Re: [RFC PATCH -next v2 0/4] arm64/ftrace: support dynamic trampoline

[Mark Rutland <mark.rutland@arm.com>]

On Wed, May 25, 2022 at 09:58:45AM -0400, Steven Rostedt wrote:
> On Wed, 25 May 2022 13:45:13 +0100
> Mark Rutland <mark.rutland@arm.com> wrote:
> 
> > ... the compiler places 3 NOPs *before* any BTI, and 2 NOPs *after* any BTI,
> > still recording the location of the first NOP. So in the two cases we get:
> > 
> > 		NOP		<--- recorded location
> > 		NOP
> > 		NOP
> > 	__func_without_bti:
> > 		NOP
> > 		NOP
> > 
> > 		NOP		<--- recorded location
> > 		NOP
> > 		NOP
> > 	__func_with_bti:
> > 		BTI
> > 		NOP
> > 		NOP
> 
> Are you saying that the above "recorded location" is what we have in
> mcount_loc section?

Yes; I'm saying that with this series, the compiler would record that into the
mcount_loc section.

Note that's not necessarily what goes into rec->ip, which we can adjust at
initialization time to be within the function. We'd need to record the
presence/absence of the BTI somewhere (I guess in dyn_arch_ftrace).

> If that's the case, we will need to modify it to point to something that
> kallsyms will recognize (ie. sym+0 or greater). Because that will cause
> set_ftrace_filter to fail as well.

Yup, understood. Like I mentioned it also wrecks the unwinder and would make it
really hard to implement RELIABLE_STACKTRACE.

Just to be clear, I don't think we should follow this specific approach. I just
wrote the examples to clarify what was being proposed.

Thanks,
Mark.