Message-ID: <CAM9d7cjNzWzq26+2F1xLV-+qrQ7WwQ-h_TY5_URqT+CymAUs+Q@mail.gmail.com>
Date:   Fri, 2 Dec 2022 15:36:24 -0800
From:   Namhyung Kim <namhyung@...nel.org>
To:     Ian Rogers <irogers@...gle.com>
Cc:     Arnaldo Carvalho de Melo <acme@...nel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        Mark Rutland <mark.rutland@....com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...nel.org>,
        Masami Hiramatsu <mhiramat@...nel.org>,
        Steven Rostedt <rostedt@...dmis.org>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Leo Yan <leo.yan@...aro.org>,
        Kan Liang <kan.liang@...ux.intel.com>,
        linux-kernel@...r.kernel.org, linux-perf-users@...r.kernel.org,
        Stephane Eranian <eranian@...gle.com>
Subject: Re: [PATCH v2 3/4] perf build: Use libtraceevent from the system

On Fri, Dec 2, 2022 at 11:45 AM Ian Rogers <irogers@...gle.com> wrote:
>
> On Fri, Dec 2, 2022 at 10:29 AM Arnaldo Carvalho de Melo
> <acme@...nel.org> wrote:
> >
> > Em Fri, Dec 02, 2022 at 10:08:04AM -0800, Namhyung Kim escreveu:
> > > On Wed, Nov 30, 2022 at 12:13 PM Ian Rogers <irogers@...gle.com> wrote:
> > > >
> > > > On Wed, Nov 30, 2022 at 11:05 AM Namhyung Kim <namhyung@...nel.org> wrote:
> > > > >
> > > > > On Tue, Nov 29, 2022 at 10:30 PM Ian Rogers <irogers@...gle.com> wrote:
> > > > > >
> > > > > > Remove the LIBTRACEEVENT_DYNAMIC and LIBTRACEFS_DYNAMIC. If
> > > > > > libtraceevent isn't installed or NO_LIBTRACEEVENT=1 is passed to the
> > > > > > build, don't compile in libtraceevent and libtracefs support. This
> > > > > > also disables CONFIG_TRACE that controls "perf
> > > > > > trace". CONFIG_TRACEEVENT is used to control enablement in
> > > > > > Build/Makefiles, HAVE_LIBTRACEEVENT is used in C code. Without
> > > > > > HAVE_LIBTRACEEVENT tracepoints are disabled and as such the commands
> > > > > > kmem, kwork, lock, sched and timechart are removed. The majority of
> > > > > > commands continue to work including "perf test".
> > > > >
> > > > > Maybe we can have a different approach.  I guess the trace data
> > > > > access is isolated then we can make dummy interfaces when there's
> > > > > no libtraceevent.  This way we don't need to touch every command
> > > > > and let it fail when it's asked.
> > > >
> > > > Sounds like a worthwhile refactor that can land on top of this change.
> > > >
> > > > > The motivation is that we should be able to run the sub-commands
> > > > > as much as possible.  In fact, we could run 'record' part only on the
> > > > > target machine and pass the data to the host for analysis with a
> > > > > full-fledged perf.  Also some commands like 'perf lock contention'
> > > > > can run with or without libtraceevent (using BPF only).
> > > >
> > > > The issue here is that perf lock contention will use evsel__new_tp and
> > > > internally that uses libtraceevent. As such it is removed without
> > > > HAVE_LIBTRACEEVENT. Without the evsel there's not much perf lock
> > > > contention can do, so rather than litter the code with
> > > > HAVE_LIBTRACEEVENT and for it to be broken, I made the choice just to
> > > > remove it from the no libtraceevent build for now.
> > >
> > > I don't think it needs evsel__new_tp() when BPF is used.
> > > The BPF program is attached to the raw tracepoint without
> > > perf_event and the result is written to the BPF map.
> > >
> > > >
> > > > I think it is worth pursuing these patches in the shape they are in so
> > > > that we can land the removal of tools/lib/traceevent and ensure the
> > > > migration away from an out-of-date version of that library.
> > >
> > > Yeah, I agree that we should remove the stale libtraceevent but
> > > I'd like to do it with minimal changes in the perf code base.
> > > Let me take a look at this.
> >
> > Ok, was going to take a look at this patchkit, will wait for you now.
> >
> > - Arnaldo
>
> So the problematic function is:
> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/evsel.c?h=perf/core#n482
> struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
>
> on success it will always initialize:
> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/evsel.c?h=perf/core#n499
> evsel->tp_format = trace_event__tp_format(sys, name);
>
> tp_format is a libtraceevent type so this patch makes the variable
> conditional on HAVE_LIBTRACEEVENT. I think this will remain true even
> with a new evsel__newtp function that isn't dependent on
> libtraceevent. I also think adding such things is best kept out of
> this patch which is trying to do the minimal (only 312 lines of
> addition) to have a build without libtraceevent.
>
> Given we expect builds to use libtraceevent and we should build off of
> this patch, I think we should look to land it for 6.2.

So I took a quick look and it seems not quite easy to isolate
libtraceevent accesses from perf trace, dynamic sort keys and
other places.  That would need some amount of work.

Now I think that we may apply this first and then work on enabling
more functionalities without libtraceevent gradually.

Also I think it'd be a major change for regular users if there's
no libtraceevent installed in their system.  Maybe we could add
a warning during the build?

Thanks,
Namhyung

Powered by blists - more mailing lists