Network File System (NFS) Direct Data Placement

The Remote Direct Memory Access (RDMA) Transport for Remote Procedure Call (RPC) allows an RPC client application to post buffers in a Chunk list for specific arguments and results from an RPC call. The RDMA transport header conveys this list of client buffer addresses to the server where the application can associate them with client data and use RDMA operations to transfer the results directly to and from the posted buffers on the client. The client and server must agree on a consistent mapping of posted buffers to RPC. This document details the mapping for each version of the NFS protocol .

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 .

The following changes will be made, relative to : References to will be replaced with references to . Corrections and updates relative to new language in will be introduced. References to obsolete RFCs will be replaced. The reference to a non-existant NFSv4 SYMLINK operation will be replaced with NFSv4 CREATE(NF4LNK). The discussion of 12KB and 36KB inline threshold will be removed. The discussion of NFSv4 COMPOUND handling will be completed. An explicit discussion of NFSv4.0 and NFSv4.1 backchannel operation will be introduced. An IANA Considerations section is required by IDNITS. Code excerpts will be modernized. Other minor changes and editorial corrections may also be made.

The RDMA Read list, in the RDMA transport header, allows an RPC client to marshal RPC call data selectively. Large chunks of data, such as the file data of an NFS WRITE request, MAY be referenced by an RDMA Read list and be moved efficiently and directly placed by an RDMA Read operation initiated by the server. The process of identifying these chunks for the RDMA Read list can be implemented entirely within the RPC layer. It is transparent to the upper-level protocol, such as NFS. For instance, the file data portion of an NFS WRITE request can be selected as an RDMA "chunk" within the eXternal Data Representation (XDR) marshaling code of RPC based on a size criterion, independently of the NFS protocol layer. The XDR unmarshaling on the receiving system can identify the correspondence between Read chunks and protocol elements via the XDR position value encoded in the Read chunk entry. RPC RDMA Read chunks are employed by this NFS mapping to convey specific NFS data to the server in a manner that may be directly placed. The following sections describe this mapping for versions of the NFS protocol.

The RDMA Write list, in the RDMA transport header, allows the client to post one or more buffers into which the server will RDMA Write designated result chunks directly. If the client sends a null Write list, then results from the RPC call will be returned either as an inline reply, as chunks in an RDMA Read list of server-posted buffers, or in a client-posted reply buffer. Each posted buffer in a Write list is represented as an array of memory segments. This allows the client some flexibility in submitting discontiguous memory segments into which the server will scatter the result. Each segment is described by a triplet consisting of the segment handle or steering tag (STag), segment length, and memory address or offset.

<CODE BEGINS> struct xdr_rdma_segment { uint32 handle; /* Registered memory handle */ uint32 length; /* Length of the chunk in bytes */ uint64 offset; /* Chunk virtual address or offset */ }; struct xdr_write_chunk { struct xdr_rdma_segment target<>; }; struct xdr_write_list { struct xdr_write_chunk entry; struct xdr_write_list *next; }; <CODE ENDS> The sum of the segment lengths yields the total size of the buffer, which MUST be large enough to accept the result. If the buffer is too small, the server MUST return an XDR encode error. The server MUST return the result data for a posted buffer by progressively filling its segments, perhaps leaving some trailing segments unfilled or partially full if the size of the result is less than the total size of the buffer segments. The server returns the RDMA Write list to the client with the segment length fields overwritten to indicate the amount of data RDMA written to each segment. Results returned by direct placement MUST NOT be returned by other methods, e.g., by Read chunk list or inline. If no result data at all is returned for the element, the server places no data in the buffer(s), but does return zeros in the segment length fields corresponding to the result. The RDMA Write list allows the client to provide multiple result buffers -- each buffer maps to a specific result in the reply. The NFS client and server implementations agree by specifying the mapping of results to buffers for each RPC procedure. The following sections describe this mapping for versions of the NFS protocol. Through the use of RDMA Write lists in NFS requests, it is not necessary to employ the RDMA Read lists in the NFS replies, as described in the RPC-over-RDMA protocol. This enables more efficient operation, by avoiding the need for the server to expose buffers for RDMA, and also avoiding "RDMA_DONE" exchanges. Clients MAY additionally employ RDMA Reply chunks to receive entire messages, as described in .

A single RDMA Write list entry MAY be posted by the client to receive either the opaque file data from a READ request or the pathname from a READLINK request. The server MUST ignore a Write list for any other NFS procedure, as well as any Write list entries beyond the first in the list. Similarly, a single RDMA Read list entry MAY be posted by the client to supply the opaque file data for a WRITE request or the pathname for a SYMLINK request. The server MUST ignore any Read list for other NFS procedures, as well as additional Read list entries beyond the first in the list. Because there are no NFS version 2 or 3 requests that transfer bulk data in both directions, it is not necessary to post requests containing both Write and Read lists. Any unneeded Read or Write lists are ignored by the server. In the case where the outgoing request or expected incoming reply is larger than the maximum size supported on the connection, it is possible for the RPC layer to post the entire message or result in a special "RDMA_NOMSG" message type that is transferred entirely by RDMA. This is implemented in RPC, below NFS, and therefore has no effect on the message contents. Non-RDMA (inline) WRITE transfers MAY OPTIONALLY employ the "RDMA_MSGP" padding method described in the RPC-over-RDMA protocol, if the appropriate value for the server is known to the client. Padding allows the opaque file data to arrive at the server in an aligned fashion, which may improve server performance. The NFS version 2 and 3 protocols are frequently limited in practice to requests containing less than or equal to 8 kilobytes and 32 kilobytes of data, respectively. In these cases, it is often practical to support basic operation without employing a configuration exchange as discussed in . The server MUST post buffers large enough to receive the largest possible incoming message (approximately 12 KB for NFS version 2, or 36 KB for NFS version 3, would be vastly sufficient), and the client can post buffers large enough to receive replies based on the "rsize" it is using to the server, plus a fixed overhead for the RPC and NFS headers. Because the server MUST NOT return data in excess of this size, the client can be assured of the adequacy of its posted buffer sizes. Flow control is handled dynamically by the RPC RDMA protocol, and write padding is OPTIONAL and therefore MAY remain unused. Alternatively, if the server is administratively configured to values appropriate for all its clients, the same assurance of interoperability within the domain can be made. The use of a configuration protocol with NFS v2 and v3 is therefore OPTIONAL. Employing a configuration exchange may allow some advantage to server resource management through accurately sizing buffers, enabling the server to know exactly how many RDMA Reads may be in progress at once on the client connection, and enabling client write padding, which may be desirable for certain servers when RDMA Read is impractical.

This specification applies to the first minor version of NFS version 4 (NFSv4.0) and any subsequent minor versions that do not override this mapping. The Write list MUST be considered only for the COMPOUND procedure. This procedure returns results from a sequence of operations. Only the opaque file data from an NFS READ operation and the pathname from a READLINK operation MUST utilize entries from the Write list. If there is no Write list, i.e., the list is null, then any READ or READLINK operations in the COMPOUND MUST return their data inline. The NFSv4.0 client MUST ensure in this case that any result of its READ and READLINK requests will fit within its receive buffers, in order to avoid a resulting RDMA transport error upon transfer. The server is not required to detect this. The first entry in the Write list MUST be used by the first READ or READLINK in the COMPOUND request. The next Write list entry is used by the next READ or READLINK, and so on. If there are more READ or READLINK operations than Write list entries, then any remaining operations MUST return their results inline. If a Write list entry is presented, then the corresponding READ or READLINK MUST return its data via an RDMA Write to the buffer indicated by the Write list entry. If the Write list entry has zero RDMA segments, or if the total size of the segments is zero, then the corresponding READ or READLINK operation MUST return its result inline. The following example shows an RDMA Write list with three posted buffers A, B, and C. The designated operations in the compound request, READ and READLINK, consume the posted buffers by writing their results back to each buffer.

RDMA Write list: A --> B --> C Compound request: PUTFH LOOKUP READ PUTFH LOOKUP READLINK PUTFH LOOKUP READ | | | v v v A B C If the client does not want to have the READLINK result returned directly, then it provides a zero-length array of segment triplets for buffer B or sets the values in the segment triplet for buffer B to zeros so that the READLINK result MUST be returned inline. The situation is similar for RDMA Read lists sent by the client and applies to the NFSv4.0 WRITE and SYMLINK procedures as for v3. Additionally, inline segments too large to fit in posted buffers MAY be transferred in special "RDMA_NOMSG" messages. Non-RDMA (inline) WRITE transfers MAY OPTIONALLY employ the "RDMA_MSGP" padding method described in the RPC-over-RDMA protocol, if the appropriate value for the server is known to the client. Padding allows the opaque file data to arrive at the server in an aligned fashion, which may improve server performance. In order to ensure accurate alignment for all data, it is likely that the client will restrict its use of OPTIONAL padding to COMPOUND requests containing only a single WRITE operation. Unlike NFS versions 2 and 3, the maximum size of an NFS version 4 COMPOUND is not bounded, even when RDMA chunks are in use. While it might appear that a configuration protocol exchange (such as the one described in ) would help, in fact the layering issues involved in building COMPOUNDs by NFS make such a mechanism unworkable. However, typical NFS version 4 clients rarely issue such problematic requests. In practice, they behave in much more predictable ways, in fact most still support the traditional rsize/wsize mount parameters. Therefore, most NFS version 4 clients function over RPC-over-RDMA in the same way as NFS versions 2 and 3, operationally. There are however advantages to allowing both client and server to operate with prearranged size constraints, for example, use of the sizes to better manage the server's response cache. An extension to NFS version 4 supporting a more comprehensive exchange of upper-layer parameters is part of .

The NFS version 4 protocols support server-initiated callbacks to selected clients, in order to notify them of events such as recalled delegations, etc. These callbacks present no particular issue to being framed over RPC-over-RDMA since such callbacks do not carry bulk data such as NFS READ or NFS WRITE. They MAY be transmitted inline via RDMA_MSG, or if the callback message or its reply overflow the negotiated buffer sizes for a callback connection, they MAY be transferred via the RDMA_NOMSG method as described above for other exchanges. One special case is noteworthy: in NFS version 4.1, the callback channel is optionally negotiated to be on the same connection as one used for client requests. In this case, and because the transaction ID (XID) is present in the RPC-over-RDMA header, the client MUST ascertain whether the message is in fact an RPC REPLY, and therefore a reply to a prior request and carrying its XID, before processing it as such. By the same token, the server MUST ascertain whether an incoming message on such a callback-eligible connection is an RPC CALL, before optionally processing the XID. In the callback case, the XID present in the RPC-over-RDMA header will potentially have any value, which may (or may not) collide with an XID used by the client for a previous or future request. The client and server MUST inspect the RPC component of the message to determine its potential disposition as either an RPC CALL or RPC REPLY, prior to processing this XID, and MUST NOT reject or accept it without also determining the proper context.

NFS use of direct data placement introduces a need for an additional NFS port number assignment for networks that share traditional UDP and TCP port spaces with RDMA services. The iWARP protocol is such an example (InfiniBand is not). NFS servers for versions 2 and 3 traditionally listen for clients on UDP and TCP port 2049, and additionally, they register these with the portmapper and/or rpcbind service. However, requires NFS servers for version 4 to listen on TCP port 2049, and they are not required to register. An NFS version 2 or version 3 server supporting RPC-over-RDMA on such a network and registering itself with the RPC portmapper MAY choose an arbitrary port, or MAY use the alternative well-known port number for its RPC-over-RDMA service. The chosen port MAY be registered with the RPC portmapper under the netid assigned by the requirement in . An NFS version 4 server supporting RPC-over-RDMA on such a network MUST use the alternative well-known port number for its RPC-over-RDMA service. Clients SHOULD connect to this well-known port without consulting the RPC portmapper (as for NFSv4/TCP). The port number assigned to an NFS service over an RPC-over-RDMA transport is available from the IANA port registry .

The RDMA transport for RPC supports all RPC security models, including RPCSEC_GSS security and link- level security. The choice of RDMA Read and RDMA Write to return RPC argument and results, respectively, does not affect this, since it only changes the method of data transfer. Specifically, the requirements of ensure that this choice does not introduce new vulnerabilities. Because this document defines only the binding of the NFS protocols atop , all relevant security considerations are therefore to be described at that layer.

The author gratefully acknowledges the work of Brent Callaghan and Tom Talpey on the original NFS Direct Data Placement specification . The author also wishes to thank Bill Baker and Greg Marsden for their support of this work.