From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from out01.mta.xmission.com ([166.70.13.231]:60480 "EHLO out01.mta.xmission.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753052AbaBOBlD (ORCPT ); Fri, 14 Feb 2014 20:41:03 -0500 From: ebiederm@xmission.com (Eric W. Biederman) To: "J. Bruce Fields" Cc: Josef Bacik , linux-btrfs@vger.kernel.org References: <1392403428-4593-1-git-send-email-jbacik@fb.com> <20140214220525.GB3506@fieldses.org> Date: Fri, 14 Feb 2014 17:40:55 -0800 In-Reply-To: <20140214220525.GB3506@fieldses.org> (J. Bruce Fields's message of "Fri, 14 Feb 2014 17:05:25 -0500") Message-ID: <8761ohnoew.fsf@xmission.com> MIME-Version: 1.0 Content-Type: text/plain Subject: Re: [PATCH] Btrfs: unset DCACHE_DISCONNECTED when mounting default subvol Sender: linux-btrfs-owner@vger.kernel.org List-ID: "J. Bruce Fields" writes: > On Fri, Feb 14, 2014 at 01:43:48PM -0500, Josef Bacik wrote: >> A user was running into errors from an NFS export of a subvolume that had a >> default subvol set. When we mount a default subvol we will use d_obtain_alias() >> to find an existing dentry for the subvolume in the case that the root subvol >> has already been mounted, or a dummy one is allocated in the case that the root >> subvol has not already been mounted. This allows us to connect the dentry later >> on if we wander into the path. However if we don't ever wander into the path we >> will keep DCACHE_DISCONNECTED set for a long time, which angers NFS. It doesn't >> appear to cause any problems but it is annoying nonetheless, so simply unset >> DCACHE_DISCONNECTED in the get_default_root case and switch btrfs_lookup() to >> use d_materialise_unique() instead which will make everything play nicely >> together and reconnect stuff if we wander into the defaul subvol path from a >> different way. With this patch I'm no longer getting the NFS errors when >> exporting a volume that has been mounted with a default subvol set. Thanks, > > Looks obviously correct, but based on a quick grep, there are four > d_obtain_alias callers outside export methods: > > - btrfs/super.c:get_default_root() > - fs/ceph/super.c:open_root_dentry() > - fs/nfs/getroot.c:nfs_get_root() > - fs/nilfs2/super.c:nilfs_get_root_dentry() > > It'd be nice to give them a common d_obtain_alias variant instead of > making them all clear this by hand. I am in favor of one small fix at a time, so that progress is made and fixing something just for btrfs seems reasonable for the short term. > Of those nilfs2 also uses d_splice_alias. I think that problem would > best be solved by fixing d_splice_alias not to require a > DCACHE_DISCONNECTED dentry; IS_ROOT() on its own should be fine. You mean by renaming d_splice_alias d_materialise_unique? Or is there a useful distinction you see that should be preserved between the two methods? Right now my inclination is that everyone should just use d_materialise_unique and we should kill d_splice_alias. And by everyone I mean all file systems that are either distributed (implementing d_revalidate) or exportable by knfsd. One of the interesting things that d_materialise_unique does is get the lazy rename case correct for a distributed filesystem. check_submounts_and_drop can drop a directory when it is found not to be accessible by that name, but later when we look it up d_materialise_uniuqe will resuscciate the existing dentry. Eric