Internet-Draft                                 Brent Callaghan
Expires: January August 2005                    Sun Microsystems, Inc.                                Tom Talpey
                                        Network Appliance, Inc.

Document: draft-ietf-nfsv4-nfsdirect-00.txt      July, 2004 draft-ietf-nfsv4-nfsdirect-01         February, 2005

                       NFS Direct Data Placement

Status of this Memo

     By submitting this Internet-Draft, I certify that any applicable
     patent or other IPR claims of which I am aware have been disclosed,
     or will be disclosed, and any of which I become aware will be dis-
     closed, in accordance with RFC 3668.

     Internet-Drafts are working documents of the Internet Engineering
     Task Force (IETF), its areas, and its working groups.  Note that
     other groups may also distribute working documents as Internet-

     Internet-Drafts are draft documents valid for a maximum of six
     months and may be updated, replaced, or obsoleted by other docu-
     ments at any time.  It is inappropriate to use Internet-Drafts as
     reference material or to cite them other than as "work in

     The list of current Internet-Drafts can be accessed at The list of Inter-
     net-Draft Shadow Directories can be accessed at

Copyright Notice

     Copyright (C) The Internet Society (2004). (2005).  All Rights Reserved.


   The RDMA transport for ONC RPC provides direct data placement for NFS
   data.  Direct data placement not only reduces the amount of data that
   needs to be copied in an NFS call, but allows much of the data
   movement over the network to be implemented in RDMA hardware. This
   draft describes the use of direct data placement by means of server-
   initiated RDMA operations into client-supplied buffers in a Chunk
   list for implementations of NFS versions 2, 3, and 4 over an RDMA

Table of Contents

   1.  Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2
   2.  RDMA Read List . . . . . . . . . . . . . . . . . . . . . . . 2
   3.  RDMA Write List  . . . . . . . . . . . . . . . . . . . . . . 3
   4.  NFS Versions 2 and 3 Mapping . . . . . . . . . . . . . . . . 4
   5.  NFS Version 4 Mapping  . . . . . . . . . . . . . . . . . . . 5
   6.  Security . . . . . . . . . . . . . . . . . . . . . . . . . . 7
   7.  IANA Considerations  . . . . . . . . . . . . . . . . . . . . 7
   8.  Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 7
   9.  Normative References . . . . . . . . . . . . . . . . . . . . 7
   10.  Informative References  . . . . . . . . . . . . . . . . . . 8
   11.  Authors' Addresses  . . . . . . . . . . . . . . . . . . . . 9
   12.  Full Copyright Statement  . . . . . . . . . . . . . . . . . 9
   Acknowledgement  . . . . . . . . . . . . . . . . . . . . . . .  10

1.  Introduction

   The RDMA Transport for ONC RPC [RPCRDMA] 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.

2.  RDMA Read List

   The RDMA Read list, in the RDMA transport header, allows an RPC
   client to marshall RPC call data selectively.  Large chunks of data,
   such as the file data of an NFS WRITE request, can be referenced from
   the RDMA Read list and be moved efficiently and direct-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 XDR marshalling code of RPC based on a size criterion,
   independently of the NFS protocol layer. The XDR unmarshalling 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.

   However, the implementation of the RDMA Write list requires some help
   from the NFS protocol layer to identify the candidate chunks.  Since
   there is no simple XDR position value to unambiguously label Write
   chunks, both client and server must agree on a mapping of write chunk
   entries to protocol elements.

3.  RDMA Write List

   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 as either an
   in-line 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.

      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;

   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 in-line.

   The RDMA Write list allows the client to provide multiple result
   buffers - each buffer must map to a specific result in the reply. The
   NFS client and server implementations must agree on 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/RDMA protocol. This enables more efficient
   operation, by avoiding the need for the server to expose buffers for
   RDMA, and also avoiding "RDMA_DONE" messages.

4.  NFS Versions 2 and 3 Mapping

   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 will 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 will 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 which 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/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 [RPCRDMA].  The server can
   post buffers large enough to receive the largest possible incoming
   message (approximately 12KB/36KB would be vastly sufficient in the
   above cases), and the client can post buffers large enough to receive
   replies based on the "rsize" it is using to the server.  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

5.  NFS Version 4 Mapping

   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 will 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 will 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 either in-
   line or via RDMA READ (using the Read list).

   The first entry in the Write list must be used by the first READ or
   READLINK in the request.  The next Write list entry by the 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 in-line or via the Read list.

   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 in-
   line or via Read list.

   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:

                       |                   |                   |
                       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 will be returned in-line.

   The situation is similar for RDMA Read lists 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/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 unbounded, even when RDMA chunks are in use.  While it
   might appear that a configuration protocol exchange (as described in
   [RPCRDMA]) would help, in fact the layering issues involved in
   building COMPOUNDs by NFS make such a mechanism unworkable.  Instead,
   an extension to NFS version 4 supporting a more comprehensive
   exchange of upper layer (NFSv4) parameters is proposed in
   [NFSSESSIONS].  This proposal also addresses other use of the sizes,
   such as in the server's response cache.

6.  Security

   The RDMA transport for ONC RPC supports RPCSEC_GSS security as well
   as link-level security.  The use of RDMA Write to return RPC results
   does not affect ONC RPC security.

7.  IANA Considerations

   NFS use of direct data placement introduces no new IANA

8.  Acknowledgements

   The authors would like to thank Dave Noveck and Chet Juszczak for
   their contributions to this document.

9.  Normative References

      R. Srinivasan, "RPC: Remote Procedure Call Protocol Specification
      Version 2",
      Standards Track RFC,

      R. Srinivasan, "XDR: External Data Representation Standard",
      Standards Track RFC,

      "NFS: Network File System Protocol Specification",
      (NFS version 2) Informational RFC,

      B. Callaghan, B. Pawlowski, P. Staubach, "NFS Version 3 Protocol
      Informational RFC,

      S. Shepler, B. Callaghan, D. Robinson, R. Thurlow, C. Beame, M.
      Eisler, D. Noveck, "NFS version 4 Protocol",
      Standards Track RFC,

10.  Informative References

      B. Callaghan, T. Talpey, "RDMA Transport for ONC RPC"
      Internet Draft Work in Progress,

      T. Talpey, S. Shepler, J. Bauman, "NFSv4 Session Extensions"
      Internet Draft Work in Progress,

11.  Authors' Addresses

     Brent Callaghan
     Sun Microsystems, Inc.
     17 Network Circle
     Menlo Park,
     1614 Montalto Dr.
     Mountain View, California 94025 94040 USA

     Phone: +1 650 786 5067 968 2333

     Tom Talpey
     Network Appliance, Inc.
     375 Totten Pond Road
     Waltham, MA 02451 USA

     Phone: +1 781 768 5329

12.  Full Copyright Statement

     Copyright (C) The Internet Society (2004). (2005).  This document is sub-
     ject to the rights, licenses and restrictions contained in BCP 78
     and except as set forth therein, the authors retain all their

     This document and the information contained herein are provided on

Intellectual Property

     The IETF takes no position regarding the validity or scope of any
     Intellectual Property Rights or other rights that might be claimed
     to pertain to the implementation or use of the technology described
     in this document or the extent to which any license under such
     rights might or might not be available; nor does it represent that
     it has made any independent effort to identify any such rights.
     Information on the procedures with respect to rights in RFC docu-
     ments can be found in BCP 78 and BCP 79.

     Copies of IPR disclosures made to the IETF Secretariat and any
     assurances of licenses to be made available, or the result of an
     attempt made to obtain a general license or permission for the use
     of such proprietary rights by implementers or users of this speci-
     fication can be obtained from the IETF on-line IPR repository at

     The IETF invites any interested party to bring to its attention any
     copyrights, patents or patent applications, or other proprietary
     rights that may cover technology that may be required to implement
     this standard.  Please address the information to the IETF at ietf-


     Funding for the RFC Editor function is currently provided by the
     Internet Society.