Using DNS SRV to Specify a Global File Name Space with NFS version 4

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

With the advent of fs_locations attributes in the NFS Version 4 protocol , NFS servers can cooperate to build a file name space that crosses server boundaries, as detailed in the description of referrals in . With NFS Version 4 referrals, a file server may indicate to its client that the file system name tree beneath a given name in the server is not present on itself, but is represented by a filesystem in some other set of servers. The mechanism is general, allowing servers to describe any filesystem as being reachable by requests to any of a set of servers. Thus, starting with a single NFS Version 4 server, using these referrals, an NFS Version 4 client might be able to see a large name space associated with a collection of interrelated NFS Version 4 file servers. An organization could use this capability to construct a uniform file name space for itself. An organization might wish to publish the starting point for this name space to its clients. In many cases, the organization will want to publish this starting point to a broader set of possible clients. At the same time, it is useful to require clients to know only the smallest amount of information in order to locate the appropriate name space. Simultaneously, that required information should be constant through the life of an organization if the clients are not to require reconfiguration as administrative events change, for instance, a server's name or address.

Providing an organization's published file system name space is a service, and it is appropriate to use the DNS to locate it. As with the AFSDB resource record type , the client need only utter the (relatively) constant domain name for an organization in order to locate its file system name space service. Once a client uses the DNS to locate one or more servers for the root of the organization's name space, it can use the standard NFS Version 4 mechanisms to navigate the remainder of the NFS servers for that organization. The use of this proposed mechanism results in a useful cross-organizational name space, just as in AFS and DCE/DFS before it. A client need know only the name of the organization in order to locate the file system name space published by that organization.

We propose the use of the DNS SRV resource record type [RFC2782] to fulfill this function. The format of the DNS SRV record is as follows: _Service._Proto.Name TTL Class SRV Priority Weight Port Target In our case, we use a Service name of "nfs4" and a conventional Protocol of "_tcp". The Target fields give the domain names of the NFS Version 4 servers that export root filesystems. An NFS Version 4 client SHOULD interpret any of the exported pseudo-root filesystems as the filesystem published by the organization with the given domain name.

Suppose a client wished to locate the root of the file system published by organization example.net. The DNS servers for the domain could publish records like _nfs4._tcp IN SRV 0 0 2049 nfs1tr.example.net _nfs4._tcp IN SRV 1 0 2049 nfs2ex.example.net The result domain names nfs1tr.example.net and nfs2ex.example.net indicate NFS Version 4 file servers that export the root of the published name space for the example.net domain. In accordance with RFC 2782, these records are to be interpreted using the Priority and Weight field values, selecting an appropriate file server with which to begin a network conversation. Subsequent accesses are carried out in accordance with ordinary NFS Version 4 protocol.

As with any DNS resource, any server installation needs to concern itself with the likely loads and effects of the presence of the resource record. The answers to requests for RRs might differ depending on what the server can tell about the client. For example, some RRs might be returned only to those clients inside some network perimeter (to provide an intranet service) and requests from other clients might be denied. As the RR directs the clients to ask for service from a given set of servers, the administrator should ensure that the identified servers can handle the expected load. Fortunately, the definition of the DNS SRV resource record offers a mechanism to distribute the load to multiple servers within a priority ordering.

There are at least two remaining questions: whether this DNS SRV record evaluation is done in the NFS server or client, and also how the domain names of the organizations are passed to client or server. A third question is how this might produce a uniform global file name space, and what prefix should be used for such file names. This specification anticipates that these SRV records will most commonly be used to define the second directory level in an inter-organizational file name space. This directory will be populated with domain names pointing to the file systems published for use under those domain names. Thus, the root directory for the file system published by example.net will effectively be mounted underneath the example.net name in a second-level directory. In general, a domain name will appear to a client as a directory name pointing to the root directory of the file system published by the organization responsible for that domain name.

For the inter-organizational name space to be a global name space, it is useful for its mount point in local systems to be uniform as well. The name /nfs4/ SHOULD be used so that names on one machine will be directly usable on any machine. Thus, the example.net published file system would be accessible as /nfs4/example.net/ on any client. Using this convention, "/nfs4/" is a mount for a special file system that is populated with the results of SRV record lookups.

SRV records are necessarily less complete than the information in the existing NFS Version 4 attributes fs_locations and the proposed fs_locations_info. For the rootpath field of fs_location, we assume that the empty string is adequate. Thus, the servers listed as targets for the SRV resource records should export the root of the organization's published file system as the pseudo-root in its exported namespace. As for the other attributes in fs_locations_info, the recommended approach is for a client to make its first possible contact with any of the referred-to servers, obtain the fs_locations_info structure from that server, and use the information from that obtained structure as the basis for its judgment of whether it would be better to use a different server representative from the set of servers for that filesystem. We recommend, though, that the process of mounting an organization's name space should permit the use of what is likely to impose the lowest cost on the server. Thus, we recommend that the client not insist on using a writable copy of the filesystem if read-only copies exist, or a zero-age copy rather than a copy that may be a little older. We presume that the organization's file name space can be navigated to provide access to higher-cost properties such as writability or currency as necessary, but that the default use when navigating to the base information for an organization ought to be as low-overhead as possible.

One extension of this rule that we might choose to inherit from AFS, though, is to give a special meaning to the domain name of an organization preceded by a period ("."). It might be reasonable to have names mounting the filesystem for a period-prefixed domain name (e.g., ".example.net") attempt to mount only a read-write instance of that organization's root filesystem, rather than permitting the use of read-only instances of that filesystem. Thus, /nfs4/example.net/users might be a directory in a read-only instance of the root filesystem of the organization "example.net", while

/nfs4/.example.net/users would be a writable form of that same directory. A small benefit of following this convention is that names with the period prefix are treated as "hidden" in many operating systems, so that the visible name remains the lowest-overhead name.

The result of the DNS search SHOULD appear as a (pseudo-)directory in the client name space, cached for a time no longer than the RR's TTL. A further refinement is advisable, and SHOULD be deployed: that only fully-qualified domain names appear as directories. That is, in many environments, DNS names may be abbreviated from their fully-qualified form. In such circumstances, multiple names might be given to file system code that all resolve to the same DNS SRV RRs. The abbreviated form SHOULD be represented in the client's name space cache as a symbolic link, pointing to the fully-qualified name, case-canonicalized when appropriate. This will allow pathnames obtained with, say, getcwd() to include the DNS name that is most likely to be usable outside the scope of any particular DNS abbreviation convention.

Another consideration is what agent should be responsible for interpreting the SRV records. It could be done just as well by the client or by the server, though we expect that most clients will include this function themselves. Using something like Automounter technology, the client would be responsible for interpreting names under a particular directory, discovering the appropriate filesystem to mount, and mounting it in the appropriate place in the client name space before returning control to the application doing a lookup. Alternatively, one could imagine the existence of an NFS version 4 server that awaited similar domain-name lookups, then consulted the DNS SRV records to determine the servers for the indicated published file system, and then returned that information via NFS Version 4 attributes as a referral in the way outlined by Noveck and Burnett . In either case, the result of the DNS lookup should be cached (obeying TTL) so that the result could be returned more quickly the next time. We strongly suggest that this functionality be implemented by NFS clients. While we recognize that it would be possible to configure clients so that they relied on a specially-configured server to do their SRV lookups for them, we feel that such a requirement would impose unusual dependencies and vulnerabilities for the deployers of such clients.

This DNS use has no obvious relationship to the NFS4ID RR. The NFS4ID RR is a mechanism to help clients and servers configure themselves with respect to the domain strings used in "who" strings in ACL entries and in owner and group names. The authentication/authorization domain string of a server need have no direct relationship to the name of the organization that is publishing a file name space of which this server's filesystems form a part. At the same time, it might be seen as straightforward or normal for such a server to refer to the ownership of most of its files using a domain string with an evident relationship to that NFS4ID-given domain name, but this document imposes no such requirement.

Naive use of the DNS may effectively give clients published server referrals that are intrusive substitutes for the servers intended by domain administrators. It may be possible to build a trust chain by using DNSSEC to implement this function on the client, or by implementing this function on an NFS Version 4 server that uses DNSSEC and maintaining a trust relationship with that server. This trust chain also breaks if the SRV interpreter accepts responses from insecure DNS zones. Thus, it would likely be prudent also to use domain-based service principal names for the servers for the root filesystems as indicated as the targets of the SRV records. The idea here is that one wants to authenticate {nfs, domainname, host.fqdn}, not simply {nfs, host.fqdn}, when the server is a domain's root file server obtained through an insecure DNS SRV RR lookup. The domain administrator can thus ensure that only domain root NFSv4 servers have credentials for such domain-based service principal names. Domain-based service principal names are being proposed in the KITTEN WG .