Re: [PATCH 1/2] perf tools: Add reference timestamp to perf header

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Mon, 2010-12-13 at 15:22 -0200, Arnaldo Carvalho de Melo wrote:
> > No, the last one already happened, you cannot postpone the last one,
> > there will always be another excuse.
> 
> Did you understand the use case? How to have multiple reference times
> when appending?

Yes, I understood it perfectly, I just detest the existence of these
pure user records, they need to die ASAP.

> "last one" means adding a multiplexor, PERF_RECORD_LAST_ONE if you will,
> and inside it we add new events if the need arises. Then never again we
> add a PERF_RECORD_ event in userspace. 

Nah, that sucks too, the whole concept of pure user-space events in that
stream sucks.

There's multiple things you can do, you could:

 - create a kernel event PERF_RECORD_NEW_BUFFER and stuff that into each
fresh buffer when its created, it could contain all kinds of 1-time
information, like:
	* this CLOCK_MONOTONIC offset (for what little good that does, since
our clock isn't strictly sync'ed to CLOCK_MONOTONIC so we can incur
arbitrary drift).
 	* architecture details, like 64/32 host info needed for the
PERF_SAMPLE_REGS stuff.

 - extend the existing header infrastructure to write a new header in
front of the new stream. The main header already has a data section that
points to the end of the stream, add a continuation header section that
points to a continuation-header used to appends and record the clock
offset data in there.

 - something else entirely.

Just stop using these stupid fake events and be somewhat creative.