From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752903Ab1GUNDo (ORCPT ); Thu, 21 Jul 2011 09:03:44 -0400 Received: from mail-wy0-f174.google.com ([74.125.82.174]:53596 "EHLO mail-wy0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751888Ab1GUNDn (ORCPT ); Thu, 21 Jul 2011 09:03:43 -0400 Date: Thu, 21 Jul 2011 15:03:38 +0200 From: Frederic Weisbecker To: Ingo Molnar Cc: "H. Peter Anvin" , LKML , Peter Zijlstra , Jason Wessel , Thomas Gleixner , Prasad , Will Deacon , Paul Mundt Subject: Re: [PATCH 3/6] x86: Allow the user not to build hw_breakpoints Message-ID: <20110721130334.GA16081@somewhere> References: <1310655837-328-1-git-send-email-fweisbec@gmail.com> <1310655837-328-4-git-send-email-fweisbec@gmail.com> <4E1F5EF3.6050601@zytor.com> <20110721072656.GF9216@elte.hu> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20110721072656.GF9216@elte.hu> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu, Jul 21, 2011 at 09:26:56AM +0200, Ingo Molnar wrote: > > * H. Peter Anvin wrote: > > > On 07/14/2011 08:03 AM, Frederic Weisbecker wrote: > > > So that hw_breakpoints and perf can be not built on > > > specific embedded systems. > > > > I want to emphasize I am very, very unhappy about this. It should > > be possible to not build perf while still have breakpoints > > available... breakpoints are way more important than perf. > > What we could indeed do is to separate out a 'core perf' portion that > is necessary for hw-breakpoints to work fine, thus allowing for > example the PMU drivers to be disabled. That would still require a big chunk of perf. > > Otherwise we have expressed hw breakpoint APIs via perf events and > that model is working well. Making hw-breakpoints a separate > subsystem again with isolated (and partly duplicated) infrastructure > would be a step back really. I actually don't think it's working well. What we have with the current design is the dependency to perf as a big midlayer that is apparently convenient but actually induce some nasty things. Just look how we need those ptrace_get_breakpoints() protection to deal with perf exit path implementation for example. Or the need for archs to translate arch ptrace breakpoint info into generic perf attrs. I think we had to try the current design just to see if that could plug nicely. But now that we have this for several releases, I can only conclude that we should revert back to the design Prasad proposed, consisting in having breakpoints a service used by perf but not the opposite. For ptrace, all it takes is a generic hook in the preempt notifiers to activate/deactivate breakpoints. I much prefer that to a big dependency on a perf core midlayer.