigned long tcp_ptr__; \
59 asm volatile("add " __percpu_arg(1) ", %0" \
61 : "m" (this_cpu_off), "0" (ptr)); \
62 (typeof(*(ptr)) __kernel __force *)tcp_ptr__; \
65 #define __percpu_prefix ""
68 #define __percpu_arg(x) __percpu_prefix "%" #x
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.
75 * There also must be an entry in vmlinux_64.lds.S
77 #define DECLARE_INIT_PER_CPU(var) \
78 extern typeof(var) init_per_cpu_var(var)
80 #ifdef CONFIG_X86_64_SMP
81 #define init_per_cpu_var(var) init_per_cpu__##var
83 #define init_per_cpu_var(var) var
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);
90 #define percpu_to_op(qual, op, var, val) \
92 typedef typeof(var) pto_T__; \
98 switch (sizeof(var)) { \
100 asm qual (op "b %1,"__percpu_arg(0) \
102 : "qi" ((pto_T__)(val))); \
105 asm qual (op "w %1,"__percpu_arg(0) \
107 : "ri" ((pto_T__)(val))); \
110 asm qual (op "l %1,"__percpu_arg(0) \
112 : "ri" ((pto_T__)(val))); \
115 asm qual (op "q %1,"__percpu_arg(0) \
117 : "re" ((pto_T__)(val))); \
119 default: __bad_percpu_size(); \
124 * Generate a percpu add to memory instruction and optimize code
125 * if one is added or subtracted.
127 #define percpu_add_op(qual, var, val) \
129 typedef typeof(var) pao_T__; \
130 const int pao_ID__ = (__builtin_constant_p(val) && \
131 ((val) == 1 || (val) == -1)) ? \
138 switch (sizeof(var)) { \
141 asm qual ("incb "__percpu_arg(0) : "+m" (var)); \
142 else if (pao_ID__ == -1) \
143 asm qual ("decb "__percpu_arg(0) : "+m" (var)); \
145 asm qual ("addb %1, "__percpu_arg(0) \
147 : "qi" ((pao_T__)(val))); \
151 asm qual ("incw "__percpu_arg(0) : "+m" (var)); \
152 else if (pao_ID__ == -1) \
153 asm qual ("decw "__percpu_arg(0) : "+m" (var)); \
155 asm qual ("addw %1, "__percpu_arg(0) \
157 : "ri" ((pao_T__)(val))); \
161 asm qual ("incl "__percpu_arg(0) : "+m" (var)); \
162 else if (pao_ID__ == -1) \
163 asm qual ("decl "__percpu_arg(0) : "+m" (var)); \
165 asm qual ("addl %1, "__percpu_arg(0) \
167 : "ri" ((pao_T__)(val))); \
171 asm qual ("incq "__percpu_arg(0) : "+m" (var)); \
172 else if (pao_ID__ == -1) \
173 asm qual ("decq "__percpu_arg(0) : "+m" (var)); \
175 asm qual ("addq %1, "__percpu_arg(0) \
177 : "re" ((pao_T__)(val))); \
179 default: __bad_percpu_size(); \
183 #define percpu_from_op(qual, op, var) \
185 typeof(var) pfo_ret__; \
186 switch (sizeof(var)) { \
188 asm qual (op "b "__percpu_arg(1)",%0" \
193 asm qual (op "w "__percpu_arg(1)",%0" \
198 asm qual (op "l "__percpu_arg(1)",%0" \
203 asm qual (op "q "__percpu_arg(1)",%0" \
207 default: __bad_percpu_size(); \
212 #define percpu_stable_op(op, var) \
214 typeof(var) pfo_ret__; \
215 switch (sizeof(var)) { \
217 asm(op "b "__percpu_arg(P1)",%0" \
222 asm(op "w "__percpu_arg(P1)",%0" \
227 asm(op "l "__percpu_arg(P1)",%0" \
232 asm(op "q "__percpu_arg(P1)",%0" \
236 default: __bad_percpu_size(); \
241 #define percpu_unary_op(qual, op, var) \
243 switch (sizeof(var)) { \
245 asm qual (op "b "__percpu_arg(0) \
249 asm qual (op "w "__percpu_arg(0) \
253 asm qual (op "l "__percpu_arg(0) \
257 asm qual (op "q "__percpu_arg(0) \
260 default: __bad_percpu_size(); \
265 * Add return operation
267 #define percpu_add_return_op(qual, var, val) \
269 typeof(var) paro_ret__ = val; \
270 switch (sizeof(var)) { \
272 asm qual ("xaddb %0, "__percpu_arg(1) \
273 : "+q" (paro_ret__), "+m" (var) \
277 asm qual ("xaddw %0, "__percpu_arg(1) \
278 : "+r" (paro_ret__), "+m" (var) \
282 asm qual ("xaddl %0, "__percpu_arg(1) \
283 : "+r" (paro_ret__), "+m" (var) \
287 asm qual ("xaddq %0, "__percpu_arg(1) \
288 : "+re" (paro_ret__), "+m" (var) \
291 default: __bad_percpu_size(); \
298 * xchg is implemented using cmpxchg without a lock prefix. xchg is
299 * expensive due to the implied lock prefix. The processor cannot prefetch
300 * cachelines if xchg is used.
302 #define percpu_xchg_op(qual, var, nval) \
304 typeof(var) pxo_ret__; \
305 typeof(var) pxo_new__ = (nval); \
306 switch (sizeof(var)) { \
308 asm qual ("\n\tmov "__percpu_arg(1)",%%al" \
309 "\n1:\tcmpxchgb %2, "__percpu_arg(1) \
311 : "=&a" (pxo_ret__), "+m" (var) \
316 asm qual ("\n\tmov "__percpu_arg(1)",%%ax" \
317 "\n1:\tcmpxchgw %2, "__percpu_arg(1) \
319 : "=&a" (pxo_ret__), "+m" (var) \
324 asm qual ("\n\tmov "__percpu_arg(1)",%%eax" \
325 "\n1:\tcmpxchgl %2, "__percpu_arg(1) \
327 : "=&a" (pxo_ret__), "+m" (var) \
332 asm qual ("\n\tmov "__percpu_arg(1)",%%rax" \
333 "\n1:\tcmpxchgq %2, "__percpu_arg(1) \
335 : "=&a" (pxo_ret__), "+m" (var) \
339 default: __bad_percpu_size(); \
345 * cmpxchg has no such implied lock semantics as a result it is much
346 * more efficient for cpu local operations.
348 #define percpu_cmpxchg_op(qual, var, oval, nval) \
350 typeof(var) pco_ret__; \
351 typeof(var) pco_old__ = (oval); \
352 typeof(var) pco_new__ = (nval); \
353 switch (sizeof(var)) { \
355 asm qual ("cmpxchgb %2, "__percpu_arg(1) \
356 : "=a" (pco_ret__), "+m" (var) \
357 : "q" (pco_new__), "0" (pco_old__) \
361 asm qual ("cmpxchgw %2, "__percpu_arg(1) \
362 : "=a" (pco_ret__), "+m" (var) \
363 : "r" (pco_new__), "0" (pco_old__) \
367 asm qual ("cmpxchgl %2, "__percpu_arg(1) \
368 : "=a" (pco_ret__), "+m" (var) \
369 : "r" (pco_new__), "0" (pco_old__) \
373 asm qual ("cmpxchgq %2, "__percpu_arg(1) \
374 : "=a" (pco_ret__), "+m" (var) \
375 : "r" (pco_new__), "0" (pco_old__) \
378 default: __bad_percpu_size(); \
384 * this_cpu_read() makes gcc load the percpu variable every time it is
385 * accessed while this_cpu_read_stable() allows the value to be cached.
386 * this_cpu_read_stable() is more efficient and can be used if its value
387 * is guaranteed to be valid across cpus. The current users include
388 * get_current() and get_thread_info() both of which are actually
389 * per-thread variables implemented as per-cpu variables and thus
390 * stable for the duration of the respective task.
392 #define this_cpu_read_stable(var) percpu_stable_op("mov", var)
394 #define raw_cpu_read_1(pcp) percpu_from_op(, "mov", pcp)
395 #define raw_cpu_read_2(pcp) percpu_from_op(, "mov", pcp)
396 #define raw_cpu_read_4(pcp) percpu_from_op(, "mov", pcp)
398 #define raw_cpu_write_1(pcp, val) percpu_to_op(, "mov", (pcp), val)
399 #define raw_cpu_write_2(pcp, val) percpu_to_op(, "mov", (pcp), val)
400 #define raw_cpu_write_4(pcp, val) percpu_to_op(, "mov", (pcp), val)
401 #define raw_cpu_add_1(pcp, val) percpu_add_op(, (pcp), val)
402 #define raw_cpu_add_2(pcp, val) percpu_add_op(, (pcp), val)
403 #define raw_cpu_add_4(pcp, val) percpu_add_op(, (pcp), val)
404 #define raw_cpu_and_1(pcp, val) percpu_to_op(, "and", (pcp), val)
405 #define raw_cpu_and_2(pcp, val) percpu_to_op(, "and", (pcp), val)
406 #define raw_cpu_and_4(pcp, val) percpu_to_op(, "and", (pcp), val)
407 #define raw_cpu_or_1(pcp, val) percpu_to_op(, "or", (pcp), val)
408 #define raw_cpu_or_2(pcp, val) percpu_to_op(, "or", (pcp), val)
409 #define raw_cpu_or_4(pcp, val) percpu_to_op(, "or", (pcp), val)
412 * raw_cpu_xchg() can use a load-store since it is not required to be
415 #define raw_percpu_xchg_op(var, nval) \
417 typeof(var) pxo_ret__ = raw_cpu_read(var); \
418 raw_cpu_write(var, (nval)); \
422 #define raw_cpu_xchg_1(pcp, val) raw_percpu_xchg_op(pcp, val)
423 #define raw_cpu_xchg_2(pcp, val) raw_percpu_xchg_op(pcp, val)
424 #define raw_cpu_xchg_4(pcp, val) raw_percpu_xchg_op(pcp, val)
426 #define this_cpu_read_1(pcp) percpu_from_op(volatile, "mov", pcp)
427 #define this_cpu_read_2(pcp) percpu_from_op(volatile, "mov", pcp)
428 #define this_cpu_read_4(pcp) percpu_from_op(volatile, "mov", pcp)
429 #define this_cpu_write_1(pcp, val) percpu_to_op(volatile, "mov", (pcp), val)
430 #define this_cpu_write_2(pcp, val) percpu_to_op(volatile, "mov", (pcp), val)
431 #define this_cpu_write_4(pcp, val) percpu_to_op(volatile, "mov", (pcp), val)
432 #define this_cpu_add_1(pcp, val) percpu_add_op(volatile, (pcp), val)
433 #define this_cpu_add_2(pcp, val) percpu_add_op(volatile, (pcp), val)
434 #define this_cpu_add_4(pcp, val) percpu_add_op(volatile, (pcp), val)
435 #define this_cpu_and_1(pcp, val) percpu_to_op(volatile, "and", (pcp), val)
436 #define this_cpu_and_2(pcp, val) percpu_to_op(volatile, "and", (pcp), val)
437 #define this_cpu_and_4(pcp, val) percpu_to_op(volatile, "and", (pcp), val)
438 #define this_cpu_or_1(pcp, val) percpu_to_op(volatile, "or", (pcp), val)
439 #define this_cpu_or_2(pcp, val) percpu_to_op(volatile, "or", (pcp), val)
440 #define this_cpu_or_4(pcp, val) percpu_to_op(volatile, "or", (pcp), val)
441 #define this_cpu_xchg_1(pcp, nval) percpu_xchg_op(volatile, pcp, nval)
442 #define this_cpu_xchg_2(pcp, nval) percpu_xchg_op(volatile, pcp, nval)
443 #define this_cpu_xchg_4(pcp, nval) percpu_xchg_op(volatile, pcp, nval)
445 #define raw_cpu_add_return_1(pcp, val) percpu_add_return_op(, pcp, val)
446 #define raw_cpu_add_return_2(pcp, val) percpu_add_return_op(, pcp, val)
447 #define raw_cpu_add_return_4(pcp, val) percpu_add_return_op(, pcp, val)
448 #define raw_cpu_cmpxchg_1(pcp, oval, nval) percpu_cmpxchg_op(, pcp, oval, nval)
449 #define raw_cpu_cmpxchg_2(pcp, oval, nval) percpu_cmpxchg_op(, pcp, oval, nval)
450 #define raw_cpu_cmpxchg_4(pcp, oval, nval) percpu_cmpxchg_op(, pcp, oval, nval)
452 #define this_cpu_add_return_1(pcp, val) percpu_add_return_op(volatile, pcp, val)
453 #define this_cpu_add_return_2(pcp, val) percpu_add_return_op(volatile, pcp, val)
454 #define this_cpu_add_return_4(pcp, val) percpu_add_return_op(volatile, pcp, val)
455 #define this_cpu_cmpxchg_1(pcp, oval, nval) percpu_cmpxchg_op(volatile, pcp, oval, nval)
456 #define this_cpu_cmpxchg_2(pcp, oval, nval) percpu_cmpxchg_op(volatile, pcp, oval, nval)
457 #define this_cpu_cmpxchg_4(pcp, oval, nval) percpu_cmpxchg_op(volatile, pcp, oval, nval)
459 #ifdef CONFIG_X86_CMPXCHG64
460 #define percpu_cmpxchg8b_double(pcp1, pcp2, o1, o2, n1, n2) \
463 typeof(pcp1) __o1 = (o1), __n1 = (n1); \
464 typeof(pcp2) __o2 = (o2), __n2 = (n2); \
465 asm volatile("cmpxchg8b "__percpu_arg(1) \
467 : CC_OUT(z) (__ret), "+m" (pcp1), "+m" (pcp2), "+a" (__o1), "+d" (__o2) \
468 : "b" (__n1), "c" (__n2)); \
472 #define raw_cpu_cmpxchg_double_4 percpu_cmpxchg8b_double
473 #define this_cpu_cmpxchg_double_4 percpu_cmpxchg8b_double
474 #endif /* CONFIG_X86_CMPXCHG64 */
477 * Per cpu atomic 64 bit operations are only available under 64 bit.
478 * 32 bit must fall back to generic operations.
481 #define raw_cpu_read_8(pcp) percpu_from_op(, "mov", pcp)
482 #define raw_cpu_write_8(pcp, val) percpu_to_op(, "mov", (pcp), val)
483 #define raw_cpu_add_8(pcp, val) percpu_add_op(, (pcp), val)
484 #define raw_cpu_and_8(pcp, val) percpu_to_op(, "and", (pcp), val)
485 #define raw_cpu_or_8(pcp, val) percpu_to_op(, "or", (pcp), val)
486 #define raw_cpu_add_return_8(pcp, val) percpu_add_return_op(, pcp, val)
487 #define raw_cpu_xchg_8(pcp, nval) raw_percpu_xchg_op(pcp, nval)
488 #define raw_cpu_cmpxchg_8(pcp, oval, nval) percpu_cmpxchg_op(, pcp, oval, nval)
490 #define this_cpu_read_8(pcp) percpu_from_op(volatile, "mov", pcp)
491 #define this_cpu_write_8(pcp, val) percpu_to_op(volatile, "mov", (pcp), val)
492 #define this_cpu_add_8(pcp, val) percpu_add_op(volatile, (pcp), val)
493 #define this_cpu_and_8(pcp, val) percpu_to_op(volatile, "and", (pcp), val)
494 #define this_cpu_or_8(pcp, val) percpu_to_op(volatile, "or", (pcp), val)
495 #define this_cpu_add_return_8(pcp, val) percpu_add_return_op(volatile, pcp, val)
496 #define this_cpu_xchg_8(pcp, nval) percpu_xchg_op(volatile, pcp, nval)
497 #define this_cpu_cmpxchg_8(pcp, oval, nval) percpu_cmpxchg_op(volatile, pcp, oval, nval)
500 * Pretty complex macro to generate cmpxchg16 instruction. The instruction
501 * is not supported on early AMD64 processors so we must be able to emulate
502 * it in software. The address used in the cmpxchg16 instruction must be
503 * aligned to a 16 byte boundary.
505 #define percpu_cmpxchg16b_double(pcp1, pcp2, o1, o2, n1, n2) \
508 typeof(pcp1) __o1 = (o1), __n1 = (n1); \
509 typeof(pcp2) __o2 = (o2), __n2 = (n2); \
510 alternative_io("leaq %P1,%%rsi\n\tcall this_cpu_cmpxchg16b_emu\n\t", \
511 "cmpxchg16b " __percpu_arg(1) "\n\tsetz %0\n\t", \
513 ASM_OUTPUT2("=a" (__ret), "+m" (pcp1), \
514 "+m" (pcp2), "+d" (__o2)), \
515 "b" (__n1), "c" (__n2), "a" (__o1) : "rsi"); \
519 #define raw_cpu_cmpxchg_double_8 percpu_cmpxchg16b_double
520 #define this_cpu_cmpxchg_double_8 percpu_cmpxchg16b_double
524 static __always_inline
bool x86_this_cpu_constant_test_bit(unsigned int nr
,
525 const unsigned long __percpu
*addr
)
527 unsigned long __percpu
*a
=
528 (unsigned long __percpu
*)addr
+ nr
/ BITS_PER_LONG
;
531 return ((1UL << (nr
% BITS_PER_LONG
)) & raw_cpu_read_8(*a
)) != 0;
533 return ((1UL << (nr
% BITS_PER_LONG
)) & raw_cpu_read_4(*a
)) != 0;
537 static inline bool x86_this_cpu_variable_test_bit(int nr
,
538 const unsigned long __percpu
*addr
)
542 asm volatile("btl "__percpu_arg(2)",%1"
545 : "m" (*(unsigned long __percpu
*)addr
), "Ir" (nr
));
550 #define x86_this_cpu_test_bit(nr, addr) \
551 (__builtin_constant_p((nr)) \
552 ? x86_this_cpu_constant_test_bit((nr), (addr)) \
553 : x86_this_cpu_variable_test_bit((nr), (addr)))
556 #include <asm-generic/percpu.h>
558 /* We can use this directly for local CPU (faster). */
559 DECLARE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off
);
561 #endif /* !__ASSEMBLY__ */
566 * Define the "EARLY_PER_CPU" macros. These are used for some per_cpu
567 * variables that are initialized and accessed before there are per_cpu
571 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue) \
572 DEFINE_PER_CPU(_type, _name) = _initvalue; \
573 __typeof__(_type) _name##_early_map[NR_CPUS] __initdata = \
574 { [0 ... NR_CPUS-1] = _initvalue }; \
575 __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
577 #define DEFINE_EARLY_PER_CPU_READ_MOSTLY(_type, _name, _initvalue) \
578 DEFINE_PER_CPU_READ_MOSTLY(_type, _name) = _initvalue; \
579 __typeof__(_type) _name##_early_map[NR_CPUS] __initdata = \
580 { [0 ... NR_CPUS-1] = _initvalue }; \
581 __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
583 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name) \
584 EXPORT_PER_CPU_SYMBOL(_name)
586 #define DECLARE_EARLY_PER_CPU(_type, _name) \
587 DECLARE_PER_CPU(_type, _name); \
588 extern __typeof__(_type) *_name##_early_ptr; \
589 extern __typeof__(_type) _name##_early_map[]
591 #define DECLARE_EARLY_PER_CPU_READ_MOSTLY(_type, _name) \
592 DECLARE_PER_CPU_READ_MOSTLY(_type, _name); \
593 extern __typeof__(_type) *_name##_early_ptr; \
594 extern __typeof__(_type) _name##_early_map[]
596 #define early_per_cpu_ptr(_name) (_name##_early_ptr)
597 #define early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
598 #define early_per_cpu(_name, _cpu) \
599 *(early_per_cpu_ptr(_name) ? \
600 &early_per_cpu_ptr(_name)[_cpu] : \
601 &per_cpu(_name, _cpu))
603 #else /* !CONFIG_SMP */
604 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue) \
605 DEFINE_PER_CPU(_type, _name) = _initvalue
607 #define DEFINE_EARLY_PER_CPU_READ_MOSTLY(_type, _name, _initvalue) \
608 DEFINE_PER_CPU_READ_MOSTLY(_type, _name) = _initvalue
610 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name) \
611 EXPORT_PER_CPU_SYMBOL(_name)
613 #define DECLARE_EARLY_PER_CPU(_type, _name) \
614 DECLARE_PER_CPU(_type, _name)
616 #define DECLARE_EARLY_PER_CPU_READ_MOSTLY(_type, _name) \
617 DECLARE_PER_CPU_READ_MOSTLY(_type, _name)
619 #define early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
620 #define early_per_cpu_ptr(_name) NULL
621 /* no early_per_cpu_map() */
623 #endif /* !CONFIG_SMP */
625 #endif /* _ASM_X86_PERCPU_H */