From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1764483AbXHNPn2 (ORCPT ); Tue, 14 Aug 2007 11:43:28 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1764206AbXHNPmr (ORCPT ); Tue, 14 Aug 2007 11:42:47 -0400 Received: from igw3.br.ibm.com ([32.104.18.26]:59953 "EHLO igw3.br.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1762516AbXHNPmp (ORCPT ); Tue, 14 Aug 2007 11:42:45 -0400 Subject: Re: SLUB doesn't work with kdump kernel on Cell From: Lucio Correia Reply-To: ljhc@br.ibm.com To: Christoph Lameter Cc: linux-kernel@vger.kernel.org, Arnd Bergmann In-Reply-To: References: <1186604151.3608.46.camel@localhost.localdomain> Content-Type: text/plain Organization: IBM Date: Tue, 14 Aug 2007 12:43:59 -0300 Message-Id: <1187106239.3298.2.camel@localhost.localdomain> Mime-Version: 1.0 X-Mailer: Evolution 2.10.3 (2.10.3-2.fc7) Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Wed, 2007-08-08 at 13:26 -0700, Christoph Lameter wrote: > On Wed, 8 Aug 2007, Lucio Correia wrote: > > > Hi Christoph, > > > > I found a problem with SLUB when trying to boot a kdump kernel on a Cell > > QS20 Blade running Fedora 7, kernel 2.6.22.5. If I use SLAB for the > > kdump kernel, everything works ok. The fact is that SLUB doesn't find a > > page frame for allocation in the current node, due to the flag > > GFP_THISNODE on a call to new_slab, and stops at a BUG_ON on line 1802 > > of slub.c. > > This is due to the node 1 having only one page. Could you just switch node > 1 off? > > SLAB boots because it falls back to node 0 for the control structures. So > it creates useless control structures for node 1. These are then never > used since any allocation attempt to node 1 falls back to node 0. Hi Christoph, Shouldn't SLUB falls back to other node also for the case it can't allocate memory? > > > I understand that this flag should not be removed, and that there is a > > better solution, but it demonstrates the problem. Could you give me some > > direction on the better way to solve this problem? > > Do not create a node that just has one page in it? > > Or make it truly empty? An empty node will cause GFP_THISNODE to fall back > and you then have the same useless control structure allocation as on > SLAB. -- Lucio Correia Software Engineer IBM LTC Brazil