From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1763116AbYEGRtk (ORCPT ); Wed, 7 May 2008 13:49:40 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1753772AbYEGRtU (ORCPT ); Wed, 7 May 2008 13:49:20 -0400 Received: from mx2.mail.elte.hu ([157.181.151.9]:33046 "EHLO mx2.mail.elte.hu" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752637AbYEGRtQ (ORCPT ); Wed, 7 May 2008 13:49:16 -0400 Date: Wed, 7 May 2008 19:49:00 +0200 From: Ingo Molnar To: Linus Torvalds Cc: Matthew Wilcox , Andrew Morton , "J. Bruce Fields" , "Zhang, Yanmin" , LKML , Alexander Viro , linux-fsdevel@vger.kernel.org Subject: Re: AIM7 40% regression with 2.6.26-rc1 Message-ID: <20080507174900.GB13591@elte.hu> References: <1210052904.3453.30.camel@ymzhang> <20080506114449.GC32591@elte.hu> <20080506120934.GH19219@parisc-linux.org> <20080506162332.GI19219@parisc-linux.org> <20080506102153.5484c6ac.akpm@linux-foundation.org> <20080507163811.GY19219@parisc-linux.org> <20080507172246.GA13262@elte.hu> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.17 (2007-11-01) X-ELTE-VirusStatus: clean X-ELTE-SpamScore: -1.5 X-ELTE-SpamLevel: X-ELTE-SpamCheck: no X-ELTE-SpamVersion: ELTE 2.0 X-ELTE-SpamCheck-Details: score=-1.5 required=5.9 tests=BAYES_00 autolearn=no SpamAssassin version=3.2.3 -1.5 BAYES_00 BODY: Bayesian spam probability is 0 to 1% [score: 0.0000] Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org * Linus Torvalds wrote: > > There's far more normal mutex fastpath use during an AIM7 run than > > any BKL use. So if it's due to any direct fastpath overhead and the > > resulting widening of the window for the real slowdown, we should > > see a severe slowdown on AIM7 with CONFIG_MUTEX_DEBUG=y. Agreed? > > Not agreed. > > The BKL is special because it is a *single* lock. ok, indeed my suggestion is wrong and this would not be a good comparison. another idea: my trial-baloon patch should test your theory too, because the generic down_trylock() is still the 'fat' version, it does: spin_lock_irqsave(&sem->lock, flags); count = sem->count - 1; if (likely(count >= 0)) sem->count = count; spin_unlock_irqrestore(&sem->lock, flags); if there is a noticeable performance difference between your trial-ballon patch and mine, then the micro-cost of the BKL very much matters to this workload. Agreed about that? but i'd be _hugely_ surprised about it. The tty code's BKL use should i think only happen when a task exits and releases the tty - and a task exit - even if this is a threaded test (which AIM7 can be - not sure which exact parameters Yanmin used) - the costs of thread creation and thread exit are just not in the same ballpark as any BKL micro-costs. Dunno, maybe i overlooked some high-freq BKL user. (but any such site would have shown up before) Even assuming a widening of the critical path and some catastrophic domino effect (that does show up as increased scheduling) i've never seen a 40% drop like this. this regression, to me, has "different scheduling behavior" written all over it - but that's just an impression. I'm not going to bet against you though ;-) Ingo