Return-Path: linux-nfs-owner@vger.kernel.org Received: from rcsinet15.oracle.com ([148.87.113.117]:32980 "EHLO rcsinet15.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932085Ab2B1Ofz convert rfc822-to-8bit (ORCPT ); Tue, 28 Feb 2012 09:35:55 -0500 Subject: Re: "Using NFS over UDP on high-speed links such as Gigabit can cause silent data corruption." Mime-Version: 1.0 (Apple Message framework v1257) Content-Type: text/plain; charset=us-ascii From: Chuck Lever In-Reply-To: <20120228124646.GA2528@umich.edu> Date: Tue, 28 Feb 2012 09:35:40 -0500 Cc: Jeff Layton , Harshula , Steve Dickson , NeilBrown , linux-nfs@vger.kernel.org Message-Id: References: <1330406521.9157.16.camel@serendib> <20120228065218.7e110936@tlielax.poochiereds.net> <20120228124646.GA2528@umich.edu> To: Jim Rees Sender: linux-nfs-owner@vger.kernel.org List-ID: On Feb 28, 2012, at 7:46 AM, Jim Rees wrote: > Jeff Layton wrote: > > On Tue, 28 Feb 2012 16:22:01 +1100 > Harshula wrote: > >> Hi Steve, >> >> The following openSUSE nfs-utils patch, warn-nfs-udp.patch, is not >> included upstream: >> >> https://build.opensuse.org/package/view_file?file=warn-nfs-udp.patch&package=nfs-utils&project=openSUSE%3AFactory&rev=8e3e60c70e8270cd4afa036e13f6b2bb >> >> Please consider including it. >> >> Thanks, >> # >> > > I think that patch looks reasonable and clearly documenting the > problems with UDP is a wonderful thing. > > It may be best to send it formally to steved and the list as a real > [PATCH] with a real description and SoB line. > > This feels like the wrong place to document this, since it affects anything > that uses udp, not just nfs. NFS has a particular sensitivity to unreliable datagram transports, and that is a well-known problem. NetApp's retired TR-3183 and many Oracle meta documents mention the problems with NFS over UDP. Most other uses of UDP do not involve such large datagrams. My comment is that if the text in the TRANSPORT METHODS section in nfs(5) about UDP reassembly is not adequate it should be updated. I would rather see the meat of the proposed text merged into that section; otherwise we have two disparate sections discussing the same topic. That section is where this kind of discussion belongs. > It also seems like this should be solved in > the network layer with an adaptive frag time. But I'm not volunteering to > do that. As above, most other uses of UDP do not involve large packets. But I wonder if it is appropriate for us to suggest a change in the default setting. -- Chuck Lever chuck[dot]lever[at]oracle[dot]com