From mboxrd@z Thu Jan 1 00:00:00 1970 From: Mimi Zohar Subject: Re: [PATCH v3 1/1] vfs: iversion truncate bug fix Date: Thu, 05 Jan 2012 14:49:54 -0500 Message-ID: <1325792995.2062.11.camel@falcor.watson.ibm.com> References: <1324560391.1964.8.camel@falcor> <20120104152801.d8f555ca.akpm@linux-foundation.org> <1325723630.13419.2.camel@falcor> <20120104164638.cc21fc2e.akpm@linux-foundation.org> <20120105020639.GA1161@kroah.com> <1325737033.13419.43.camel@falcor> <20120105045354.GE24466@dastard> <20120105181401.GA418@infradead.org> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: Dave Chinner , Greg KH , Andrew Morton , Dmitry Kasatkin , linux-fsdevel@vger.kernel.org, linux-security-module@vger.kernel.org, viro@zeniv.linux.org.uk, linux-kernel@vger.kernel.org To: Christoph Hellwig Return-path: In-Reply-To: <20120105181401.GA418@infradead.org> Sender: linux-security-module-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org On Thu, 2012-01-05 at 13:14 -0500, Christoph Hellwig wrote: > On Thu, Jan 05, 2012 at 03:53:54PM +1100, Dave Chinner wrote: > > That's seems like a rather unreliable way of detecting that a file > > has changed to me. I mean, only ext4 uses inode_inc_version() when > > it internally dirties an inode, and only ext4 sets the MS_I_VERSION > > so that inode_inc_version is only called for ext4 inodes when > > timestamps change. > > And even ext4 only does it when using the non-default "i_version" > mount option. > > > Hence just adding an increment to the truncate case doesn't seem to > > be sufficient to me. e.g. what about the equivalent case of having a > > hole punched in the file via fallocate? The file has definitely > > changed, but i_version won't change.... > > > > Perhaps bumping i_version in __mark_inode_dirty() might be the best > > way to capture all changes (other than timestamp updates) to any > > inode regardless of the filesystem type? > > It has the same problem as the timestamp updates doing that right now - > the fs can't do locking around it, and it can't return errors. That's > something affecting at least btrfs, xfs and IIRC ubifs, and probably > the cluster filesystems as well. The right answer is to replace the > timestmap updates which are the only places doing that with a method > as Josef had planned to do, and then we can include the i_version > updates in there. > > That assumes we figure out a coherent way to do it - except for the > conditional abuse in file_updates_times it's currently entirely under > fs control. So the best way to fix it would be to: > > - move the fs-private use into those filesystems actually using it. > Note that a lot less actually check for it rather than just updating > it based on some cargo cult, and most only do so for directories. > - figure a why what exact change count semantics NFS (and IMA) want > and find a way to implement them so that the fs can tell the callers > that they don't exist. > > Btw, does IMA care about these beeing persistent? By 'persistent' I assume you mean across boots. IMA (and IMA-appraisal) measure and appraise files the first time they're accessed/executed. So no, it does not need to be persistent. IMA/IMA-appraisal just need some way to detect file change in order to know whether the file needs to be re-measured/appraised on subsequent access. Mimi