Re: [PATCH v2 2/9] bpf/btf: tracing: Move finding func-proto API and getting func-param API to BTF

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

On Tue, 18 Jul 2023 16:12:55 -0700
Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:

> On Tue, Jul 18, 2023 at 4:03 PM Masami Hiramatsu <mhiramat@kernel.org> wrote:
> >
> > On Tue, 18 Jul 2023 10:11:01 -0700
> > Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
> >
> > > On Tue, Jul 18, 2023 at 6:56 AM Masami Hiramatsu <mhiramat@kernel.org> wrote:
> > > >
> > > > On Tue, 18 Jul 2023 19:44:31 +0900
> > > > Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:
> > > >
> > > > > > >  static const struct btf_param *find_btf_func_param(const char *funcname, s32 *nr,
> > > > > > >                                                    bool tracepoint)
> > > > > > >  {
> > > > > > > +       struct btf *btf = traceprobe_get_btf();
> > > > > >
> > > > > > I found that traceprobe_get_btf() only returns the vmlinux's btf. But
> > > > > > if the function is
> > > > > > defined in a kernel module, we should get the module's btf.
> > > > > >
> > > > >
> > > > > Good catch! That should be a separated fix (or improvement?)
> > > > > I think it's better to use btf_get() and btf_put(), and pass btf via
> > > > > traceprobe_parse_context.
> > > >
> > > > Hmm, it seems that there is no exposed API to get the module's btf.
> > > > Should I use btf_idr and btf_idr_lock directly to find the corresponding
> > > > btf? If there isn't yet, I will add it too.
> > >
> > > There is bpf_find_btf_id.
> > > Probably drop 'static' from it and use it.
> >
> > Thanks! BTW, that API seems to search BTF type info by name. If user want to
> > specify a module name, do we need a new API? (Or expand the function to parse
> > a module name in given name?)
> 
> We can allow users specify module name, but how would it help?
> Do you want to allow static func names ?
> But module name won't help. There can be many statics with the same name
> in the module. Currently pahole filters out all ambiguous things in BTF.
> Alan is working on better representation of statics in BTF.
> The work is still in progress.

Ah, got it. So currently we don't have to worry about that case.

> 
> For now I don't see a need for an api to specify module, since it's not
> a modifier that can be relied upon to disambiguate.
> Hence bpf_find_btf_id that transparently searches across all should be enough.
> At least it was enough for all of bpf use cases.

OK. After updating the BTF I will revisit here.

Thank you!

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