Re: [PATCH v3 0/8] bpf: fprobe: rethook: Use ftrace_regs instead of pt_regs

[Google]

On Thu, 17 Aug 2023 10:57:13 +0200
Florent Revest <revest@xxxxxxxxxxxx> wrote:

> On Sat, Aug 12, 2023 at 7:36 AM Masami Hiramatsu (Google)
> <mhiramat@xxxxxxxxxx> wrote:
> >
> > Hi,
> >
> > Here is the 3rd version of RFC series to use ftrace_regs instead of pt_regs.
> > The previous version is here;
> >
> > https://lore.kernel.org/all/169139090386.324433.6412259486776991296.stgit@devnote2/
> >
> > This also includes the generic part and minimum modifications of arch
> > dependent code. (e.g. not including rethook for arm64.)
> 
> I think that one aspect that's missing from the discussion (and maybe
> the series) so far is plans to actually save partial registers in the
> existing rethook trampolines.

Yes, it is arch-dependent part. We have to recheck what registers are
required for the rethook, and that is saved correctly on partial pt_regs
on each architecture.

> For now the series makes everything called by the rethook trampolines
> handle the possibility of having a sparse ftrace_regs but the rethook
> trampolines still save full ftrace_regs. I think that to rip the full
> benefits of this series, we should have the rethook trampolines save
> the equivalent ftrace_regs as the light "args" version of the ftrace
> trampoline.

I think this part depends on the architecture implementation, but yes.
Arm64 can *add* the rethook implementation but not enable KRETPROBE_ON_RETHOOK.
(do not remove kretprobe trampoline)
For this perpose, we need HAVE_RETHOOK_WITH_REGS;

 config KRETPROBE_ON_RETHOOK
         def_bool y
-        depends on HAVE_RETHOOK
+        depends on HAVE_RETHOOK_WITH_REGS
         depends on KRETPROBES
         select RETHOOK

So there will be pt_regs rethook and ftrace_regs (partial regs) rethook.

I would like to replace rethook's pt_regs with ftrace_regs too. However the
most problematic part is kretprobe. If CONFIG_KRETPROBE_ON_RETHOOK=y, the 
rethook must use pt_regs instead of ftrace_regs for API compatibility.
But it makes hard to integrate the rethook and function-graph trace return
hook. (I will discuss this in LPC)

Thank you,

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