18c2ecf20Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */
28c2ecf20Sopenharmony_ci/*
38c2ecf20Sopenharmony_ci * Author: Huacai Chen <chenhuacai@loongson.cn>
48c2ecf20Sopenharmony_ci * Copyright (C) 2020 Loongson Technology Corporation Limited
58c2ecf20Sopenharmony_ci */
68c2ecf20Sopenharmony_ci#ifndef _ASM_FPU_H
78c2ecf20Sopenharmony_ci#define _ASM_FPU_H
88c2ecf20Sopenharmony_ci
98c2ecf20Sopenharmony_ci#include <linux/sched.h>
108c2ecf20Sopenharmony_ci#include <linux/sched/task_stack.h>
118c2ecf20Sopenharmony_ci#include <linux/ptrace.h>
128c2ecf20Sopenharmony_ci#include <linux/thread_info.h>
138c2ecf20Sopenharmony_ci#include <linux/bitops.h>
148c2ecf20Sopenharmony_ci
158c2ecf20Sopenharmony_ci#include <asm/cpu.h>
168c2ecf20Sopenharmony_ci#include <asm/cpu-features.h>
178c2ecf20Sopenharmony_ci#include <asm/current.h>
188c2ecf20Sopenharmony_ci#include <asm/inst.h>
198c2ecf20Sopenharmony_ci#include <asm/loongarchregs.h>
208c2ecf20Sopenharmony_ci#include <asm/ptrace.h>
218c2ecf20Sopenharmony_ci#include <asm/processor.h>
228c2ecf20Sopenharmony_ci
238c2ecf20Sopenharmony_cistruct sigcontext;
248c2ecf20Sopenharmony_ci
258c2ecf20Sopenharmony_ciextern void kernel_fpu_begin(void);
268c2ecf20Sopenharmony_ciextern void kernel_fpu_end(void);
278c2ecf20Sopenharmony_ci
288c2ecf20Sopenharmony_ciextern void _init_fpu(unsigned int);
298c2ecf20Sopenharmony_ciextern void _save_fp(struct loongarch_fpu *);
308c2ecf20Sopenharmony_ciextern void _restore_fp(struct loongarch_fpu *);
318c2ecf20Sopenharmony_ci
328c2ecf20Sopenharmony_ciextern void _save_lsx(struct loongarch_fpu *fpu);
338c2ecf20Sopenharmony_ciextern void _restore_lsx(struct loongarch_fpu *fpu);
348c2ecf20Sopenharmony_ciextern void _init_lsx_upper(void);
358c2ecf20Sopenharmony_ciextern void _restore_lsx_upper(struct loongarch_fpu *fpu);
368c2ecf20Sopenharmony_ci
378c2ecf20Sopenharmony_ciextern void _save_lasx(struct loongarch_fpu *fpu);
388c2ecf20Sopenharmony_ciextern void _restore_lasx(struct loongarch_fpu *fpu);
398c2ecf20Sopenharmony_ciextern void _init_lasx_upper(void);
408c2ecf20Sopenharmony_ciextern void _restore_lasx_upper(struct loongarch_fpu *fpu);
418c2ecf20Sopenharmony_ci
428c2ecf20Sopenharmony_cistatic inline void enable_lsx(void);
438c2ecf20Sopenharmony_cistatic inline void disable_lsx(void);
448c2ecf20Sopenharmony_cistatic inline void save_lsx(struct task_struct *t);
458c2ecf20Sopenharmony_cistatic inline void restore_lsx(struct task_struct *t);
468c2ecf20Sopenharmony_ci
478c2ecf20Sopenharmony_cistatic inline void enable_lasx(void);
488c2ecf20Sopenharmony_cistatic inline void disable_lasx(void);
498c2ecf20Sopenharmony_cistatic inline void save_lasx(struct task_struct *t);
508c2ecf20Sopenharmony_cistatic inline void restore_lasx(struct task_struct *t);
518c2ecf20Sopenharmony_ci
528c2ecf20Sopenharmony_ci/*
538c2ecf20Sopenharmony_ci * Mask the FCSR Cause bits according to the Enable bits, observing
548c2ecf20Sopenharmony_ci * that Unimplemented is always enabled.
558c2ecf20Sopenharmony_ci */
568c2ecf20Sopenharmony_cistatic inline unsigned long mask_fcsr_x(unsigned long fcsr)
578c2ecf20Sopenharmony_ci{
588c2ecf20Sopenharmony_ci	return fcsr & ((fcsr & FPU_CSR_ALL_E) <<
598c2ecf20Sopenharmony_ci			(ffs(FPU_CSR_ALL_X) - ffs(FPU_CSR_ALL_E)));
608c2ecf20Sopenharmony_ci}
618c2ecf20Sopenharmony_ci
628c2ecf20Sopenharmony_cistatic inline int is_fp_enabled(void)
638c2ecf20Sopenharmony_ci{
648c2ecf20Sopenharmony_ci	return (csr_read32(LOONGARCH_CSR_EUEN) & CSR_EUEN_FPEN) ?
658c2ecf20Sopenharmony_ci		1 : 0;
668c2ecf20Sopenharmony_ci}
678c2ecf20Sopenharmony_ci
688c2ecf20Sopenharmony_cistatic inline int is_lsx_enabled(void)
698c2ecf20Sopenharmony_ci{
708c2ecf20Sopenharmony_ci	if (!cpu_has_lsx)
718c2ecf20Sopenharmony_ci		return 0;
728c2ecf20Sopenharmony_ci
738c2ecf20Sopenharmony_ci	return (csr_read32(LOONGARCH_CSR_EUEN) & CSR_EUEN_LSXEN) ?
748c2ecf20Sopenharmony_ci		1 : 0;
758c2ecf20Sopenharmony_ci}
768c2ecf20Sopenharmony_ci
778c2ecf20Sopenharmony_cistatic inline int is_lasx_enabled(void)
788c2ecf20Sopenharmony_ci{
798c2ecf20Sopenharmony_ci	if (!cpu_has_lasx)
808c2ecf20Sopenharmony_ci		return 0;
818c2ecf20Sopenharmony_ci
828c2ecf20Sopenharmony_ci	return (csr_read32(LOONGARCH_CSR_EUEN) & CSR_EUEN_LASXEN) ?
838c2ecf20Sopenharmony_ci		1 : 0;
848c2ecf20Sopenharmony_ci}
858c2ecf20Sopenharmony_ci
868c2ecf20Sopenharmony_cistatic inline int is_simd_enabled(void)
878c2ecf20Sopenharmony_ci{
888c2ecf20Sopenharmony_ci	return is_lsx_enabled() | is_lasx_enabled();
898c2ecf20Sopenharmony_ci}
908c2ecf20Sopenharmony_ci
918c2ecf20Sopenharmony_ci#define enable_fpu()						\
928c2ecf20Sopenharmony_cido {								\
938c2ecf20Sopenharmony_ci	set_csr_euen(CSR_EUEN_FPEN);				\
948c2ecf20Sopenharmony_ci} while (0)
958c2ecf20Sopenharmony_ci
968c2ecf20Sopenharmony_ci#define disable_fpu()						\
978c2ecf20Sopenharmony_cido {								\
988c2ecf20Sopenharmony_ci	clear_csr_euen(CSR_EUEN_FPEN);				\
998c2ecf20Sopenharmony_ci} while (0)
1008c2ecf20Sopenharmony_ci
1018c2ecf20Sopenharmony_ci#define clear_fpu_owner()	clear_thread_flag(TIF_USEDFPU)
1028c2ecf20Sopenharmony_ci
1038c2ecf20Sopenharmony_cistatic inline int is_fpu_owner(void)
1048c2ecf20Sopenharmony_ci{
1058c2ecf20Sopenharmony_ci	return test_thread_flag(TIF_USEDFPU);
1068c2ecf20Sopenharmony_ci}
1078c2ecf20Sopenharmony_ci
1088c2ecf20Sopenharmony_cistatic inline void __own_fpu(void)
1098c2ecf20Sopenharmony_ci{
1108c2ecf20Sopenharmony_ci	enable_fpu();
1118c2ecf20Sopenharmony_ci	set_thread_flag(TIF_USEDFPU);
1128c2ecf20Sopenharmony_ci	KSTK_EUEN(current) |= CSR_EUEN_FPEN;
1138c2ecf20Sopenharmony_ci}
1148c2ecf20Sopenharmony_ci
1158c2ecf20Sopenharmony_cistatic inline void own_fpu_inatomic(int restore)
1168c2ecf20Sopenharmony_ci{
1178c2ecf20Sopenharmony_ci	if (cpu_has_fpu && !is_fpu_owner()) {
1188c2ecf20Sopenharmony_ci		__own_fpu();
1198c2ecf20Sopenharmony_ci		if (restore)
1208c2ecf20Sopenharmony_ci			_restore_fp(&current->thread.fpu);
1218c2ecf20Sopenharmony_ci	}
1228c2ecf20Sopenharmony_ci}
1238c2ecf20Sopenharmony_ci
1248c2ecf20Sopenharmony_cistatic inline void own_fpu(int restore)
1258c2ecf20Sopenharmony_ci{
1268c2ecf20Sopenharmony_ci	preempt_disable();
1278c2ecf20Sopenharmony_ci	own_fpu_inatomic(restore);
1288c2ecf20Sopenharmony_ci	preempt_enable();
1298c2ecf20Sopenharmony_ci}
1308c2ecf20Sopenharmony_ci
1318c2ecf20Sopenharmony_cistatic inline void lose_fpu_inatomic(int save, struct task_struct *tsk)
1328c2ecf20Sopenharmony_ci{
1338c2ecf20Sopenharmony_ci	if (is_fpu_owner()) {
1348c2ecf20Sopenharmony_ci		if (!is_simd_enabled()) {
1358c2ecf20Sopenharmony_ci			if (save)
1368c2ecf20Sopenharmony_ci				_save_fp(&tsk->thread.fpu);
1378c2ecf20Sopenharmony_ci			disable_fpu();
1388c2ecf20Sopenharmony_ci		} else {
1398c2ecf20Sopenharmony_ci			if (save) {
1408c2ecf20Sopenharmony_ci				if (!is_lasx_enabled())
1418c2ecf20Sopenharmony_ci					save_lsx(tsk);
1428c2ecf20Sopenharmony_ci				else
1438c2ecf20Sopenharmony_ci					save_lasx(tsk);
1448c2ecf20Sopenharmony_ci			}
1458c2ecf20Sopenharmony_ci			disable_fpu();
1468c2ecf20Sopenharmony_ci			disable_lsx();
1478c2ecf20Sopenharmony_ci			disable_lasx();
1488c2ecf20Sopenharmony_ci			clear_tsk_thread_flag(tsk, TIF_USEDSIMD);
1498c2ecf20Sopenharmony_ci		}
1508c2ecf20Sopenharmony_ci		clear_tsk_thread_flag(tsk, TIF_USEDFPU);
1518c2ecf20Sopenharmony_ci	}
1528c2ecf20Sopenharmony_ci	KSTK_EUEN(tsk) &= ~(CSR_EUEN_FPEN | CSR_EUEN_LSXEN | CSR_EUEN_LASXEN);
1538c2ecf20Sopenharmony_ci}
1548c2ecf20Sopenharmony_ci
1558c2ecf20Sopenharmony_cistatic inline void lose_fpu(int save)
1568c2ecf20Sopenharmony_ci{
1578c2ecf20Sopenharmony_ci	preempt_disable();
1588c2ecf20Sopenharmony_ci	lose_fpu_inatomic(save, current);
1598c2ecf20Sopenharmony_ci	preempt_enable();
1608c2ecf20Sopenharmony_ci}
1618c2ecf20Sopenharmony_ci
1628c2ecf20Sopenharmony_cistatic inline void init_fpu(void)
1638c2ecf20Sopenharmony_ci{
1648c2ecf20Sopenharmony_ci	unsigned int fcsr = current->thread.fpu.fcsr;
1658c2ecf20Sopenharmony_ci
1668c2ecf20Sopenharmony_ci	__own_fpu();
1678c2ecf20Sopenharmony_ci	_init_fpu(fcsr);
1688c2ecf20Sopenharmony_ci	set_used_math();
1698c2ecf20Sopenharmony_ci}
1708c2ecf20Sopenharmony_ci
1718c2ecf20Sopenharmony_cistatic inline void save_fp(struct task_struct *tsk)
1728c2ecf20Sopenharmony_ci{
1738c2ecf20Sopenharmony_ci	if (cpu_has_fpu)
1748c2ecf20Sopenharmony_ci		_save_fp(&tsk->thread.fpu);
1758c2ecf20Sopenharmony_ci}
1768c2ecf20Sopenharmony_ci
1778c2ecf20Sopenharmony_cistatic inline void restore_fp(struct task_struct *tsk)
1788c2ecf20Sopenharmony_ci{
1798c2ecf20Sopenharmony_ci	if (cpu_has_fpu)
1808c2ecf20Sopenharmony_ci		_restore_fp(&tsk->thread.fpu);
1818c2ecf20Sopenharmony_ci}
1828c2ecf20Sopenharmony_ci
1838c2ecf20Sopenharmony_cistatic inline void save_fpu_regs(struct task_struct *tsk)
1848c2ecf20Sopenharmony_ci{
1858c2ecf20Sopenharmony_ci	unsigned int euen;
1868c2ecf20Sopenharmony_ci
1878c2ecf20Sopenharmony_ci	if (tsk == current) {
1888c2ecf20Sopenharmony_ci		preempt_disable();
1898c2ecf20Sopenharmony_ci
1908c2ecf20Sopenharmony_ci		euen = csr_read32(LOONGARCH_CSR_EUEN);
1918c2ecf20Sopenharmony_ci
1928c2ecf20Sopenharmony_ci#ifdef CONFIG_CPU_HAS_LASX
1938c2ecf20Sopenharmony_ci		if (euen & CSR_EUEN_LASXEN)
1948c2ecf20Sopenharmony_ci			_save_lasx(&current->thread.fpu);
1958c2ecf20Sopenharmony_ci		else
1968c2ecf20Sopenharmony_ci#endif
1978c2ecf20Sopenharmony_ci#ifdef CONFIG_CPU_HAS_LSX
1988c2ecf20Sopenharmony_ci		if (euen & CSR_EUEN_LSXEN)
1998c2ecf20Sopenharmony_ci			_save_lsx(&current->thread.fpu);
2008c2ecf20Sopenharmony_ci		else
2018c2ecf20Sopenharmony_ci#endif
2028c2ecf20Sopenharmony_ci		if (euen & CSR_EUEN_FPEN)
2038c2ecf20Sopenharmony_ci			_save_fp(&current->thread.fpu);
2048c2ecf20Sopenharmony_ci
2058c2ecf20Sopenharmony_ci		preempt_enable();
2068c2ecf20Sopenharmony_ci	}
2078c2ecf20Sopenharmony_ci}
2088c2ecf20Sopenharmony_ci
2098c2ecf20Sopenharmony_cistatic inline int is_simd_owner(void)
2108c2ecf20Sopenharmony_ci{
2118c2ecf20Sopenharmony_ci	return test_thread_flag(TIF_USEDSIMD);
2128c2ecf20Sopenharmony_ci}
2138c2ecf20Sopenharmony_ci
2148c2ecf20Sopenharmony_ci#ifdef CONFIG_CPU_HAS_LSX
2158c2ecf20Sopenharmony_ci
2168c2ecf20Sopenharmony_cistatic inline void enable_lsx(void)
2178c2ecf20Sopenharmony_ci{
2188c2ecf20Sopenharmony_ci	if (cpu_has_lsx)
2198c2ecf20Sopenharmony_ci		csr_xchg32(CSR_EUEN_LSXEN, CSR_EUEN_LSXEN, LOONGARCH_CSR_EUEN);
2208c2ecf20Sopenharmony_ci}
2218c2ecf20Sopenharmony_ci
2228c2ecf20Sopenharmony_cistatic inline void disable_lsx(void)
2238c2ecf20Sopenharmony_ci{
2248c2ecf20Sopenharmony_ci	if (cpu_has_lsx)
2258c2ecf20Sopenharmony_ci		csr_xchg32(0, CSR_EUEN_LSXEN, LOONGARCH_CSR_EUEN);
2268c2ecf20Sopenharmony_ci}
2278c2ecf20Sopenharmony_ci
2288c2ecf20Sopenharmony_cistatic inline void save_lsx(struct task_struct *t)
2298c2ecf20Sopenharmony_ci{
2308c2ecf20Sopenharmony_ci	if (cpu_has_lsx)
2318c2ecf20Sopenharmony_ci		_save_lsx(&t->thread.fpu);
2328c2ecf20Sopenharmony_ci}
2338c2ecf20Sopenharmony_ci
2348c2ecf20Sopenharmony_cistatic inline void restore_lsx(struct task_struct *t)
2358c2ecf20Sopenharmony_ci{
2368c2ecf20Sopenharmony_ci	if (cpu_has_lsx)
2378c2ecf20Sopenharmony_ci		_restore_lsx(&t->thread.fpu);
2388c2ecf20Sopenharmony_ci}
2398c2ecf20Sopenharmony_ci
2408c2ecf20Sopenharmony_cistatic inline void init_lsx_upper(void)
2418c2ecf20Sopenharmony_ci{
2428c2ecf20Sopenharmony_ci	if (cpu_has_lsx)
2438c2ecf20Sopenharmony_ci		_init_lsx_upper();
2448c2ecf20Sopenharmony_ci}
2458c2ecf20Sopenharmony_ci
2468c2ecf20Sopenharmony_cistatic inline void restore_lsx_upper(struct task_struct *t)
2478c2ecf20Sopenharmony_ci{
2488c2ecf20Sopenharmony_ci	if (cpu_has_lsx)
2498c2ecf20Sopenharmony_ci		_restore_lsx_upper(&t->thread.fpu);
2508c2ecf20Sopenharmony_ci}
2518c2ecf20Sopenharmony_ci
2528c2ecf20Sopenharmony_ci#else
2538c2ecf20Sopenharmony_cistatic inline void enable_lsx(void) {}
2548c2ecf20Sopenharmony_cistatic inline void disable_lsx(void) {}
2558c2ecf20Sopenharmony_cistatic inline void save_lsx(struct task_struct *t) {}
2568c2ecf20Sopenharmony_cistatic inline void restore_lsx(struct task_struct *t) {}
2578c2ecf20Sopenharmony_cistatic inline void init_lsx_upper(void) {}
2588c2ecf20Sopenharmony_cistatic inline void restore_lsx_upper(struct task_struct *t) {}
2598c2ecf20Sopenharmony_ci#endif
2608c2ecf20Sopenharmony_ci
2618c2ecf20Sopenharmony_ci#ifdef CONFIG_CPU_HAS_LASX
2628c2ecf20Sopenharmony_ci
2638c2ecf20Sopenharmony_cistatic inline void enable_lasx(void)
2648c2ecf20Sopenharmony_ci{
2658c2ecf20Sopenharmony_ci
2668c2ecf20Sopenharmony_ci	if (cpu_has_lasx)
2678c2ecf20Sopenharmony_ci		csr_xchg32(CSR_EUEN_LASXEN, CSR_EUEN_LASXEN, LOONGARCH_CSR_EUEN);
2688c2ecf20Sopenharmony_ci}
2698c2ecf20Sopenharmony_ci
2708c2ecf20Sopenharmony_cistatic inline void disable_lasx(void)
2718c2ecf20Sopenharmony_ci{
2728c2ecf20Sopenharmony_ci	if (cpu_has_lasx)
2738c2ecf20Sopenharmony_ci		csr_xchg32(0, CSR_EUEN_LASXEN, LOONGARCH_CSR_EUEN);
2748c2ecf20Sopenharmony_ci}
2758c2ecf20Sopenharmony_ci
2768c2ecf20Sopenharmony_cistatic inline void save_lasx(struct task_struct *t)
2778c2ecf20Sopenharmony_ci{
2788c2ecf20Sopenharmony_ci	if (cpu_has_lasx)
2798c2ecf20Sopenharmony_ci		_save_lasx(&t->thread.fpu);
2808c2ecf20Sopenharmony_ci}
2818c2ecf20Sopenharmony_ci
2828c2ecf20Sopenharmony_cistatic inline void restore_lasx(struct task_struct *t)
2838c2ecf20Sopenharmony_ci{
2848c2ecf20Sopenharmony_ci	if (cpu_has_lasx)
2858c2ecf20Sopenharmony_ci		_restore_lasx(&t->thread.fpu);
2868c2ecf20Sopenharmony_ci}
2878c2ecf20Sopenharmony_ci
2888c2ecf20Sopenharmony_cistatic inline void init_lasx_upper(void)
2898c2ecf20Sopenharmony_ci{
2908c2ecf20Sopenharmony_ci	if (cpu_has_lasx)
2918c2ecf20Sopenharmony_ci		_init_lasx_upper();
2928c2ecf20Sopenharmony_ci}
2938c2ecf20Sopenharmony_ci
2948c2ecf20Sopenharmony_cistatic inline void restore_lasx_upper(struct task_struct *t)
2958c2ecf20Sopenharmony_ci{
2968c2ecf20Sopenharmony_ci	if (cpu_has_lasx)
2978c2ecf20Sopenharmony_ci		_restore_lasx_upper(&t->thread.fpu);
2988c2ecf20Sopenharmony_ci}
2998c2ecf20Sopenharmony_ci
3008c2ecf20Sopenharmony_ci#else
3018c2ecf20Sopenharmony_cistatic inline void enable_lasx(void) {}
3028c2ecf20Sopenharmony_cistatic inline void disable_lasx(void) {}
3038c2ecf20Sopenharmony_cistatic inline void save_lasx(struct task_struct *t) {}
3048c2ecf20Sopenharmony_cistatic inline void restore_lasx(struct task_struct *t) {}
3058c2ecf20Sopenharmony_cistatic inline void init_lasx_upper(void) {}
3068c2ecf20Sopenharmony_cistatic inline void restore_lasx_upper(struct task_struct *t) {}
3078c2ecf20Sopenharmony_ci#endif
3088c2ecf20Sopenharmony_ci
3098c2ecf20Sopenharmony_cistatic inline int thread_lsx_context_live(void)
3108c2ecf20Sopenharmony_ci{
3118c2ecf20Sopenharmony_ci	if (!cpu_has_lsx)
3128c2ecf20Sopenharmony_ci		return 0;
3138c2ecf20Sopenharmony_ci
3148c2ecf20Sopenharmony_ci	return test_thread_flag(TIF_LSX_CTX_LIVE);
3158c2ecf20Sopenharmony_ci}
3168c2ecf20Sopenharmony_ci
3178c2ecf20Sopenharmony_cistatic inline int thread_lasx_context_live(void)
3188c2ecf20Sopenharmony_ci{
3198c2ecf20Sopenharmony_ci	if (!cpu_has_lasx)
3208c2ecf20Sopenharmony_ci		return 0;
3218c2ecf20Sopenharmony_ci
3228c2ecf20Sopenharmony_ci	return test_thread_flag(TIF_LASX_CTX_LIVE);
3238c2ecf20Sopenharmony_ci}
3248c2ecf20Sopenharmony_ci
3258c2ecf20Sopenharmony_ci#endif /* _ASM_FPU_H */
326