sw64: remove hmcall swpctx from context switch

Sunway inclusion
category: feature
bugzilla: https://gitee.com/openeuler/kernel/issues/I56OLG

--------------------------------

It used to switch context in hmcall swpctx, which is not flexible
enough. We try to make it happen in kernel without hmcall swpctx.

To achieve this end, not only fpu state but also usp and tls pointer
have to be saved and restored. For process creation and hibernation,
the current tls pointer has to be read from CSR:TID as it may be out-
of-sync with the saved value. For suspend, it's better to be saved
and restored because there is no guarantee that WAKEUP interrupt will
be used. To do this, we add hmcall fixup to access CSR:TID and obtain
backward compatibility for user.

Besides, the old `unique` is too obscure to be understood. To make it
clear, we rename it to `tp` which is short for tls pointer, and then
retain HMC_rdunique/wrunique as alias of HMC_rdtp/wrtp.
Signed-off-by: NHe Sheng <hesheng@wxiat.com>
Signed-off-by: NGu Zitao <guzitao@wxiat.com>

arch/sw_64/include/asm/hmcall.h

@@ -45,20 +45,19 @@
 
 
 /* 0x80  - 0xBF : User Level HMC routine */
-#define HMC_bpt			0x80
-#define HMC_callsys		0x83
-#define HMC_imb			0x86
+#include <uapi/asm/hmcall.h>
+
+/* Following will be deprecated from user level invocation */
 #define HMC_rwreg		0x87
-#define HMC_rdunique		0x9E
-#define HMC_wrunique		0x9F
 #define HMC_sz_uflush		0xA8
-#define HMC_gentrap		0xAA
-#define HMC_wrperfmon		0xB0
 #define HMC_longtime		0xB1
 
 #ifdef __KERNEL__
 #ifndef __ASSEMBLY__
 
+#include <linux/init.h>
+extern void __init fixup_hmcall(void);
+
 extern void halt(void) __attribute__((noreturn));
 #define __halt() __asm__ __volatile__ ("sys_call %0 #halt" : : "i" (HMC_halt))
 
@@ -183,6 +182,7 @@ __CALL_HMC_W1(wrtimer, unsigned long);
 __CALL_HMC_RW3(tbivpn, unsigned long, unsigned long, unsigned long, unsigned long);
 __CALL_HMC_RW2(cpuid, unsigned long, unsigned long, unsigned long);
 
+__CALL_HMC_W1(wrtp, unsigned long);
 /*
  * TB routines..
  */

arch/sw_64/include/asm/mmu_context.h

@@ -73,8 +73,7 @@ switch_mm(struct mm_struct *prev_mm, struct mm_struct *next_mm,
 	  struct task_struct *next)
 {
 	/* Check if our ASN is of an older version, and thus invalid. */
-	unsigned long asn;
-	unsigned long mmc;
+	unsigned long asn, mmc, ptbr;
 	long cpu = smp_processor_id();
 
 #ifdef CONFIG_SMP
@@ -94,17 +93,13 @@ switch_mm(struct mm_struct *prev_mm, struct mm_struct *next_mm,
 #endif
 
 	/*
-	 * Always update the PCB ASN.  Another thread may have allocated
-	 * a new mm->context (via flush_tlb_mm) without the ASN serial
+	 * Update CSR:UPN and CSR:PTBR. Another thread may have allocated
+	 * a new mm->context[asid] (via flush_tlb_mm) without the ASN serial
 	 * number wrapping.  We have no way to detect when this is needed.
 	 */
-	task_thread_info(next)->pcb.asn = mmc & HARDWARE_ASN_MASK;
-	/*
-	 * Always update the PCB PTBR. If next is kernel thread, it must
-	 * update PTBR. If next is user process, it's ok to update PTBR.
-	 */
-	task_thread_info(next)->pcb.ptbr = virt_to_pfn(next_mm->pgd);
-	load_asn_ptbr(task_thread_info(next)->pcb.asn, task_thread_info(next)->pcb.ptbr);
+	asn = mmc & HARDWARE_ASN_MASK;
+	ptbr = virt_to_pfn(next_mm->pgd);
+	load_asn_ptbr(asn, ptbr);
 }
 
 extern void __load_new_mm_context(struct mm_struct *);
@@ -141,8 +136,6 @@ static inline int init_new_context(struct task_struct *tsk,
 
 	for_each_possible_cpu(i)
 		mm->context.asid[i] = 0;
-	if (tsk != current)
-		task_thread_info(tsk)->pcb.ptbr = virt_to_pfn(mm->pgd);
 	return 0;
 }
 
@@ -154,7 +147,6 @@ static inline void destroy_context(struct mm_struct *mm)
 static inline void enter_lazy_tlb(struct mm_struct *mm,
 				  struct task_struct *tsk)
 {
-	task_thread_info(tsk)->pcb.ptbr = virt_to_pfn(mm->pgd);
 }
 
 static inline int arch_dup_mmap(struct mm_struct *oldmm,

arch/sw_64/include/asm/switch_to.h

@@ -6,27 +6,39 @@
 
 extern void __fpstate_save(struct task_struct *save_to);
 extern void __fpstate_restore(struct task_struct *restore_from);
-extern struct task_struct *__switch_to(unsigned long pcb,
-		struct task_struct *prev, struct task_struct *next);
+extern struct task_struct *__switch_to(struct task_struct *prev,
+				       struct task_struct *next);
 extern void restore_da_match_after_sched(void);
 
-static inline void fpstate_save(struct task_struct *task)
+static inline void aux_save(struct task_struct *task)
 {
-	if (likely(!(task->flags & PF_KTHREAD)))
+	struct pcb_struct *pcb;
+
+	if (likely(!(task->flags & PF_KTHREAD))) {
+		pcb = &task_thread_info(task)->pcb;
+		pcb->usp = rdusp();
+		pcb->tp = rtid();
 		__fpstate_save(task);
+	}
 }
 
-static inline void fpstate_restore(struct task_struct *task)
+static inline void aux_restore(struct task_struct *task)
 {
-	if (likely(!(task->flags & PF_KTHREAD)))
+	struct pcb_struct *pcb;
+
+	if (likely(!(task->flags & PF_KTHREAD))) {
+		pcb = &task_thread_info(task)->pcb;
+		wrusp(pcb->usp);
+		wrtp(pcb->tp);
 		__fpstate_restore(task);
+	}
 }
 
 static inline void __switch_to_aux(struct task_struct *prev,
 				   struct task_struct *next)
 {
-	fpstate_save(prev);
-	fpstate_restore(next);
+	aux_save(prev);
+	aux_restore(next);
 }
 
 
@@ -34,9 +46,8 @@ static inline void __switch_to_aux(struct task_struct *prev,
 do {									\
 	struct task_struct *__prev = (prev);				\
 	struct task_struct *__next = (next);				\
-	__u64 __nextpcb = virt_to_phys(&task_thread_info(__next)->pcb);	\
 	__switch_to_aux(__prev, __next);				\
-	(last) = __switch_to(__nextpcb, __prev, __next);		\
+	(last) = __switch_to(__prev, __next);				\
 	check_mmu_context();						\
 } while (0)
 

arch/sw_64/include/asm/thread_info.h

@@ -20,7 +20,7 @@ struct pcb_struct {
 	unsigned long ptbr;
 	unsigned int pcc;
 	unsigned int asn;
-	unsigned long unique;
+	unsigned long tp;
 	unsigned long flags;
 	unsigned long da_match, da_mask;
 	unsigned long dv_match, dv_mask;
@@ -47,6 +47,14 @@ struct thread_info {
 #endif
 };
 
+static __always_inline u64 rtid(void)
+{
+	u64 val;
+
+	asm volatile("rtid %0" : "=r" (val) : :);
+	return val;
+}
+
 /*
  * Macros/functions for gaining access to the thread information structure.
  */

arch/sw_64/include/asm/vcpu.h

@@ -32,7 +32,7 @@ struct vcpucb {
 	unsigned long vcpu_irq_disabled;
 	unsigned long vcpu_irq;
 	unsigned long ptbr;
-	unsigned long int_stat0;
+	unsigned long tid;
 	unsigned long int_stat1;
 	unsigned long int_stat2;
 	unsigned long int_stat3;

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

@@ -7,8 +7,10 @@
 #define HMC_bpt			0x80
 #define HMC_callsys		0x83
 #define HMC_imb			0x86
-#define HMC_rdunique		0x9E
-#define HMC_wrunique		0x9F
+#define HMC_rdtp		0x9E
+#define HMC_wrtp		0x9F
+#define HMC_rdunique		HMC_rdtp
+#define HMC_wrunique		HMC_wrtp
 #define HMC_gentrap		0xAA
 #define HMC_wrperfmon		0xB0
 

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

@@ -36,7 +36,8 @@ struct user_fpsimd_state {
 #define FPREG_END		62
 #define FPCR			63
 #define PC			64
-#define UNIQUE			65
+#define TP			65
+#define UNIQUE			TP
 #define VECREG_BASE		67
 #define VECREG_END		161
 #define F31_V1			98

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/
+	    vdso.o vdso/ hmcall.o
 
 obj-$(CONFIG_ACPI) += acpi.o
 obj-$(CONFIG_STACKTRACE) += stacktrace.o

arch/sw_64/kernel/early_init.c

@@ -23,6 +23,7 @@ static void __init sw64_setup_platform_ops(void)
 
 asmlinkage __visible void __init sw64_start_kernel(void)
 {
+	fixup_hmcall();
 	sw64_setup_chip_ops();
 	sw64_setup_platform_ops();
 	sw64_platform->ops_fixup();

arch/sw_64/kernel/entry.S

@@ -384,11 +384,10 @@ $syscall_trace_failed:
  * Integer register context switch
  * The callee-saved registers must be saved and restored.
  *
- *   a0: physical address of next task's pcb, used by hmcode
- *   a1: previous task_struct (must be preserved across the switch)
- *   a2: next task_struct
+ *   a0: previous task_struct (must be preserved across the switch)
+ *   a1: next task_struct
  *
- * The value of a1  must be preserved by this function, as that's how
+ * The value of a0  must be preserved by this function, as that's how
  * arguments are passed to schedule_tail.
  */
 	.align 4
@@ -397,33 +396,28 @@ $syscall_trace_failed:
 __switch_to:
 	.prologue 0
 	/* Save context into prev->thread */
-	stl	$26, TASK_THREAD_RA($17)
-	stl	$30, TASK_THREAD_SP($17)
-	stl	$9, TASK_THREAD_S0($17)
-	stl	$10, TASK_THREAD_S1($17)
-	stl	$11, TASK_THREAD_S2($17)
-	stl	$12, TASK_THREAD_S3($17)
-	stl	$13, TASK_THREAD_S4($17)
-	stl	$14, TASK_THREAD_S5($17)
-	stl	$15, TASK_THREAD_S6($17)
+	stl	$26, TASK_THREAD_RA($16)
+	stl	$30, TASK_THREAD_SP($16)
+	stl	$9, TASK_THREAD_S0($16)
+	stl	$10, TASK_THREAD_S1($16)
+	stl	$11, TASK_THREAD_S2($16)
+	stl	$12, TASK_THREAD_S3($16)
+	stl	$13, TASK_THREAD_S4($16)
+	stl	$14, TASK_THREAD_S5($16)
+	stl	$15, TASK_THREAD_S6($16)
 	/* Restore context from next->thread */
-	ldl	$26, TASK_THREAD_RA($18)
-	ldl	$9, TASK_THREAD_S0($18)
-	ldl	$10, TASK_THREAD_S1($18)
-	ldl	$11, TASK_THREAD_S2($18)
-	ldl	$12, TASK_THREAD_S3($18)
-	ldl	$13, TASK_THREAD_S4($18)
-	ldl	$14, TASK_THREAD_S5($18)
-	ldl	$15, TASK_THREAD_S6($18)
-	sys_call HMC_swpctx
-	/*
-	 * SP has been saved and restored by HMC_swpctx,
-	 * and restore it again here for future expansion.
-	 */
-	ldl     $30, TASK_THREAD_SP($18)
+	ldl	$26, TASK_THREAD_RA($17)
+	ldl     $30, TASK_THREAD_SP($17)
+	ldl	$9, TASK_THREAD_S0($17)
+	ldl	$10, TASK_THREAD_S1($17)
+	ldl	$11, TASK_THREAD_S2($17)
+	ldl	$12, TASK_THREAD_S3($17)
+	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, $0
+	mov	$16, $0
 	ret
 	.end __switch_to
 
@@ -436,7 +430,6 @@ __switch_to:
 	.ent ret_from_fork
 ret_from_fork:
 	ldi	$26, ret_from_sys_call
-	mov	$17, $16
 	jmp	$31, schedule_tail
 	.end ret_from_fork
 
@@ -447,7 +440,6 @@ ret_from_fork:
 	.globl ret_from_kernel_thread
 	.ent ret_from_kernel_thread
 ret_from_kernel_thread:
-	mov	$17, $16
 	call	$26, schedule_tail
 	mov	$9, $27
 	mov	$10, $16

arch/sw_64/kernel/hibernate.c

@@ -14,7 +14,7 @@ void save_processor_state(void)
 
 	vcb->ksp = rdksp();
 	vcb->usp = rdusp();
-	vcb->pcbb = rdpcbb();
+	vcb->tid = rtid();
 	vcb->ptbr = rdptbr();
 }
 
@@ -24,7 +24,7 @@ void restore_processor_state(void)
 
 	wrksp(vcb->ksp);
 	wrusp(vcb->usp);
-	wrpcbb(vcb->pcbb);
+	wrtp(vcb->tid);
 	wrptbr(vcb->ptbr);
 	sflush();
 	tbiv();

arch/sw_64/kernel/hmcall.c

+// SPDX-License-Identifier: GPL-2.0
+/*
+ * arch/sw_64/kernel/hmcall.c
+ *
+ * Copyright (C) 2022 WXIAT
+ * Author: He Sheng
+ */
+
+#include <asm/hmcall.h>
+#include <asm/page.h>
+
+#define A0(func)	(((HMC_##func & 0xFF) >> 6) & 0x1)
+#define A1(func)	((((HMC_##func & 0xFF)>>6) & 0x2) >> 1)
+#define A2(func)	((HMC_##func & 0x3F) << 7)
+
+#define T(func)		((A0(func) ^ A1(func)) & 0x1)
+#define B0(func)	((T(func) | A0(func)) << 13)
+#define B1(func)	(((~T(func) & 1) | A1(func)) << 14)
+
+#define PRI_BASE	0x10000UL
+
+#define HMCALL_ENTRY(func)	(PRI_BASE | B1(func) | B0(func) | A2(func))
+
+
+static inline void fixup_rdtp(void)
+{
+	unsigned int *entry = __va(HMCALL_ENTRY(rdtp));
+
+	entry[0] = 0x181ffec7;	/* pri_rcsr $0, CSR__TID */
+	entry[1] = 0x1ee00000;	/* pri_ret $23 */
+}
+
+static inline void fixup_wrtp(void)
+{
+	unsigned int *entry = __va(HMCALL_ENTRY(wrtp));
+
+	entry[0] = 0x1a1fffc7;	/* pri_wcsr $16, CSR__TID */
+	entry[1] = 0x1ee00000;	/* pri_ret $23 */
+}
+
+void __init fixup_hmcall(void)
+{
+#if defined(CONFIG_SUBARCH_C3A) || defined(CONFIG_SUBARCH_C3B)
+	fixup_rdtp();
+	fixup_wrtp();
+#endif
+}
+
+#undef A0
+#undef A1
+#undef A2
+#undef T
+#undef B0
+#undef B1

arch/sw_64/kernel/kgdb.c

@@ -95,7 +95,7 @@ struct dbg_reg_def_t dbg_reg_def[DBG_MAX_REG_NUM] = {
 
 	{ "pc", 8, offsetof(struct pt_regs, pc)},
 	{ "", 8, -1 },
-	{ "unique", 8, -1},
+	{ "tp", 8, -1},
 };
 
 char *dbg_get_reg(int regno, void *mem, struct pt_regs *regs)

arch/sw_64/kernel/process.c

@@ -125,7 +125,7 @@ flush_thread(void)
 	wrfpcr(FPCR_DYN_NORMAL | ieee_swcr_to_fpcr(0));
 
 	/* Clean slate for TLS.  */
-	current_thread_info()->pcb.unique = 0;
+	current_thread_info()->pcb.tp = 0;
 }
 
 void
@@ -135,7 +135,11 @@ release_thread(struct task_struct *dead_task)
 
 int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 {
-	fpstate_save(src);
+	/*
+	 * aux_save() has to read the current TLS pointer from CSR:TID as it
+	 * may be out-of-sync with the saved value.
+	 */
+	aux_save(src);
 	*dst = *src;
 	return 0;
 }
@@ -168,6 +172,7 @@ copy_thread(unsigned long clone_flags, unsigned long usp,
 		childti->pcb.usp = 0;
 		return 0;
 	}
+
 	/*
 	 * Note: if CLONE_SETTLS is not set, then we must inherit the
 	 * value from the parent, which will have been set by the block
@@ -176,10 +181,11 @@ copy_thread(unsigned long clone_flags, unsigned long usp,
 	 * application calling fork.
 	 */
 	if (clone_flags & CLONE_SETTLS)
-		childti->pcb.unique = tls;
+		childti->pcb.tp = regs->r20;
 	else
 		regs->r20 = 0;
-	childti->pcb.usp = usp ?: rdusp();
+	if (usp)
+		childti->pcb.usp = usp;
 	*childregs = *regs;
 	childregs->r0 = 0;
 	childregs->r19 = 0;
@@ -202,7 +208,7 @@ void sw64_elf_core_copy_regs(elf_greg_t *dest, struct pt_regs *regs)
 		dest[i] = *(__u64 *)((void *)regs + regoffsets[i]);
 	dest[30] = ti == current_thread_info() ? rdusp() : ti->pcb.usp;
 	dest[31] = regs->pc;
-	dest[32] = ti->pcb.unique;
+	dest[32] = ti->pcb.tp;
 }
 EXPORT_SYMBOL(sw64_elf_core_copy_regs);
 

arch/sw_64/kernel/ptrace.c

@@ -72,7 +72,7 @@ short regoffsets[32] = {
 
 static int pcboff[] = {
 	[USP] = PCB_OFF(usp),
-	[UNIQUE] = PCB_OFF(unique),
+	[TP] = PCB_OFF(tp),
 	[DA_MATCH] = PCB_OFF(da_match),
 	[DA_MASK] = PCB_OFF(da_mask),
 	[DV_MATCH] = PCB_OFF(dv_match),

arch/sw_64/kernel/smp.c

@@ -111,7 +111,7 @@ void smp_callin(void)
 	mmgrab(&init_mm);
 	current->active_mm = &init_mm;
 	/* update csr:ptbr */
-	wrptbr(PFN_PHYS(current_thread_info()->pcb.ptbr));
+	wrptbr(virt_to_phys(init_mm.pgd));
 
 	/* inform the notifiers about the new cpu */
 	notify_cpu_starting(cpuid);

arch/sw_64/kernel/suspend.c

@@ -33,6 +33,7 @@ void sw64_suspend_enter(void)
 	 */
 
 	disable_local_timer();
+	current_thread_info()->pcb.tp = rtid();
 
 #ifdef CONFIG_SW64_SUSPEND_DEEPSLEEP_BOOTCORE
 	sw64_suspend_deep_sleep(&suspend_state);
@@ -40,6 +41,7 @@ void sw64_suspend_enter(void)
 	mtinten();
 	asm("halt");
 #endif
+	wrtp(current_thread_info()->pcb.tp);
 
 	disable_local_timer();
 }

arch/sw_64/mm/fault.c

@@ -66,17 +66,15 @@ void show_all_vma(void)
  */
 void __load_new_mm_context(struct mm_struct *next_mm)
 {
-	unsigned long mmc;
-	struct pcb_struct *pcb;
+	unsigned long mmc, asn, ptbr;
 
 	mmc = __get_new_mm_context(next_mm, smp_processor_id());
 	next_mm->context.asid[smp_processor_id()] = mmc;
 
-	pcb = &current_thread_info()->pcb;
-	pcb->asn = mmc & HARDWARE_ASN_MASK;
-	pcb->ptbr = virt_to_pfn(next_mm->pgd);
+	asn = mmc & HARDWARE_ASN_MASK;
+	ptbr = virt_to_pfn(next_mm->pgd);
 
-	load_asn_ptbr(pcb->asn, pcb->ptbr);
+	load_asn_ptbr(asn, ptbr);
 }
 
 /*

arch/sw_64/mm/init.c

@@ -82,19 +82,8 @@ pgd_alloc(struct mm_struct *mm)
 static inline void
 switch_to_system_map(void)
 {
-	unsigned long newptbr;
-
-	/*
-	 * Initialize the kernel's page tables.  Linux puts the vptb in
-	 * the last slot of the L1 page table.
-	 */
 	memset(swapper_pg_dir, 0, PAGE_SIZE);
-	newptbr = virt_to_pfn(swapper_pg_dir);
-
-	/* Also set up the real kernel PCB while we're at it.  */
-	init_thread_info.pcb.ptbr = newptbr;
-	init_thread_info.pcb.flags = 1;	/* set FEN, clear everything else */
-	wrptbr(PFN_PHYS(newptbr));
+	wrptbr(virt_to_phys(swapper_pg_dir));
 	tbiv();
 }