[RFC 2/3] arm64: refactor save_stack_trace()

[Steven Rostedt]

On Wed, 15 Jul 2015 09:20:42 +0900
AKASHI Takahiro <takahiro.akashi at linaro.org> wrote:

> On 07/14/2015 10:31 PM, Steven Rostedt wrote:
> > On Tue, 14 Jul 2015 21:47:10 +0900
> > Jungseok Lee <jungseoklee85 at gmail.com> wrote:
> >
> >> Is the below example an unexpected result?
> >> Entry 17 and 18 are ftrace_call and ftrace_ops_no_ops, respectively.
> 
> [snip]
> 
> > Note, function tracing does not disable interrupts. This looks to be
> > that an interrupt came in while __aloc_skb() was being traced.
> 
> Yeah, I think so, too. But if my insight is correct, it's not __alloc_skb()
> but one of functions that it calls. As I said in the commit log message
> of patch[1/3], the exact traced function will not be listed by
> save_stack_trace() because we don't create a stack frame at mcount().
> I think this is a flaw in the current implementation (on x86).
> 
> what do you think, Steve?
> 

mcount (well ftrace_call actually) does indeed create a stack frame for
itself *and* for what called it. At least on x86_64. See mcount_64.S.

With -pg -mfentry, it creates a stack frame. Without -mfentry, mcount
is called after the current function's frame is made so we don't need
to do much.

Here's what the -mfentry version does:

	pushq %rbp
	pushq 8*2(%rsp)  /* this is the parent pointer */
	pushq %rbp
	movq %rsp, %rbp
	pushq 8*3(%rsp)   /* Return address to ftrace_call */
	pushq %rbp
	movq %rsp, %rbp


Thus the stack looks like this:

                                         <---+
	|                              |     |
	+------------------------------+     |
	| return address for func      |     |
	| return address for func_call |     |
	| original %rbp                |     |
	+------------------------------+     |
	| return address for func      |     |
	| ptr to parent frame (%rbp)   | ----+
 	+------------------------------| <-----+
	| return address for func_call |       |
        | ptr to next frame (%rbp)     | ------+
	+------------------------------+ <---+
                                             |
                                             |
 Current %rbp points to func_call frame -----+

 The first box isn't used as a frame, but is used by ftrace_call to save
 information to restore everything properly.

Thus, __alloc_skb() is what is currently being traced.


-- Steve