From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751808Ab1GOOK3 (ORCPT ); Fri, 15 Jul 2011 10:10:29 -0400 Received: from cantor2.suse.de ([195.135.220.15]:34343 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750969Ab1GOOK1 (ORCPT ); Fri, 15 Jul 2011 10:10:27 -0400 Date: Fri, 15 Jul 2011 15:10:21 +0100 From: Mel Gorman To: Christoph Hellwig Cc: Andrew Morton , Linux-MM , Linux-Netdev , LKML , David Miller , Neil Brown , Peter Zijlstra Subject: Re: [PATCH 00/14] Swap-over-NBD without deadlocking v5 Message-ID: <20110715141021.GZ7529@suse.de> References: <1308575540-25219-1-git-send-email-mgorman@suse.de> <20110706165146.be7ab61b.akpm@linux-foundation.org> <20110707094737.GG15285@suse.de> <20110707125831.GA15412@infradead.org> MIME-Version: 1.0 Content-Type: text/plain; charset=iso-8859-15 Content-Disposition: inline In-Reply-To: <20110707125831.GA15412@infradead.org> 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 07, 2011 at 08:58:31AM -0400, Christoph Hellwig wrote: > On Thu, Jul 07, 2011 at 10:47:37AM +0100, Mel Gorman wrote: > > Additional complexity is required for swap-over-NFS but affects the > > core kernel far less than this series. I do not have a series prepared > > but from what's in a distro kernel, supporting NFS requires extending > > address_space_operations for swapfile activation/deactivation with > > some minor helpers and the bulk of the remaining complexity within > > NFS itself. > > The biggest addition for swap over NFS is to add proper support for > a filesystem interface to do I/O on random kernel pages instead of > the current nasty bmap hack the swapfile code is using. Splitting > that work from all the required VM infrastructure should make life > easier for everyone involved and allows merging it independeny as > both bits have other uses case as well. > The swap-over-nfs patches allows this possibility. There is a swapon interface added that could be used by the filesystem to ensure the underlying blocks are allocated and a swap_out/swap_in interface that takes a struct file, struct page and writeback_control. This would be an alternative to bmap being used to record the blocks backing each extent. Any objection to the swap-over-NBD stuff going ahead to get part of the complexity out of the way? -- Mel Gorman SUSE Labs