RE: [patch, resend] x86, perf_counter, bts: optimize BTS overflow handling

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

On Tue, 2009-09-15 at 12:35 +0100, Metzger, Markus T wrote:
> >-----Original Message-----
> >From: Peter Zijlstra [mailto:a.p.zijlstra@chello.nl]
> >Sent: Tuesday, September 15, 2009 1:25 PM
> >To: Metzger, Markus T
> >Cc: mingo@elte.hu; tglx@linutronix.de; hpa@zytor.com; markus.t.metzger@gmail.com; linux-
> >kernel@vger.kernel.org
> >Subject: RE: [patch, resend] x86, perf_counter, bts: optimize BTS overflow handling
> >
> >On Tue, 2009-09-15 at 12:18 +0100, Metzger, Markus T wrote:
> >
> >> >> Restructure perf_counter sampling into sample creation and sample
> >> >> output.
> >> >> Prepare a single reference sample for BTS sampling and update the from
> >> >> and to address fields when draining the BTS buffer.
> >> >> Drain the entire BTS buffer between a single perf_output_begin() /
> >> >> perf_output_end() pair.
> >> >
> >> >Generally looks very nice, one thing though, why did you take regs out
> >> >of perf_sample_data, now you get to pass around one extra param..
> >>
> >>
> >> That is to make perf_sample_data describe the sample directly.
> >>
> >> The regs are used to create the sample; they're not needed once
> >> the sample has been created.
> >
> >True, but I'd have been lazy and not added that extra parameter all over
> >the place, but sure ;-)
> >
> >This is now fast enough to BTS trace the kernel too?
> 
> It is. At least on my box.
> 
> Once that patch has been accepted, I will ask Ingo to drop
> 1653192f510bd8114b7b133d7289e6e5c3e95046.

Patch looks good,

Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>

You could also look at throttling the BTS interrupt when you notice
you're doing them back-to-back..