Re: [PATCH 2/4] perf, tools: Implement duration_time as a proper event

From: Jiri Olsa
Date: Wed Mar 27 2019 - 04:58:18 EST


On Tue, Mar 26, 2019 at 03:18:21PM -0700, Andi Kleen wrote:
> From: Andi Kleen <ak@xxxxxxxxxxxxxxx>
>
> The perf metric expression use duration_time internally to normalize
> events. Normal perf stat without -x also prints the duration time.
> But when using -x, the interval is not output anywhere, which
> is inconvenient for any post processing which often wants to
> normalize values to time.
>
> So implement duration_time as a proper perf event that
> can be specified explicitely with -e.
>
> The previous implementation of duration_time only worked
> for metric processing. This adds the concept of a tool
> event that is handled by the tool. On the kernel level
> it is still mapped to the dummy software event, but the
> values are not read anymore, but instead computed by the tool.
>
> Add proper plumbing to handle this in the event parser,
> and display it in perf stat. We don't want duration_time to be added up,
> so it's only printed for the first CPU.
>
> % perf stat -e duration_time,cycles true
>
> Performance counter stats for 'true':
>
> 555,476 ns duration_time
> 771,958 cycles
>
> 0.000555476 seconds time elapsed
>
> 0.000644000 seconds user
> 0.000000000 seconds sys
>
> Signed-off-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> ---
> tools/perf/builtin-stat.c | 28 ++++++++++++++++++-------
> tools/perf/util/evsel.h | 6 ++++++
> tools/perf/util/parse-events.c | 38 +++++++++++++++++++++++++++++-----
> tools/perf/util/parse-events.h | 4 ++++
> tools/perf/util/parse-events.l | 11 +++++++++-
> tools/perf/util/parse-events.y | 12 +++++++++++
> 6 files changed, 86 insertions(+), 13 deletions(-)
>
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index 49ee3c2033ec..7f9c4b7f5d69 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -244,11 +244,25 @@ perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
> process_synthesized_event, NULL);
> }
>
> +static int read_single_counter(struct perf_evsel *counter, int cpu,
> + int thread, struct timespec *rs)
> +{
> + if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
> + u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
> + struct perf_counts_values *count =
> + perf_counts(counter->counts, cpu, thread);
> + count->ena = count->run = val;
> + count->val = val;
> + return 0;
> + }
> + return perf_evsel__read_counter(counter, cpu, thread);
> +}

so now that we have time in a separate event, we could
get rid of the isolated update_stats(&walltime_nsecs_stats) calls
and move them to perf_stat__update_shadow_stats?

jirka