Re: Unsupported CONFIG_FPROBE and CONFIG_RETHOOK on ARM64

[Masami Hiramatsu (Google) <mhiramat@kernel.org>]

On Tue, 10 Sep 2024 11:23:29 -0700
Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:

> + arm ML and maintainers
> 
> On Wed, Sep 4, 2024 at 6:02 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
> >
> > Hey,
> >
> > I just recently realized that we are still missing multi-kprobe
> > support for ARM64, which depends on CONFIG_FPROBE. And CONFIG_FPROBE
> > seems to require CONFIG_HAVE_RETHOOK, which, it turns out, is not
> > implemented for ARM64.
> >
> > It took me a while to realize what's going on, as I roughly remembered
> > (and confirmed through lore search) that Masami's original rethook
> > patches had arm64-specific bits. Long story short:
> >
> > 0f8f8030038a Revert "arm64: rethook: Add arm64 rethook implementation"
> > 83acdce68949 arm64: rethook: Add arm64 rethook implementation
> >
> > The patch was landed and then reverted. I found some discussion online
> > and it seems like the plan was to land arch-specific bits shortly
> > after bpf-next PR.
> >
> > But it seems like that never happened. Why?
> >
> > I see s390x, RISC-V, loongarch (I'm not even mentioning x86-64) all
> > have CONFIG_HAVE_RETHOOK, even powerpc is getting one (see [0]), it
> > seems. How come ARM64 is the one left out?
> >
> > Can anyone please provide some context? And if that's just an
> > oversight, can we prioritize landing this for ARM64 ASAP?
> >
> >   [0] https://lore.kernel.org/bpf/20240830113131.7597-1-adubey@linux.ibm.com/
> >
> 
> Masami, Steven,
> 
> Does Linus have to be in CC to get any reply here? Come on, it's been
> almost a full week.

Sorry about bothering you, let me check that. But I think we eventually
need my fprobe-on-fgraph patch which allows all architecture uses ftrace_regs
instead of pt_regs for ftrace/fgraph users. That allows arm64 to implement
fprobe.

> 
> Maybe ARM64 folks have some context?... And hopefully desire to see
> this through so that ARM64 doesn't stick out as a lesser-supported
> platform as far as tracing goes compared to loongarch, s390x, and
> powerpc (which just landed rethook support, see [2]).

I think lesser-supported or not is not a matter, but they need to keep 
their architecutre healthy. Mark said that the current rethook
implementation is not acceptable because arm64 can not manually generate
pt_regs. So we need to use ftrace_regs for that.
So eventually, we need my fprobe series.

https://lore.kernel.org/bpf/164338038439.2429999.17564843625400931820.stgit@devnote2/

Thank you,

> 
> Note that there was already an implementation (see [1]), but for some
> reason it never made it.
> 
>   [1] https://lore.kernel.org/bpf/164338038439.2429999.17564843625400931820.stgit@devnote2/
>   [2] https://lore.kernel.org/bpf/172562357215.467568.2172858907419105155.b4-ty@ellerman.id.au/
> 
> >
> > -- Andrii


-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>