Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932146Ab0ATC2b (ORCPT ); Tue, 19 Jan 2010 21:28:31 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1756030Ab0ATC2a (ORCPT ); Tue, 19 Jan 2010 21:28:30 -0500 Received: from ozlabs.org ([203.10.76.45]:35532 "EHLO ozlabs.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753341Ab0ATC23 (ORCPT ); Tue, 19 Jan 2010 21:28:29 -0500 Date: Wed, 20 Jan 2010 13:28:02 +1100 From: Anton Blanchard To: Ingo Molnar Cc: Paul Mackerras , Peter Zijlstra , linux-kernel@vger.kernel.org, Michael Neuling Subject: Re: [PATCH 2/2] perf tools: Cope with sparsely-numbered CPUs Message-ID: <20100120022802.GW12666@kryten> References: <20091215093437.GB18661@brick.ozlabs.ibm.com> <20091215120851.GA21780@elte.hu> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20091215120851.GA21780@elte.hu> User-Agent: Mutt/1.5.20 (2009-06-14) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 2495 Lines: 53 Hi Ingo, > > For system-wide monitoring, the perf tools currently ask how many CPUs are > > online, and then instantiate perf_events for CPUs 0 ... N-1. This doesn't > > work correctly when CPUs are numbered sparsely. For example, a four-core > > POWER6 in single-thread mode will have CPUs 0, 2, 4 and 6. The perf tools > > will try to open counters on CPUs 0, 1, 2 and 3, and either fail with an > > error message or silently ignore CPUs 4 and 6. > > > > This fixes the problem by making perf stat, perf record and perf top > > create counters for increasing CPU numbers until they have a counter > > for each online CPU. If the attempt to create a counter on a given > > CPU fails, we get an ENODEV error and we just move on to the next CPU. > > To avoid an infinite loop in case the number of online CPUs gets > > reduced while we are creating counters, we re-read the number of > > online CPUs when we fail to create a counter on some CPU. > > > > Reported-by: Michael Neuling > > Tested-by: Michael Neuling > > Cc: stable@kernel.org > > Signed-off-by: Paul Mackerras > > --- > > tools/perf/builtin-record.c | 36 ++++++++++++++++++++++++++++-------- > > tools/perf/builtin-stat.c | 15 +++++++++++++-- > > tools/perf/builtin-top.c | 27 +++++++++++++++++++-------- > > 3 files changed, 60 insertions(+), 18 deletions(-) > > nice fix! > > The linecount bloat is a bit worrying though. I'm wondering, since we have 3 > loops now (and possibly more upcoming), wouldnt it be a cleaner fix to have > some generic idiom of 'loop through all online cpus' somewhere in lib/*.c? > > This would work better in the long run than spreading all the sysconf calls > and special cases across all those callsites. (new tools will inevitably get > it wrong as well) > > As a practical matter we can commit your fix and do the cleanup/consolidation > as a separate patch, to not hold up your fix (and to help fix/bisect any > problems that would happen due to the consolidation) - as long as a > consolidation patch is forthcoming as well. It looks like this hasn't made it to mainline. Any chance we could get it in and look at a cleanup post 2.6.33? Thanks, Anton -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/