Re: [PATCH 13/15] perf_counter: provide generic callchain bits

[Corey Ashford <cjashfor@linux.vnet.ibm.com>]

Peter Zijlstra wrote:
> On Tue, 2009-03-31 at 20:12 +1100, Paul Mackerras wrote:
>> Corey Ashford writes:
>>
>>> If this is the case, what is the exact meaning of PERF_COUNTER_SIMPLE 
>>> now? and PERF_COUNTER_GROUP?  One simplification would be that reading 
>>> the fd of a group leader would always read up all of the counters in the 
>>> group (along with their enabled and running times if those bits are 
>>> set), and that reading a single counter's fd would yield only that 
>>> counter's values and times (if enabled).  In effect, these two values, 
>>> PERF_COUNTER_GROUP and PERF_COUNTER_SIMPLE would no longer be necessary 
>>> at all.  Other bits would be used to determine what's in the mmap'd samples.
>> Now that events are only read through mmap, it's quite simple -
>> hw_event.read_format controls what read() gives you, and
>> hw_event.record_type controls what gets put into the pages that you
>> get with mmap.
>>
>> Currently read_format and record_type don't use the same set of bit
>> definitions.  Maybe they should?  I don't have a strong feeling about
>> it, but that might be a nice simplification.
> 
> Something like the below? 
> 
> That still has the record and read things separate, but as one unified
> overflow output.
> 
> I reduced record and read fields to 32bits, as the output type only has
> 32bits.
> 
> diff did a wonderful job making the patch unreadable :/
>

This looks great!  I like the idea of two bit vectors to describe what 
you get when you read and the format of the mmap'd data.  And I like the 
ability not to include the header on each record if the user space 
program knows the exact format of the sample records.

Paul is right, too, that having a read_format bit to determine whether 
reading the leader gives all of the counters or just the leader is a 
nice addition.  I can't think of when I'd use it, but it's a good 
capability.

- Corey