Re: [PATCH V2 08/13] perf tools: show kernel overhead

[Jiri Olsa]

On Fri, Dec 02, 2016 at 04:19:16PM -0500, kan.li...@intel.com wrote:

SNIP

>  --show-profiling-cost:
>       Show extra time cost during perf profiling
> +     Sort the extra time cost by CPU
> +     If CPU information is not available in perf_sample, using -1 instead.
> +     The time cost include:
> +     - SAM: sample overhead. For x86, it's the time cost in perf NMI handler.
> +     - MUX: multiplexing overhead. The time cost spends on rotate context.
> +     - SB: side-band events overhead. The time cost spends on iterating all
> +           events that need to receive side-band events.
>  
>  include::callchain-overhead-calculation.txt[]
>  
> diff --git a/tools/perf/util/event.h b/tools/perf/util/event.h
> index d96e215..dd4ec5c 100644
> --- a/tools/perf/util/event.h
> +++ b/tools/perf/util/event.h
> @@ -245,6 +245,12 @@ enum auxtrace_error_type {
>       PERF_AUXTRACE_ERROR_MAX
>  };
>  
> +struct total_overhead {
> +     struct perf_overhead_entry      total_sample[MAX_NR_CPUS + 1];
> +     struct perf_overhead_entry      total_mux[MAX_NR_CPUS + 1];
> +     struct perf_overhead_entry      total_sb[MAX_NR_CPUS + 1];
> +};

I think this should be either:
   - dynamically allocated (there's cpu count available in the session)
   - or made static within perf report (as in shadow stats) and counted
     in report's overhead tool callback

I also don't like that you do the process-related counts
in the xxx[MAX_NR_CPUS] member, we should have separated
struct perf_overhead_entry for that

jirla