Re: [PATCH] Make Function Duration Tracer work with __gnu_mcount_nc

[Tim Bird <tim.bird@am.sony.com>]

On 03/19/2010 02:43 PM, Ashwin Chaugule wrote:
> Is there a better way to fix this in the original ftrace_graph_caller ?
> The lr isn't pushed on the stack before "bl mcount".
> 
> --
> 
> From: Ashwin Chaugule <ashwinc@quicinc.com>
> 
> Newer compilers use the __gnu_mcount_nc stub in every function
> prologue. The lr of the instrumented function is pushed on the
> stack before branching to __gnu_mcount_nc. Pop it before returning
> back to instrumented function so that it can return to its parent
> from its epilogue.
> 
> 
> Signed-off-by: Ashwin Chaugule <ashwinc@quicinc.com>
> ---
>  arch/arm/kernel/entry-common.S |   20 ++++++++++++++++++++
>  1 files changed, 20 insertions(+), 0 deletions(-)
> 
> diff --git a/arch/arm/kernel/entry-common.S b/arch/arm/kernel/entry-common.S
> index b8a9e47..30feac3 100644
> --- a/arch/arm/kernel/entry-common.S
> +++ b/arch/arm/kernel/entry-common.S
> @@ -144,6 +144,14 @@ ENTRY(__gnu_mcount_nc)
>      adr r0, ftrace_stub
>      cmp r0, r2
>      bne gnu_trace
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +    ldr r1, =ftrace_graph_return
> +    ldr r2, [r1]
> +    cmp r0, r2        @ if *ftrace_graph_return != ftrace_stub
> +    bne new_ftrace_graph_caller
> +#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
> +
>      ldmia sp!, {r0-r3, ip, lr}
>      bx ip
>  
> @@ -155,6 +163,18 @@ gnu_trace:
>      mov pc, r2
>      ldmia sp!, {r0-r3, ip, lr}
>      bx ip
> +
> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> +ENTRY(new_ftrace_graph_caller)        @ for use with __gnu_mcount_nc
> +    sub r0, fp, #4                    @ &lr of instrumented routine 
> (&parent)
> +    mov r1, lr                        @ instrumented routine (func)
> +    sub r1, r1, #MCOUNT_INSN_SIZE
> +    mov r2, fp                        @ frame pointer
> +    bl prepare_ftrace_return
> +    ldmia sp!, {r0-r3, ip, lr}        @ need to pop lr, pushed before
> +    mov pc, ip                        @ bl __gnu_mcount_nc
> +
> +#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
>     
>  ENTRY(mcount)
>      stmdb sp!, {r0-r3, lr}

Wow, interesting timing.  I just finished yesterday my own
version of this patch (for 2.6.29) - it looks amazingingly similar.

What's up with putting fp in r2 - is that required in the latest
prepare_ftrace_return()?
 -- Tim

------------- patch follows ----------------
Add support to __gnu_mcount_nc, used by newer (4.x?) Gnu compilers,
for ftrace function_graph support on ARM.

Signed-off-by: Tim Bird <tim.bird@am.sony.com>
---
 arch/arm/kernel/entry-common.S |   13 +++++++++++++
 1 file changed, 13 insertions(+)

--- a/arch/arm/kernel/entry-common.S
+++ b/arch/arm/kernel/entry-common.S
@@ -140,6 +140,12 @@ ENTRY(__gnu_mcount_nc)
 	adr r0, ftrace_stub
 	cmp r0, r2
 	bne gnu_trace
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+	ldr r1, =ftrace_graph_return
+	ldr r2, [r1]
+	cmp r0, r2		@ if *ftrace_graph_return != ftrace_stub
+	bne gnu_ftrace_graph_caller
+#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
 	ldmia sp!, {r0-r3, ip, lr}
 	bx ip

@@ -196,6 +202,13 @@ return_to_handler:
 	ldmia sp!, {r0-r3}
 	mov pc, lr

+ENTRY(gnu_ftrace_graph_caller)
+	sub r0, fp, #4			@ &lr of instrumented routine (&parent)
+	sub r1, lr, #MCOUNT_INSN_SIZE	@ instrumented routine (func)
+	bl prepare_ftrace_return
+	ldmia sp!, {r0-r3, ip, lr}
+	bx ip
+
 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */

 #endif /* CONFIG_DYNAMIC_FTRACE */