Re: [PATCH RFC bpf-next v3 4/6] bpf: report hot simulated callchains when 1M instructions limit is met

[Jiri Olsa <olsajiri@gmail.com>]

On Fri, May 29, 2026 at 11:44:39AM -0700, Eduard Zingerman wrote:
> On Fri, 2026-05-29 at 12:23 +0200, Jiri Olsa wrote:
> > On Wed, May 27, 2026 at 12:29:52AM -0700, Eduard Zingerman wrote:
> > 
> > SNIP
> > 
> > > +
> > >  static int do_check_insn(struct bpf_verifier_env *env, bool *do_print_state)
> > >  {
> > >  	int err;
> > > @@ -17381,6 +17527,7 @@ static int do_check(struct bpf_verifier_env *env)
> > >  		insn_aux = &env->insn_aux_data[env->insn_idx];
> > >  
> > >  		if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
> > > +			print_hotspots(env);
> > 
> > hi,
> > this is really cool feature, I've already used it to fix issues :)
> 
> Hi Jiri, thank you for the feedback!
> Anything you find lacking?

so in my case the first 2 stacktraces identified data copy deep in
the code, which was wrongly done with standard loop instead of bpf_iter,
so it led me directly to the problem

In another run I got the 'Most varying ...' output and it wasn't
that obvious for me how helpful that could be, but I did not check
on that in detail yet

> 
> After discussion with Alexei (who is somewhat skeptical of the whole idea),
> I plan to add:
> - samples for values residing in the diverging registers
> - walk back from iter_next call using jump history and
>   print 5-10 instructions modifying the register in question
>   (in a manner similar to backtracking done for mark chain precision).
> 
> > Would it be possible to make this configurable? like to be able
> > to get this output for program that passed and display more than
> > 3 hottest callchains?
> 
> For sure, there are several options for a way to toggle this print out:
> - unconditionally make it a part of v1 log;
> - add it as a bit in bpf_attr->log_level.
> 
> I'd say that v1 log is a better option here.

sounds good, I think it's negligible compared to the rest of the log

> 
> As for number of callchains printed, I'd like to avoid tweaking

I guess it's better to have it hardcoded, no need to configure that

> bpf_attr. So this leaves several options:
> - make it top 5-7-10, just hardcode a bigger number.
> - add a threshold value specifying how many times a callchain should
>   be visited to get printed. E.g. only print callchains visited more
>   than 1000 times, if such don't exist then put threshold to 100,
>   if such don't exist, print top 3.

+1 to just print top X stacks

thanks,
jirka