Re: [RFC/PATCHSET 0/5] perf ftrace: Implement function latency histogram (v1)

From: Namhyung Kim <namhyung@kernel.org>
To: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Stephane Eranian <eranian@google.com>,
	Jiri Olsa <jolsa@redhat.com>, Ingo Molnar <mingo@kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	LKML <linux-kernel@vger.kernel.org>,
	Andi Kleen <ak@linux.intel.com>, Ian Rogers <irogers@google.com>,
	Song Liu <songliubraving@fb.com>,
	Changbin Du <changbin.du@gmail.com>
Subject: Re: [RFC/PATCHSET 0/5] perf ftrace: Implement function latency histogram (v1)
Date: Fri, 3 Dec 2021 10:11:34 -0800	[thread overview]
Message-ID: <CAM9d7cgJXVofGqcbDbvL8feKHkJVxNNYiPUQ2zDGFzzbcFnk3w@mail.gmail.com> (raw)
In-Reply-To: <Yaoe7YWV0m4RYXnd@kernel.org>

On Fri, Dec 3, 2021 at 5:43 AM Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
> Em Wed, Dec 01, 2021 at 09:21:52AM -0800, Namhyung Kim escreveu:
> > On Wed, Dec 1, 2021 at 3:59 AM Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
> > > Em Tue, Nov 30, 2021 at 04:36:49PM -0800, Stephane Eranian escreveu:
> > > > I am not too fond of the flat option because as we had more bpf tools
> > > > like function latency, then we keep extending the list of commands
> > > > each with a small span which is different
> > > > from what we have right now.
>
> > > I think we should focus on the tool end result, not on how it is
> > > implemented, i.e. in this specific "function latency" tool ftrace is
> > > used with BPF, but we could perhaps have used some other mechanism.
>
> > Agreed, but I think function latency belongs to function tracing
> > conceptually.  So I added it as a subcommand in perf ftrace
> > not just because of the implementation.
>
> Fair enough, I think we can go as-is then, whoever finds this overly
> long can do:
>
> alias flat="perf ftrace latency"
>
> And go:
>
> flat -p `pidof firefox`
>
> etc.

Looks good!  Thanks.

>
> > > I think all these tools should have as much as possible a common set of
> > > options, like the targets (cpu, cgroup, pid, tid, etc) so that one can
> > > go from different views for those targets by just changing the name of
> > > the tool.
>
> > Currently, perf ftrace shares the target options with both subcommands.
> > Please see common_options in cmd_ftrace().
>
> Sure, but I was alluding to all perf tools, not just 'perf ftrace'
> subcommands, i.e. one can go from:
>
> perf ftrace trace --pid `pidof firefox`
>
> to:
>
> perf trace --pid `pidof firefox`
>
> to:
>
> perf stat --pid `pidof firefox`
>
> to:
>
> perf top --pid `pidof firefox`
>
> and get different views of that workload.

Right, that's what I want too.  Unfortunately, perf ftrace uses
the -t option for a different purpose.  But it follows other
conventions like -a for system-wide, -p for pid already.

>
> Have you thought about userspace function latencies? What tech would you
> use for that, since ftrace doesn't cover those?
>
> Would be nice that a single tool could be used to obtain userspace and
> kernel space function latencies, just like 'perf probe' can be used for
> kernel and userspace, choosing, behind the scenes, kprobes or uprobes.

Yep, that's the next step.  I think we can use uprobes
like 'perf probe' does.

Thanks,
Namhyung