diff mbox series

[v2,1/3] arm64: implement ftrace with regs

Message ID 20180817102724.022FB68EF4@newverein.lst.de (mailing list archive)
State New, archived
Headers show
Series arm64 live patching | expand

Commit Message

Torsten Duwe Aug. 17, 2018, 10:27 a.m. UTC
Check for compiler support of -fpatchable-function-entry and use it
to intercept functions immediately on entry, saving the LR in x9.
Disable ftracing in efi/libstub, because this triggers cross-section
linker errors now (-pg is disabled already for those files).
Add an ftrace_caller which can handle LR in x9, as well as an
ftrace_regs_caller that additionally writes out a set of pt_regs
for inspection.

Signed-off-by: Torsten Duwe <duwe@suse.de>

Comments

Julien Thierry Aug. 17, 2018, 5:20 p.m. UTC | #1
Hi Torsten,

On 17/08/18 11:27, Torsten Duwe wrote:
> Check for compiler support of -fpatchable-function-entry and use it
> to intercept functions immediately on entry, saving the LR in x9.
> Disable ftracing in efi/libstub, because this triggers cross-section
> linker errors now (-pg is disabled already for those files).
> Add an ftrace_caller which can handle LR in x9, as well as an
> ftrace_regs_caller that additionally writes out a set of pt_regs
> for inspection.
> 
> Signed-off-by: Torsten Duwe <duwe@suse.de>
> 
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -110,6 +110,7 @@ config ARM64
>   	select HAVE_DEBUG_KMEMLEAK
>   	select HAVE_DMA_CONTIGUOUS
>   	select HAVE_DYNAMIC_FTRACE
> +	select HAVE_DYNAMIC_FTRACE_WITH_REGS

Actually there's an issue here. CONFIG_DYNAMIC_FTRACE_WITH_REGS is 
automatically selected if the arch has HAVE_DYNAMIC_FTRACE_WITH_REGS and 
DYNAMIC_FTRACE. Meaning we can no longer enable DYNAMIC_FTRACE without 
CONFIG_DYNAMIC_FTRACE_WITH_REGS.

So we wouldn't build an arm64 kernel with dynamic function tracing 
anymore without a recent enough compiler :( .

Is there a way to have "HAVE_DYNAMIC_FTRACE_WITH_REGS" depending on 
whether the compiler supports the patchable function entry or not?

(Not familiar enough with the build system to suggest something here).

>   	select HAVE_EFFICIENT_UNALIGNED_ACCESS
>   	select HAVE_FTRACE_MCOUNT_RECORD
>   	select HAVE_FUNCTION_TRACER
> --- a/arch/arm64/Makefile
> +++ b/arch/arm64/Makefile
> @@ -78,6 +78,15 @@ ifeq ($(CONFIG_ARM64_MODULE_PLTS),y)
>   KBUILD_LDFLAGS_MODULE	+= -T $(srctree)/arch/arm64/kernel/module.lds
>   endif
>   
> +ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS

use "ifeq($(CONFIG_DYNAMIC_FTRACE_WITH_REGS),y)", 
CONFIG_DYNAMIC_FTRACE_WITH_REGS is defined even if it is set to 'n' and 
will still try to use the compiler option.

> +  CC_FLAGS_FTRACE := -fpatchable-function-entry=2
> +  KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
> +  ifeq ($(call cc-option,-fpatchable-function-entry=2),)
> +    $(error Cannot use CONFIG_DYNAMIC_FTRACE_WITH_REGS: \
> +             -fpatchable-function-entry not supported by compiler)
> +  endif
> +endif
> +
>   # Default value
>   head-y		:= arch/arm64/kernel/head.o
>   
> --- a/arch/arm64/include/asm/ftrace.h
> +++ b/arch/arm64/include/asm/ftrace.h
> @@ -16,6 +16,13 @@
>   #define MCOUNT_ADDR		((unsigned long)_mcount)
>   #define MCOUNT_INSN_SIZE	AARCH64_INSN_SIZE
>   
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +#define ARCH_SUPPORTS_FTRACE_OPS 1
> +#define REC_IP_BRANCH_OFFSET 4
> +#else
> +#define REC_IP_BRANCH_OFFSET 0
> +#endif
> +
>   #ifndef __ASSEMBLY__
>   #include <linux/compat.h>
>   
> --- a/arch/arm64/kernel/Makefile
> +++ b/arch/arm64/kernel/Makefile
> @@ -7,9 +7,9 @@ CPPFLAGS_vmlinux.lds	:= -DTEXT_OFFSET=$(
>   AFLAGS_head.o		:= -DTEXT_OFFSET=$(TEXT_OFFSET)
>   CFLAGS_armv8_deprecated.o := -I$(src)
>   
> -CFLAGS_REMOVE_ftrace.o = -pg
> -CFLAGS_REMOVE_insn.o = -pg
> -CFLAGS_REMOVE_return_address.o = -pg
> +CFLAGS_REMOVE_ftrace.o = -pg $(CC_FLAGS_FTRACE)
> +CFLAGS_REMOVE_insn.o = -pg $(CC_FLAGS_FTRACE)
> +CFLAGS_REMOVE_return_address.o = -pg $(CC_FLAGS_FTRACE)
>   
>   # Object file lists.
>   arm64-obj-y		:= debug-monitors.o entry.o irq.o fpsimd.o		\
> --- a/drivers/firmware/efi/libstub/Makefile
> +++ b/drivers/firmware/efi/libstub/Makefile
> @@ -11,7 +11,8 @@ cflags-$(CONFIG_X86)		+= -m$(BITS) -D__K
>   				   -fPIC -fno-strict-aliasing -mno-red-zone \
>   				   -mno-mmx -mno-sse -fshort-wchar
>   
> -cflags-$(CONFIG_ARM64)		:= $(subst -pg,,$(KBUILD_CFLAGS)) -fpie
> +cflags-$(CONFIG_ARM64)		:= $(filter-out -pg $(CC_FLAGS_FTRACE)\
> +				  ,$(KBUILD_CFLAGS)) -fpie
>   cflags-$(CONFIG_ARM)		:= $(subst -pg,,$(KBUILD_CFLAGS)) \
>   				   -fno-builtin -fpic -mno-single-pic-base
>   
> --- a/arch/arm64/kernel/entry-ftrace.S
> +++ b/arch/arm64/kernel/entry-ftrace.S
> @@ -13,6 +13,8 @@
>   #include <asm/assembler.h>
>   #include <asm/ftrace.h>
>   #include <asm/insn.h>
> +#include <asm/asm-offsets.h>
> +#include <asm/assembler.h>
>   
>   /*
>    * Gcc with -pg will put the following code in the beginning of each function:
> @@ -123,6 +125,7 @@ skip_ftrace_call:			// }
>   ENDPROC(_mcount)
>   
>   #else /* CONFIG_DYNAMIC_FTRACE */
> +#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY
>   /*
>    * _mcount() is used to build the kernel with -pg option, but all the branch
>    * instructions to _mcount() are replaced to NOP initially at kernel start up,
> @@ -162,6 +165,88 @@ ftrace_graph_call:			// ftrace_graph_cal
>   
>   	mcount_exit
>   ENDPROC(ftrace_caller)
> +#else /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
> +ENTRY(_mcount)
> +	mov     x10, lr
> +	mov     lr, x9
> +	ret     x10
> +ENDPROC(_mcount)
> +
> +ENTRY(ftrace_regs_caller)
> +	stp	x29, x9, [sp, #-16]!
> +	sub	sp, sp, #S_FRAME_SIZE
> +
> +	stp	x10, x11, [sp, #80]
> +	stp	x12, x13, [sp, #96]
> +	stp	x14, x15, [sp, #112]
> +	stp	x16, x17, [sp, #128]
> +	stp	x18, x19, [sp, #144]
> +	stp	x20, x21, [sp, #160]
> +	stp	x22, x23, [sp, #176]
> +	stp	x24, x25, [sp, #192]
> +	stp	x26, x27, [sp, #208]
> +
> +	b	ftrace_common
> +ENDPROC(ftrace_regs_caller)
> +
> +ENTRY(ftrace_caller)
> +	stp	x29, x9, [sp, #-16]!
> +	sub	sp, sp, #S_FRAME_SIZE
> +
> +ftrace_common:
> +	stp	x28, x29, [sp, #224]
> +
> +	/* save function arguments */
> +	stp	x0, x1, [sp]
> +	stp	x2, x3, [sp, #16]
> +	stp	x4, x5, [sp, #32]
> +	stp	x6, x7, [sp, #48]
> +	stp	x8, x9, [sp, #64]
> +
> +	/* The link Register at callee entry */
> +	str	x9, [sp, #S_LR]
> +	/* The program counter just after the ftrace call site */
> +	str	lr, [sp, #S_PC]
> +	/* The stack pointer as it was on ftrace_caller entry... */
> +	add	x29, sp, #S_FRAME_SIZE+16	/* ...is also our new FP */
> +	str	x29, [sp, #S_SP]
> +
> +	adrp    x0, function_trace_op
> +	ldr     x2, [x0, #:lo12:function_trace_op]
> +	mov	x1, x9		/* saved LR == parent IP */
> +	sub	x0, lr, #8	/* function entry == IP */
> +	mov	x3, sp		/* pt_regs are @sp */
> +
> +	.global ftrace_call
> +ftrace_call:
> +
> +	bl	ftrace_stub

I should've asked this in my initial review. But why are we always 
calling ftrace_stub rather than have a patchable "nop" like in the 
!CC_USING_PATCHABLE_FUNCTION_ENTRY version?

> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +	.global ftrace_graph_call
> +ftrace_graph_call:			// ftrace_graph_caller();
> +	nop				// If enabled, this will be replaced
> +					// "b ftrace_graph_caller"
> +#endif
> +
> +ftrace_common_return:
> +	ldp	x0, x1, [sp]
> +	ldp	x2, x3, [sp, #16]
> +	ldp	x4, x5, [sp, #32]
> +	ldp	x6, x7, [sp, #48]
> +	ldp	x8, x9, [sp, #64]
> +
> +	ldp	x28, x29, [sp, #224]
> +
> +	ldr	x9, [sp, #S_PC]
> +	ldr	lr, [sp, #S_LR]
> +	add	sp, sp, #S_FRAME_SIZE+16
> +
> +	ret	x9
> +
> +ENDPROC(ftrace_caller)
> +
> +#endif /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
>   #endif /* CONFIG_DYNAMIC_FTRACE */
>   
>   ENTRY(ftrace_stub)
> @@ -197,12 +284,20 @@ ENDPROC(ftrace_stub)
>    * and run return_to_handler() later on its exit.
>    */
>   ENTRY(ftrace_graph_caller)
> +#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY
>   	mcount_get_lr_addr	  x0	//     pointer to function's saved lr
>   	mcount_get_pc		  x1	//     function's pc
>   	mcount_get_parent_fp	  x2	//     parent's fp
>   	bl	prepare_ftrace_return	// prepare_ftrace_return(&lr, pc, fp)
>   
>   	mcount_exit
> +#else
> +	add	x0, sp, #S_LR	/* address of (LR pointing into caller) */
> +	ldr	x1, [sp, #S_PC]
> +	ldr	x2, [sp, #S_X28 + 8]	/* caller's frame pointer */
> +	bl	prepare_ftrace_return
> +	b	ftrace_common_return
> +#endif
>   ENDPROC(ftrace_graph_caller)
>   
>   /*
> --- a/arch/arm64/kernel/ftrace.c
> +++ b/arch/arm64/kernel/ftrace.c
> @@ -70,7 +70,8 @@ int ftrace_update_ftrace_func(ftrace_fun
>    */
>   int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
>   {
> -	unsigned long pc = rec->ip;
> +	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
> +	int ret;
>   	u32 old, new;
>   	long offset = (long)pc - (long)addr;
>   
> @@ -128,22 +129,54 @@ int ftrace_make_call(struct dyn_ftrace *
>   	}
>   
>   	old = aarch64_insn_gen_nop();
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +#define INSN_MOV_X9_X30 aarch64_insn_gen_add_sub_imm(AARCH64_INSN_REG_9, \
> +		AARCH64_INSN_REG_30, 0, \
> +		AARCH64_INSN_VARIANT_64BIT, AARCH64_INSN_ADSB_ADD)

It might just be me, but since this macro is used in two different 
functions I'm not overly fond of defining in the middle of the function 
code.

Thanks,

> +	new = INSN_MOV_X9_X30;
> +	ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
> +	if (ret)
> +		return ret;
> +	smp_wmb();	/* ensure LR saver is in place before ftrace call */
> +#endif
>   	new = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
>   
>   	return ftrace_modify_code(pc, old, new, true);
>   }
>   
> +int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
> +			unsigned long addr)
> +{
> +	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
> +	u32 old, new;
> +
> +	old = aarch64_insn_gen_branch_imm(pc, old_addr, true);
> +	new = aarch64_insn_gen_branch_imm(pc, addr, true);
> +
> +	return ftrace_modify_code(pc, old, new, true);
> +}
> +
> +
>   /*
>    * Turn off the call to ftrace_caller() in instrumented function
>    */
>   int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
>   		    unsigned long addr)
>   {
> -	unsigned long pc = rec->ip;
> +	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
>   	bool validate = true;
> +	int ret;
>   	u32 old = 0, new;
>   	long offset = (long)pc - (long)addr;
>   
> +#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY
> +	/* -fpatchable-function-entry= does not generate a profiling call
> +	 *  initially; the NOPs are already there.
> +	 */
> +	if (addr == MCOUNT_ADDR)
> +		return 0;
> +#endif
> +
>   	if (offset < -SZ_128M || offset >= SZ_128M) {
>   #ifdef CONFIG_ARM64_MODULE_PLTS
>   		u32 replaced;
> @@ -188,7 +221,15 @@ int ftrace_make_nop(struct module *mod,
>   
>   	new = aarch64_insn_gen_nop();
>   
> -	return ftrace_modify_code(pc, old, new, validate);
> +	ret = ftrace_modify_code(pc, old, new, validate);
> +	if (ret)
> +		return ret;
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +	smp_wmb();	/* ftrace call must not remain without LR saver. */
> +	old = INSN_MOV_X9_X30;
> +	ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
> +#endif
> +	return ret;
>   }
>   
>   void arch_ftrace_update_code(int command)
> --- a/include/asm-generic/vmlinux.lds.h
> +++ b/include/asm-generic/vmlinux.lds.h
> @@ -115,6 +115,7 @@
>   #define MCOUNT_REC()	. = ALIGN(8);				\
>   			__start_mcount_loc = .;			\
>   			KEEP(*(__mcount_loc))			\
> +			KEEP(*(__patchable_function_entries))	\
>   			__stop_mcount_loc = .;
>   #else
>   #define MCOUNT_REC()
> --- a/include/linux/compiler_types.h
> +++ b/include/linux/compiler_types.h
> @@ -61,8 +61,12 @@ extern void __chk_io_ptr(const volatile
>   #if defined(CC_USING_HOTPATCH) && !defined(__CHECKER__)
>   #define notrace __attribute__((hotpatch(0,0)))
>   #else
> +#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY
> +#define notrace __attribute__((patchable_function_entry(0)))
> +#else
>   #define notrace __attribute__((no_instrument_function))
>   #endif
> +#endif
>   
>   /* Intel compiler defines __GNUC__. So we will overwrite implementations
>    * coming from above header files here
>
Ard Biesheuvel Aug. 20, 2018, 6:55 a.m. UTC | #2
On 17 August 2018 at 12:27, Torsten Duwe <duwe@lst.de> wrote:
> Check for compiler support of -fpatchable-function-entry and use it
> to intercept functions immediately on entry, saving the LR in x9.

Could you please add a note that this is safe because IPA register
allocation is turned off as well by that option? I had to go and find
the GCC patches to confirm that this is the case.

> Disable ftracing in efi/libstub, because this triggers cross-section
> linker errors now (-pg is disabled already for those files).
> Add an ftrace_caller which can handle LR in x9, as well as an
> ftrace_regs_caller that additionally writes out a set of pt_regs
> for inspection.
>
> Signed-off-by: Torsten Duwe <duwe@suse.de>
>
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -110,6 +110,7 @@ config ARM64
>         select HAVE_DEBUG_KMEMLEAK
>         select HAVE_DMA_CONTIGUOUS
>         select HAVE_DYNAMIC_FTRACE
> +       select HAVE_DYNAMIC_FTRACE_WITH_REGS
>         select HAVE_EFFICIENT_UNALIGNED_ACCESS
>         select HAVE_FTRACE_MCOUNT_RECORD
>         select HAVE_FUNCTION_TRACER
> --- a/arch/arm64/Makefile
> +++ b/arch/arm64/Makefile
> @@ -78,6 +78,15 @@ ifeq ($(CONFIG_ARM64_MODULE_PLTS),y)
>  KBUILD_LDFLAGS_MODULE  += -T $(srctree)/arch/arm64/kernel/module.lds
>  endif
>
> +ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +  CC_FLAGS_FTRACE := -fpatchable-function-entry=2
> +  KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
> +  ifeq ($(call cc-option,-fpatchable-function-entry=2),)
> +    $(error Cannot use CONFIG_DYNAMIC_FTRACE_WITH_REGS: \
> +             -fpatchable-function-entry not supported by compiler)
> +  endif
> +endif
> +
>  # Default value
>  head-y         := arch/arm64/kernel/head.o
>
> --- a/arch/arm64/include/asm/ftrace.h
> +++ b/arch/arm64/include/asm/ftrace.h
> @@ -16,6 +16,13 @@
>  #define MCOUNT_ADDR            ((unsigned long)_mcount)
>  #define MCOUNT_INSN_SIZE       AARCH64_INSN_SIZE
>
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +#define ARCH_SUPPORTS_FTRACE_OPS 1
> +#define REC_IP_BRANCH_OFFSET 4
> +#else
> +#define REC_IP_BRANCH_OFFSET 0
> +#endif
> +
>  #ifndef __ASSEMBLY__
>  #include <linux/compat.h>
>
> --- a/arch/arm64/kernel/Makefile
> +++ b/arch/arm64/kernel/Makefile
> @@ -7,9 +7,9 @@ CPPFLAGS_vmlinux.lds    := -DTEXT_OFFSET=$(
>  AFLAGS_head.o          := -DTEXT_OFFSET=$(TEXT_OFFSET)
>  CFLAGS_armv8_deprecated.o := -I$(src)
>
> -CFLAGS_REMOVE_ftrace.o = -pg
> -CFLAGS_REMOVE_insn.o = -pg
> -CFLAGS_REMOVE_return_address.o = -pg
> +CFLAGS_REMOVE_ftrace.o = -pg $(CC_FLAGS_FTRACE)
> +CFLAGS_REMOVE_insn.o = -pg $(CC_FLAGS_FTRACE)
> +CFLAGS_REMOVE_return_address.o = -pg $(CC_FLAGS_FTRACE)
>
>  # Object file lists.
>  arm64-obj-y            := debug-monitors.o entry.o irq.o fpsimd.o              \
> --- a/drivers/firmware/efi/libstub/Makefile
> +++ b/drivers/firmware/efi/libstub/Makefile
> @@ -11,7 +11,8 @@ cflags-$(CONFIG_X86)          += -m$(BITS) -D__K
>                                    -fPIC -fno-strict-aliasing -mno-red-zone \
>                                    -mno-mmx -mno-sse -fshort-wchar
>
> -cflags-$(CONFIG_ARM64)         := $(subst -pg,,$(KBUILD_CFLAGS)) -fpie
> +cflags-$(CONFIG_ARM64)         := $(filter-out -pg $(CC_FLAGS_FTRACE)\
> +                                 ,$(KBUILD_CFLAGS)) -fpie
>  cflags-$(CONFIG_ARM)           := $(subst -pg,,$(KBUILD_CFLAGS)) \
>                                    -fno-builtin -fpic -mno-single-pic-base
>
> --- a/arch/arm64/kernel/entry-ftrace.S
> +++ b/arch/arm64/kernel/entry-ftrace.S
> @@ -13,6 +13,8 @@
>  #include <asm/assembler.h>
>  #include <asm/ftrace.h>
>  #include <asm/insn.h>
> +#include <asm/asm-offsets.h>
> +#include <asm/assembler.h>
>
>  /*
>   * Gcc with -pg will put the following code in the beginning of each function:
> @@ -123,6 +125,7 @@ skip_ftrace_call:                   // }
>  ENDPROC(_mcount)
>
>  #else /* CONFIG_DYNAMIC_FTRACE */
> +#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY
>  /*
>   * _mcount() is used to build the kernel with -pg option, but all the branch
>   * instructions to _mcount() are replaced to NOP initially at kernel start up,
> @@ -162,6 +165,88 @@ ftrace_graph_call:                 // ftrace_graph_cal
>
>         mcount_exit
>  ENDPROC(ftrace_caller)
> +#else /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
> +ENTRY(_mcount)
> +       mov     x10, lr
> +       mov     lr, x9
> +       ret     x10
> +ENDPROC(_mcount)
> +
> +ENTRY(ftrace_regs_caller)
> +       stp     x29, x9, [sp, #-16]!
> +       sub     sp, sp, #S_FRAME_SIZE
> +
> +       stp     x10, x11, [sp, #80]
> +       stp     x12, x13, [sp, #96]
> +       stp     x14, x15, [sp, #112]
> +       stp     x16, x17, [sp, #128]
> +       stp     x18, x19, [sp, #144]
> +       stp     x20, x21, [sp, #160]
> +       stp     x22, x23, [sp, #176]
> +       stp     x24, x25, [sp, #192]
> +       stp     x26, x27, [sp, #208]
> +

Any reason in particular this stack frame is the wrong way around? We
usually keep the frame pointer at the top.

> +       b       ftrace_common
> +ENDPROC(ftrace_regs_caller)
> +
> +ENTRY(ftrace_caller)
> +       stp     x29, x9, [sp, #-16]!
> +       sub     sp, sp, #S_FRAME_SIZE
> +
> +ftrace_common:
> +       stp     x28, x29, [sp, #224]
> +
> +       /* save function arguments */
> +       stp     x0, x1, [sp]
> +       stp     x2, x3, [sp, #16]
> +       stp     x4, x5, [sp, #32]
> +       stp     x6, x7, [sp, #48]
> +       stp     x8, x9, [sp, #64]
> +
> +       /* The link Register at callee entry */
> +       str     x9, [sp, #S_LR]
> +       /* The program counter just after the ftrace call site */
> +       str     lr, [sp, #S_PC]
> +       /* The stack pointer as it was on ftrace_caller entry... */
> +       add     x29, sp, #S_FRAME_SIZE+16       /* ...is also our new FP */
> +       str     x29, [sp, #S_SP]
> +
> +       adrp    x0, function_trace_op
> +       ldr     x2, [x0, #:lo12:function_trace_op]

Please use the ldr_l macro here

> +       mov     x1, x9          /* saved LR == parent IP */
> +       sub     x0, lr, #8      /* function entry == IP */
> +       mov     x3, sp          /* pt_regs are @sp */
> +
> +       .global ftrace_call
> +ftrace_call:
> +
> +       bl      ftrace_stub
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +       .global ftrace_graph_call
> +ftrace_graph_call:                     // ftrace_graph_caller();
> +       nop                             // If enabled, this will be replaced
> +                                       // "b ftrace_graph_caller"
> +#endif
> +
> +ftrace_common_return:
> +       ldp     x0, x1, [sp]
> +       ldp     x2, x3, [sp, #16]
> +       ldp     x4, x5, [sp, #32]
> +       ldp     x6, x7, [sp, #48]
> +       ldp     x8, x9, [sp, #64]
> +
> +       ldp     x28, x29, [sp, #224]
> +
> +       ldr     x9, [sp, #S_PC]
> +       ldr     lr, [sp, #S_LR]
> +       add     sp, sp, #S_FRAME_SIZE+16
> +
> +       ret     x9
> +
> +ENDPROC(ftrace_caller)
> +
> +#endif /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
>  #endif /* CONFIG_DYNAMIC_FTRACE */
>
>  ENTRY(ftrace_stub)
> @@ -197,12 +284,20 @@ ENDPROC(ftrace_stub)
>   * and run return_to_handler() later on its exit.
>   */
>  ENTRY(ftrace_graph_caller)
> +#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY
>         mcount_get_lr_addr        x0    //     pointer to function's saved lr
>         mcount_get_pc             x1    //     function's pc
>         mcount_get_parent_fp      x2    //     parent's fp
>         bl      prepare_ftrace_return   // prepare_ftrace_return(&lr, pc, fp)
>
>         mcount_exit
> +#else
> +       add     x0, sp, #S_LR   /* address of (LR pointing into caller) */
> +       ldr     x1, [sp, #S_PC]
> +       ldr     x2, [sp, #S_X28 + 8]    /* caller's frame pointer */
> +       bl      prepare_ftrace_return
> +       b       ftrace_common_return
> +#endif
>  ENDPROC(ftrace_graph_caller)
>
>  /*
> --- a/arch/arm64/kernel/ftrace.c
> +++ b/arch/arm64/kernel/ftrace.c
> @@ -70,7 +70,8 @@ int ftrace_update_ftrace_func(ftrace_fun
>   */
>  int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
>  {
> -       unsigned long pc = rec->ip;
> +       unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;

We usually put spaces around arithmetic operators.

> +       int ret;
>         u32 old, new;
>         long offset = (long)pc - (long)addr;
>
> @@ -128,22 +129,54 @@ int ftrace_make_call(struct dyn_ftrace *
>         }
>
>         old = aarch64_insn_gen_nop();
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS

Please use IS_ENABLED() so that the code below is visible to the
compiler regardless of whether it is actually enabled.

> +#define INSN_MOV_X9_X30 aarch64_insn_gen_add_sub_imm(AARCH64_INSN_REG_9, \
> +               AARCH64_INSN_REG_30, 0, \
> +               AARCH64_INSN_VARIANT_64BIT, AARCH64_INSN_ADSB_ADD)

Please don't define the macro inline like that, and in fact, I think
you can drop it altogether and add a comment instead to clarify that
add x9, x30, #0 == mov x9, x30

> +       new = INSN_MOV_X9_X30;
> +       ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
> +       if (ret)
> +               return ret;
> +       smp_wmb();      /* ensure LR saver is in place before ftrace call */
> +#endif
>         new = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
>
>         return ftrace_modify_code(pc, old, new, true);
>  }
>
> +int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
> +                       unsigned long addr)
> +{
> +       unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
> +       u32 old, new;
> +
> +       old = aarch64_insn_gen_branch_imm(pc, old_addr, true);
> +       new = aarch64_insn_gen_branch_imm(pc, addr, true);
> +
> +       return ftrace_modify_code(pc, old, new, true);
> +}
> +
> +
>  /*
>   * Turn off the call to ftrace_caller() in instrumented function
>   */
>  int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
>                     unsigned long addr)
>  {
> -       unsigned long pc = rec->ip;
> +       unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
>         bool validate = true;
> +       int ret;
>         u32 old = 0, new;
>         long offset = (long)pc - (long)addr;
>
> +#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY

Use IS_ENABLED()

> +       /* -fpatchable-function-entry= does not generate a profiling call
> +        *  initially; the NOPs are already there.
> +        */
> +       if (addr == MCOUNT_ADDR)
> +               return 0;
> +#endif
> +
>         if (offset < -SZ_128M || offset >= SZ_128M) {
>  #ifdef CONFIG_ARM64_MODULE_PLTS

No changes needed at all for modules that are more than 128 MB away
from the core kernel? That seems odd to me, given that we currently
have a single PLT entry per module, and so you cannot mix ordinary
ftrace and ftrace with regs in a single module. Or is that not a
problem?

>                 u32 replaced;
> @@ -188,7 +221,15 @@ int ftrace_make_nop(struct module *mod,
>
>         new = aarch64_insn_gen_nop();
>
> -       return ftrace_modify_code(pc, old, new, validate);
> +       ret = ftrace_modify_code(pc, old, new, validate);
> +       if (ret)
> +               return ret;
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS

Use IS_ENABLED()

> +       smp_wmb();      /* ftrace call must not remain without LR saver. */
> +       old = INSN_MOV_X9_X30;

Uhm, ok, you're using this multiple times. You can still keep the
comment and use the add/sub version directly, or introduce a mov
variant in a separate patch.

> +       ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
> +#endif
> +       return ret;
>  }
>
>  void arch_ftrace_update_code(int command)
> --- a/include/asm-generic/vmlinux.lds.h
> +++ b/include/asm-generic/vmlinux.lds.h
> @@ -115,6 +115,7 @@
>  #define MCOUNT_REC()   . = ALIGN(8);                           \
>                         __start_mcount_loc = .;                 \
>                         KEEP(*(__mcount_loc))                   \
> +                       KEEP(*(__patchable_function_entries))   \
>                         __stop_mcount_loc = .;
>  #else
>  #define MCOUNT_REC()
> --- a/include/linux/compiler_types.h
> +++ b/include/linux/compiler_types.h
> @@ -61,8 +61,12 @@ extern void __chk_io_ptr(const volatile
>  #if defined(CC_USING_HOTPATCH) && !defined(__CHECKER__)
>  #define notrace __attribute__((hotpatch(0,0)))
>  #else
> +#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY
> +#define notrace __attribute__((patchable_function_entry(0)))
> +#else
>  #define notrace __attribute__((no_instrument_function))
>  #endif
> +#endif
>
>  /* Intel compiler defines __GNUC__. So we will overwrite implementations
>   * coming from above header files here
AKASHI Takahiro Aug. 29, 2018, 7:11 a.m. UTC | #3
On Fri, Aug 17, 2018 at 12:27:24PM +0200, Torsten Duwe wrote:
> Check for compiler support of -fpatchable-function-entry and use it
> to intercept functions immediately on entry, saving the LR in x9.
> Disable ftracing in efi/libstub, because this triggers cross-section
> linker errors now (-pg is disabled already for those files).
> Add an ftrace_caller which can handle LR in x9, as well as an

I think that we have a bit detailed descriptions about what a function's
initial prologue looks like and how it will be patched to enable or
disable ftrace on that function for the sake of better understandings.
(in entry-ftrace.S or ftrace.c?)

> ftrace_regs_caller that additionally writes out a set of pt_regs
> for inspection.
> 
> Signed-off-by: Torsten Duwe <duwe@suse.de>
> 
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -110,6 +110,7 @@ config ARM64
>  	select HAVE_DEBUG_KMEMLEAK
>  	select HAVE_DMA_CONTIGUOUS
>  	select HAVE_DYNAMIC_FTRACE
> +	select HAVE_DYNAMIC_FTRACE_WITH_REGS
>  	select HAVE_EFFICIENT_UNALIGNED_ACCESS
>  	select HAVE_FTRACE_MCOUNT_RECORD
>  	select HAVE_FUNCTION_TRACER
> --- a/arch/arm64/Makefile
> +++ b/arch/arm64/Makefile
> @@ -78,6 +78,15 @@ ifeq ($(CONFIG_ARM64_MODULE_PLTS),y)
>  KBUILD_LDFLAGS_MODULE	+= -T $(srctree)/arch/arm64/kernel/module.lds
>  endif
>  
> +ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +  CC_FLAGS_FTRACE := -fpatchable-function-entry=2
> +  KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
> +  ifeq ($(call cc-option,-fpatchable-function-entry=2),)
> +    $(error Cannot use CONFIG_DYNAMIC_FTRACE_WITH_REGS: \
> +             -fpatchable-function-entry not supported by compiler)
> +  endif
> +endif
> +
>  # Default value
>  head-y		:= arch/arm64/kernel/head.o
>  
> --- a/arch/arm64/include/asm/ftrace.h
> +++ b/arch/arm64/include/asm/ftrace.h
> @@ -16,6 +16,13 @@
>  #define MCOUNT_ADDR		((unsigned long)_mcount)
>  #define MCOUNT_INSN_SIZE	AARCH64_INSN_SIZE
>  
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +#define ARCH_SUPPORTS_FTRACE_OPS 1
> +#define REC_IP_BRANCH_OFFSET 4

Some explanation about "4" will be helpful here.

> +#else
> +#define REC_IP_BRANCH_OFFSET 0
> +#endif
> +
>  #ifndef __ASSEMBLY__
>  #include <linux/compat.h>
>  
> --- a/arch/arm64/kernel/Makefile
> +++ b/arch/arm64/kernel/Makefile
> @@ -7,9 +7,9 @@ CPPFLAGS_vmlinux.lds	:= -DTEXT_OFFSET=$(
>  AFLAGS_head.o		:= -DTEXT_OFFSET=$(TEXT_OFFSET)
>  CFLAGS_armv8_deprecated.o := -I$(src)
>  
> -CFLAGS_REMOVE_ftrace.o = -pg
> -CFLAGS_REMOVE_insn.o = -pg
> -CFLAGS_REMOVE_return_address.o = -pg
> +CFLAGS_REMOVE_ftrace.o = -pg $(CC_FLAGS_FTRACE)
> +CFLAGS_REMOVE_insn.o = -pg $(CC_FLAGS_FTRACE)
> +CFLAGS_REMOVE_return_address.o = -pg $(CC_FLAGS_FTRACE)

You don't need to remove "-pg" explicitly here because it won't be added to
XX_CFLAGS when you defines CC_FLAGS_FTRACE.

>  # Object file lists.
>  arm64-obj-y		:= debug-monitors.o entry.o irq.o fpsimd.o		\
> --- a/drivers/firmware/efi/libstub/Makefile
> +++ b/drivers/firmware/efi/libstub/Makefile
> @@ -11,7 +11,8 @@ cflags-$(CONFIG_X86)		+= -m$(BITS) -D__K
>  				   -fPIC -fno-strict-aliasing -mno-red-zone \
>  				   -mno-mmx -mno-sse -fshort-wchar
>  
> -cflags-$(CONFIG_ARM64)		:= $(subst -pg,,$(KBUILD_CFLAGS)) -fpie
> +cflags-$(CONFIG_ARM64)		:= $(filter-out -pg $(CC_FLAGS_FTRACE)\
> +				  ,$(KBUILD_CFLAGS)) -fpie
>  cflags-$(CONFIG_ARM)		:= $(subst -pg,,$(KBUILD_CFLAGS)) \
>  				   -fno-builtin -fpic -mno-single-pic-base
>  
> --- a/arch/arm64/kernel/entry-ftrace.S
> +++ b/arch/arm64/kernel/entry-ftrace.S
> @@ -13,6 +13,8 @@
>  #include <asm/assembler.h>
>  #include <asm/ftrace.h>
>  #include <asm/insn.h>
> +#include <asm/asm-offsets.h>
> +#include <asm/assembler.h>
>  
>  /*
>   * Gcc with -pg will put the following code in the beginning of each function:
> @@ -123,6 +125,7 @@ skip_ftrace_call:			// }
>  ENDPROC(_mcount)
>  
>  #else /* CONFIG_DYNAMIC_FTRACE */
> +#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY

I think that using CONFIG_DYNAMIC_FTRACE_WITH_REG here sounds more consistent.

>  /*
>   * _mcount() is used to build the kernel with -pg option, but all the branch
>   * instructions to _mcount() are replaced to NOP initially at kernel start up,
> @@ -162,6 +165,88 @@ ftrace_graph_call:			// ftrace_graph_cal
>  
>  	mcount_exit
>  ENDPROC(ftrace_caller)
> +#else /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
> +ENTRY(_mcount)
> +	mov     x10, lr
> +	mov     lr, x9
> +	ret     x10
> +ENDPROC(_mcount)

I don't think we need a definition of _mcount because patchable-function-entry
won't generate a call site of _mcount nor other code does.
You can simply define MCOUNT_ADDR as ULONG_MAX.

> +ENTRY(ftrace_regs_caller)
> +	stp	x29, x9, [sp, #-16]!
> +	sub	sp, sp, #S_FRAME_SIZE
> +
> +	stp	x10, x11, [sp, #80]
> +	stp	x12, x13, [sp, #96]
> +	stp	x14, x15, [sp, #112]
> +	stp	x16, x17, [sp, #128]
> +	stp	x18, x19, [sp, #144]
> +	stp	x20, x21, [sp, #160]
> +	stp	x22, x23, [sp, #176]
> +	stp	x24, x25, [sp, #192]
> +	stp	x26, x27, [sp, #208]
> +
> +	b	ftrace_common
> +ENDPROC(ftrace_regs_caller)
> +
> +ENTRY(ftrace_caller)
> +	stp	x29, x9, [sp, #-16]!
> +	sub	sp, sp, #S_FRAME_SIZE
> +
> +ftrace_common:
> +	stp	x28, x29, [sp, #224]
> +
> +	/* save function arguments */
> +	stp	x0, x1, [sp]
> +	stp	x2, x3, [sp, #16]
> +	stp	x4, x5, [sp, #32]
> +	stp	x6, x7, [sp, #48]
> +	stp	x8, x9, [sp, #64]
> +
> +	/* The link Register at callee entry */
> +	str	x9, [sp, #S_LR]
> +	/* The program counter just after the ftrace call site */
> +	str	lr, [sp, #S_PC]
> +	/* The stack pointer as it was on ftrace_caller entry... */
> +	add	x29, sp, #S_FRAME_SIZE+16	/* ...is also our new FP */
> +	str	x29, [sp, #S_SP]
> +
> +	adrp    x0, function_trace_op
> +	ldr     x2, [x0, #:lo12:function_trace_op]
> +	mov	x1, x9		/* saved LR == parent IP */
> +	sub	x0, lr, #8	/* function entry == IP */
> +	mov	x3, sp		/* pt_regs are @sp */
> +
> +	.global ftrace_call
> +ftrace_call:
> +
> +	bl	ftrace_stub
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +	.global ftrace_graph_call
> +ftrace_graph_call:			// ftrace_graph_caller();
> +	nop				// If enabled, this will be replaced
> +					// "b ftrace_graph_caller"
> +#endif
> +
> +ftrace_common_return:
> +	ldp	x0, x1, [sp]
> +	ldp	x2, x3, [sp, #16]
> +	ldp	x4, x5, [sp, #32]
> +	ldp	x6, x7, [sp, #48]
> +	ldp	x8, x9, [sp, #64]
> +
> +	ldp	x28, x29, [sp, #224]
> +
> +	ldr	x9, [sp, #S_PC]
> +	ldr	lr, [sp, #S_LR]
> +	add	sp, sp, #S_FRAME_SIZE+16
> +
> +	ret	x9
> +
> +ENDPROC(ftrace_caller)
> +
> +#endif /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
>  #endif /* CONFIG_DYNAMIC_FTRACE */
>  
>  ENTRY(ftrace_stub)
> @@ -197,12 +284,20 @@ ENDPROC(ftrace_stub)
>   * and run return_to_handler() later on its exit.
>   */
>  ENTRY(ftrace_graph_caller)
> +#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY

ditto.

>  	mcount_get_lr_addr	  x0	//     pointer to function's saved lr
>  	mcount_get_pc		  x1	//     function's pc
>  	mcount_get_parent_fp	  x2	//     parent's fp
>  	bl	prepare_ftrace_return	// prepare_ftrace_return(&lr, pc, fp)
>  
>  	mcount_exit
> +#else
> +	add	x0, sp, #S_LR	/* address of (LR pointing into caller) */
> +	ldr	x1, [sp, #S_PC]
> +	ldr	x2, [sp, #S_X28 + 8]	/* caller's frame pointer */
> +	bl	prepare_ftrace_return
> +	b	ftrace_common_return
> +#endif
>  ENDPROC(ftrace_graph_caller)
>  
>  /*
> --- a/arch/arm64/kernel/ftrace.c
> +++ b/arch/arm64/kernel/ftrace.c
> @@ -70,7 +70,8 @@ int ftrace_update_ftrace_func(ftrace_fun
>   */
>  int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
>  {
> -	unsigned long pc = rec->ip;
> +	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
> +	int ret;
>  	u32 old, new;
>  	long offset = (long)pc - (long)addr;
>  
> @@ -128,22 +129,54 @@ int ftrace_make_call(struct dyn_ftrace *
>  	}
>  
>  	old = aarch64_insn_gen_nop();
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +#define INSN_MOV_X9_X30 aarch64_insn_gen_add_sub_imm(AARCH64_INSN_REG_9, \
> +		AARCH64_INSN_REG_30, 0, \
> +		AARCH64_INSN_VARIANT_64BIT, AARCH64_INSN_ADSB_ADD)
> +	new = INSN_MOV_X9_X30;
> +	ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
> +	if (ret)
> +		return ret;
> +	smp_wmb();	/* ensure LR saver is in place before ftrace call */
> +#endif
>  	new = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
>  
>  	return ftrace_modify_code(pc, old, new, true);
>  }
>  
> +int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
> +			unsigned long addr)
> +{
> +	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
> +	u32 old, new;
> +
> +	old = aarch64_insn_gen_branch_imm(pc, old_addr, true);
> +	new = aarch64_insn_gen_branch_imm(pc, addr, true);
> +
> +	return ftrace_modify_code(pc, old, new, true);
> +}
> +
> +
>  /*
>   * Turn off the call to ftrace_caller() in instrumented function
>   */
>  int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
>  		    unsigned long addr)
>  {
> -	unsigned long pc = rec->ip;
> +	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
>  	bool validate = true;
> +	int ret;
>  	u32 old = 0, new;
>  	long offset = (long)pc - (long)addr;
>  
> +#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY

ditto.

> +	/* -fpatchable-function-entry= does not generate a profiling call
> +	 *  initially; the NOPs are already there.
> +	 */
> +	if (addr == MCOUNT_ADDR)
> +		return 0;
> +#endif
> +
>  	if (offset < -SZ_128M || offset >= SZ_128M) {
>  #ifdef CONFIG_ARM64_MODULE_PLTS
>  		u32 replaced;
> @@ -188,7 +221,15 @@ int ftrace_make_nop(struct module *mod,
>  
>  	new = aarch64_insn_gen_nop();
>  
> -	return ftrace_modify_code(pc, old, new, validate);
> +	ret = ftrace_modify_code(pc, old, new, validate);
> +	if (ret)
> +		return ret;
> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> +	smp_wmb();	/* ftrace call must not remain without LR saver. */
> +	old = INSN_MOV_X9_X30;

Given a minimum of performance impact, we don't necessarily have to
remove this instruction.

Thanks,
-Takahiro AKASHI

> +	ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
> +#endif
> +	return ret;
>  }
>  
>  void arch_ftrace_update_code(int command)
> --- a/include/asm-generic/vmlinux.lds.h
> +++ b/include/asm-generic/vmlinux.lds.h
> @@ -115,6 +115,7 @@
>  #define MCOUNT_REC()	. = ALIGN(8);				\
>  			__start_mcount_loc = .;			\
>  			KEEP(*(__mcount_loc))			\
> +			KEEP(*(__patchable_function_entries))	\
>  			__stop_mcount_loc = .;
>  #else
>  #define MCOUNT_REC()
> --- a/include/linux/compiler_types.h
> +++ b/include/linux/compiler_types.h
> @@ -61,8 +61,12 @@ extern void __chk_io_ptr(const volatile
>  #if defined(CC_USING_HOTPATCH) && !defined(__CHECKER__)
>  #define notrace __attribute__((hotpatch(0,0)))
>  #else
> +#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY
> +#define notrace __attribute__((patchable_function_entry(0)))
> +#else
>  #define notrace __attribute__((no_instrument_function))
>  #endif
> +#endif
>  
>  /* Intel compiler defines __GNUC__. So we will overwrite implementations
>   * coming from above header files here
diff mbox series

Patch

--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -110,6 +110,7 @@  config ARM64
 	select HAVE_DEBUG_KMEMLEAK
 	select HAVE_DMA_CONTIGUOUS
 	select HAVE_DYNAMIC_FTRACE
+	select HAVE_DYNAMIC_FTRACE_WITH_REGS
 	select HAVE_EFFICIENT_UNALIGNED_ACCESS
 	select HAVE_FTRACE_MCOUNT_RECORD
 	select HAVE_FUNCTION_TRACER
--- a/arch/arm64/Makefile
+++ b/arch/arm64/Makefile
@@ -78,6 +78,15 @@  ifeq ($(CONFIG_ARM64_MODULE_PLTS),y)
 KBUILD_LDFLAGS_MODULE	+= -T $(srctree)/arch/arm64/kernel/module.lds
 endif
 
+ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
+  CC_FLAGS_FTRACE := -fpatchable-function-entry=2
+  KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
+  ifeq ($(call cc-option,-fpatchable-function-entry=2),)
+    $(error Cannot use CONFIG_DYNAMIC_FTRACE_WITH_REGS: \
+             -fpatchable-function-entry not supported by compiler)
+  endif
+endif
+
 # Default value
 head-y		:= arch/arm64/kernel/head.o
 
--- a/arch/arm64/include/asm/ftrace.h
+++ b/arch/arm64/include/asm/ftrace.h
@@ -16,6 +16,13 @@ 
 #define MCOUNT_ADDR		((unsigned long)_mcount)
 #define MCOUNT_INSN_SIZE	AARCH64_INSN_SIZE
 
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
+#define ARCH_SUPPORTS_FTRACE_OPS 1
+#define REC_IP_BRANCH_OFFSET 4
+#else
+#define REC_IP_BRANCH_OFFSET 0
+#endif
+
 #ifndef __ASSEMBLY__
 #include <linux/compat.h>
 
--- a/arch/arm64/kernel/Makefile
+++ b/arch/arm64/kernel/Makefile
@@ -7,9 +7,9 @@  CPPFLAGS_vmlinux.lds	:= -DTEXT_OFFSET=$(
 AFLAGS_head.o		:= -DTEXT_OFFSET=$(TEXT_OFFSET)
 CFLAGS_armv8_deprecated.o := -I$(src)
 
-CFLAGS_REMOVE_ftrace.o = -pg
-CFLAGS_REMOVE_insn.o = -pg
-CFLAGS_REMOVE_return_address.o = -pg
+CFLAGS_REMOVE_ftrace.o = -pg $(CC_FLAGS_FTRACE)
+CFLAGS_REMOVE_insn.o = -pg $(CC_FLAGS_FTRACE)
+CFLAGS_REMOVE_return_address.o = -pg $(CC_FLAGS_FTRACE)
 
 # Object file lists.
 arm64-obj-y		:= debug-monitors.o entry.o irq.o fpsimd.o		\
--- a/drivers/firmware/efi/libstub/Makefile
+++ b/drivers/firmware/efi/libstub/Makefile
@@ -11,7 +11,8 @@  cflags-$(CONFIG_X86)		+= -m$(BITS) -D__K
 				   -fPIC -fno-strict-aliasing -mno-red-zone \
 				   -mno-mmx -mno-sse -fshort-wchar
 
-cflags-$(CONFIG_ARM64)		:= $(subst -pg,,$(KBUILD_CFLAGS)) -fpie
+cflags-$(CONFIG_ARM64)		:= $(filter-out -pg $(CC_FLAGS_FTRACE)\
+				  ,$(KBUILD_CFLAGS)) -fpie
 cflags-$(CONFIG_ARM)		:= $(subst -pg,,$(KBUILD_CFLAGS)) \
 				   -fno-builtin -fpic -mno-single-pic-base
 
--- a/arch/arm64/kernel/entry-ftrace.S
+++ b/arch/arm64/kernel/entry-ftrace.S
@@ -13,6 +13,8 @@ 
 #include <asm/assembler.h>
 #include <asm/ftrace.h>
 #include <asm/insn.h>
+#include <asm/asm-offsets.h>
+#include <asm/assembler.h>
 
 /*
  * Gcc with -pg will put the following code in the beginning of each function:
@@ -123,6 +125,7 @@  skip_ftrace_call:			// }
 ENDPROC(_mcount)
 
 #else /* CONFIG_DYNAMIC_FTRACE */
+#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY
 /*
  * _mcount() is used to build the kernel with -pg option, but all the branch
  * instructions to _mcount() are replaced to NOP initially at kernel start up,
@@ -162,6 +165,88 @@  ftrace_graph_call:			// ftrace_graph_cal
 
 	mcount_exit
 ENDPROC(ftrace_caller)
+#else /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
+ENTRY(_mcount)
+	mov     x10, lr
+	mov     lr, x9
+	ret     x10
+ENDPROC(_mcount)
+
+ENTRY(ftrace_regs_caller)
+	stp	x29, x9, [sp, #-16]!
+	sub	sp, sp, #S_FRAME_SIZE
+
+	stp	x10, x11, [sp, #80]
+	stp	x12, x13, [sp, #96]
+	stp	x14, x15, [sp, #112]
+	stp	x16, x17, [sp, #128]
+	stp	x18, x19, [sp, #144]
+	stp	x20, x21, [sp, #160]
+	stp	x22, x23, [sp, #176]
+	stp	x24, x25, [sp, #192]
+	stp	x26, x27, [sp, #208]
+
+	b	ftrace_common
+ENDPROC(ftrace_regs_caller)
+
+ENTRY(ftrace_caller)
+	stp	x29, x9, [sp, #-16]!
+	sub	sp, sp, #S_FRAME_SIZE
+
+ftrace_common:
+	stp	x28, x29, [sp, #224]
+
+	/* save function arguments */
+	stp	x0, x1, [sp]
+	stp	x2, x3, [sp, #16]
+	stp	x4, x5, [sp, #32]
+	stp	x6, x7, [sp, #48]
+	stp	x8, x9, [sp, #64]
+
+	/* The link Register at callee entry */
+	str	x9, [sp, #S_LR]
+	/* The program counter just after the ftrace call site */
+	str	lr, [sp, #S_PC]
+	/* The stack pointer as it was on ftrace_caller entry... */
+	add	x29, sp, #S_FRAME_SIZE+16	/* ...is also our new FP */
+	str	x29, [sp, #S_SP]
+
+	adrp    x0, function_trace_op
+	ldr     x2, [x0, #:lo12:function_trace_op]
+	mov	x1, x9		/* saved LR == parent IP */
+	sub	x0, lr, #8	/* function entry == IP */
+	mov	x3, sp		/* pt_regs are @sp */
+
+	.global ftrace_call
+ftrace_call:
+
+	bl	ftrace_stub
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+	.global ftrace_graph_call
+ftrace_graph_call:			// ftrace_graph_caller();
+	nop				// If enabled, this will be replaced
+					// "b ftrace_graph_caller"
+#endif
+
+ftrace_common_return:
+	ldp	x0, x1, [sp]
+	ldp	x2, x3, [sp, #16]
+	ldp	x4, x5, [sp, #32]
+	ldp	x6, x7, [sp, #48]
+	ldp	x8, x9, [sp, #64]
+
+	ldp	x28, x29, [sp, #224]
+
+	ldr	x9, [sp, #S_PC]
+	ldr	lr, [sp, #S_LR]
+	add	sp, sp, #S_FRAME_SIZE+16
+
+	ret	x9
+
+ENDPROC(ftrace_caller)
+
+#endif /* CC_USING_PATCHABLE_FUNCTION_ENTRY */
 #endif /* CONFIG_DYNAMIC_FTRACE */
 
 ENTRY(ftrace_stub)
@@ -197,12 +284,20 @@  ENDPROC(ftrace_stub)
  * and run return_to_handler() later on its exit.
  */
 ENTRY(ftrace_graph_caller)
+#ifndef CC_USING_PATCHABLE_FUNCTION_ENTRY
 	mcount_get_lr_addr	  x0	//     pointer to function's saved lr
 	mcount_get_pc		  x1	//     function's pc
 	mcount_get_parent_fp	  x2	//     parent's fp
 	bl	prepare_ftrace_return	// prepare_ftrace_return(&lr, pc, fp)
 
 	mcount_exit
+#else
+	add	x0, sp, #S_LR	/* address of (LR pointing into caller) */
+	ldr	x1, [sp, #S_PC]
+	ldr	x2, [sp, #S_X28 + 8]	/* caller's frame pointer */
+	bl	prepare_ftrace_return
+	b	ftrace_common_return
+#endif
 ENDPROC(ftrace_graph_caller)
 
 /*
--- a/arch/arm64/kernel/ftrace.c
+++ b/arch/arm64/kernel/ftrace.c
@@ -70,7 +70,8 @@  int ftrace_update_ftrace_func(ftrace_fun
  */
 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 {
-	unsigned long pc = rec->ip;
+	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
+	int ret;
 	u32 old, new;
 	long offset = (long)pc - (long)addr;
 
@@ -128,22 +129,54 @@  int ftrace_make_call(struct dyn_ftrace *
 	}
 
 	old = aarch64_insn_gen_nop();
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
+#define INSN_MOV_X9_X30 aarch64_insn_gen_add_sub_imm(AARCH64_INSN_REG_9, \
+		AARCH64_INSN_REG_30, 0, \
+		AARCH64_INSN_VARIANT_64BIT, AARCH64_INSN_ADSB_ADD)
+	new = INSN_MOV_X9_X30;
+	ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
+	if (ret)
+		return ret;
+	smp_wmb();	/* ensure LR saver is in place before ftrace call */
+#endif
 	new = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
 
 	return ftrace_modify_code(pc, old, new, true);
 }
 
+int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
+			unsigned long addr)
+{
+	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
+	u32 old, new;
+
+	old = aarch64_insn_gen_branch_imm(pc, old_addr, true);
+	new = aarch64_insn_gen_branch_imm(pc, addr, true);
+
+	return ftrace_modify_code(pc, old, new, true);
+}
+
+
 /*
  * Turn off the call to ftrace_caller() in instrumented function
  */
 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
 		    unsigned long addr)
 {
-	unsigned long pc = rec->ip;
+	unsigned long pc = rec->ip+REC_IP_BRANCH_OFFSET;
 	bool validate = true;
+	int ret;
 	u32 old = 0, new;
 	long offset = (long)pc - (long)addr;
 
+#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY
+	/* -fpatchable-function-entry= does not generate a profiling call
+	 *  initially; the NOPs are already there.
+	 */
+	if (addr == MCOUNT_ADDR)
+		return 0;
+#endif
+
 	if (offset < -SZ_128M || offset >= SZ_128M) {
 #ifdef CONFIG_ARM64_MODULE_PLTS
 		u32 replaced;
@@ -188,7 +221,15 @@  int ftrace_make_nop(struct module *mod,
 
 	new = aarch64_insn_gen_nop();
 
-	return ftrace_modify_code(pc, old, new, validate);
+	ret = ftrace_modify_code(pc, old, new, validate);
+	if (ret)
+		return ret;
+#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
+	smp_wmb();	/* ftrace call must not remain without LR saver. */
+	old = INSN_MOV_X9_X30;
+	ret = ftrace_modify_code(pc-REC_IP_BRANCH_OFFSET, old, new, true);
+#endif
+	return ret;
 }
 
 void arch_ftrace_update_code(int command)
--- a/include/asm-generic/vmlinux.lds.h
+++ b/include/asm-generic/vmlinux.lds.h
@@ -115,6 +115,7 @@ 
 #define MCOUNT_REC()	. = ALIGN(8);				\
 			__start_mcount_loc = .;			\
 			KEEP(*(__mcount_loc))			\
+			KEEP(*(__patchable_function_entries))	\
 			__stop_mcount_loc = .;
 #else
 #define MCOUNT_REC()
--- a/include/linux/compiler_types.h
+++ b/include/linux/compiler_types.h
@@ -61,8 +61,12 @@  extern void __chk_io_ptr(const volatile
 #if defined(CC_USING_HOTPATCH) && !defined(__CHECKER__)
 #define notrace __attribute__((hotpatch(0,0)))
 #else
+#ifdef CC_USING_PATCHABLE_FUNCTION_ENTRY
+#define notrace __attribute__((patchable_function_entry(0)))
+#else
 #define notrace __attribute__((no_instrument_function))
 #endif
+#endif
 
 /* Intel compiler defines __GNUC__. So we will overwrite implementations
  * coming from above header files here