Re: [PATCH] Add function graph tracer support for ARM

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

Uwe � wrote:
>>>> --- a/arch/arm/kernel/entry-common.S
>>>> +++ b/arch/arm/kernel/entry-common.S
>>>> @@ -135,8 +135,16 @@ ENTRY(mcount)
>>>>  	adr r0, ftrace_stub
>>>>  	cmp r0, r2
>>>>  	bne trace
>>>> +
>>>> +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
>>>> +	ldr r1, =ftrace_graph_return
>>>> +	ldr r2, [r1]
>>>> +	cmp r0, r2		@ if *ftrace_graph_return != ftrace_stub
>>>> +	bne ftrace_graph_caller
>>>> +#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
>>>> +
>>>>  	ldr lr, [fp, #-4]			@ restore lr
>>>> -	ldmia sp!, {r0-r3, pc}
>>>> +	ldmia sp!, {r0-r3, pc}			@ return doing nothing
>>> If ftrace_trace_function != ftrace_stub then ftrace_graph_caller isn't
>>> called.  Is this correct?
...
> Yes, the comments are as good as the comments for ftrace_trace_function.
> What I meant is that with your code using both ftrace_trace_function and
> ftrace_graph_caller doesn't work, because if ftrace_trace_function !=
> ftrace_stub then the check for ftrace_trace_function is simply skipped.

OK - I understand now.  I didn't think of this case, but you
are correct.  This has the same behavior as the x86 case.
Whether that's good or not is hard to say.  I wouldn't want
to add extra tests (since this is very much a hot path), but
maybe this should be noted somewhere.  I'll test to see
how this impacts the user-visible operation of the tracers.
(I assume this means you would have to reset the function
tracer, if it was active, before invoking a function graph
tracer.)

Alternatively, maybe the logic of installing the function graph
tracer should automatically reset the function tracer?

Steve,

Should both a function trace and a function graph tracer
be allowed to be active at the same time?  I would think
not, but maybe there's a use case I'm missing.

> And I just noticed something else:  There is a 2nd implementation of
> mcount in arch/arm/kernel/entry-common.S for CONFIG_DYNAMIC_FTRACE
> (which is currently not selectable for ARM).  Maybe add a note to this
> mcount that it needs fixing for graph tracing when it is revived?

Good catch.  I'll put something in.

Thanks!!
 -- Tim

=============================
Tim Bird
Architecture Group Chair, CE Linux Forum
Senior Staff Engineer, Sony Corporation of America
=============================