Re: [PATCH v8 02/14] perf: Add AUX area to ring buffer for raw data streams

[Andi Kleen <ak@linux.intel.com>]

On Mon, Nov 17, 2014 at 01:24:00PM -0800, Sukadev Bhattiprolu wrote:
> Alexander Shishkin [alexander.shishkin@linux.intel.com] wrote:
> | From: Peter Zijlstra <peterz@infradead.org>
> | 
> | This patch introduces "AUX space" in the perf mmap buffer, intended for
> | exporting high bandwidth data streams to userspace, such as instruction
> | flow traces.
> | 
> | AUX space is a ring buffer, defined by aux_{offset,size} fields in the
> | user_page structure, and read/write pointers aux_{head,tail}, which abide
> | by the same rules as data_* counterparts of the main perf buffer.
> 
> The "format" of this raw stream of data can differ across processors
> (and architecutres) correct ? i.e the perf tool processing of this
> aux data will also differ across processors ?

Yes.

The main caveat is that you need special handling code in perf.

> 
> Power8 processors support what we call 24x7 counters that collect info
> on a large number of events. The current 24x7 support in
> (arch/powerpc/perf/hv-24x7.c) currently uses reading one counter at a
> time using the pmu->read() interface.
> 
> We are looking for ways to export much larger number of counters through
> one pmu->read() or another call. I am trying to see if the hv-24x7 pmu
> could use this aux interface and then have/implement a helper in perf
> tool to extract the counter values.

It's probably only worth it if your amount of data is extremely
high and has to be to compressed (like in PT -- no choice, cannot do it
otherwise)

If your data amounts are more reasonable it will be a lot
simpler to stay with the current perf interfaces, even
though they have a bit higher overhead.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only