Re: [RFC PATCHv3] perf tools: add event grouping capability to "perf stat"

[Stephane Eranian <eranian@google.com>]

On Thu, Nov 25, 2010 at 2:19 PM, Stephane Eranian <eranian@google.com> wrote:
> On Thu, Nov 25, 2010 at 7:32 AM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
>> On Wed, 2010-11-24 at 17:54 -0800, Corey Ashford wrote:
>>> Add the ability to create multiple event groups, each with their own leader
>>> using the existing "-e <event>[,<event> ...] [-e <event>[,<event>]]"
>>> syntax.  Each additional -e switch creates a new group, and each event
>>> listed within a -e switch is within that group.
>>>
>>> Changes since v1:
>>> - Because of a flub, v2 did not contain the changes I had intended to make,
>>> and instead, v2 had the same patch contents as v1.
>>> - When perf stat is not supplied any events on the command line, put
>>> each default event in its own group.
>>
>> I like this, but could you also extend this to perf-record? its a bit
>> odd to diverge between the two.
>>
>> Using Stephane's latest syntax changes you could actually do something
>> like:
>>
>> perf record -e task-clock:freq=1000,cycles:period=0
>>
>> Which would create a group with 1 sampling counter and a counting
>> counter (at which point we should probably start flipping
>> PERF_SAMPLE_READ).
>>
>
> I think using PERF_SAMPLE_READ may expose a problem in the
> perf.data format. To correctly parse a sample created with SAMPLE_READ,
> you need to know the attr.read_format. But for that you need to know the
> event which caused the sample, but for that you need the SAMPLE_ID,
> and you don't know if it's there or not. In other words, there is a chicken
> and egg problem.
>
> I think the issue is that PERF_RECORD_SAMPLE is missing a mandatory
> piece of information: overflow event ID. This must a mandatory field, not
> optional as it is today. It is okay when you have only one group, but we'd
> like to go beyond that.
>
Second thought on this, I think the problem is in the kernel and not so much
in the perf.data file. Kernel must provide enough information to correlate
samples to events. It must do in a way that is not optional. Otherwise, as
soon as you have multiple groups, you won't be able to parse SAMPLE_READ.