Re: [RESEND PATCH V3 3/8] perf/x86/intel: Support hardware TopDown metrics
From: Peter Zijlstra
Date: Thu Aug 29 2019 - 05:17:24 EST
On Wed, Aug 28, 2019 at 08:11:51PM -0700, Andi Kleen wrote:
> On Wed, Aug 28, 2019 at 06:28:57PM +0200, Peter Zijlstra wrote:
> > On Wed, Aug 28, 2019 at 09:17:54AM -0700, Andi Kleen wrote:
> > > > This really doesn't make sense to me; if you set FIXED_CTR3 := 0, you'll
> > > > never trigger the overflow there; this then seems to suggest the actual
> > >
> > > The 48bit counter might overflow in a few hours.
> > Sure; the point is? Kan said it should not be too big; a full 48bit wrap
> > around must be too big or nothing is.
> We expect users to avoid making it too big, but we cannot rule it out.
> Actually for the common perf stat for a long time case you can make it fairly
> big because the percentages still work out over the complete run time.
> The problem with letting it accumulate too much is mainly if you
> want to use a continuously running metrics counter to time smaller
> regions by taking deltas, for example using RDPMC.
> In this case the small regions would be too inaccurate after some time.
> But to make the first case work absolutely need to handle the overflow
> case. Otherwise the metrics would just randomly stop at some
But then you need -max_period, not 0. By using half the period, there is
no ambiguity on overflow.