From mboxrd@z Thu Jan 1 00:00:00 1970 From: Al Viro Subject: Re: [PATCH] reiserfs: Expand i_mutex to enclose lookup_one_len Date: Sun, 3 May 2009 09:52:36 +0100 Message-ID: <20090503085236.GT8633@ZenIV.linux.org.uk> References: <49FB1F20.8040400@suse.com> Mime-Version: 1.0 Return-path: Content-Disposition: inline In-Reply-To: <49FB1F20.8040400@suse.com> Sender: reiserfs-devel-owner@vger.kernel.org List-ID: Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Jeff Mahoney Cc: Linux Kernel Mailing List , ReiserFS Mailing List , Andrew Morton , Al Viro , Alexander Beregalov , David On Fri, May 01, 2009 at 12:11:12PM -0400, Jeff Mahoney wrote: > 2.6.30-rc3 introduced some sanity checks in the VFS code to avoid NFS > bugs by ensuring that lookup_one_len is always called under i_mutex. > > This patch expands the i_mutex locking to enclose lookup_one_len. This was > always required, but not not enforced in the reiserfs code since it > does locking around the xattr interactions with the xattr_sem. > > This is obvious enough, but it survived an overnight 50 thread ACL test. It's not enough, unfortunately ;-/ It deals with the warning, but it leaves an actual hole in there. Look: what happens if we mount it r/o without that directory and then remount r/w? We get dentry for privroot, hash it (negative at that point), then do actual mkdir, unlock root and modify the ->d_compare() of root to reject lookups on that sucker. Too late - in the meanwhile lookups might very well come and find privroot in dcache. BTW, the way ->d_compare() is done in there is rather dumb - if (q1 == &priv_root->d_name) return -ENOENT; ... would do just as well. Why don't we do that lookup *once* (on ->get_sb(), before anything can come and race with us), and then just keep negative dentry if the directory hadn't been around? And set d_compare() for root immediately after that lookup... I've applied your patch as-is, and unless you have objections to the variant above I'll do that as incremental. Comments?