Re: [PATCH] perf: collect multiplexing timing information in perf record

[David Ahern <dsahern@gmail.com>]

On 08/26/2011 12:02 PM, Stephane Eranian wrote:

>>> +static int
>>> +sample_read2u64(const u64 *array, u64 fmt)
>>> +{
>>> +     u64 nr = 1;
>>> +     int ret = 1; /* nr or value */
>>> +
>>> +     if (fmt & PERF_FORMAT_TOTAL_TIME_ENABLED)
>>> +             ret++;
>>> +
>>> +     if (fmt & PERF_FORMAT_TOTAL_TIME_RUNNING)
>>> +             ret++;
>>> +
>>> +     if (fmt & PERF_FORMAT_GROUP) {
>>> +             nr = *(u64 *)array;
>>> +             ret += nr;
>>> +     }
>>> +
>>> +     if (fmt & PERF_FORMAT_ID)
>>> +             ret += nr;
>>
>> Why not add
>> struct read_format {
>>        u64     value;
>>        u64     time_enabled;
>>        u64     time_running;
>>        u64     id;
>> };
>>
>> to perf_sample and save the data there?
>>
> I am not following you here.
> Are you talking about the kernel API or perf tool internals?
> 
> The only way to have timing saved in each sample is via
> PERF_SAMPLE_READ and a read_format which includes
> time_enabled + time_running.

I meant in that function you added above -- sample_read2u64. You are
moving the array pointer forward and essentially throwing away data that
was pushed to userspace. Why not add a struct and save that data for use
elsewhere in the builtin code? And the order of the data is a function
of whether PERF_FORMAT_GROUP is set (see perf_event.h file).

David