Re: [RFC PATCH 24/32] x86/ftrace: Enable HAVE_FUNCTION_GRAPH_FREGS

[Google]

On Mon, 6 Nov 2023 11:37:10 -0500
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> On Mon, 6 Nov 2023 10:05:49 +0900
> Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx> wrote:
> 
> > So for x86-64,
> > 
> >  - rdi, rsi, rcx, rdx, r8, r9, and rsp
> >  - rax and rdx
> >  - rbp
> > 
> > (BTW, why orig_rax is cleared?)
> 
> You mean from ftrace_caller?
> 
> That's a "hack" to determine if we need to call the direct trampoline or
> not. When you have both a direct trampoline and ftrace functions on the
> same function, it will call ftrace_ops_list_func() to iterate all the
> registered ftrace callbacks. The direct callback helper will set "orig_rax"
> to let the return of the ftrace trampoline call the direct callback.

Got it. So does ftrace_regs need a placeholder for direct trampoline?
(Or, can we use a register to pass it?)
I think we don't need to clear it for return_to_handler() but if
`ftrace_regs` spec requires it, it is better to do so.

Thank you,

> 
> Remember if a direct callback is by itself, the fentry will call that
> direct trampoline without going through the ftrace trampoline. This is used
> to tell the ftrace trampoline that it's attached to a direct caller and
> needs to call that and not return back to the function it is tracing.
> 
> See later down in that file we have:
> 
> 	/*
> 	 * If ORIG_RAX is anything but zero, make this a call to that.
> 	 * See arch_ftrace_set_direct_caller().
> 	 */
> 	testq	%rax, %rax
> 
> -- Steve


-- 
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>