From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754780AbYIYPJ6 (ORCPT ); Thu, 25 Sep 2008 11:09:58 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752919AbYIYPJt (ORCPT ); Thu, 25 Sep 2008 11:09:49 -0400 Received: from smtp1.linux-foundation.org ([140.211.169.13]:39429 "EHLO smtp1.linux-foundation.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752823AbYIYPJs (ORCPT ); Thu, 25 Sep 2008 11:09:48 -0400 Date: Thu, 25 Sep 2008 08:05:46 -0700 (PDT) From: Linus Torvalds To: Peter Zijlstra cc: Martin Bligh , Martin Bligh , Steven Rostedt , linux-kernel@vger.kernel.org, Ingo Molnar , Thomas Gleixner , Andrew Morton , prasad@linux.vnet.ibm.com, Mathieu Desnoyers , "Frank Ch. Eigler" , David Wilder , hch@lst.de, Tom Zanussi , Steven Rostedt Subject: Re: [RFC PATCH 1/3] Unified trace buffer In-Reply-To: <1222354409.16700.215.camel@lappy.programming.kicks-ass.net> Message-ID: References: <20080924051056.650388887@goodmis.org> <33307c790809240847r31c8b683na15ff5488b60d25b@mail.gmail.com> <1222272686.16700.162.camel@lappy.programming.kicks-ass.net> <33307c790809240949i3026170i8f9ac1d67a0fcf00@mail.gmail.com> <33307c790809241403w236f2242y18ba44982d962287@mail.gmail.com> <1222339303.16700.197.camel@lappy.programming.kicks-ass.net> <8f3aa8d60809250733q70561e6agfa3b00da83773e9f@mail.gmail.com> <1222354409.16700.215.camel@lappy.programming.kicks-ass.net> User-Agent: Alpine 1.10 (LFD 962 2008-03-14) MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu, 25 Sep 2008, Peter Zijlstra wrote: > > Right - if you use raw tsc you're dependent on clock speed, if we'd > normalize that on ns instead you'd need at least: [...] Please don't normalize to ns. It's really quite hard, and it's rather _expensive_ on many CPU's. It involves a non-constant 64-bit divide, after all. I bet it can be optimized to be a multiply-by-inverse instead, but it would be a 128-bit (or maybe just 96-bit?) multiply, and the code would be nasty, and likely rather more expensive than the TSC reading itself. Sure, you have to normalize at _some_ point, and normalizing early might make some things simpler, but the main thing that would become easier is people messing about in the raw log buffer on their own directly, which would hopefully be something that we'd discourage _anyway_ (ie we should try to use helper functions for people to do things like "get the next event data", not only because the headers are going to be odd due to trying to pack things together, but because maybe we can more easily extend on them later that way when nobody accesses the headers by hand). And I don't think normalizing later is in any way more fundamentally hard. It just means that you do part of the expensive things after you have gathered the trace, rather than during. Linus