From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756596AbbESPOL (ORCPT ); Tue, 19 May 2015 11:14:11 -0400 Received: from mail-pa0-f54.google.com ([209.85.220.54]:33748 "EHLO mail-pa0-f54.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756578AbbESPOI (ORCPT ); Tue, 19 May 2015 11:14:08 -0400 Date: Wed, 20 May 2015 00:12:45 +0900 From: Namhyung Kim To: Arnaldo Carvalho de Melo Cc: Ingo Molnar , Peter Zijlstra , Jiri Olsa , LKML , David Ahern , Adrian Hunter , Andi Kleen , Frederic Weisbecker , Stephane Eranian Subject: Re: [PATCH 35/40] perf record: Synthesize COMM event for a command line workload Message-ID: <20150519151245.GC29162@danjae.kornet> References: <1431909055-21442-1-git-send-email-namhyung@kernel.org> <1431909055-21442-36-git-send-email-namhyung@kernel.org> <20150518124535.GD15972@kernel.org> <20150519074643.GM21663@sejong> <20150519140220.GH13946@kernel.org> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline In-Reply-To: <20150519140220.GH13946@kernel.org> User-Agent: Mutt/1.5.23 (2014-03-12) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, May 19, 2015 at 11:02:20AM -0300, Arnaldo Carvalho de Melo wrote: > Em Tue, May 19, 2015 at 04:46:43PM +0900, Namhyung Kim escreveu: > > On Mon, May 18, 2015 at 09:45:35AM -0300, Arnaldo Carvalho de Melo wrote: > > > Em Mon, May 18, 2015 at 09:30:50AM +0900, Namhyung Kim escreveu: > > > > When perf creates a new child to profile, the events are enabled on > > > > exec(). And in this case, it doesn't synthesize any event for the > > > > child since they'll be generated during exec(). But there's an window > > > > between the enabling and the event generation. > > > > > > > > It used to be overcome since samples are only in kernel (so we always > > > > have the map) and the comm is overridden by a later COMM event. > > > > However it won't work anymore since those samples will go to a missing > > > > thread now but the COMM event will create a (current) thread. This > > > > leads to those early samples (like native_write_msr_safe) not having a > > > > comm but pid (like ':15328'). > > > > > > > So it needs to synthesize COMM event for the child explicitly before > > > > enabling so that it can have a correct comm. But at this time, the > > > > comm will be "perf" since it's not exec-ed yet. > > > > > > This looks reasonable, but I think it probably needs to be done > > > somewhere in perf_evlist__prepare_workload() or > > > perf_evlist__start_workload(), as this affects other tools as well, like > > > 'top', 'trace' and any other that may want to do this start-workload use > > > case. > > > > Hmm.. I need to look at this again as it only affects on processing > > indexed data files which used to have a separate missing threads tree. > > Humm, you're thinking about where you managed to reproduce the problem, > I am thinking outside indexing, etc, i.e. by definition we either enable > the event before we fork, so that we get the PERF_RECORD_FORK/COMM or we > synthesize it either from /proc or directly (preferred) if we decide to > do it after the fork/exec, right? But as I said before, later COMM event will override thread->comm to a proper string as long as it can find a matching thread. So I think it has no problem in the current code. In the old version of this patchset (v3), indexing made it impossible for COMM event to find a matching thread since it used to have a separate tree for threads that have sampled before any FORK/COMM event came. I think it doesn't apply to the current version anymore, I will check it tomorrow. Thanks, Namhyung > > - Arnaldo > > > That's the reason why I didn't put it in a generic place like you > > said. > > > > However I changed not to use the separate tree - the purpose of the > > tree was to reduce lock acquisition on thread searching but it already > > grabs a rwlock with thread refcounting change. > > > > Will check whether this is still needed.. > > > > Thanks, > > Namhyung > > > > > > > > > > I also wonder if we can't overcome this without using /proc, i.e. > > > actually moving the "start the workload" to just before the fork, so > > > that the kernel covers that as well. > > > > > > Or, alternatively, the thread can be created without having to look at > > > /proc at all, but by directly creating the struct thread, with the > > > correct COMM, pid, etc, that we know, since we forked it, etc.