linux-kernel - Re: [PATCH 4/4] perf python: Add counting.py as example for counting perf events

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAP-5=fVqRrO2pQx6he9N2R1A11RzVTFb-0KBvg4ydK4Eh7qSFw@mail.gmail.com>
Date: Thu, 1 May 2025 09:11:42 -0700
From: Ian Rogers <irogers@...gle.com>
To: Gautam Menghani <gautam@...ux.ibm.com>
Cc: peterz@...radead.org, mingo@...hat.com, acme@...nel.org, 
	namhyung@...nel.org, mark.rutland@....com, alexander.shishkin@...ux.intel.com, 
	jolsa@...nel.org, adrian.hunter@...el.com, kan.liang@...ux.intel.com, 
	linux-perf-users@...r.kernel.org, linux-kernel@...r.kernel.org, 
	maddy@...ux.ibm.com
Subject: Re: [PATCH 4/4] perf python: Add counting.py as example for counting
 perf events

On Thu, May 1, 2025 at 9:04 AM Ian Rogers <irogers@...gle.com> wrote:
>
> On Thu, May 1, 2025 at 2:37 AM Gautam Menghani <gautam@...ux.ibm.com> wrote:
> >
> > Add counting.py - a python version of counting.c to demonstrate
> > measuring and reading of counts for given perf events.
> >
> > Signed-off-by: Gautam Menghani <gautam@...ux.ibm.com>
> > ---
> >  tools/perf/python/counting.py | 41 +++++++++++++++++++++++++++++++++++
> >  1 file changed, 41 insertions(+)
> >  create mode 100755 tools/perf/python/counting.py
> >
> > diff --git a/tools/perf/python/counting.py b/tools/perf/python/counting.py
> > new file mode 100755
> > index 000000000000..0c58907bd8bf
> > --- /dev/null
> > +++ b/tools/perf/python/counting.py
> > @@ -0,0 +1,41 @@
> > +#!/usr/bin/env python3
> > +# SPDX-License-Identifier: GPL-2.0
> > +# -*- python -*-
> > +# -*- coding: utf-8 -*-
> > +
> > +import perf
> > +
> > +def main():
> > +        cpus = perf.cpu_map()
> > +        thread_map = perf.thread_map(-1)
> > +        evlist = perf.evlist(cpus, thread_map)
> > +
> > +        evsel1 = perf.evsel(type = perf.TYPE_SOFTWARE,
> > +                 config = perf.COUNT_SW_CPU_CLOCK,
> > +                 read_format = perf.FORMAT_TOTAL_TIME_ENABLED | perf.FORMAT_TOTAL_TIME_RUNNING,
> > +                 disabled=1)
> > +        evlist.add(evsel1)
> > +
> > +        evsel2 = perf.evsel(type = perf.TYPE_SOFTWARE,
> > +                 config = perf.COUNT_SW_TASK_CLOCK,
> > +                 read_format = perf.FORMAT_TOTAL_TIME_ENABLED | perf.FORMAT_TOTAL_TIME_RUNNING,
> > +                 disabled=1)
> > +        evlist.add(evsel2)
>
> Nice example! Would this be better as:
> ```
>     cpus    = perf.cpu_map()
>     threads = perf.thread_map(-1)
>     evlist = perf.parse_events("cpu-clock,task-clock", cpus, threads)
> ```
> If you run `perf stat -vv -e 'cpu-clock,task-clock' .. ` you can
> double check the perf event attribute bits. For example in
> tracepoint.py we remove the SAMPLE_IP:
> https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/python/tracepoint.py?h=perf-tools-next#n27
>
> > +
> > +        evlist.open()
> > +        evlist.enable()
> > +
> > +        count = 100000
> > +        while count > 0:
> > +            count -= 1
> > +
> > +        evlist.disable()
> > +        evsel = evlist.next(None)
> > +        while evsel != None:
> > +            counts = evsel.read(0, 0)
>
> Rather than just reading on the first CPU and thread, perhaps change
> to iterate over the cpus and threads? Something like:
> ```
>     for evsel in evlist:
>         for cpu in cpus:
>             for thread in threads:
>                 counts = evsel.read(cpu, thread)
>                 print(f"For {evsel} read val={counts.val}
> enable={counts.ena} run ={counts.run}")
> ```

Oh note, it is easy to confuse a thread or a CPU with the index into
the thread_map or cpu_map. For CPUs the two values of CPU and index
are often the same making it very easy to write code that works only
on CPU events (not events on things like memory controllers that
typically list a CPU per-socket). We added a struct perf_cpu to try to
avoid this kind of issue in the C code but in the python code
everything is back to being ints :-(

Thanks,
Ian

> Thanks,
> Ian
>
> > +            print(counts.val, counts.ena, counts.run)
> > +            evsel = evlist.next(evsel)
> > +        evlist.close()
> > +
> > +if __name__ == '__main__':
> > +    main()
> > --
> > 2.49.0
> >