From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mx2.suse.de ([195.135.220.15]:59118 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932375AbcIVIyG (ORCPT ); Thu, 22 Sep 2016 04:54:06 -0400 Date: Thu, 22 Sep 2016 10:52:05 +0200 From: David Sterba To: Omar Sandoval Cc: "Austin S. Hemmelgarn" , Holger =?iso-8859-1?Q?Hoffst=E4tte?= , linux-btrfs@vger.kernel.org, quwenruo@cn.fujitsu.com, osandov@fb.com Subject: Re: Status of free-space-tree feature Message-ID: <20160922085205.GM16983@suse.cz> Reply-To: dsterba@suse.cz References: <20160921092406.GD16983@twin.jikos.cz> <57E2602C.7050703@applied-asynchrony.com> <85fb7ee0-7f6a-e8a0-46c6-dac12c3e69d2@gmail.com> <20160921203152.GA7694@vader.DHCP.thefacebook.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <20160921203152.GA7694@vader.DHCP.thefacebook.com> Sender: linux-btrfs-owner@vger.kernel.org List-ID: On Wed, Sep 21, 2016 at 01:31:52PM -0700, Omar Sandoval wrote: > > > I'm not sure I understand - can you explain why this is was so wrong? > > > Or Omar maybe? > > > > > > If btrfsck wants to correct something (write), it can simply always > > > and unconditionally invalidate the fst instead of trying to "repair" > > > it..and I think that's what happens at the moment (at least I think > > > it did for me recently). That seems like a safe and simple way. > > I know this is what it does with the regular FSC, but I'm not sure if it > > does so with the FST. If it doesn't, it probably should though. > > It doesn't. The free space cache is easy to invalidate because we can > just compare the generation number of the superblock to that of the > space cache, but as it exists now, the free space tree doesn't have > anything equivalent. That means that any modifications that btrfs-progs > made to a space_cache=v2 filesystem could potentially have caused > filesystem corruption :/ > > However, I talked this through with Chris, and he came up with a great > idea that will help us deal with both this issue and the endianness > issue [1] in one fell swoop. Basically, my objection to adding a compat > bit for the endianness bug was that it would unnecessarily affect the > vast majority of users on x86; forcing those users to recreate their > free space tree seemed silly. However, because of the btrfs-progs bug, > just to be safe, we might as well force everyone to recreate their free > space tree. > > The solution is to add a FREE_SPACE_TREE_VALID compat_ro bit. If the bit > isn't set, then we destroy and rebuild the free space tree. This covers > the case of big-endian kernels which created broken free space trees and > filesystems which could have possibly gone through btrfs-progs. > > This time we'll make sure not to make btrfs-progs think it can mount > FREE_SPACE_TREE_VALID filesystems read-write. We can even have > btrfs-progs check for that bit and clear it if it's mounting read-write. > The next time it gets mounted, the kernel will recreate the tree. It's > not pretty, but it'll work. Sounds like a good plan to me. The bit is a form of 'clear_cache' mount. We need to to a coordinated fix (kernel, progs), if the patches are ready soon, 4.9 is feasible target.