Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id ; Tue, 23 Apr 2002 10:45:27 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id ; Tue, 23 Apr 2002 10:45:26 -0400 Received: from leibniz.math.psu.edu ([146.186.130.2]:54947 "EHLO math.psu.edu") by vger.kernel.org with ESMTP id ; Tue, 23 Apr 2002 10:45:26 -0400 Date: Tue, 23 Apr 2002 10:45:22 -0400 (EDT) From: Alexander Viro To: Alvaro Figueroa cc: LKML Subject: Re: Adding snapshot capability to Linux In-Reply-To: <1019572222.800.6.camel@lucy> Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On 23 Apr 2002, Alvaro Figueroa wrote: > > Instead of changing VFS you can probably make a generic stackable FS module > > .....that can stack on top of the physical filesystems and happily take > > snapshots at "FS" level :) ! and you can use the FIST to create a basic > > stackable FS and then modify it to take care of snapshoting ! > > Since this solution doens't solve Lisbor's request of using it on smb > filesystems, well, you could as well save up all of the programmer > cycles and use EVMS. > > It has a pluggin for treating normal partitions as EVMS objets, so you > don't need to translate them or so; and with EVMS you can even use RW > snapshots. You _can't_ get consistent snapshots without cooperation from fs. LVM, EVMS, whatever. Only filesystem knows what IO needs to be pushed to make what we have on device consistent and what IO needs to be held back. Neither VFS nor device driver do not and can not have such knowledge - it depends both on fs layout and on implementation details. - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/