From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from rcsinet10.oracle.com ([148.87.113.121]:25987 "EHLO rcsinet10.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753836Ab0ICPC0 convert rfc822-to-8bit (ORCPT ); Fri, 3 Sep 2010 11:02:26 -0400 Subject: Re: [PATCH 1/1] mount: RDMA processing in the mount command is broken Content-Type: text/plain; charset=us-ascii From: Chuck Lever In-Reply-To: <4C801F67.50604@RedHat.com> Date: Fri, 3 Sep 2010 11:01:12 -0400 Cc: Linux NFS Mailing list Message-Id: References: <1283457651-12010-1-git-send-email-steved@redhat.com> <4C801F67.50604@RedHat.com> To: Steve Dickson Sender: linux-nfs-owner@vger.kernel.org List-ID: MIME-Version: 1.0 On Sep 2, 2010, at 6:04 PM, Steve Dickson wrote: > > > On 09/02/2010 04:35 PM, Chuck Lever wrote: >> >> On Sep 2, 2010, at 4:00 PM, Steve Dickson wrote: >> >>> The mounting code that process RMDA mounts is broken in a few places >>> >>> First with '-o proto=rdma' was broken because nfs_get_proto() >>> did not how to convert a netid of 'rdma' in to a AF_INET >>> address family. >>> >>> Secondly, '-o rdma' was broken because po_get() was being using >>> to detect the existence of 'rdma' in the options. With '-o rdma' >>> there is no value associated with that option so po_get() >>> was always return NULL. >> >> Looking at nfs(5), "rdma" as a stand-alone option isn't documented. Only "proto=rdma" appears to be mentioned. Thus I would argue that this is already behaving correctly. >> >> Are you proposing to add support for "-o rdma" ? If so what would that mean? > See Documentation/filesystems/nfs/nfs-rdma.txt > > I see no problem in handing this same why we handle 'udp' and 'tcp' I used nfs(5) as the specification for the text-based option parsing support in the mount.nfs command, since that's generally what administrators and packagers look to when trying to figure out how to use mount. I don't see any evidence in utils/mount/* that a stand-alone "rdma" option was ever supported. The kernel documentation could be wrong, or could be based on the kernel's mount option parser, which does support a stand-alone "rdma" mount option. So, my opinion is that, today, mount.nfs is not misbehaving by rejecting the "rdma" mount option. This is not a bug, it's working as designed. To add an "rdma" mount option, therefore, is really introducing a new feature to mount.nfs. So, I think such a patch needs appropriate changes to user documentation, and this mode of invoking rdma needs testing, since clearly no-one noticed it was missing until now. As with the rpc.nfsd changes, these mount option changes are two logically distinct but related changes that belong in separate patches. -- chuck[dot]lever[at]oracle[dot]com