From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753498AbZGJJgR (ORCPT ); Fri, 10 Jul 2009 05:36:17 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752068AbZGJJgE (ORCPT ); Fri, 10 Jul 2009 05:36:04 -0400 Received: from brick.kernel.dk ([93.163.65.50]:59969 "EHLO kernel.dk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752021AbZGJJgD (ORCPT ); Fri, 10 Jul 2009 05:36:03 -0400 Date: Fri, 10 Jul 2009 11:36:01 +0200 From: Jens Axboe To: Li Zefan Cc: Linux Kernel , Alan.Brunelle@hp.com, Ingo Molnar , Steven Rostedt , Frederic Weisbecker Subject: Re: [PATCH] Fix blktrace unaligned memory access Message-ID: <20090710093601.GF23611@kernel.dk> References: <20090710075339.GX23611@kernel.dk> <4A5704EA.5000807@cn.fujitsu.com> <20090710091337.GD23611@kernel.dk> <4A570A34.5090002@cn.fujitsu.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <4A570A34.5090002@cn.fujitsu.com> Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Fri, Jul 10 2009, Li Zefan wrote: > Jens Axboe wrote: > > On Fri, Jul 10 2009, Li Zefan wrote: > >> Jens Axboe wrote: > >>> Hi, > >>> > >>> It seems that relay_reserve() (or the ring_buffer_event_data(), that one > >>> still needs some love) can return unaligned memory, there's no way > >>> around that when you have pdu lengths that aren't a nice size. This can > >>> cause unaligned access warnings on platforms that care about alignment. > >>> > >> Seems relay_reserve() does nothing for alignment..On the other hand, > >> ring_buffer_event_data() returns a ptr which is 32bit-aligned, but > >> this still means it can cause unaligned accesses on 64bits arch, while > >> I think it's fixable in ring buffer, it's certainly not an easy job. > > > > Right, it's a bit nasty... > > > > Lai Jiangshan noticed this issue long ago and had some ideas in mind > how to fix ring buffer, but never try it out for it will probably be > frustrating.. > > >>> This is an RFC, perhaps we can fix this in some other way. This one takes > >>> the simple approach, use an on-stack copy and memcpy() that to the > >>> destination. > >>> > >> or get_unaligned() ? > > > > put_unaligned(), you mean? The big question is then which is faster, using > > put_unaligned() or doing the memcpy() of the structure... > > > > Ah, I meant put_unaligned(). > > I think the patch you posted can be a workaround at least for now, and > can be improved by detecting HAVE_EFFICIENT_UNALIGNED_ACCESS We either do one or the other, I don't want to clutter the code with both methods implemented and some ifdef checking and deciding. I'll check the unaligned put here on x86/ppc/sparc and then decide which one is faster. -- Jens Axboe