LoongArch/ftrace: Add dynamic function tracer support

The compiler has inserted 2 NOPs before the regular function prologue.
T series registers are available and safe because of LoongArch's psABI.

At runtime, we can replace nop with bl to enable ftrace call and replace
bl with nop to disable ftrace call. The bl instruction requires us to
save the original RA value, so it saves RA at t0 here.

Details are:

| Compiled   |       Disabled         |        Enabled         |
+------------+------------------------+------------------------+
| nop        | move     t0, ra        | move    t0, ra         |
| nop        | nop                    | bl      ftrace_caller  |
| func_body  | func_body              | func_body              |

The RA value will be recovered by ftrace_regs_entry, and restored into
RA before returning to the regular function prologue. When a function is
not being traced, the "move t0, ra" is not harmful.

1) ftrace_make_call, ftrace_make_nop (in kernel/ftrace.c)
   The two functions turn each recorded call site of filtered functions
   into a call to ftrace_caller or nops.

2) ftracce_update_ftrace_func (in kernel/ftrace.c)
   turns the nops at ftrace_call into a call to a generic entry for
   function tracers.

3) ftrace_caller (in kernel/mcount_dyn.S)
   The entry where each _mcount call sites calls to once they are
   filtered to be traced.

Co-developed-by: Jinyang He <hejinyang@loongson.cn>
Signed-off-by: Jinyang He <hejinyang@loongson.cn>
Signed-off-by: Qing Zhang <zhangqing@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>

arch/loongarch/Kconfig

@@ -88,6 +88,7 @@ config LOONGARCH
 	select HAVE_C_RECORDMCOUNT
 	select HAVE_DEBUG_STACKOVERFLOW
 	select HAVE_DMA_CONTIGUOUS
+	select HAVE_DYNAMIC_FTRACE
 	select HAVE_EBPF_JIT
 	select HAVE_EXIT_THREAD
 	select HAVE_FAST_GUP

arch/loongarch/Makefile

@@ -25,6 +25,11 @@ endif
 32bit-emul		= elf32loongarch
 64bit-emul		= elf64loongarch
 
+ifdef CONFIG_DYNAMIC_FTRACE
+KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
+CC_FLAGS_FTRACE := -fpatchable-function-entry=2
+endif
+
 ifdef CONFIG_64BIT
 tool-archpref		= $(64bit-tool-archpref)
 UTS_MACHINE		:= loongarch64

arch/loongarch/include/asm/ftrace.h

@@ -11,9 +11,30 @@
 #define MCOUNT_INSN_SIZE 4		/* sizeof mcount call */
 
 #ifndef __ASSEMBLY__
+
+#ifndef CONFIG_DYNAMIC_FTRACE
+
 #define mcount _mcount
 extern void _mcount(void);
 extern void prepare_ftrace_return(unsigned long self_addr, unsigned long callsite_sp, unsigned long old);
+
+#else
+
+struct dyn_ftrace;
+struct dyn_arch_ftrace { };
+
+#define ftrace_init_nop ftrace_init_nop
+int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec);
+
+static inline unsigned long ftrace_call_adjust(unsigned long addr)
+{
+	return addr;
+}
+
+void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent);
+
+#endif /* CONFIG_DYNAMIC_FTRACE */
+
 #endif /* __ASSEMBLY__ */
 
 #endif /* CONFIG_FUNCTION_TRACER */

arch/loongarch/include/asm/inst.h

@@ -349,6 +349,17 @@ static inline bool is_stack_alloc_ins(union loongarch_instruction *ip)
 		is_imm12_negative(ip->reg2i12_format.immediate);
 }
 
+int larch_insn_read(void *addr, u32 *insnp);
+int larch_insn_write(void *addr, u32 insn);
+int larch_insn_patch_text(void *addr, u32 insn);
+
+u32 larch_insn_gen_nop(void);
+u32 larch_insn_gen_b(unsigned long pc, unsigned long dest);
+u32 larch_insn_gen_bl(unsigned long pc, unsigned long dest);
+
+u32 larch_insn_gen_or(enum loongarch_gpr rd, enum loongarch_gpr rj, enum loongarch_gpr rk);
+u32 larch_insn_gen_move(enum loongarch_gpr rd, enum loongarch_gpr rj);
+
 u32 larch_insn_gen_lu32id(enum loongarch_gpr rd, int imm);
 u32 larch_insn_gen_lu52id(enum loongarch_gpr rd, enum loongarch_gpr rj, int imm);
 u32 larch_insn_gen_jirl(enum loongarch_gpr rd, enum loongarch_gpr rj, unsigned long pc, unsigned long dest);

arch/loongarch/include/asm/unwind.h

@@ -20,7 +20,7 @@ struct unwind_state {
 	char type; /* UNWINDER_XXX */
 	struct stack_info stack_info;
 	struct task_struct *task;
-	bool first, error;
+	bool first, error, is_ftrace;
 	unsigned long sp, pc, ra;
 };
 

arch/loongarch/kernel/Makefile

@@ -16,11 +16,16 @@ obj-$(CONFIG_EFI) 		+= efi.o
 obj-$(CONFIG_CPU_HAS_FPU)	+= fpu.o
 
 ifdef CONFIG_FUNCTION_TRACER
-obj-y += mcount.o ftrace.o
-CFLAGS_REMOVE_ftrace.o = $(CC_FLAGS_FTRACE)
-CFLAGS_REMOVE_inst.o = $(CC_FLAGS_FTRACE)
-CFLAGS_REMOVE_time.o = $(CC_FLAGS_FTRACE)
-CFLAGS_REMOVE_perf_event.o = $(CC_FLAGS_FTRACE)
+  ifndef CONFIG_DYNAMIC_FTRACE
+    obj-y += mcount.o ftrace.o
+    CFLAGS_REMOVE_ftrace.o = $(CC_FLAGS_FTRACE)
+  else
+    obj-y += mcount_dyn.o ftrace_dyn.o
+    CFLAGS_REMOVE_ftrace_dyn.o = $(CC_FLAGS_FTRACE)
+  endif
+  CFLAGS_REMOVE_inst.o = $(CC_FLAGS_FTRACE)
+  CFLAGS_REMOVE_time.o = $(CC_FLAGS_FTRACE)
+  CFLAGS_REMOVE_perf_event.o = $(CC_FLAGS_FTRACE)
 endif
 
 obj-$(CONFIG_MODULES)		+= module.o module-sections.o

arch/loongarch/kernel/ftrace_dyn.c

@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Based on arch/arm64/kernel/ftrace.c
+ *
+ * Copyright (C) 2022 Loongson Technology Corporation Limited
+ */
+
+#include <linux/ftrace.h>
+#include <linux/uaccess.h>
+
+#include <asm/inst.h>
+
+static int ftrace_modify_code(unsigned long pc, u32 old, u32 new, bool validate)
+{
+	u32 replaced;
+
+	if (validate) {
+		if (larch_insn_read((void *)pc, &replaced))
+			return -EFAULT;
+
+		if (replaced != old)
+			return -EINVAL;
+	}
+
+	if (larch_insn_patch_text((void *)pc, new))
+		return -EPERM;
+
+	return 0;
+}
+
+int ftrace_update_ftrace_func(ftrace_func_t func)
+{
+	u32 new;
+	unsigned long pc;
+
+	pc = (unsigned long)&ftrace_call;
+	new = larch_insn_gen_bl(pc, (unsigned long)func);
+
+	return ftrace_modify_code(pc, 0, new, false);
+}
+
+/*
+ * The compiler has inserted 2 NOPs before the regular function prologue.
+ * T series registers are available and safe because of LoongArch's psABI.
+ *
+ * At runtime, we can replace nop with bl to enable ftrace call and replace bl
+ * with nop to disable ftrace call. The bl requires us to save the original RA
+ * value, so it saves RA at t0 here.
+ *
+ * Details are:
+ *
+ * | Compiled   |       Disabled         |        Enabled         |
+ * +------------+------------------------+------------------------+
+ * | nop        | move     t0, ra        | move     t0, ra        |
+ * | nop        | nop                    | bl       ftrace_caller |
+ * | func_body  | func_body              | func_body              |
+ *
+ * The RA value will be recovered by ftrace_regs_entry, and restored into RA
+ * before returning to the regular function prologue. When a function is not
+ * being traced, the "move t0, ra" is not harmful.
+ */
+
+int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec)
+{
+	u32 old, new;
+	unsigned long pc;
+
+	pc = rec->ip;
+	old = larch_insn_gen_nop();
+	new = larch_insn_gen_move(LOONGARCH_GPR_T0, LOONGARCH_GPR_RA);
+
+	return ftrace_modify_code(pc, old, new, true);
+}
+
+int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
+{
+	u32 old, new;
+	unsigned long pc;
+
+	pc = rec->ip + LOONGARCH_INSN_SIZE;
+
+	old = larch_insn_gen_nop();
+	new = larch_insn_gen_bl(pc, addr);
+
+	return ftrace_modify_code(pc, old, new, true);
+}
+
+int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr)
+{
+	u32 old, new;
+	unsigned long pc;
+
+	pc = rec->ip + LOONGARCH_INSN_SIZE;
+
+	new = larch_insn_gen_nop();
+	old = larch_insn_gen_bl(pc, addr);
+
+	return ftrace_modify_code(pc, old, new, true);
+}
+
+void arch_ftrace_update_code(int command)
+{
+	command |= FTRACE_MAY_SLEEP;
+	ftrace_modify_all_code(command);
+}
+
+int __init ftrace_dyn_arch_init(void)
+{
+	return 0;
+}

arch/loongarch/kernel/inst.c

@@ -2,8 +2,100 @@
 /*
  * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
  */
+#include <linux/sizes.h>
+#include <linux/uaccess.h>
+
+#include <asm/cacheflush.h>
 #include <asm/inst.h>
 
+static DEFINE_RAW_SPINLOCK(patch_lock);
+
+int larch_insn_read(void *addr, u32 *insnp)
+{
+	int ret;
+	u32 val;
+
+	ret = copy_from_kernel_nofault(&val, addr, LOONGARCH_INSN_SIZE);
+	if (!ret)
+		*insnp = val;
+
+	return ret;
+}
+
+int larch_insn_write(void *addr, u32 insn)
+{
+	int ret;
+	unsigned long flags = 0;
+
+	raw_spin_lock_irqsave(&patch_lock, flags);
+	ret = copy_to_kernel_nofault(addr, &insn, LOONGARCH_INSN_SIZE);
+	raw_spin_unlock_irqrestore(&patch_lock, flags);
+
+	return ret;
+}
+
+int larch_insn_patch_text(void *addr, u32 insn)
+{
+	int ret;
+	u32 *tp = addr;
+
+	if ((unsigned long)tp & 3)
+		return -EINVAL;
+
+	ret = larch_insn_write(tp, insn);
+	if (!ret)
+		flush_icache_range((unsigned long)tp,
+				   (unsigned long)tp + LOONGARCH_INSN_SIZE);
+
+	return ret;
+}
+
+u32 larch_insn_gen_nop(void)
+{
+	return INSN_NOP;
+}
+
+u32 larch_insn_gen_bl(unsigned long pc, unsigned long dest)
+{
+	long offset = dest - pc;
+	unsigned int immediate_l, immediate_h;
+	union loongarch_instruction insn;
+
+	if ((offset & 3) || offset < -SZ_128M || offset >= SZ_128M) {
+		pr_warn("The generated bl instruction is out of range.\n");
+		return INSN_BREAK;
+	}
+
+	offset >>= 2;
+
+	immediate_l = offset & 0xffff;
+	offset >>= 16;
+	immediate_h = offset & 0x3ff;
+
+	insn.reg0i26_format.opcode = bl_op;
+	insn.reg0i26_format.immediate_l = immediate_l;
+	insn.reg0i26_format.immediate_h = immediate_h;
+
+	return insn.word;
+}
+
+u32 larch_insn_gen_or(enum loongarch_gpr rd, enum loongarch_gpr rj, enum loongarch_gpr rk)
+{
+	union loongarch_instruction insn;
+
+	insn.reg3_format.opcode = or_op;
+	insn.reg3_format.rd = rd;
+	insn.reg3_format.rj = rj;
+	insn.reg3_format.rk = rk;
+
+	return insn.word;
+}
+
+u32 larch_insn_gen_move(enum loongarch_gpr rd, enum loongarch_gpr rj)
+{
+	return larch_insn_gen_or(rd, rj, 0);
+}
+
 u32 larch_insn_gen_lu32id(enum loongarch_gpr rd, int imm)
 {
 	union loongarch_instruction insn;

arch/loongarch/kernel/mcount_dyn.S

@@ -0,0 +1,85 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) 2022 Loongson Technology Corporation Limited
+ */
+
+#include <asm/export.h>
+#include <asm/ftrace.h>
+#include <asm/regdef.h>
+#include <asm/stackframe.h>
+
+	.text
+/*
+ * Due to -fpatchable-function-entry=2: the compiler inserted 2 NOPs before the
+ * regular C function prologue. When PC arrived here, the last 2 instructions
+ * are as follows:
+ * 	move		t0, ra
+ * 	bl		callsite (for modules, callsite is a tramplione)
+ *
+ * modules trampoline is as follows:
+ * 	lu12i.w		t1, callsite[31:12]
+ * 	lu32i.d		t1, callsite[51:32]
+ * 	lu52i.d		t1, t1, callsite[63:52]
+ * 	jirl		zero, t1, callsite[11:0] >> 2
+ *
+ * See arch/loongarch/kernel/ftrace_dyn.c for details. Here, pay attention to
+ * that the T series regs are available and safe because each C functions
+ * follows the LoongArch's psABI as well.
+ */
+
+	.macro  ftrace_regs_entry
+	PTR_ADDI	sp, sp, -PT_SIZE
+	PTR_S		t0, sp, PT_R1  /* Save parent ra at PT_R1(RA) */
+	PTR_S		a0, sp, PT_R4
+	PTR_S		a1, sp, PT_R5
+	PTR_S		a2, sp, PT_R6
+	PTR_S		a3, sp, PT_R7
+	PTR_S		a4, sp, PT_R8
+	PTR_S		a5, sp, PT_R9
+	PTR_S		a6, sp, PT_R10
+	PTR_S		a7, sp, PT_R11
+	PTR_S		fp, sp, PT_R22
+	PTR_S		ra, sp, PT_ERA /* Save trace function ra at PT_ERA */
+	PTR_ADDI	t8, sp, PT_SIZE
+	PTR_S		t8, sp, PT_R3
+	.endm
+
+SYM_FUNC_START(ftrace_stub)
+	jr		ra
+SYM_FUNC_END(ftrace_stub)
+
+SYM_CODE_START(ftrace_common)
+	PTR_ADDI	a0, ra, -8	/* arg0: ip */
+	move		a1, t0		/* arg1: parent_ip */
+	la.pcrel	t1, function_trace_op
+	PTR_L		a2, t1, 0	/* arg2: op */
+	move		a3, sp		/* arg3: regs */
+
+SYM_INNER_LABEL(ftrace_call, SYM_L_GLOBAL)
+	bl		ftrace_stub
+/*
+ * As we didn't use S series regs in this assmembly code and all calls
+ * are C function which will save S series regs by themselves, there is
+ * no need to restore S series regs. The T series is available and safe
+ * at the callsite, so there is no need to restore the T series regs.
+ */
+ftrace_common_return:
+	PTR_L		ra, sp, PT_R1
+	PTR_L		a0, sp, PT_R4
+	PTR_L		a1, sp, PT_R5
+	PTR_L		a2, sp, PT_R6
+	PTR_L		a3, sp, PT_R7
+	PTR_L		a4, sp, PT_R8
+	PTR_L		a5, sp, PT_R9
+	PTR_L		a6, sp, PT_R10
+	PTR_L		a7, sp, PT_R11
+	PTR_L		fp, sp, PT_R22
+	PTR_L		t0, sp, PT_ERA
+	PTR_ADDI	sp, sp, PT_SIZE
+	jr		t0
+SYM_CODE_END(ftrace_common)
+
+SYM_CODE_START(ftrace_caller)
+	ftrace_regs_entry
+	b		ftrace_common
+SYM_CODE_END(ftrace_caller)

arch/loongarch/kernel/unwind_prologue.c

@@ -8,6 +8,16 @@
 #include <asm/ptrace.h>
 #include <asm/unwind.h>
 
+static inline void unwind_state_fixup(struct unwind_state *state)
+{
+#ifdef CONFIG_DYNAMIC_FTRACE
+	static unsigned long ftrace = (unsigned long)ftrace_call + 4;
+
+	if (state->pc == ftrace)
+		state->is_ftrace = true;
+#endif
+}
+
 unsigned long unwind_get_return_address(struct unwind_state *state)
 {
 
@@ -41,15 +51,30 @@ static bool unwind_by_guess(struct unwind_state *state)
 
 static bool unwind_by_prologue(struct unwind_state *state)
 {
-	struct stack_info *info = &state->stack_info;
-	union loongarch_instruction *ip, *ip_end;
 	long frame_ra = -1;
 	unsigned long frame_size = 0;
 	unsigned long size, offset, pc = state->pc;
+	struct pt_regs *regs;
+	struct stack_info *info = &state->stack_info;
+	union loongarch_instruction *ip, *ip_end;
 
 	if (state->sp >= info->end || state->sp < info->begin)
 		return false;
 
+	if (state->is_ftrace) {
+		/*
+		 * As we meet ftrace_regs_entry, reset first flag like first doing
+		 * tracing. Prologue analysis will stop soon because PC is at entry.
+		 */
+		regs = (struct pt_regs *)state->sp;
+		state->first = true;
+		state->is_ftrace = false;
+		state->pc = regs->csr_era;
+		state->ra = regs->regs[1];
+		state->sp = regs->regs[3];
+		return true;
+	}
+
 	if (!kallsyms_lookup_size_offset(pc, &size, &offset))
 		return false;
 
@@ -95,7 +120,7 @@ static bool unwind_by_prologue(struct unwind_state *state)
 
 	state->pc = *(unsigned long *)(state->sp + frame_ra);
 	state->sp = state->sp + frame_size;
-	return !!__kernel_text_address(state->pc);
+	goto out;
 
 first:
 	state->first = false;
@@ -104,7 +129,9 @@ first:
 
 	state->pc = state->ra;
 
-	return !!__kernel_text_address(state->ra);
+out:
+	unwind_state_fixup(state);
+	return !!__kernel_text_address(state->pc);
 }
 
 void unwind_start(struct unwind_state *state, struct task_struct *task,