Re: [PATCH 7/7] perf tools: add feature header record to pipe-mode

[David Carrillo-Cisneros]

On Thu, May 18, 2017 at 9:13 AM, Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> On Wed, May 17, 2017 at 09:16:02PM -0700, David Carrillo-Cisneros wrote:
>
> SNIP
>
>>  static const char *perf_ns__names[] = {
>> diff --git a/tools/perf/util/event.h b/tools/perf/util/event.h
>> index db2de6413518..d404f50260f8 100644
>> --- a/tools/perf/util/event.h
>> +++ b/tools/perf/util/event.h
>> @@ -244,6 +244,19 @@ enum perf_user_event_type { /* above any possible kernel type */
>>       PERF_RECORD_STAT_ROUND                  = 77,
>>       PERF_RECORD_EVENT_UPDATE                = 78,
>>       PERF_RECORD_TIME_CONV                   = 79,
>> +     PERF_RECORD_HEADER_HOSTNAME             = 80,
>> +     PERF_RECORD_HEADER_OSRELEASE            = 81,
>> +     PERF_RECORD_HEADER_VERSION              = 82,
>> +     PERF_RECORD_HEADER_ARCH                 = 83,
>> +     PERF_RECORD_HEADER_NRCPUS               = 84,
>> +     PERF_RECORD_HEADER_CPUDESC              = 85,
>> +     PERF_RECORD_HEADER_CPUID                = 86,
>> +     PERF_RECORD_HEADER_TOTAL_MEM            = 87,
>> +     PERF_RECORD_HEADER_CMDLINE              = 88,
>> +     PERF_RECORD_HEADER_EVENT_DESC           = 89,
>> +     PERF_RECORD_HEADER_CPU_TOPOLOGY         = 90,
>> +     PERF_RECORD_HEADER_NUMA_TOPOLOGY        = 91,
>> +     PERF_RECORD_HEADER_PMU_MAPPINGS         = 92,
>>       PERF_RECORD_HEADER_MAX
>>  };
>
> SNIP
>
>> +
>> +/*
>> + * we use a mapping table to go from record type to feature header
>> + * because we have no guarantee that new record types may not be added
>> + * after the feature header.
>> + */
>> +#define REC2FEAT(a)  [PERF_RECORD_HEADER_##a] = HEADER_##a
>> +
>> +static const int rec2feat[PERF_RECORD_HEADER_MAX] = {
>> +     REC2FEAT(HOSTNAME),
>> +     REC2FEAT(OSRELEASE),
>> +     REC2FEAT(VERSION),
>> +     REC2FEAT(ARCH),
>> +     REC2FEAT(NRCPUS),
>> +     REC2FEAT(CPUDESC),
>> +     REC2FEAT(CPUID),
>> +     REC2FEAT(TOTAL_MEM),
>> +     REC2FEAT(EVENT_DESC),
>> +     REC2FEAT(CMDLINE),
>> +     REC2FEAT(CPU_TOPOLOGY),
>> +     REC2FEAT(NUMA_TOPOLOGY),
>> +     REC2FEAT(PMU_MAPPINGS),
>>  };
>
> hum, how about instead of adding all those new events we add
> just one: PERF_RECORD_FEATURE, which would carry the id of the
> feature, like:
>
> struct feature_event {
>         struct perf_event_header        header;
>         u64     id;
>         char    data[]; /* size bytes of raw data specific to the feature */
> };
>
>
> this way we could also omit the reverse event to feature map

I think that'd work. I'll try that.