18c2ecf20Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */
28c2ecf20Sopenharmony_ci/*
38c2ecf20Sopenharmony_ci * Copyright (C) 2020 Loongson Technology Ltd.
48c2ecf20Sopenharmony_ci */
58c2ecf20Sopenharmony_ci#ifndef __ASM_PERCPU_H
68c2ecf20Sopenharmony_ci#define __ASM_PERCPU_H
78c2ecf20Sopenharmony_ci
88c2ecf20Sopenharmony_ci#include <asm/cmpxchg.h>
98c2ecf20Sopenharmony_ci
108c2ecf20Sopenharmony_ci/*
118c2ecf20Sopenharmony_ci * The "address" (in fact, offset from $r21) of a per-CPU variable is close to
128c2ecf20Sopenharmony_ci * the loading address of main kernel image, but far from where the modules are
138c2ecf20Sopenharmony_ci * loaded. Tell the compiler this fact when using explicit relocs.
148c2ecf20Sopenharmony_ci */
158c2ecf20Sopenharmony_ci#if defined(MODULE) && defined(CONFIG_AS_HAS_EXPLICIT_RELOCS)
168c2ecf20Sopenharmony_ci# if __has_attribute(model)
178c2ecf20Sopenharmony_ci#  define PER_CPU_ATTRIBUTES __attribute__((model("extreme")))
188c2ecf20Sopenharmony_ci# else
198c2ecf20Sopenharmony_ci#  error compiler support for the model attribute is necessary when a recent assembler is used
208c2ecf20Sopenharmony_ci# endif
218c2ecf20Sopenharmony_ci#endif
228c2ecf20Sopenharmony_ci
238c2ecf20Sopenharmony_ci/* Use r21 for fast access */
248c2ecf20Sopenharmony_ciregister unsigned long __my_cpu_offset __asm__("$r21");
258c2ecf20Sopenharmony_ci
268c2ecf20Sopenharmony_cistatic inline void set_my_cpu_offset(unsigned long off)
278c2ecf20Sopenharmony_ci{
288c2ecf20Sopenharmony_ci	__my_cpu_offset = off;
298c2ecf20Sopenharmony_ci	csr_write64(off, PERCPU_BASE_KS);
308c2ecf20Sopenharmony_ci}
318c2ecf20Sopenharmony_ci#define __my_cpu_offset __my_cpu_offset
328c2ecf20Sopenharmony_ci
338c2ecf20Sopenharmony_ci#define PERCPU_OP(op, asm_op, c_op)					\
348c2ecf20Sopenharmony_cistatic __always_inline unsigned long __percpu_##op(void *ptr,		\
358c2ecf20Sopenharmony_ci			unsigned long val, int size)			\
368c2ecf20Sopenharmony_ci{									\
378c2ecf20Sopenharmony_ci	unsigned long ret;						\
388c2ecf20Sopenharmony_ci									\
398c2ecf20Sopenharmony_ci	switch (size) {							\
408c2ecf20Sopenharmony_ci	case 4:								\
418c2ecf20Sopenharmony_ci		__asm__ __volatile__(					\
428c2ecf20Sopenharmony_ci		"am"#asm_op".w"	" %[ret], %[val], %[ptr]	\n"	\
438c2ecf20Sopenharmony_ci		: [ret] "=&r" (ret), [ptr] "+ZB"(*(u32 *)ptr)		\
448c2ecf20Sopenharmony_ci		: [val] "r" (val));					\
458c2ecf20Sopenharmony_ci		break;							\
468c2ecf20Sopenharmony_ci	case 8:								\
478c2ecf20Sopenharmony_ci		__asm__ __volatile__(					\
488c2ecf20Sopenharmony_ci		"am"#asm_op".d" " %[ret], %[val], %[ptr]	\n"	\
498c2ecf20Sopenharmony_ci		: [ret] "=&r" (ret), [ptr] "+ZB"(*(u64 *)ptr)		\
508c2ecf20Sopenharmony_ci		: [val] "r" (val));					\
518c2ecf20Sopenharmony_ci		break;							\
528c2ecf20Sopenharmony_ci	default:							\
538c2ecf20Sopenharmony_ci		ret = 0;						\
548c2ecf20Sopenharmony_ci		BUILD_BUG();						\
558c2ecf20Sopenharmony_ci	}								\
568c2ecf20Sopenharmony_ci									\
578c2ecf20Sopenharmony_ci	return ret c_op val;						\
588c2ecf20Sopenharmony_ci}
598c2ecf20Sopenharmony_ci
608c2ecf20Sopenharmony_ciPERCPU_OP(add, add, +)
618c2ecf20Sopenharmony_ciPERCPU_OP(and, and, &)
628c2ecf20Sopenharmony_ciPERCPU_OP(or, or, |)
638c2ecf20Sopenharmony_ci#undef PERCPU_OP
648c2ecf20Sopenharmony_ci
658c2ecf20Sopenharmony_cistatic __always_inline unsigned long __percpu_read(void __percpu *ptr, int size)
668c2ecf20Sopenharmony_ci{
678c2ecf20Sopenharmony_ci	unsigned long ret;
688c2ecf20Sopenharmony_ci
698c2ecf20Sopenharmony_ci	switch (size) {
708c2ecf20Sopenharmony_ci	case 1:
718c2ecf20Sopenharmony_ci		__asm__ __volatile__ ("ldx.b %[ret], $r21, %[ptr]	\n"
728c2ecf20Sopenharmony_ci		: [ret] "=&r"(ret)
738c2ecf20Sopenharmony_ci		: [ptr] "r"(ptr)
748c2ecf20Sopenharmony_ci		: "memory");
758c2ecf20Sopenharmony_ci		break;
768c2ecf20Sopenharmony_ci	case 2:
778c2ecf20Sopenharmony_ci		__asm__ __volatile__ ("ldx.h %[ret], $r21, %[ptr]	\n"
788c2ecf20Sopenharmony_ci		: [ret] "=&r"(ret)
798c2ecf20Sopenharmony_ci		: [ptr] "r"(ptr)
808c2ecf20Sopenharmony_ci		: "memory");
818c2ecf20Sopenharmony_ci		break;
828c2ecf20Sopenharmony_ci	case 4:
838c2ecf20Sopenharmony_ci		__asm__ __volatile__ ("ldx.w %[ret], $r21, %[ptr]	\n"
848c2ecf20Sopenharmony_ci		: [ret] "=&r"(ret)
858c2ecf20Sopenharmony_ci		: [ptr] "r"(ptr)
868c2ecf20Sopenharmony_ci		: "memory");
878c2ecf20Sopenharmony_ci		break;
888c2ecf20Sopenharmony_ci	case 8:
898c2ecf20Sopenharmony_ci		__asm__ __volatile__ ("ldx.d %[ret], $r21, %[ptr]	\n"
908c2ecf20Sopenharmony_ci		: [ret] "=&r"(ret)
918c2ecf20Sopenharmony_ci		: [ptr] "r"(ptr)
928c2ecf20Sopenharmony_ci		: "memory");
938c2ecf20Sopenharmony_ci		break;
948c2ecf20Sopenharmony_ci	default:
958c2ecf20Sopenharmony_ci		ret = 0;
968c2ecf20Sopenharmony_ci		BUILD_BUG();
978c2ecf20Sopenharmony_ci	}
988c2ecf20Sopenharmony_ci
998c2ecf20Sopenharmony_ci	return ret;
1008c2ecf20Sopenharmony_ci}
1018c2ecf20Sopenharmony_ci
1028c2ecf20Sopenharmony_cistatic __always_inline void __percpu_write(void __percpu *ptr, unsigned long val, int size)
1038c2ecf20Sopenharmony_ci{
1048c2ecf20Sopenharmony_ci	switch (size) {
1058c2ecf20Sopenharmony_ci	case 1:
1068c2ecf20Sopenharmony_ci		__asm__ __volatile__("stx.b %[val], $r21, %[ptr]	\n"
1078c2ecf20Sopenharmony_ci		:
1088c2ecf20Sopenharmony_ci		: [val] "r" (val), [ptr] "r" (ptr)
1098c2ecf20Sopenharmony_ci		: "memory");
1108c2ecf20Sopenharmony_ci		break;
1118c2ecf20Sopenharmony_ci	case 2:
1128c2ecf20Sopenharmony_ci		__asm__ __volatile__("stx.h %[val], $r21, %[ptr]	\n"
1138c2ecf20Sopenharmony_ci		:
1148c2ecf20Sopenharmony_ci		: [val] "r" (val), [ptr] "r" (ptr)
1158c2ecf20Sopenharmony_ci		: "memory");
1168c2ecf20Sopenharmony_ci		break;
1178c2ecf20Sopenharmony_ci	case 4:
1188c2ecf20Sopenharmony_ci		__asm__ __volatile__("stx.w %[val], $r21, %[ptr]	\n"
1198c2ecf20Sopenharmony_ci		:
1208c2ecf20Sopenharmony_ci		: [val] "r" (val), [ptr] "r" (ptr)
1218c2ecf20Sopenharmony_ci		: "memory");
1228c2ecf20Sopenharmony_ci		break;
1238c2ecf20Sopenharmony_ci	case 8:
1248c2ecf20Sopenharmony_ci		__asm__ __volatile__("stx.d %[val], $r21, %[ptr]	\n"
1258c2ecf20Sopenharmony_ci		:
1268c2ecf20Sopenharmony_ci		: [val] "r" (val), [ptr] "r" (ptr)
1278c2ecf20Sopenharmony_ci		: "memory");
1288c2ecf20Sopenharmony_ci		break;
1298c2ecf20Sopenharmony_ci	default:
1308c2ecf20Sopenharmony_ci		BUILD_BUG();
1318c2ecf20Sopenharmony_ci	}
1328c2ecf20Sopenharmony_ci}
1338c2ecf20Sopenharmony_ci
1348c2ecf20Sopenharmony_cistatic __always_inline unsigned long __percpu_xchg(void *ptr, unsigned long val, int size)
1358c2ecf20Sopenharmony_ci{
1368c2ecf20Sopenharmony_ci	switch (size) {
1378c2ecf20Sopenharmony_ci	case 1:
1388c2ecf20Sopenharmony_ci	case 2:
1398c2ecf20Sopenharmony_ci		return __xchg_small((volatile void *)ptr, val, size);
1408c2ecf20Sopenharmony_ci
1418c2ecf20Sopenharmony_ci	case 4:
1428c2ecf20Sopenharmony_ci		return __xchg_asm("amswap.w", (volatile u32 *)ptr, (u32)val);
1438c2ecf20Sopenharmony_ci
1448c2ecf20Sopenharmony_ci	case 8:
1458c2ecf20Sopenharmony_ci		return __xchg_asm("amswap.d", (volatile u64 *)ptr, (u64)val);
1468c2ecf20Sopenharmony_ci
1478c2ecf20Sopenharmony_ci	default:
1488c2ecf20Sopenharmony_ci		BUILD_BUG();
1498c2ecf20Sopenharmony_ci	}
1508c2ecf20Sopenharmony_ci
1518c2ecf20Sopenharmony_ci	return 0;
1528c2ecf20Sopenharmony_ci}
1538c2ecf20Sopenharmony_ci
1548c2ecf20Sopenharmony_ci/* this_cpu_cmpxchg */
1558c2ecf20Sopenharmony_ci#define _protect_cmpxchg_local(pcp, o, n)			\
1568c2ecf20Sopenharmony_ci({								\
1578c2ecf20Sopenharmony_ci	typeof(*raw_cpu_ptr(&(pcp))) __ret;			\
1588c2ecf20Sopenharmony_ci	preempt_disable_notrace();				\
1598c2ecf20Sopenharmony_ci	__ret = cmpxchg_local(raw_cpu_ptr(&(pcp)), o, n);	\
1608c2ecf20Sopenharmony_ci	preempt_enable_notrace();				\
1618c2ecf20Sopenharmony_ci	__ret;							\
1628c2ecf20Sopenharmony_ci})
1638c2ecf20Sopenharmony_ci
1648c2ecf20Sopenharmony_ci#define _percpu_read(pcp)						\
1658c2ecf20Sopenharmony_ci({									\
1668c2ecf20Sopenharmony_ci	typeof(pcp) __retval;						\
1678c2ecf20Sopenharmony_ci	__retval = (typeof(pcp))__percpu_read(&(pcp), sizeof(pcp));	\
1688c2ecf20Sopenharmony_ci	__retval;							\
1698c2ecf20Sopenharmony_ci})
1708c2ecf20Sopenharmony_ci
1718c2ecf20Sopenharmony_ci#define _percpu_write(pcp, val)						\
1728c2ecf20Sopenharmony_cido {									\
1738c2ecf20Sopenharmony_ci	__percpu_write(&(pcp), (unsigned long)(val), sizeof(pcp));	\
1748c2ecf20Sopenharmony_ci} while (0)								\
1758c2ecf20Sopenharmony_ci
1768c2ecf20Sopenharmony_ci#define _pcp_protect(operation, pcp, val)			\
1778c2ecf20Sopenharmony_ci({								\
1788c2ecf20Sopenharmony_ci	typeof(pcp) __retval;					\
1798c2ecf20Sopenharmony_ci	preempt_disable_notrace();				\
1808c2ecf20Sopenharmony_ci	__retval = (typeof(pcp))operation(raw_cpu_ptr(&(pcp)),	\
1818c2ecf20Sopenharmony_ci					  (val), sizeof(pcp));	\
1828c2ecf20Sopenharmony_ci	preempt_enable_notrace();				\
1838c2ecf20Sopenharmony_ci	__retval;						\
1848c2ecf20Sopenharmony_ci})
1858c2ecf20Sopenharmony_ci
1868c2ecf20Sopenharmony_ci#define _percpu_add(pcp, val) \
1878c2ecf20Sopenharmony_ci	_pcp_protect(__percpu_add, pcp, val)
1888c2ecf20Sopenharmony_ci
1898c2ecf20Sopenharmony_ci#define _percpu_add_return(pcp, val) _percpu_add(pcp, val)
1908c2ecf20Sopenharmony_ci
1918c2ecf20Sopenharmony_ci#define _percpu_and(pcp, val) \
1928c2ecf20Sopenharmony_ci	_pcp_protect(__percpu_and, pcp, val)
1938c2ecf20Sopenharmony_ci
1948c2ecf20Sopenharmony_ci#define _percpu_or(pcp, val) \
1958c2ecf20Sopenharmony_ci	_pcp_protect(__percpu_or, pcp, val)
1968c2ecf20Sopenharmony_ci
1978c2ecf20Sopenharmony_ci#define _percpu_xchg(pcp, val) ((typeof(pcp)) \
1988c2ecf20Sopenharmony_ci	_pcp_protect(__percpu_xchg, pcp, (unsigned long)(val)))
1998c2ecf20Sopenharmony_ci
2008c2ecf20Sopenharmony_ci#define this_cpu_add_4(pcp, val) _percpu_add(pcp, val)
2018c2ecf20Sopenharmony_ci#define this_cpu_add_8(pcp, val) _percpu_add(pcp, val)
2028c2ecf20Sopenharmony_ci
2038c2ecf20Sopenharmony_ci#define this_cpu_add_return_4(pcp, val) _percpu_add_return(pcp, val)
2048c2ecf20Sopenharmony_ci#define this_cpu_add_return_8(pcp, val) _percpu_add_return(pcp, val)
2058c2ecf20Sopenharmony_ci
2068c2ecf20Sopenharmony_ci#define this_cpu_and_4(pcp, val) _percpu_and(pcp, val)
2078c2ecf20Sopenharmony_ci#define this_cpu_and_8(pcp, val) _percpu_and(pcp, val)
2088c2ecf20Sopenharmony_ci
2098c2ecf20Sopenharmony_ci#define this_cpu_or_4(pcp, val) _percpu_or(pcp, val)
2108c2ecf20Sopenharmony_ci#define this_cpu_or_8(pcp, val) _percpu_or(pcp, val)
2118c2ecf20Sopenharmony_ci
2128c2ecf20Sopenharmony_ci#define this_cpu_read_1(pcp) _percpu_read(pcp)
2138c2ecf20Sopenharmony_ci#define this_cpu_read_2(pcp) _percpu_read(pcp)
2148c2ecf20Sopenharmony_ci#define this_cpu_read_4(pcp) _percpu_read(pcp)
2158c2ecf20Sopenharmony_ci#define this_cpu_read_8(pcp) _percpu_read(pcp)
2168c2ecf20Sopenharmony_ci
2178c2ecf20Sopenharmony_ci#define this_cpu_write_1(pcp, val) _percpu_write(pcp, val)
2188c2ecf20Sopenharmony_ci#define this_cpu_write_2(pcp, val) _percpu_write(pcp, val)
2198c2ecf20Sopenharmony_ci#define this_cpu_write_4(pcp, val) _percpu_write(pcp, val)
2208c2ecf20Sopenharmony_ci#define this_cpu_write_8(pcp, val) _percpu_write(pcp, val)
2218c2ecf20Sopenharmony_ci
2228c2ecf20Sopenharmony_ci#define this_cpu_xchg_1(pcp, val) _percpu_xchg(pcp, val)
2238c2ecf20Sopenharmony_ci#define this_cpu_xchg_2(pcp, val) _percpu_xchg(pcp, val)
2248c2ecf20Sopenharmony_ci#define this_cpu_xchg_4(pcp, val) _percpu_xchg(pcp, val)
2258c2ecf20Sopenharmony_ci#define this_cpu_xchg_8(pcp, val) _percpu_xchg(pcp, val)
2268c2ecf20Sopenharmony_ci
2278c2ecf20Sopenharmony_ci#define this_cpu_cmpxchg_1(ptr, o, n) _protect_cmpxchg_local(ptr, o, n)
2288c2ecf20Sopenharmony_ci#define this_cpu_cmpxchg_2(ptr, o, n) _protect_cmpxchg_local(ptr, o, n)
2298c2ecf20Sopenharmony_ci#define this_cpu_cmpxchg_4(ptr, o, n) _protect_cmpxchg_local(ptr, o, n)
2308c2ecf20Sopenharmony_ci#define this_cpu_cmpxchg_8(ptr, o, n) _protect_cmpxchg_local(ptr, o, n)
2318c2ecf20Sopenharmony_ci
2328c2ecf20Sopenharmony_ci#include <asm-generic/percpu.h>
2338c2ecf20Sopenharmony_ci
2348c2ecf20Sopenharmony_ci#endif /* __ASM_PERCPU_H */
235