netid in the NFS mount command

http://article.gmane.org/gmane.linux.nfs/16950 has Chuck Lever's patch to the nfs man page, which includes:

-The default value depends on whether
-.IR proto=udp
[...]
+.BI proto= netid
+The transport protocol used by the RPC client
+to transmit requests to the NFS server for this mount point. +The value of
+.I netid
+can be either
+.B udp
or
-.IR proto=tcp
-is in effect (see below).

After talking to Chuck yesterday, it appears that he is being literal:
the netid can only be either udp or tcp, and not also udp6 or tcp6, which are the netids IETF defines (see RFC3530) for ONC RPC over UDP/IPv6 and TCP/IPv6.

If proto= is going to mean transport protocol in the Linux mount command (as opposed to meaning a netid as it does in the Solaris mount command), then please fix the man page to say that.

But I think there is value in following the Solaris convention
(which I established when I was at Sun in the 1990s working on the
Solaris NFS/TCP stack) because by specifying "proto=tcp6" in the
mount command, this can force the use of TCP over the IPv6 networking
protocol in cases where the same host name is resolvable to both an
IPv4 or IPv6 host address. Of course, one could edit nsswitch.conf and /etc/hosts or ipnodes to insert an IPv6-only host name, but that is cumbersome.

BTW, the terse netid stuff in RFC3530 is being codified in a separate to-be-approved RFC. You can look at the internet draft at:
http://tools.ietf.org/wg/nfsv4/draft-ietf-nfsv4-rpc-netid/

Netids for all the IETF transports that I am aware of are specified. Additional ones can be added after the RFC is published by registering with IANA via the process specified in the internet-draft.

Note that rdma and rdma6, corresponding to the rdma over TCP/IP and rdma over TCP/IPv6 protocols are among the netids. If people want a netid for a raw, non-IP-based RDMA transport, that will have to be registered with IANA.

The man page should be corrected.

However, before we do that, we need to revisit the issue of how the kernel selects an RPC transport capability, and resolve as much as possible how this should work.

On the other hand, if a user mounts "server" with "proto=tcp" on Solaris, and the server drops its IPv4 addresses in favor of IPv6 addresses, client sysadmins will have to be aware of this change, and modify their mount options. On Linux, the server can change to IPv6-only, and the clients will adjust without intervention.

Perhaps for one user that operates both the client and server, this is not an issue. However, this does become a problem if a large number of clients access a server that is being changed in another administrative domain. Naturally this can be hidden with automounter, but we wander straight back into the realm of diddling with configuration files, which is cumbersome. Not to mention that not everyone chooses to use an automounter.

It seems more natural to me to use an AF_INET6 address to choose NFS over IPv6, rather than a netid. One of my interests is to keep the details of AF_INET v. AF_INET6 inconsequential to NFS. It would be easy, however, to add a "netid=" mount option in addition to "proto=" where a true netid could be used instead of a protocol name.

And finally, we had to do something to allow NFS/RDMA on InfiniBand, which still doesn't have a netid, as you point out. According to the man page I have on my OpenSolaris system, it actually does allow "proto=rdma".

I personally am glad of any clarification around netids.

One thing I noticed was this in RFC 1833 (August 1995):

/*
* A mapping of (program, version, network ID) to address *
* The network identifier (r_netid):
* This is a string that represents a local identification for a
* network. This is defined by a system administrator based on local
* conventions, and cannot be depended on to have the same value on
* every system. */

From this, I assumed that netids were not standardized on purpose. Mike, does your draft address and clarify this statement? Will your draft supercede RFC 1833 and subsequent related RFCs?

--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Right, well the mount would fail in the Solaris case, so awareness will come quickly. Optimizing for the infrequent case of transition to IPv6 doesn't see like the better choice to me.

And if the server supports both IPv4 and IPv6, maybe the client's system administrator wants to make sure IPv4 is used, because, he's discovered that the IPv6 checksum support for the ethernet chips being used on the client and/or server does not exist, or it does exist, but the TCP/IP stacks in the client and/or server don't take advantage of it, and thus CPU utilization jumps by 25% and performance by say 35%.

I think is is unnatural to tell a customer that if he wants to force the use IPv6 or IPv4 on the command line, he has to specify a IPv4 or IPv6 address.

I hope you at least do that. It would be nice if the Solaris and Linux mount commands could agree on this in order to avoid enraging customers that use both, but so be it.

If "rdma" does not have an entry in /etc/netconfig then apparently OpenSolaris is not consistent with the TI-RPC architecture of Solaris.

The netid draft updates RFC 1833, and so "cannot be depended on to have the same value on every system." will no longer be true.