[PATCHv4 5/5] arm64: Trace emulation of AArch32 legacy instructions

[Steven Rostedt]

On Wed, 12 Nov 2014 11:44:52 +0000
Punit Agrawal <punit.agrawal at arm.com> wrote:

> Introduce an event to trace the usage of emulated instructions. The
> trace event is intended to help identify and encourage the migration
> of legacy software using the emulation features.
> 
> Use this event to trace usage of swp and CP15 barrier emulation.
> 
> Cc: Steven Rostedt <rostedt at goodmis.org>
> Signed-off-by: Punit Agrawal <punit.agrawal at arm.com>
> ---
>  arch/arm64/kernel/Makefile                 |  1 +
>  arch/arm64/kernel/armv8_deprecated.c       | 19 ++++++++++++--
>  arch/arm64/kernel/trace-events-emulation.h | 40 ++++++++++++++++++++++++++++++
>  3 files changed, 58 insertions(+), 2 deletions(-)
>  create mode 100644 arch/arm64/kernel/trace-events-emulation.h
> 
> diff --git a/arch/arm64/kernel/Makefile b/arch/arm64/kernel/Makefile
> index 84e9e51..b36ebd0 100644
> --- a/arch/arm64/kernel/Makefile
> +++ b/arch/arm64/kernel/Makefile
> @@ -5,6 +5,7 @@
>  CPPFLAGS_vmlinux.lds	:= -DTEXT_OFFSET=$(TEXT_OFFSET)
>  AFLAGS_head.o		:= -DTEXT_OFFSET=$(TEXT_OFFSET)
>  CFLAGS_efi-stub.o 	:= -DTEXT_OFFSET=$(TEXT_OFFSET)
> +CFLAGS_armv8_deprecated.o := -I$(src)
>  
>  CFLAGS_REMOVE_ftrace.o = -pg
>  CFLAGS_REMOVE_insn.o = -pg
> diff --git a/arch/arm64/kernel/armv8_deprecated.c b/arch/arm64/kernel/armv8_deprecated.c
> index 248b222..62d0672 100644
> --- a/arch/arm64/kernel/armv8_deprecated.c
> +++ b/arch/arm64/kernel/armv8_deprecated.c
> @@ -15,6 +15,9 @@
>  #include <linux/slab.h>
>  #include <linux/sysctl.h>
>  
> +#define CREATE_TRACE_POINTS
> +#include "trace-events-emulation.h"
> +

Note, if possible, please put this include after all other includes.
Sometimes a header file may include a trace events file and the
CREATE_TRACE_POINTS define might cause issues.

I'm sure it's fine now, but we don't want to debug this if someone does
it in the future.


>  #include <asm/insn.h>
>  #include <asm/opcodes.h>
>  #include <asm/system_misc.h>
> @@ -203,6 +206,11 @@ static int swp_handler(struct pt_regs *regs, u32 instr)
>  		regs->user_regs.regs[destreg] = data;
>  
>  ret:
> +	if (type == TYPE_SWPB)
> +		trace_instruction_emulation("swpb", regs->pc);
> +	else
> +		trace_instruction_emulation("swp", regs->pc);
> +

I don't know how critical a path the above is, but if you want to micro
optimize this, you can make it so that the branch conditional only does
the test if the tracepoint is enabled.

Now, this makes a difference if jump labels are implemented in arm64 or
not.

	if (trace_instruction_emulation_enabled()) {
		if (type == TYPE_SWPB)
			...
		else
			...
	}

That trace_instruction_emulation_enabled() is a static key (jump label)
that can turn the if into a unlikely "nop", making the code within it
in the very slow path.


>  	pr_warn_ratelimited("\"%s\" (%ld) uses obsolete SWP{B} instruction at 0x%llx\n",
>  			current->comm, (unsigned long)current->pid, regs->pc);
>  
> @@ -260,10 +268,15 @@ static int cp15barrier_handler(struct pt_regs *regs, u32 instr)
>  		 * dmb - mcr p15, 0, Rt, c7, c10, 5
>  		 * dsb - mcr p15, 0, Rt, c7, c10, 4
>  		 */
> -		if (aarch32_insn_mcr_extract_opc2(instr) == 5)
> +		if (aarch32_insn_mcr_extract_opc2(instr) == 5) {
>  			dmb(sy);
> -		else
> +			trace_instruction_emulation(
> +				"mcr p15, 0, Rt, c7, c10, 5 ; dmb", regs->pc);
> +		} else {
>  			dsb(sy);
> +			trace_instruction_emulation(
> +				"mcr p15, 0, Rt, c7, c10, 4 ; dsb", regs->pc);
> +		}
>  		break;
>  	case 5:
>  		/*
> @@ -272,6 +285,8 @@ static int cp15barrier_handler(struct pt_regs *regs, u32 instr)
>  		 * Taking an exception or returning from one acts as an
>  		 * instruction barrier. So no explicit barrier needed here.
>  		 */
> +		trace_instruction_emulation(
> +			"mcr p15, 0, Rt, c7, c5, 4 ; isb", regs->pc);
>  		break;
>  	}
>  
> diff --git a/arch/arm64/kernel/trace-events-emulation.h b/arch/arm64/kernel/trace-events-emulation.h
> new file mode 100644
> index 0000000..2ba3882
> --- /dev/null
> +++ b/arch/arm64/kernel/trace-events-emulation.h
> @@ -0,0 +1,40 @@
> +#undef TRACE_SYSTEM
> +#define TRACE_SYSTEM emulation
> +
> +#if !defined(_TRACE_EMULATION_H) || defined(TRACE_HEADER_MULTI_READ)
> +#define _TRACE_EMULATION_H
> +
> +#include <linux/tracepoint.h>
> +
> +TRACE_EVENT(instruction_emulation,
> +
> +	TP_PROTO(const char *instr, u64 addr),
> +	TP_ARGS(instr, addr),
> +
> +	TP_STRUCT__entry(
> +		__array(char, comm, TASK_COMM_LEN)

Why do you save the comm? It gets saved by the tracer (most of the
time, but we got better at saving it).

> +		__field(pid_t, pid)

Same here. The pid is already saved by the tracing infrastructure (the
comm is stored in a table that matches pids with comms to which
explains my comment above).


-- Steve

> +		__string(instr, instr)
> +		__field(u64, addr)
> +	),
> +
> +	TP_fast_assign(
> +		memcpy(__entry->comm, current->comm, TASK_COMM_LEN);
> +		__entry->pid = current->pid;
> +		__assign_str(instr, instr);
> +		__entry->addr = addr;
> +	),
> +
> +	TP_printk("instr=\"%s\" comm=%s pid=%d addr=0x%llx", __get_str(instr),
> +		__entry->comm, __entry->pid, __entry->addr)
> +);
> +
> +#endif /* _TRACE_EMULATION_H */
> +
> +/* This part must be outside protection */
> +#undef TRACE_INCLUDE_PATH
> +#undef TRACE_INCLUDE_FILE
> +#define TRACE_INCLUDE_PATH .
> +
> +#define TRACE_INCLUDE_FILE trace-events-emulation
> +#include <trace/define_trace.h>