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

[Google]

On Sun, 5 Nov 2023 18:33:01 -0500
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> On Mon, 6 Nov 2023 00:17:34 +0100
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > Changelog nor code made it clear this was partial anything. So this is
> > still the partial thing?
> > 
> > Can we then pretty clear clarify all that, and make it clear which regs
> > are in there? Because when I do 'vim -t ftrace_regs' it just gets me a
> > seemingly pointless wrapper struct, no elucidating comments nothingses.
> 
> I agree it should be better documented (like everything else). The
> ftrace_regs must have all the registers needed to produce a function's
> arguments. For x86_64, that would be:
> 
>   rdi, rsi, rdx, r8, r9, rsp
> 
> Basically anything that is needed to call mcount/fentry.

Oops, I found I missed to save rsp. let me update it.

Anyway, this will be defined clearly. ftrace_regs needs to be a partial
set of registers related to the (kernel) function call.

 - registers which is used for passing the function parameters in
   integer registers and stack pointer (for parameters on memory).

 - registers which is used for passing the return values.

 - call-frame-pointer register if exists.

So for x86-64,

 - rdi, rsi, rcx, rdx, r8, r9, and rsp
 - rax and rdx
 - rbp

(BTW, why orig_rax is cleared?)

> But yes, it's still partial registers but for archs that support
> FTRACE_WITH_REGS, it can also hold all pt_regs which can be retrieved
> by the arch_ftrace_get_regs(), which is why there's a pt_regs struct in
> the x86 version. But that's not the case for arm64, as
> arch_ftrace_get_regs() will always return NULL.

The major reason of the DYNAMIC_FTRACE_WITH_REGS is livepatch and
kprobe on ftrace (if kprobe puts probe on the ftrace address, it uses
ftrace instead of breakpoint).

Thank you,

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