Messages in this thread | | | Date | Tue, 30 Nov 2010 12:02:13 -0200 | From | Arnaldo Carvalho de Melo <> | Subject | Re: [PATCH 3/6] perf: Add timestamp to COMM and MMAP events |
| |
Em Tue, Nov 30, 2010 at 10:45:40AM +1100, Ian Munsie escreveu: > Excerpts from Peter Zijlstra's message of Mon Nov 29 22:54:50 +1100 2010: > > On Mon, 2010-11-29 at 17:06 +1100, Ian Munsie wrote: > > > This goes away if PERF_SAMPLE_CPU is added to the mix so that the > > > timestamps are offset from the first event of *their* CPU, > > > otherwise time-epoch can go negative, as it has done here.
> > How does that happen, I though the power7 sched_clock() was fully > > synchronized and monotonic across all cores?
> Oh the timestamps coming from the kernel are fine, I should have been > clearer in my email. This bug is purely restricted to the userspace code > that prints them out - in the perf_session__print_tstamp function. It > tries to print the timestamps out so that they start at 0 and if it > doesn't know what CPU the events came from it uses the very first event > it sees as epoch, but at the moment that is not necessarily going to be > the earliest event.
> Thinking about it overnight, I realise that this should also go away if > we sort them before printing things out at all, which is exactly what > I'm proposing to do in my patches.
I removed that in the current version of this patch series:
http://git.kernel.org/?p=linux/kernel/git/acme/linux-2.6.git;a=shortlog;h=refs/heads/perf/sample_id_all.v3
Now I'm just testing if a new perf.data file can be processed just fine by older tools (not possible in the sample_id_all.v1 series) and other combinations of older kernels with the new tool.
Probably I'll fold two patches and then submit for review.
- Arnaldo
| |