Re: [PATCH v22 00/20] tracing: fprobe: function_graph: Multi-function graph and fprobe on fgraph

[Jiri Olsa <olsajiri@gmail.com>]

On Fri, Jan 10, 2025 at 04:04:37PM -0800, Andrii Nakryiko wrote:
> On Thu, Jan 2, 2025 at 5:21 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Thu, Dec 26, 2024 at 02:11:16PM +0900, Masami Hiramatsu (Google) wrote:
> > > Hi,
> > >
> > > Here is the 22nd version of the series to re-implement the fprobe on
> > > function-graph tracer. The previous version is;
> > >
> > > https://lore.kernel.org/all/173379652547.973433.2311391879173461183.stgit@devnote2/
> > >
> > > This version is rebased on v6.13-rc4 with fixes on [3/20] for x86-32 and
> > > [5/20] for build error.
> >
> >
> > hi,
> > I ran the bench and I'm seeing native_sched_clock being used
> > again kretprobe_multi bench:
> >
> >      5.85%  bench            [kernel.kallsyms]                                        [k] native_sched_clock
> >             |
> >             ---native_sched_clock
> >                sched_clock
> >                |
> >                 --5.83%--trace_clock_local
> >                           ftrace_return_to_handler
> >                           return_to_handler
> >                           syscall
> >                           bpf_prog_test_run_opts
> 
> completely unrelated, Jiri, but we should stop using
> bpf_prog_test_run_opts() for benchmarking. It goes through FD
> refcounting, which is unnecessary tiny overhead, but more importantly
> it causes cache line bouncing between multiple CPUs (when doing
> multi-threaded benchmarks), which skews and limits results.

so you mean to switch directly to attaching/hitting kernel functions
or perhaps better have kernel module for that?

jirka

> 
> >                           trigger_producer_batch
> >                           start_thread
> >                           __GI___clone3
> >
> > I recall we tried to fix that before with [1] change, but that replaced
> > later with [2] changes
> >
> > When I remove the trace_clock_local call in __ftrace_return_to_handler
> > than the kretprobe-multi gets much faster (see last block below), so it
> > seems worth to make it optional
> >
> > there's some decrease in kprobe_multi benchmark compared to base numbers,
> > which I'm not sure yet why, but other than that it seems ok
> >
> > base:
> >         kprobe         :   12.873 ± 0.011M/s
> >         kprobe-multi   :   13.088 ± 0.052M/s
> >         kretprobe      :    6.339 ± 0.003M/s
> >         kretprobe-multi:    7.240 ± 0.002M/s
> >
> > fprobe_on_fgraph:
> >         kprobe         :   12.816 ± 0.002M/s
> >         kprobe-multi   :   12.126 ± 0.004M/s
> >         kretprobe      :    6.305 ± 0.018M/s
> >         kretprobe-multi:    7.740 ± 0.003M/s
> >
> > removed native_sched_clock call:
> >         kprobe         :   12.850 ± 0.006M/s
> >         kprobe-multi   :   12.115 ± 0.006M/s
> >         kretprobe      :    6.270 ± 0.017M/s
> >         kretprobe-multi:    9.190 ± 0.005M/s
> >
> >
> > happy new year ;-) thanks,
> >
> > jirka
> >
> >
> > [1] https://lore.kernel.org/bpf/172615389864.133222.14452329708227900626.stgit@devnote2/
> > [2] https://lore.kernel.org/all/20240914214805.779822616@goodmis.org/
> >
> 
> [...]