[PATCH v6 5/7] arm64: ftrace: Add dynamic ftrace support

[Will Deacon]

On Thu, Mar 13, 2014 at 10:13:48AM +0000, AKASHI Takahiro wrote:
> This patch allows "dynamic ftrace" if CONFIG_DYNAMIC_FTRACE is enabled.
> Here we can turn on and off tracing dynamically per-function base.
> 
> On arm64, this is done by patching single branch instruction to _mcount()
> inserted by gcc -pg option. The branch is replaced to NOP initially at
> kernel start up, and later on, NOP to branch to ftrace_caller() when
> enabled or branch to NOP when disabled.
> Please note that ftrace_caller() is a counterpart of _mcount() in case of
> 'static' ftrace.
> 
> More details on architecture specific requirements are described in
> Documentation/trace/ftrace-design.txt.
> 
> Signed-off-by: AKASHI Takahiro <takahiro.akashi at linaro.org>
> ---
>  arch/arm64/Kconfig               |    1 +
>  arch/arm64/include/asm/ftrace.h  |   15 +++++
>  arch/arm64/kernel/entry-ftrace.S |   43 +++++++++++++++
>  arch/arm64/kernel/ftrace.c       |  113 ++++++++++++++++++++++++++++++++++++++
>  4 files changed, 172 insertions(+)
> 
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 6b3fef6..6954959 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -33,6 +33,7 @@ config ARM64
>  	select HAVE_DMA_API_DEBUG
>  	select HAVE_DMA_ATTRS
>  	select HAVE_DMA_CONTIGUOUS
> +	select HAVE_DYNAMIC_FTRACE
>  	select HAVE_EFFICIENT_UNALIGNED_ACCESS
>  	select HAVE_FTRACE_MCOUNT_RECORD
>  	select HAVE_FUNCTION_TRACER
> diff --git a/arch/arm64/include/asm/ftrace.h b/arch/arm64/include/asm/ftrace.h
> index 58ea595..ed5c448 100644
> --- a/arch/arm64/include/asm/ftrace.h
> +++ b/arch/arm64/include/asm/ftrace.h
> @@ -18,6 +18,21 @@
>  
>  #ifndef __ASSEMBLY__
>  extern void _mcount(unsigned long);
> +
> +struct dyn_arch_ftrace {
> +	/* No extra data needed for arm64 */
> +};
> +
> +extern unsigned long ftrace_graph_call;
> +
> +static inline unsigned long ftrace_call_adjust(unsigned long addr)
> +{
> +	/*
> +	 * addr is the address of the mcount call instruction.
> +	 * recordmcount does the necessary offset calculation.
> +	 */
> +	return addr;
> +}

You could just as easily implement this as a dummy macro, but I guess it
doesn't matter either way.

>  #endif /* __ASSEMBLY__ */
>  
>  #endif /* __ASM_FTRACE_H */
> diff --git a/arch/arm64/kernel/entry-ftrace.S b/arch/arm64/kernel/entry-ftrace.S
> index 0ac31c8..c0fbe10 100644
> --- a/arch/arm64/kernel/entry-ftrace.S
> +++ b/arch/arm64/kernel/entry-ftrace.S
> @@ -86,6 +86,7 @@
>  	add	\reg, \reg, #8
>  	.endm
>  
> +#ifndef CONFIG_DYNAMIC_FTRACE
>  /*
>   * void _mcount(unsigned long return_address)
>   * @return_address: return address to instrumented function
> @@ -134,6 +135,48 @@ skip_ftrace_call:
>  #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
>  ENDPROC(_mcount)
>  
> +#else /* CONFIG_DYNAMIC_FTRACE */
> +/*
> + * _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,
> + * and later on, NOP to branch to ftrace_caller() when enabled or branch to
> + * NOP when disabled per-function base.
> + */
> +ENTRY(_mcount)
> +	ret
> +ENDPROC(_mcount)

Judging by your comment then, this should never be called. Is that right? If
so, we could add a BUG-equivalent so we know if we missed an mcount during
patching.

> +/*
> + * void ftrace_caller(unsigned long return_address)
> + * @return_address: return address to instrumented function
> + *
> + * This function is a counterpart of _mcount() in 'static' ftrace, and
> + * makes calls to:
> + *     - tracer function to probe instrumented function's entry,
> + *     - ftrace_graph_caller to set up an exit hook
> + */
> +ENTRY(ftrace_caller)
> +	mcount_enter
> +
> +	mcount_get_pc0	x0		//     function's pc
> +	mcount_get_lr	x1		//     function's lr
> +
> +	.global ftrace_call
> +ftrace_call:				// tracer(pc, lr);
> +	nop				// This will be replaced with "bl xxx"
> +					// where xxx can be any kind of tracer.
> +
> +#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
> +
> +	mcount_exit
> +ENDPROC(ftrace_caller)
> +#endif /* CONFIG_DYNAMIC_FTRACE */
> +
>  ENTRY(ftrace_stub)
>  	ret
>  ENDPROC(ftrace_stub)
> diff --git a/arch/arm64/kernel/ftrace.c b/arch/arm64/kernel/ftrace.c
> index a559ab8..8c26476 100644
> --- a/arch/arm64/kernel/ftrace.c
> +++ b/arch/arm64/kernel/ftrace.c
> @@ -17,6 +17,89 @@
>  #include <asm/ftrace.h>
>  #include <asm/insn.h>
>  
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +/*
> + * Replace a single instruction, which may be a branch or NOP.
> + * If @validate == true, a replaced instruction is checked against 'old'.
> + */
> +static int ftrace_modify_code(unsigned long pc, unsigned int old,
> +			      unsigned int new, bool validate)
> +{
> +	unsigned int replaced;

u32 is a bit clearer for instructions.

> +	/*
> +	 * Note:
> +	 * Due to modules and __init, code can disappear and change,
> +	 * we need to protect against faulting as well as code changing.
> +	 * We do this by aarch64_insn_*() which use the probe_kernel_*().
> +	 *
> +	 * No lock is held here because all the modifications are run
> +	 * through stop_machine().
> +	 */
> +	if (validate) {
> +		if (aarch64_insn_read((void *)pc, &replaced))
> +			return -EFAULT;
> +
> +		if (replaced != old)
> +			return -EINVAL;
> +	}
> +	if (aarch64_insn_patch_text_nosync((void *)pc, new))
> +		return -EPERM;

I think you're better off propagating the errors here, rather than
overriding them with EFAULT/EINVAL/EPERM.

> +
> +	return 0;
> +}
> +
> +/*
> + * Replace tracer function in ftrace_caller()
> + */
> +int ftrace_update_ftrace_func(ftrace_func_t func)
> +{
> +	unsigned long pc;
> +	unsigned int new;
> +
> +	pc = (unsigned long)&ftrace_call;
> +	new = aarch64_insn_gen_branch_imm(pc, (unsigned long)func, true);
> +
> +	return ftrace_modify_code(pc, 0, new, false);
> +}
> +
> +/*
> + * Turn on the call to ftrace_caller() in instrumented function
> + */
> +int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
> +{
> +	unsigned long pc = rec->ip;
> +	unsigned int old, new;
> +
> +	old = aarch64_insn_gen_nop();
> +	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 int old, new;
> +
> +	old = aarch64_insn_gen_branch_imm(pc, addr, true);
> +	new = aarch64_insn_gen_nop();
> +
> +	return ftrace_modify_code(pc, old, new, true);
> +}
> +
> +int __init ftrace_dyn_arch_init(void *data)
> +{
> +	*(unsigned long *)data = 0;
> +
> +	return 0;
> +}
> +#endif /* CONFIG_DYNAMIC_FTRACE */
> +
>  #ifdef CONFIG_FUNCTION_GRAPH_TRACER
>  /*
>   * function_graph tracer expects ftrace_return_to_handler() to be called
> @@ -61,4 +144,34 @@ void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
>  		return;
>  	}
>  }
> +
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +/*
> + * Turn on/off the call to ftrace_graph_caller() in ftrace_caller()
> + * depending on @enable.
> + */
> +static int ftrace_modify_graph_caller(bool enable)
> +{
> +	unsigned long pc = (unsigned long)&ftrace_graph_call;
> +	unsigned int branch, nop, old, new;
> +
> +	branch = aarch64_insn_gen_branch_imm(pc,
> +			(unsigned long)ftrace_graph_caller, false);
> +	nop = aarch64_insn_gen_nop();
> +	old = enable ? nop : branch;
> +	new = enable ? branch : nop;
> +
> +	return ftrace_modify_code(pc, old, new, true);

You could rewrite this as:

	if (enable)
		return ftrace_modify_code(pc, nop, branch, true);
	else
		return ftrace_modify_code(pc, branch, nop, true);

which I find easier to read.

Will