Re: [RFC PATCH v1 1/1] tracing/kprobe: Add multi-probe support for 'perf_kprobe' PMU

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

On Sat, 19 Aug 2023 08:22:50 -0700
Song Liu <song@kernel.org> wrote:

> On Fri, Aug 18, 2023 at 6:16 PM Masami Hiramatsu <mhiramat@kernel.org> wrote:
> >
> > On Fri, 18 Aug 2023 14:20:33 -0400
> > Steven Rostedt <rostedt@goodmis.org> wrote:
> >
> > > On Fri, 18 Aug 2023 20:13:43 +0200
> > > Francis Laniel <flaniel@linux.microsoft.com> wrote:
> > >
> > > > Hi.
> > > >
> > > > Le vendredi 18 août 2023, 17:41:41 CEST Steven Rostedt a écrit :
> > > > > On Fri, 18 Aug 2023 21:37:05 +0900
> > > > >
> > > > > Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:
> > > > > > That's why perf probe uses the offset from '_text'. Normal KASLR will just
> > > > > > moves all symbols. (Finer one will move all symbols randomely)
> > > > > > This should not need to access /proc/kallsyms but vmlinux or SystemMap.
> > > > >
> > > > > We could just pass in: "_text+offset" too.
> > > >
> > > > So, the idea would be to change the existing create_local_trace_kprobe() and
> > > > above functions to indicate the user's offset is to be used against _text and
> > > > not address?
> > >
> > > No, not to modify that function, but if you know the offset from _text (via
> > > the vmlinux), you can easily calculate it for that function.
> >
> > Note that the kprobe-event PMU interface itself allows you to specify
> > FUNC+OFFSET style;
> >
> > https://lore.kernel.org/lkml/20171206224518.3598254-5-songliubraving@fb.com/
> >
> > perf_event_attr::kprobe_func = "_text";
> > perf_event_attr::probe_offset = OFFSET;
> >
> > Then, it should be able to specify the correct one. Of course you can use
> > other unique symbols around the target symbol.
> 
> Trying to catch up with the thread.

Thanks for your reply :)

> 
> Besides the CAP_* issue, we can do this with
> 
> perf_event_attr::kprobe_func = NULL;
> perf_event_attr::kprobe_addr = address;

As I pointed, you don't need actual address, instead, you can specify the
probe point via "unique symbol" + offset. 

> 
> Then for the CAP_*, I think we should give CAP_PERFMON access to
> /proc/kallsyms. Would this work?

For the "unique symbol" + offset, you don't need the kallsyms, but need to
access the System.map or vmlinux image. In this case, we don't need to expand
the CAP_PERFMON capabilities.

Thank you,

> 
> Thanks,
> Song


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