!425 update patches for sw64 architecture

Merge Pull Request from: @guzitao 
 
These patches generally cover the following tasks:

1.optimize kernel cores, remove unused codes, fix compile errors
2.fixes for perf, iommu
3.add support, add CONFIG_THREAD_INFO_IN_TASK support, support SIMD IO access in guest, add DVFS support 
 
Link:https://gitee.com/openeuler/kernel/pulls/425 

Reviewed-by: Zheng Zengkai <zhengzengkai@huawei.com> 
Signed-off-by: Zheng Zengkai <zhengzengkai@huawei.com> 

arch/sw_64/Kconfig

@@ -107,6 +107,7 @@ config SW64
 	select SET_FS
 	select SPARSEMEM_EXTREME if SPARSEMEM
 	select SWIOTLB
+	select THREAD_INFO_IN_TASK
 
 config LOCKDEP_SUPPORT
 	def_bool y

arch/sw_64/include/asm/clock.h

@@ -11,8 +11,6 @@
 
 struct clk;
 
-extern struct cpufreq_frequency_table sw64_clockmod_table[];
-
 extern char curruent_policy[CPUFREQ_NAME_LEN];
 
 struct clk_ops {
@@ -44,7 +42,7 @@ struct clk {
 
 int clk_init(void);
 
-void sw64_set_rate(unsigned long rate);
+void sw64_set_rate(unsigned int index);
 
 struct clk *sw64_clk_get(struct device *dev, const char *id);
 

arch/sw_64/include/asm/current.h

@@ -2,9 +2,18 @@
 #ifndef _ASM_SW64_CURRENT_H
 #define _ASM_SW64_CURRENT_H
 
-#include <linux/thread_info.h>
+#ifndef __ASSEMBLY__
 
-#define get_current()	(current_thread_info()->task)
-#define current		get_current()
+struct task_struct;
+static __always_inline struct task_struct *get_current(void)
+{
+	register struct task_struct *tp __asm__("$8");
+
+	return tp;
+}
+
+#define current get_current()
+
+#endif /* __ASSEMBLY__ */
 
 #endif /* _ASM_SW64_CURRENT_H */

arch/sw_64/include/asm/efi.h

@@ -6,6 +6,10 @@
 #include <asm/early_ioremap.h>
 #ifdef CONFIG_EFI
 extern void efi_init(void);
+extern unsigned long entSuspend;
+
+#define SLEEP_ENTRY_GUID        EFI_GUID(0x59cb76bb, 0x9c3a, 0x4c8f, 0xbd, 0x5c, 0xc0, 0x0f, 0x20, 0x61, 0x18, 0x4b)
+
 #else
 #define efi_init()
 #define efi_idmap_init()

arch/sw_64/include/asm/hmcall.h

@@ -13,6 +13,8 @@
 #define HMC_sleepen		0x05
 #define HMC_rdksp		0x06
 #define HMC_wrasid		0x08
+#define HMC_rdktp		0x09
+#define HMC_wrktp		0x0A
 #define HMC_rdptbr		0x0B
 #define HMC_wrptbr		0x0C
 #define HMC_wrksp		0x0E
@@ -150,6 +152,11 @@ __CALL_HMC_VOID(wrfen);
 __CALL_HMC_VOID(sleepen);
 __CALL_HMC_VOID(mtinten);
 
+__CALL_HMC_VOID(rdktp);
+#define restore_ktp()	rdktp()
+__CALL_HMC_VOID(wrktp);
+#define save_ktp()	wrktp()
+
 __CALL_HMC_R0(rdps, unsigned long);
 
 __CALL_HMC_R0(rdusp, unsigned long);

arch/sw_64/include/asm/idle.h

+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_SW64_IDLE_H
+#define _ASM_SW64_IDLE_H
+
+extern void arch_cpu_idle(void);
+
+#endif /* _ASM_SW64_IDLE_H  */

arch/sw_64/include/asm/platform.h

@@ -2,8 +2,10 @@
 #ifndef _ASM_SW64_PLATFORM_H
 #define _ASM_SW64_PLATFORM_H
 
+#include <linux/types.h>
+#include <asm/chip3_io.h>
+
 struct sw64_platform_ops {
-	void (*kill_arch)(int mode);
 	void __iomem *(*ioportmap)(unsigned long);
 	void (*register_platform_devices)(void);
 	void (*ops_fixup)(void);
@@ -13,5 +15,15 @@ struct sw64_platform_ops {
 extern struct sw64_platform_ops *sw64_platform;
 
 extern struct sw64_platform_ops xuelang_ops;
+extern struct boot_params *sunway_boot_params;
+
+extern void sw64_halt(void);
+extern void sw64_poweroff(void);
+extern void sw64_restart(void);
+extern void (*pm_restart)(void);
+extern void (*pm_halt)(void);
+extern int i2c_set_adapter(void);
+extern void cpld_write(uint8_t slave_addr, uint8_t reg, uint8_t data);
+extern void fix_jm585_reset(void);
 
 #endif /* _ASM_SW64_PLATFORM_H */

arch/sw_64/include/asm/processor.h

@@ -12,7 +12,7 @@
 #include <asm/ptrace.h>
 
 #define task_pt_regs(task) \
-	((struct pt_regs *) (task_stack_page(task) + 2 * PAGE_SIZE) - 1)
+	((struct pt_regs *) (task->stack + THREAD_SIZE) - 1)
 
 /*
  * Returns current instruction pointer ("program counter").

arch/sw_64/include/asm/ptrace.h

@@ -4,7 +4,6 @@
 
 #include <uapi/asm/ptrace.h>
 #include <asm/hmcall.h>
-#include <asm/thread_info.h>
 #include <asm/page.h>
 
 /*
@@ -58,11 +57,6 @@ struct pt_regs {
 #define kernel_stack_pointer(regs) ((unsigned long)((regs) + 1))
 #define instruction_pointer_set(regs, val) ((regs)->pc = val)
 
-
-#define current_pt_regs() \
-	((struct pt_regs *) ((char *)current_thread_info() + 2 * PAGE_SIZE) - 1)
-#define signal_pt_regs current_pt_regs
-
 #define force_successful_syscall_return() (current_pt_regs()->r0 = 0)
 
 #define MAX_REG_OFFSET (offsetof(struct pt_regs, r18))

arch/sw_64/include/asm/smp.h

@@ -2,14 +2,16 @@
 #ifndef _ASM_SW64_SMP_H
 #define _ASM_SW64_SMP_H
 
-#include <linux/threads.h>
-#include <linux/cpumask.h>
-#include <linux/bitops.h>
-#include <asm/hmcall.h>
-#include <asm/hcall.h>
 #include <asm/core.h>
+#include <asm/current.h>
+#include <asm/hcall.h>
+#include <asm/hmcall.h>
 #include <asm/hw_init.h>
 
+#include <linux/bitops.h>
+#include <linux/cpumask.h>
+#include <linux/threads.h>
+
 /* HACK: Cabrio WHAMI return value is bogus if more than 8 bits used.. :-( */
 
 extern cpumask_t core_start;
@@ -55,7 +57,13 @@ struct smp_rcb_struct {
 #define INIT_SMP_RCB ((struct smp_rcb_struct *) __va(0x820000UL))
 
 #define hard_smp_processor_id()	__hard_smp_processor_id()
-#define raw_smp_processor_id()	(current_thread_info()->cpu)
+
+#ifdef GENERATING_ASM_OFFSETS
+#define raw_smp_processor_id() (0)
+#else
+#include <asm/asm-offsets.h>
+#define raw_smp_processor_id() (*((unsigned int *)((void *)current + TASK_CPU)))
+#endif
 
 /* The map from sequential logical cpu number to hard cid.  */
 extern int __cpu_to_rcid[NR_CPUS];

arch/sw_64/include/asm/thread_info.h

@@ -5,7 +5,6 @@
 #ifdef __KERNEL__
 
 #ifndef __ASSEMBLY__
-#include <asm/processor.h>
 #include <asm/types.h>
 #include <asm/sysinfo.h>
 
@@ -25,12 +24,11 @@ struct pcb_struct {
 struct thread_info {
 	struct pcb_struct	pcb;		/* hmcode state */
 
-	struct task_struct	*task;		/* main task structure */
 	unsigned int		flags;		/* low level flags */
 	unsigned int		ieee_state;	/* see fpu.h */
 
 	mm_segment_t		addr_limit;	/* thread address space */
-	unsigned int		cpu;		/* current CPU */
+	unsigned int            cpu;            /* current CPU */
 	int			preempt_count;	/* 0 => preemptible, <0 => BUG */
 	unsigned int		status;		/* thread-synchronous flags */
 
@@ -58,14 +56,10 @@ static __always_inline u64 rtid(void)
  */
 #define INIT_THREAD_INFO(tsk)				\
 {							\
-	.task		= &tsk,				\
 	.addr_limit	= KERNEL_DS,			\
 	.preempt_count	= INIT_PREEMPT_COUNT,		\
 }
 
-/* How to get the thread information struct from C.  */
-register struct thread_info *__current_thread_info __asm__("$8");
-#define current_thread_info()	 __current_thread_info
 
 #endif /* __ASSEMBLY__ */
 

arch/sw_64/include/uapi/asm/perf_regs.h

@@ -20,6 +20,9 @@ enum perf_event_sw64_regs {
 	PERF_REG_SW64_R13,
 	PERF_REG_SW64_R14,
 	PERF_REG_SW64_R15,
+	PERF_REG_SW64_R16,
+	PERF_REG_SW64_R17,
+	PERF_REG_SW64_R18,
 	PERF_REG_SW64_R19,
 	PERF_REG_SW64_R20,
 	PERF_REG_SW64_R21,
@@ -30,16 +33,9 @@ enum perf_event_sw64_regs {
 	PERF_REG_SW64_R26,
 	PERF_REG_SW64_R27,
 	PERF_REG_SW64_R28,
-	PERF_REG_SW64_HAE,
-	PERF_REG_SW64_TRAP_A0,
-	PERF_REG_SW64_TRAP_A1,
-	PERF_REG_SW64_TRAP_A2,
-	PERF_REG_SW64_PS,
-	PERF_REG_SW64_PC,
 	PERF_REG_SW64_GP,
-	PERF_REG_SW64_R16,
-	PERF_REG_SW64_R17,
-	PERF_REG_SW64_R18,
+	PERF_REG_SW64_SP,
+	PERF_REG_SW64_PC,
 	PERF_REG_SW64_MAX,
 };
 #endif /* _UAPI_ASM_SW64_PERF_REGS_H */

arch/sw_64/kernel/Makefile

@@ -17,7 +17,7 @@ obj-y    := entry.o fpu.o traps.o process.o sys_sw64.o irq.o \
 	    irq_sw64.o signal.o setup.o ptrace.o time.o \
 	    systbls.o dup_print.o tc.o timer.o \
 	    insn.o early_init.o topology.o cacheinfo.o \
-	    vdso.o vdso/ hmcall.o stacktrace.o
+	    vdso.o vdso/ hmcall.o stacktrace.o idle.o reset.o
 
 obj-$(CONFIG_ACPI) += acpi.o
 obj-$(CONFIG_SMP)	+= smp.o

arch/sw_64/kernel/asm-offsets.c

@@ -5,6 +5,7 @@
  * and format the required data.
  */
 
+#define GENERATING_ASM_OFFSETS  /* asm/smp.h */
 #include <linux/stddef.h>
 #include <linux/sched.h>
 #include <linux/kbuild.h>
@@ -15,11 +16,11 @@
 
 #include "traps.c"
 
+
 void foo(void)
 {
-	DEFINE(TI_TASK, offsetof(struct thread_info, task));
+	DEFINE(ASM_THREAD_SIZE, THREAD_SIZE);
 	DEFINE(TI_FLAGS, offsetof(struct thread_info, flags));
-	DEFINE(TI_CPU, offsetof(struct thread_info, cpu));
 	BLANK();
 
 	DEFINE(TASK_BLOCKED, offsetof(struct task_struct, blocked));
@@ -27,6 +28,10 @@ void foo(void)
 	DEFINE(TASK_REAL_PARENT, offsetof(struct task_struct, real_parent));
 	DEFINE(TASK_GROUP_LEADER, offsetof(struct task_struct, group_leader));
 	DEFINE(TASK_TGID, offsetof(struct task_struct, tgid));
+	DEFINE(TASK_STACK, offsetof(struct task_struct, stack));
+#ifdef CONFIG_SMP
+	DEFINE(TASK_CPU, offsetof(struct task_struct, thread_info.cpu));
+#endif
 	BLANK();
 
 	OFFSET(PSTATE_REGS, processor_state, regs);

arch/sw_64/kernel/clock.c

 // SPDX-License-Identifier: GPL-2.0
-/*
- *  linux/arch/sw/kernel/setup.c
- *
- *  Copyright (C) 1995  Linus Torvalds
- */
 
 #include <linux/clk.h>
 #include <linux/cpufreq.h>
@@ -11,6 +6,7 @@
 #include <linux/export.h>
 #include <linux/delay.h>
 
+#include <asm/sw64_init.h>
 #include <asm/sw64io.h>
 #include <asm/hw_init.h>
 #include <asm/debug.h>
@@ -29,73 +25,7 @@
 #define CORE_PLL0_CFG_SHIFT     4
 #define CORE_PLL2_CFG_SHIFT     18
 
-char   curruent_policy[CPUFREQ_NAME_LEN];
-
-/* Minimum CLK support */
-enum {
-	DC_0, DC_1, DC_2, DC_3, DC_4, DC_5, DC_6, DC_7, DC_8,
-	DC_9, DC_10, DC_11, DC_12, DC_13, DC_14, DC_15, DC_16, DC_RESV
-};
-
-static int cpu_freq[14] = {
-	0,	1200,	1800,	1900,
-	1950,	2000,	2050,	2100,
-	2150,	2200,	2250,	2300,
-	2350,	2400	};
-
-struct cpufreq_frequency_table sw64_clockmod_table[] = {
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{0, DC_1, 0},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{0, DC_2, 0},
-	{-1, DC_RESV, CPUFREQ_ENTRY_INVALID},
-	{0, DC_3, 0},
-	{0, DC_4, 0},
-	{0, DC_5, 0},
-	{0, DC_6, 0},
-	{0, DC_7, 0},
-	{0, DC_8, 0},
-	{0, DC_9, 0},
-	{0, DC_10, 0},
-	{0, DC_11, 0},
-	{0, DC_12, 0},
-	{0, DC_13, 0},
-{-1, DC_RESV, CPUFREQ_TABLE_END},
-};
-EXPORT_SYMBOL_GPL(sw64_clockmod_table);
+char curruent_policy[CPUFREQ_NAME_LEN];
 
 static struct clk cpu_clk = {
 	.name = "cpu_clk",
@@ -113,13 +43,13 @@ unsigned int __sw64_cpufreq_get(struct cpufreq_policy *policy)
 {
 	int i;
 	u64 val;
+	struct cpufreq_frequency_table *ft = policy->freq_table;
 
-	val = sw64_io_read(0, CLK_CTL);
-	val = val >> CORE_PLL2_CFG_SHIFT;
+	val = sw64_io_read(0, CLK_CTL) >> CORE_PLL2_CFG_SHIFT;
 
-	for (i = 0; i < sizeof(cpu_freq)/sizeof(int); i++) {
-		if (cpu_freq[val] == cpu_freq[i])
-			return cpu_freq[i];
+	for (i = 0; ft[i].frequency != CPUFREQ_TABLE_END; i++) {
+		if (val == ft[i].driver_data)
+			return ft[i].frequency;
 	}
 	return 0;
 }
@@ -131,61 +61,41 @@ void sw64_store_policy(struct cpufreq_policy *policy)
 }
 EXPORT_SYMBOL_GPL(sw64_store_policy);
 
-void sw64_set_rate(unsigned long rate)
+void sw64_set_rate(unsigned int index)
 {
 	unsigned int i, val;
-	int index = -1;
+	int cpu_num;
 
-	rate /= 1000000;
+	cpu_num = sw64_chip->get_cpu_num();
 
-	for (i = 0; i < sizeof(cpu_freq)/sizeof(int); i++) {
-		if (rate == cpu_freq[i]) {
-			index = i;
-			update_cpu_freq(cpu_freq[i]);
-			break;
-		}
-	}
-
-	if (index < 0)
-		return -EINVAL;
+	for (i = 0; i < cpu_num; i++) {
+		sw64_io_write(i, CLK_CTL, CORE_CLK2_R | CORE_CLK2_V | CLK_PRT);
+		val = sw64_io_read(i, CLK_CTL);
 
-	sw64_io_write(0, CLK_CTL, CORE_CLK2_R | CORE_CLK2_V | CLK_PRT);
-	sw64_io_write(1, CLK_CTL, CORE_CLK2_R | CORE_CLK2_V | CLK_PRT);
-	val = sw64_io_read(0, CLK_CTL);
+		sw64_io_write(i, CLK_CTL, val | index << CORE_PLL2_CFG_SHIFT);
 
-	sw64_io_write(0, CLK_CTL, val | index << CORE_PLL2_CFG_SHIFT);
-	sw64_io_write(1, CLK_CTL, val | index << CORE_PLL2_CFG_SHIFT);
+		udelay(1);
 
-	udelay(1);
+		sw64_io_write(i, CLK_CTL, CORE_CLK2_V | CLK_PRT
+				| index << CORE_PLL2_CFG_SHIFT);
+		val = sw64_io_read(i, CLK_CTL);
 
-	sw64_io_write(0, CLK_CTL, CORE_CLK2_V | CLK_PRT
-			| index << CORE_PLL2_CFG_SHIFT);
-	sw64_io_write(1, CLK_CTL, CORE_CLK2_V | CLK_PRT
-			| index << CORE_PLL2_CFG_SHIFT);
-	val = sw64_io_read(0, CLK_CTL);
+		/* LV1 select PLL1/PLL2 */
+		sw64_io_write(i, CLU_LV1_SEL, CLK_LV1_SEL_MUXA | CLK_LV1_SEL_PRT);
 
-	/* LV1 select PLL1/PLL2 */
-	sw64_io_write(0, CLU_LV1_SEL, CLK_LV1_SEL_MUXA | CLK_LV1_SEL_PRT);
-	sw64_io_write(1, CLU_LV1_SEL, CLK_LV1_SEL_MUXA | CLK_LV1_SEL_PRT);
+		/* Set CLK_CTL PLL0 */
+		sw64_io_write(i, CLK_CTL, val | CORE_CLK0_R | CORE_CLK0_V);
 
-	/* Set CLK_CTL PLL0 */
-	sw64_io_write(0, CLK_CTL, val | CORE_CLK0_R | CORE_CLK0_V);
-	sw64_io_write(1, CLK_CTL, val | CORE_CLK0_R | CORE_CLK0_V);
+		sw64_io_write(i, CLK_CTL, val | CORE_CLK0_R | CORE_CLK0_V
+				| index << CORE_PLL0_CFG_SHIFT);
 
-	sw64_io_write(0, CLK_CTL, val | CORE_CLK0_R | CORE_CLK0_V
-			| index << CORE_PLL0_CFG_SHIFT);
-	sw64_io_write(1, CLK_CTL, val | CORE_CLK0_R | CORE_CLK0_V
-			| index << CORE_PLL0_CFG_SHIFT);
+		udelay(1);
 
-	udelay(1);
+		sw64_io_write(i, CLK_CTL, val | CORE_CLK0_V
+				| index << CORE_PLL0_CFG_SHIFT);
 
-	sw64_io_write(0, CLK_CTL, val | CORE_CLK0_V
-			| index << CORE_PLL0_CFG_SHIFT);
-	sw64_io_write(1, CLK_CTL, val | CORE_CLK0_V
-			| index << CORE_PLL0_CFG_SHIFT);
-
-	/* LV1 select PLL0/PLL1 */
-	sw64_io_write(0, CLU_LV1_SEL, CLK_LV1_SEL_MUXB | CLK_LV1_SEL_PRT);
-	sw64_io_write(1, CLU_LV1_SEL, CLK_LV1_SEL_MUXB | CLK_LV1_SEL_PRT);
+		/* LV1 select PLL0/PLL1 */
+		sw64_io_write(i, CLU_LV1_SEL, CLK_LV1_SEL_MUXB | CLK_LV1_SEL_PRT);
+	}
 }
 EXPORT_SYMBOL_GPL(sw64_set_rate);

arch/sw_64/kernel/dup_print.c

@@ -3,7 +3,7 @@
 #include <linux/smp.h>
 #include <linux/spinlock.h>
 
-#include <asm/chip3_io.h>
+#include <asm/platform.h>
 #include <asm/io.h>
 
 #ifdef CONFIG_SW64_RRK

arch/sw_64/kernel/early_init.c

@@ -24,6 +24,7 @@ static void __init sw64_setup_platform_ops(void)
 asmlinkage __visible void __init sw64_start_kernel(void)
 {
 	fixup_hmcall();
+	save_ktp();
 	sw64_setup_chip_ops();
 	sw64_setup_platform_ops();
 	sw64_platform->ops_fixup();

arch/sw_64/kernel/entry.S

@@ -20,7 +20,7 @@
  * regs 29-30 saved and set up by HMcode
  */
 
-	.macro SAVE_COMMON_REGS
+	.macro SAVE_ALL
 	ldi	$sp, -PT_REGS_PS($sp)
 	stl	$0, PT_REGS_R0($sp)
 	stl	$1, PT_REGS_R1($sp)
@@ -32,6 +32,13 @@
 	stl	$6, PT_REGS_R6($sp)
 	stl	$7, PT_REGS_R7($sp)
 	stl	$8, PT_REGS_R8($sp)
+	stl	$9, PT_REGS_R9($sp)
+	stl	$10, PT_REGS_R10($sp)
+	stl	$11, PT_REGS_R11($sp)
+	stl	$12, PT_REGS_R12($sp)
+	stl	$13, PT_REGS_R13($sp)
+	stl	$14, PT_REGS_R14($sp)
+	stl	$15, PT_REGS_R15($sp)
 	stl	$19, PT_REGS_R19($sp)
 	stl	$20, PT_REGS_R20($sp)
 	stl	$21, PT_REGS_R21($sp)
@@ -41,9 +48,10 @@
 	stl	$25, PT_REGS_R25($sp)
 	stl	$26, PT_REGS_R26($sp)
 	stl	$27, PT_REGS_R27($sp)
+	sys_call HMC_rdktp
 	.endm
 
-	.macro RESTORE_COMMON_REGS
+	.macro RESTORE_ALL
 	ldl	$0, PT_REGS_R0($sp)
 	ldl	$1, PT_REGS_R1($sp)
 	ldl	$2, PT_REGS_R2($sp)
@@ -53,6 +61,13 @@
 	ldl	$6, PT_REGS_R6($sp)
 	ldl	$7, PT_REGS_R7($sp)
 	ldl	$8, PT_REGS_R8($sp)
+	ldl	$9, PT_REGS_R9($sp)
+	ldl	$10, PT_REGS_R10($sp)
+	ldl	$11, PT_REGS_R11($sp)
+	ldl	$12, PT_REGS_R12($sp)
+	ldl	$13, PT_REGS_R13($sp)
+	ldl	$14, PT_REGS_R14($sp)
+	ldl	$15, PT_REGS_R15($sp)
 	ldl	$19, PT_REGS_R19($sp)
 	ldl	$20, PT_REGS_R20($sp)
 	ldl	$21, PT_REGS_R21($sp)
@@ -66,36 +81,6 @@
 	ldi	$sp, PT_REGS_PS($sp)
 	.endm
 
-	.macro SAVE_CALLEE_REGS
-	stl	$9, PT_REGS_R9($sp)
-	stl	$10, PT_REGS_R10($sp)
-	stl	$11, PT_REGS_R11($sp)
-	stl	$12, PT_REGS_R12($sp)
-	stl	$13, PT_REGS_R13($sp)
-	stl	$14, PT_REGS_R14($sp)
-	stl	$15, PT_REGS_R15($sp)
-	.endm
-
-	.macro RESTORE_CALLEE_REGS
-	ldl	$9, PT_REGS_R9($sp)
-	ldl	$10, PT_REGS_R10($sp)
-	ldl	$11, PT_REGS_R11($sp)
-	ldl	$12, PT_REGS_R12($sp)
-	ldl	$13, PT_REGS_R13($sp)
-	ldl	$14, PT_REGS_R14($sp)
-	ldl	$15, PT_REGS_R15($sp)
-	.endm
-
-	.macro SAVE_ALL
-	SAVE_COMMON_REGS
-	SAVE_CALLEE_REGS
-	.endm
-
-	.macro RESTORE_ALL
-	RESTORE_CALLEE_REGS
-	RESTORE_COMMON_REGS
-	.endm
-
 /*
  * Non-syscall kernel entry points.
  */
@@ -105,9 +90,7 @@
 	.ent entInt
 entInt:
 	SAVE_ALL
-	ldi	$8, 0x3fff
 	ldi	$26, ret_from_sys_call
-	bic	$sp, $8, $8
 	mov	$sp, $19
 	call	$31, do_entInt
 	.end entInt
@@ -117,9 +100,7 @@ entInt:
 	.ent entArith
 entArith:
 	SAVE_ALL
-	ldi	$8, 0x3fff
 	ldi	$26, ret_from_sys_call
-	bic	$sp, $8, $8
 	mov	$sp, $18
 	call	$31, do_entArith
 	.end entArith
@@ -129,9 +110,7 @@ entArith:
 	.ent entMM
 entMM:
 	SAVE_ALL
-	ldi	$8, 0x3fff
 	ldi	$26, ret_from_sys_call
-	bic	$sp, $8, $8
 	mov	$sp, $19
 	call	$31, do_page_fault
 	.end entMM
@@ -141,10 +120,8 @@ entMM:
 	.ent entIF
 entIF:
 	SAVE_ALL
-	ldi	$8, 0x3fff
 	ldi	$26, ret_from_sys_call
-	bic	$sp, $8, $8
-	mov	$sp, $17
+	mov	$sp, $18
 	call	$31, do_entIF
 	.end entIF
 
@@ -159,8 +136,6 @@ entIF:
 	.ent entUna
 entUna:
 	SAVE_ALL
-	ldi	$8, 0x3fff
-	bic	$sp, $8, $8
 	mov	$sp, $19
 	ldl	$0, PT_REGS_PS($sp)
 	and	$0, 8, $0		/* user mode ? */
@@ -192,8 +167,6 @@ entUna:
 entSys:
 
 	SAVE_ALL
-	ldi	$8, 0x3fff
-	bic	$sp, $8, $8
 	ldi	$4, NR_SYSCALLS($31)
 	stl	$16, PT_REGS_R16($sp)
 	ldi	$5, sys_call_table
@@ -411,8 +384,8 @@ __switch_to:
 	ldl	$13, TASK_THREAD_S4($17)
 	ldl	$14, TASK_THREAD_S5($17)
 	ldl	$15, TASK_THREAD_S6($17)
-	ldi	$8, 0x3fff
-	bic	$sp, $8, $8
+	mov	$17, $8
+	sys_call HMC_wrktp
 	mov	$16, $0
 	ret
 	.end __switch_to

arch/sw_64/kernel/ftrace.c

@@ -100,10 +100,12 @@ void arch_ftrace_update_code(int command)
 
 int __init ftrace_dyn_arch_init(void)
 {
-	init_thread_info.dyn_ftrace_addr = FTRACE_ADDR;
+	struct thread_info *ti = task_thread_info(&init_task);
+
+	ti->dyn_ftrace_addr = FTRACE_ADDR;
 
 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
-	init_thread_info.dyn_ftrace_regs_addr = FTRACE_REGS_ADDR;
+	ti->dyn_ftrace_regs_addr = FTRACE_REGS_ADDR;
 #endif
 	return 0;
 }

arch/sw_64/kernel/head.S

@@ -22,9 +22,11 @@ __start:
 	br	$27, 1f
 1:	ldgp	$29, 0($27)
 	/* We need to get current_task_info loaded up...  */
-	ldi	$8, init_thread_union
+	ldi	$8, init_task
+	ldl	$30, TASK_STACK($8)
 	/* ... and find our stack ... */
-	ldi	$30, ASM_THREAD_SIZE($8)
+	ldi	$30, ASM_THREAD_SIZE($30)
+
 	/* ... and then we can clear bss data.  */
 	ldi	$16, __bss_start
 	ldi	$18, __bss_stop
@@ -38,8 +40,10 @@ __start:
 	call	$26, relocate_kernel
 	ldl	$29, 0($30)
 	addl	$29, $0, $29
+	addl	$8, $0, $8
+	ldi	$30, 8($30)
 	/* Repoint the sp into the new kernel image */
-	ldi	$30, ASM_THREAD_SIZE($8)
+	addl	$30, $0, $30
 #endif
 	/* ... and then we can start the kernel.  */
 	call	$26, sw64_start_kernel
@@ -68,11 +72,13 @@ __smp_callin:
 	s4addl	$0, $1, $1
 	ldw	$0, 0($1)	# Get logical cpu number
 
-	ldi	$2, tidle_ksp
+	ldi	$2, idle_task_pointer
 	s8addl	$0, $2, $2
-	ldl	$30, 0($2)	# Get ksp of idle thread
+	ldl	$8, 0($2)	# Get ksp of idle thread
+	sys_call HMC_wrktp
 
-	ldi	$8, -ASM_THREAD_SIZE($30) # Find "current"
+	ldl	$30, TASK_STACK($8)
+	ldi	$30, ASM_THREAD_SIZE($30)
 
 	call	$26, smp_callin
 	sys_call HMC_halt

arch/sw_64/kernel/hmcall.c

@@ -76,6 +76,22 @@ static inline void fixup_wrasid(void)
 	entry[9] = 0x1ef00000;	/* pri_ret/b p23 */
 }
 
+static inline void fixup_rdktp(void)
+{
+	unsigned int *entry = __va(HMCALL_ENTRY(rdktp));
+
+	entry[0] = 0x95161000;	/* pri_ldl/p $8, VC__KTP(vcpucb) */
+	entry[1] = 0x1ee00000;	/* pri_ret $23 */
+}
+
+static inline void fixup_wrktp(void)
+{
+	unsigned int *entry = __va(HMCALL_ENTRY(wrktp));
+
+	entry[0] = 0xb5161000;	/* pri_stl/p $8, VC__KTP(vcpucb) */
+	entry[1] = 0x1ee00000;	/* pri_ret $23 */
+}
+
 void __init fixup_hmcall(void)
 {
 #if defined(CONFIG_SUBARCH_C3B)
@@ -83,6 +99,8 @@ void __init fixup_hmcall(void)
 	fixup_wrtp();
 	fixup_tbiasid();
 	fixup_wrasid();
+	fixup_rdktp();
+	fixup_wrktp();
 #endif
 }
 

arch/sw_64/kernel/idle.c

+// SPDX-License-Identifier: GPL-2.0
+/*
+ * sw64 idle loop support.
+ *
+ */
+#include <linux/cpu.h>
+#include <linux/irqflags.h>
+#include <asm/cpu.h>
+#include <asm/idle.h>
+
+#ifdef CONFIG_HOTPLUG_CPU
+void arch_cpu_idle_dead(void)
+{
+	play_dead();
+}
+#endif
+
+void cpu_idle(void)
+{
+	int i;
+
+	local_irq_enable();
+	cpu_relax();
+
+	if (is_in_guest())
+		hcall(HCALL_HALT, 0, 0, 0);
+	else {
+		for (i = 0; i < 16; i++)
+			asm("nop");
+		asm("halt");
+	}
+}
+
+void arch_cpu_idle(void)
+{
+	cpu_idle();
+}

arch/sw_64/kernel/perf_regs.c

@@ -8,7 +8,24 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)
 	if (WARN_ON_ONCE((u32)idx >= PERF_REG_SW64_MAX))
 		return 0;
 
-	return ((unsigned long *)regs)[idx];
+	switch (idx) {
+	case PERF_REG_SW64_R16:
+		return regs->r16;
+	case PERF_REG_SW64_R17:
+		return regs->r17;
+	case PERF_REG_SW64_R18:
+		return regs->r18;
+	case PERF_REG_SW64_R19 ... PERF_REG_SW64_R28:
+		return ((unsigned long *)regs)[idx - 3];
+	case PERF_REG_SW64_GP:
+		return regs->gp;
+	case PERF_REG_SW64_SP:
+		return (user_mode(regs) ? rdusp() : (u64)(regs + 1));
+	case PERF_REG_SW64_PC:
+		return regs->pc;
+	default:
+		return ((unsigned long *)regs)[idx];
+	}
 }
 
 #define REG_RESERVED (~((1ULL << PERF_REG_SW64_MAX) - 1))

arch/sw_64/kernel/relocate.c

@@ -239,9 +239,6 @@ unsigned int __init relocate_kernel(void)
 		if (plat_post_relocation(offset))
 			goto out;
 
-		/* The current thread is now within the relocated image */
-		__current_thread_info = RELOCATED(&init_thread_union);
-
 		/* Return the new kernel's offset */
 		return offset;
 	}

arch/sw_64/kernel/smp.c

@@ -34,7 +34,7 @@ EXPORT_SYMBOL(__cpu_to_rcid);
 int __rcid_to_cpu[NR_CPUS];		/* Map physical to logical */
 EXPORT_SYMBOL(__rcid_to_cpu);
 
-void *tidle_ksp[NR_CPUS];
+void *idle_task_pointer[NR_CPUS];
 
 /* State of each CPU */
 DEFINE_PER_CPU(int, cpu_state) = { 0 };
@@ -130,7 +130,7 @@ static int secondary_cpu_start(int cpuid, struct task_struct *idle)
 	/*
 	 * Precalculate the target ksp.
 	 */
-	tidle_ksp[cpuid] = idle->stack + THREAD_SIZE;
+	idle_task_pointer[cpuid] = idle;
 
 	DBGS("Starting secondary cpu %d: state 0x%lx\n", cpuid, idle->state);
 
@@ -237,7 +237,6 @@ void __init native_smp_prepare_cpus(unsigned int max_cpus)
 	memset(ipi_data, 0, sizeof(ipi_data));
 
 	init_cpu_topology();
-	current_thread_info()->cpu = 0;
 	store_cpu_topology(smp_processor_id());
 	numa_add_cpu(smp_processor_id());
 
@@ -366,10 +365,7 @@ void handle_ipi(struct pt_regs *regs)
 
 			case IPI_CPU_STOP:
 				local_irq_disable();
-				pr_crit("other core panic, now halt...\n");
-				while (1)
-					asm("nop");
-				halt();
+				asm("halt");
 
 			default:
 				pr_crit("Unknown IPI on CPU %d: %lu\n", this_cpu, which);

drivers/firmware/efi/sunway-init.c

@@ -25,6 +25,8 @@
 
 #include <asm/efi.h>
 
+unsigned long entSuspend;
+
 static int __init is_memory(efi_memory_desc_t *md)
 {
 	if (md->attribute & (EFI_MEMORY_WB|EFI_MEMORY_WT|EFI_MEMORY_WC))
@@ -32,7 +34,8 @@ static int __init is_memory(efi_memory_desc_t *md)
 	return 0;
 }
 static efi_config_table_type_t arch_tables[] __initdata = {
-	{SMBIOS3_TABLE_GUID, NULL, NULL}
+	{SMBIOS3_TABLE_GUID, NULL, NULL},
+	{SLEEP_ENTRY_GUID, &entSuspend, "SLEEP ENTRY"}
 };
 
 static int __init uefi_init(u64 efi_system_table)

drivers/platform/sw64/Makefile

+# SPDX-License-Identifier: GPL-2.0
+obj-$(CONFIG_PLATFORM_XUELANG)	+= legacy_xuelang.o