This specification defines the NFS version 3 protocol, that is the over-the-wire protocol by which a client accesses a server. The protocol provides a well-defined interface to a server's file resources. A client or server implements the protocol and provides a mapping of the local file system semantics and actions into those defined in the NFS version 3 protocol. Implementations may differ to varying degrees, depending on the extent to which a given environment can support all the operations and semantics defined in the NFS version 3 protocol. Although implementations exist and are used to illustrate various aspects of the NFS version 3 protocol, the protocol specification itself is the final description of how clients access server resources.
Because the NFS version 3 protocol is designed to be operating-system independent, it does not necessarily match the semantics of any existing system. Server implementations are expected to make a best effort at supporting the protocol. If a server cannot support a particular protocol procedure, it may return the error, NFS3ERR_NOTSUP, that indicates that the operation is not supported. For example, many operating systems do not support the notion of a hard link. A server that cannot support hard links should return NFS3ERR_NOTSUP in response to a LINK request. FSINFO describes the most commonly unsupported procedures in the properties bit map. Alternatively, a server may not natively support a given operation, but can emulate it in the NFS version 3 protocol implementation to provide greater functionality.
In some cases, a server can support most of the semantics described by the protocol but not all. For example, the ctime field in the fattr structure gives the time that a file's attributes were last modified. Many systems do not keep this information. In this case, rather than not support the GETATTR operation, a server could simulate it by returning the last modified time in place of ctime. Servers must be careful when simulating attribute information because of possible side effects on clients. For example, many clients use file modification times as a basis for their cache consistency scheme.
NFS servers are dumb and NFS clients are smart. It is the clients that do the work required to convert the generalized file access that servers provide into a file access method that is useful to applications and users. In the LINK example given above, a UNIX client that received an NFS3ERR_NOTSUP error from a server would do the recovery necessary to either make it look to the application like the link request had succeeded or return a reasonable error. In general, it is the burden of the client to recover.
The NFS version 3 protocol assumes a stateless server implementation. Statelessness means that the server does not need to maintain state about any of its clients in order to function correctly. Stateless servers have a distinct advantage over stateful servers in the event of a crash. With stateless servers, a client need only retry a request until the server responds; the client does not even need to know that the server has crashed. See additional comments in Duplicate request cache on page 99.
For a server to be useful, it holds nonvolatile state: data stored in the file system. Design assumptions in the NFS version 3 protocol regarding flushing of modified data to stable storage reduce the number of failure modes in which data loss can occur. In this way, NFS version 3 protocol implementations can tolerate transient failures, including transient failures of the network. In general, server implementations of the NFS version 3 protocol cannot tolerate a non-transient failure of the stable storage itself. However, there exist fault tolerant implementations which attempt to address such problems.
That is not to say that an NFS version 3 protocol server can't maintain noncritical state. In many cases, servers will maintain state (cache) about previous operations to increase performance. For example, a client READ request might trigger a read-ahead of the next block of the file into the server's data cache in the anticipation that the client is doing a sequential read and the next client READ request will be satisfied from the server's data cache instead of from the disk. Read-ahead on the server increases performance by overlapping server disk I/O with client requests. The important point here is that the read-ahead block is not necessary for correct server behavior. If the server crashes and loses its memory cache of read buffers, recovery is simple on reboot - clients will continue read operations retrieving data from the server disk.
Most data-modifying operations in the NFS protocol are synchronous. That is, when a data modifying procedure returns to the client, the client can assume that the operation has completed and any modified data associated with the request is now on stable storage. For example, a synchronous client WRITE request may cause the server to update data blocks, file system information blocks, and file attribute information - the latter information is usually referred to as metadata. When the WRITE operation completes, the client can assume that the write data is safe and discard it. This is a very important part of the stateless nature of the server. If the server did not flush dirty data to stable storage before returning to the client, the client would have no way of knowing when it was safe to discard modified data. The following data modifying procedures are synchronous: WRITE (with stable flag set to FILE_SYNC), CREATE, MKDIR, SYMLINK, MKNOD, REMOVE, RMDIR, RENAME, LINK, and COMMIT.
The NFS version 3 protocol introduces safe asynchronous writes on the server, when the WRITE procedure is used in conjunction with the COMMIT procedure. The COMMIT procedure provides a way for the client to flush data from previous asynchronous WRITE requests on the server to stable storage and to detect whether it is necessary to retransmit the data. See the procedure descriptions of WRITE on page 49 and COMMIT on page 92.
The LOOKUP procedure is used by the client to traverse multicomponent file names (pathnames). Each call to LOOKUP is used to resolve one segment of a pathname. There are two reasons for restricting LOOKUP to a single segment: it is hard to standardize a common format for hierarchical file names and the client and server may have different mappings of pathnames to file systems. This would imply that either the client must break the path name at file system attachment points, or the server must know about the client's file system attachment points. In NFS version 3 protocol implementations, it is the client that constructs the hierarchical file name space using mounts to build a hierarchy. Support utilities, such as the Automounter, provide a way to manage a shared, consistent image of the file name space while still being driven by the client mount process.
Clients can perform caching in varied manner. The general practice with the NFS version 2 protocol was to implement a time-based client-server cache consistency mechanism. It is expected NFS version 3 protocol implementations will use a similar mechanism. The NFS version 3 protocol has some explicit support, in the form of additional attribute information to eliminate explicit attribute checks. However, caching is not required, nor is any caching policy defined by the protocol. Neither the NFS version 2 protocol nor the NFS version 3 protocol provide a means of maintaining strict client-server consistency (and, by implication, consistency across client caches).