From: Amir Goldstein Subject: Re: Atomic file data replace API Date: Wed, 29 Dec 2010 11:20:17 +0200 Message-ID: References: <20101228025937.GE10149@thunk.org> <4D1A351B.5080604@redhat.com> <4D1A6663.8000307@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Cc: Ric Wheeler , "Ted Ts'o" , linux-fsdevel , linux-ext4@vger.kernel.org To: Olaf van der Spek Return-path: In-Reply-To: Sender: linux-fsdevel-owner@vger.kernel.org List-Id: linux-ext4.vger.kernel.org On Wed, Dec 29, 2010 at 12:58 AM, Olaf van der Spek wrote: > On Tue, Dec 28, 2010 at 11:36 PM, Ric Wheeler wrote: >> I think that various developers have answered this for you several times. > > Not really, unfortunately. Haven't seen a single link to code that > shows how to do it properly. > Temp file, fsync, rename is often mentioned but that skips the > preserving meta-data part and this part, which you also skipped: > One use case would be updating a file in a safe way when you have > write access to that file but not to anything else. > I think it is safe to say that the *only* option you have now is "temp file, fsync, rename". There is no "generic atomic file data replace API in Linux", though it is available via private ioctl for XFS and EXT4. You have started a bit of a storm with your previous thread, which doesn't help you much in moving forward in the current thread (previous thread is still more popular). I suggest that you humbly swallow you need to know WHY is it hard to implement non-durable atomic API and focus your attention on the very achievable data replace API. IMHO, implementing atomic swap_inodes_data operation shouldn't be difficult in most file systems (only implementation is simple, but testing and maintaining is not to be taken lightly). Something along the lines of: 1. aquire inodes write/truncate locks 2. start transaction 3. check/update quota limits 4. swap inodes i_data content 5. invalidate (or swap?) inodes page caches 6. mark inodes dirty 7. end transaction & release locks The real challenge would be to get everyone to agree on a common API and carve it in stone to the kernel's ABI (is it just swap_inodes_data? maybe also swap_inode_data_ranges? what about some options?) Also, as wacky and (some say) faulty the UNIX permissions models is, current systems have grown old with it, and even 'improving' the behavior of some applications, may wake up sleeping monsters, so it will not be done until enough people have pointed out security or usability issues, which could not be solved otherwise. In other words, until you find an *application* that wants to allow other user to modify the content of a file and preserve it's metadata and ownership. And unless that application cannot find a better way to achieve what it wanted to do in the first place, or unless that application already has a large install base which suffers from *a problem*, you will not have proven *the need*. Maybe preserving privileged extended attributes is *a need*. I wouldn't know myself. Amir.