RFC 5666
This document is obsolete. Please refer to RFC 8166.






Internet Engineering Task Force (IETF)                         T. Talpey
Request for Comments: 5666                                  Unaffiliated
Category: Standards Track                                   B. Callaghan
ISSN: 2070-1721                                                    Apple
                                                            January 2010


    Remote Direct Memory Access Transport for Remote Procedure Call

Abstract



   This document describes a protocol providing Remote Direct Memory
   Access (RDMA) as a new transport for Remote Procedure Call (RPC).
   The RDMA transport binding conveys the benefits of efficient, bulk-
   data transport over high-speed networks, while providing for minimal
   change to RPC applications and with no required revision of the
   application RPC protocol, or the RPC protocol itself.

Status of This Memo



   This is an Internet Standards Track document.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Further information on
   Internet Standards is available in Section 2 of RFC 5741.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   http://www.rfc-editor.org/info/rfc5666.

Copyright Notice



   Copyright (c) 2010 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.





Talpey & Callaghan           Standards Track                    [Page 1]

RFC 5666                 RDMA Transport for RPC             January 2010


   This document may contain material from IETF Documents or IETF
   Contributions published or made publicly available before November
   10, 2008.  The person(s) controlling the copyright in some of this
   material may not have granted the IETF Trust the right to allow
   modifications of such material outside the IETF Standards Process.
   Without obtaining an adequate license from the person(s) controlling
   the copyright in such materials, this document may not be modified
   outside the IETF Standards Process, and derivative works of it may
   not be created outside the IETF Standards Process, except to format
   it for publication as an RFC or to translate it into languages other
   than English.

Table of Contents



   1. Introduction ....................................................3
      1.1. Requirements Language ......................................4
   2. Abstract RDMA Requirements ......................................4
   3. Protocol Outline ................................................5
      3.1. Short Messages .............................................6
      3.2. Data Chunks ................................................6
      3.3. Flow Control ...............................................7
      3.4. XDR Encoding with Chunks ...................................8
      3.5. XDR Decoding with Read Chunks .............................11
      3.6. XDR Decoding with Write Chunks ............................12
      3.7. XDR Roundup and Chunks ....................................13
      3.8. RPC Call and Reply ........................................14
      3.9. Padding ...................................................17
   4. RPC RDMA Message Layout ........................................18
      4.1. RPC-over-RDMA Header ......................................18
      4.2. RPC-over-RDMA Header Errors ...............................20
      4.3. XDR Language Description ..................................20
   5. Long Messages ..................................................22
      5.1. Message as an RDMA Read Chunk .............................23
      5.2. RDMA Write of Long Replies (Reply Chunks) .................24
   6. Connection Configuration Protocol ..............................25
      6.1. Initial Connection State ..................................26
      6.2. Protocol Description ......................................26
   7. Memory Registration Overhead ...................................28
   8. Errors and Error Recovery ......................................28
   9. Node Addressing ................................................28
   10. RPC Binding ...................................................29
   11. Security Considerations .......................................30
   12. IANA Considerations ...........................................31
   13. Acknowledgments ...............................................32
   14. References ....................................................33
      14.1. Normative References .....................................33
      14.2. Informative References ...................................33




Talpey & Callaghan           Standards Track                    [Page 2]

RFC 5666                 RDMA Transport for RPC             January 2010


1.  Introduction



   Remote Direct Memory Access (RDMA) [RFC5040, RFC5041], [IB] is a
   technique for efficient movement of data between end nodes, which
   becomes increasingly compelling over high-speed transports.  By
   directing data into destination buffers as it is sent on a network,
   and placing it via direct memory access by hardware, the double
   benefit of faster transfers and reduced host overhead is obtained.

   Open Network Computing Remote Procedure Call (ONC RPC, or simply,
   RPC) [RFC5531] is a remote procedure call protocol that has been run
   over a variety of transports.  Most RPC implementations today use UDP
   or TCP.  RPC messages are defined in terms of an eXternal Data
   Representation (XDR) [RFC4506], which provides a canonical data
   representation across a variety of host architectures.  An XDR data
   stream is conveyed differently on each type of transport.  On UDP,
   RPC messages are encapsulated inside datagrams, while on a TCP byte
   stream, RPC messages are delineated by a record marking protocol.  An
   RDMA transport also conveys RPC messages in a unique fashion that
   must be fully described if client and server implementations are to
   interoperate.

   RDMA transports present new semantics unlike the behaviors of either
   UDP or TCP alone.  They retain message delineations like UDP while
   also providing a reliable, sequenced data transfer like TCP.  Also,
   they provide the new efficient, bulk-transfer service of RDMA.  RDMA
   transports are therefore naturally viewed as a new transport type by
   RPC.

   RDMA as a transport will benefit the performance of RPC protocols
   that move large "chunks" of data, since RDMA hardware excels at
   moving data efficiently between host memory and a high-speed network
   with little or no host CPU involvement.  In this context, the Network
   File System (NFS) protocol, in all its versions [RFC1094] [RFC1813]
   [RFC3530] [RFC5661], is an obvious beneficiary of RDMA.  A complete
   problem statement is discussed in [RFC5532], and related NFSv4 issues
   are discussed in [RFC5661].  Many other RPC-based protocols will also
   benefit.

   Although the RDMA transport described here provides relatively
   transparent support for any RPC application, the proposal goes
   further in describing mechanisms that can optimize the use of RDMA
   with more active participation by the RPC application.








Talpey & Callaghan           Standards Track                    [Page 3]

RFC 5666                 RDMA Transport for RPC             January 2010


1.1.  Requirements Language



   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 [RFC2119].

2.  Abstract RDMA Requirements



   An RPC transport is responsible for conveying an RPC message from a
   sender to a receiver.  An RPC message is either an RPC call from a
   client to a server, or an RPC reply from the server back to the
   client.  An RPC message contains an RPC call header followed by
   arguments if the message is an RPC call, or an RPC reply header
   followed by results if the message is an RPC reply.  The call header
   contains a transaction ID (XID) followed by the program and procedure
   number as well as a security credential.  An RPC reply header begins
   with an XID that matches that of the RPC call message, followed by a
   security verifier and results.  All data in an RPC message is XDR
   encoded.  For a complete description of the RPC protocol and XDR
   encoding, see [RFC5531] and [RFC4506].

   This protocol assumes the following abstract model for RDMA
   transports.  These terms, common in the RDMA lexicon, are used in
   this document.  A more complete glossary of RDMA terms can be found
   in [RFC5040].

   o Registered Memory
        All data moved via tagged RDMA operations is resident in
        registered memory at its destination.  This protocol assumes
        that each segment of registered memory MUST be identified with a
        steering tag of no more than 32 bits and memory addresses of up
        to 64 bits in length.

   o RDMA Send
        The RDMA provider supports an RDMA Send operation with
        completion signaled at the receiver when data is placed in a
        pre-posted buffer.  The amount of transferred data is limited
        only by the size of the receiver's buffer.  Sends complete at
        the receiver in the order they were issued at the sender.

   o RDMA Write
        The RDMA provider supports an RDMA Write operation to directly
        place data in the receiver's buffer.  An RDMA Write is initiated
        by the sender and completion is signaled at the sender.  No
        completion is signaled at the receiver.  The sender uses a
        steering tag, memory address, and length of the remote
        destination buffer.  RDMA Writes are not necessarily ordered
        with respect to one another, but are ordered with respect to



Talpey & Callaghan           Standards Track                    [Page 4]

RFC 5666                 RDMA Transport for RPC             January 2010


        RDMA Sends; a subsequent RDMA Send completion obtained at the
        receiver guarantees that prior RDMA Write data has been
        successfully placed in the receiver's memory.

   o RDMA Read
        The RDMA provider supports an RDMA Read operation to directly
        place peer source data in the requester's buffer.  An RDMA Read
        is initiated by the receiver and completion is signaled at the
        receiver.  The receiver provides steering tags, memory
        addresses, and a length for the remote source and local
        destination buffers.  Since the peer at the data source receives
        no notification of RDMA Read completion, there is an assumption
        that on receiving the data, the receiver will signal completion
        with an RDMA Send message, so that the peer can free the source
        buffers and the associated steering tags.

   This protocol is designed to be carried over all RDMA transports
   meeting the stated requirements.  This protocol conveys to the RPC
   peer information sufficient for that RPC peer to direct an RDMA layer
   to perform transfers containing RPC data and to communicate their
   result(s).  For example, it is readily carried over RDMA transports
   such as Internet Wide Area RDMA Protocol (iWARP) [RFC5040, RFC5041],
   or InfiniBand [IB].

3.  Protocol Outline



   An RPC message can be conveyed in identical fashion, whether it is a
   call or reply message.  In each case, the transmission of the message
   proper is preceded by transmission of a transport-specific header for
   use by RPC-over-RDMA transports.  This header is analogous to the
   record marking used for RPC over TCP, but is more extensive, since
   RDMA transports support several modes of data transfer; it is
   important to allow the upper-layer protocol to specify the most
   efficient mode for each of the segments in a message.  Multiple
   segments of a message may thereby be transferred in different ways to
   different remote memory destinations.

   All transfers of a call or reply begin with an RDMA Send that
   transfers at least the RPC-over-RDMA header, usually with the call or
   reply message appended, or at least some part thereof.  Because the
   size of what may be transmitted via RDMA Send is limited by the size
   of the receiver's pre-posted buffer, the RPC-over-RDMA transport
   provides a number of methods to reduce the amount transferred by
   means of the RDMA Send, when necessary, by transferring various parts
   of the message using RDMA Read and RDMA Write.






Talpey & Callaghan           Standards Track                    [Page 5]

RFC 5666                 RDMA Transport for RPC             January 2010


   RPC-over-RDMA framing replaces all other RPC framing (such as TCP
   record marking) when used atop an RPC/RDMA association, even though
   the underlying RDMA protocol may itself be layered atop a protocol
   with a defined RPC framing (such as TCP).  It is however possible for
   RPC/RDMA to be dynamically enabled, in the course of negotiating the
   use of RDMA via an upper-layer exchange.  Because RPC framing
   delimits an entire RPC request or reply, the resulting shift in
   framing must occur between distinct RPC messages, and in concert with
   the transport.

3.1.  Short Messages



   Many RPC messages are quite short.  For example, the NFS version 3
   GETATTR request, is only 56 bytes: 20 bytes of RPC header, plus a
   32-byte file handle argument and 4 bytes of length.  The reply to
   this common request is about 100 bytes.

   There is no benefit in transferring such small messages with an RDMA
   Read or Write operation.  The overhead in transferring steering tags
   and memory addresses is justified only by large transfers.  The
   critical message size that justifies RDMA transfer will vary
   depending on the RDMA implementation and network, but is typically of
   the order of a few kilobytes.  It is appropriate to transfer a short
   message with an RDMA Send to a pre-posted buffer.  The RPC-over-RDMA
   header with the short message (call or reply) immediately following
   is transferred using a single RDMA Send operation.

   Short RPC messages over an RDMA transport:

        RPC Client                           RPC Server
            |               RPC Call              |
       Send |   ------------------------------>   |
            |                                     |
            |               RPC Reply             |
            |   <------------------------------   | Send

3.2.  Data Chunks



   Some protocols, like NFS, have RPC procedures that can transfer very
   large chunks of data in the RPC call or reply and would cause the
   maximum send size to be exceeded if one tried to transfer them as
   part of the RDMA Send.  These large chunks typically range from a
   kilobyte to a megabyte or more.  An RDMA transport can transfer large
   chunks of data more efficiently via the direct placement of an RDMA
   Read or RDMA Write operation.  Using direct placement instead of
   inline transfer not only avoids expensive data copies, but provides
   correct data alignment at the destination.




Talpey & Callaghan           Standards Track                    [Page 6]

RFC 5666                 RDMA Transport for RPC             January 2010


3.3.  Flow Control



   It is critical to provide RDMA Send flow control for an RDMA
   connection.  RDMA receive operations will fail if a pre-posted
   receive buffer is not available to accept an incoming RDMA Send, and
   repeated occurrences of such errors can be fatal to the connection.
   This is a departure from conventional TCP/IP networking where buffers
   are allocated dynamically on an as-needed basis, and where
   pre-posting is not required.

   It is not practical to provide for fixed credit limits at the RPC
   server.  Fixed limits scale poorly, since posted buffers are
   dedicated to the associated connection until consumed by receive
   operations.  Additionally, for protocol correctness, the RPC server
   must always be able to reply to client requests, whether or not new
   buffers have been posted to accept future receives.  (Note that the
   RPC server may in fact be a client at some other layer.  For example,
   NFSv4 callbacks are processed by the NFSv4 client, acting as an RPC
   server.  The credit discussions apply equally in either case.)

   Flow control for RDMA Send operations is implemented as a simple
   request/grant protocol in the RPC-over-RDMA header associated with
   each RPC message.  The RPC-over-RDMA header for RPC call messages
   contains a requested credit value for the RPC server, which MAY be
   dynamically adjusted by the caller to match its expected needs.  The
   RPC-over-RDMA header for the RPC reply messages provides the granted
   result, which MAY have any value except it MUST NOT be zero when no
   in-progress operations are present at the server, since such a value
   would result in deadlock.  The value MAY be adjusted up or down at
   each opportunity to match the server's needs or policies.

   The RPC client MUST NOT send unacknowledged requests in excess of
   this granted RPC server credit limit.  If the limit is exceeded, the
   RDMA layer may signal an error, possibly terminating the connection.
   Even if an error does not occur, it is OPTIONAL that the server
   handle the excess request(s), and it MAY return an RPC error to the
   client.  Also note that the never-zero requirement implies that an
   RPC server MUST always provide at least one credit to each connected
   RPC client from which no requests are outstanding.  The client would
   deadlock otherwise, unable to send another request.

   While RPC calls complete in any order, the current flow control limit
   at the RPC server is known to the RPC client from the Send ordering
   properties.  It is always the most recent server-granted credit value
   minus the number of requests in flight.






Talpey & Callaghan           Standards Track                    [Page 7]

RFC 5666                 RDMA Transport for RPC             January 2010


   Certain RDMA implementations may impose additional flow control
   restrictions, such as limits on RDMA Read operations in progress at
   the responder.  Because these operations are outside the scope of
   this protocol, they are not addressed and SHOULD be provided for by
   other layers.  For example, a simple upper-layer RPC consumer might
   perform single-issue RDMA Read requests, while a more sophisticated,
   multithreaded RPC consumer might implement its own First In, First
   Out (FIFO) queue of such operations.  For further discussion of
   possible protocol implementations capable of negotiating these
   values, see Section 6 "Connection Configuration Protocol" of this
   document, or [RFC5661].

3.4.  XDR Encoding with Chunks



   The data comprising an RPC call or reply message is marshaled or
   serialized into a contiguous stream by an XDR routine.  XDR data
   types such as integers, strings, arrays, and linked lists are
   commonly implemented over two very simple functions that encode
   either an XDR data unit (32 bits) or an array of bytes.

   Normally, the separate data items in an RPC call or reply are encoded
   as a contiguous sequence of bytes for network transmission over UDP
   or TCP.  However, in the case of an RDMA transport, local routines
   such as XDR encode can determine that (for instance) an opaque byte
   array is large enough to be more efficiently moved via an RDMA data
   transfer operation like RDMA Read or RDMA Write.

   Semantically speaking, the protocol has no restriction regarding data
   types that may or may not be represented by a read or write chunk.
   In practice however, efficiency considerations lead to the conclusion
   that certain data types are not generally "chunkable".  Typically,
   only those opaque and aggregate data types that may attain
   substantial size are considered to be eligible.  With today's
   hardware, this size may be a kilobyte or more.  However, any object
   MAY be chosen for chunking in any given message.

   The eligibility of XDR data items to be candidates for being moved as
   data chunks (as opposed to being marshaled inline) is not specified
   by the RPC-over-RDMA protocol.  Chunk eligibility criteria MUST be
   determined by each upper-layer in order to provide for an
   interoperable specification.  One such example with rationale, for
   the NFS protocol family, is provided in [RFC5667].

   The interface by which an upper-layer implementation communicates the
   eligibility of a data item locally to RPC for chunking is out of
   scope for this specification.  In many implementations, it is
   possible to implement a transparent RPC chunking facility.  However,
   such implementations may lead to inefficiencies, either because they



Talpey & Callaghan           Standards Track                    [Page 8]

RFC 5666                 RDMA Transport for RPC             January 2010


   require the RPC layer to perform expensive registration and
   de-registration of memory "on the fly", or they may require using
   RDMA chunks in reply messages, along with the resulting additional
   handshaking with the RPC-over-RDMA peer.  However, these issues are
   internal and generally confined to the local interface between RPC
   and its upper layers, one in which implementations are free to
   innovate.  The only requirement is that the resulting RPC RDMA
   protocol sent to the peer is valid for the upper layer.  See, for
   example, [RFC5667].

   When sending any message (request or reply) that contains an eligible
   large data chunk, the XDR encoding routine avoids moving the data
   into the XDR stream.  Instead, it does not encode the data portion,
   but records the address and size of each chunk in a separate "read
   chunk list" encoded within RPC RDMA transport-specific headers.  Such
   chunks will be transferred via RDMA Read operations initiated by the
   receiver.

   When the read chunks are to be moved via RDMA, the memory for each
   chunk is registered.  This registration may take place within XDR
   itself, providing for full transparency to upper layers, or it may be
   performed by any other specific local implementation.

   Additionally, when making an RPC call that can result in bulk data
   transferred in the reply, write chunks MAY be provided to accept the
   data directly via RDMA Write.  These write chunks will therefore be
   pre-filled by the RPC server prior to responding, and XDR decode of
   the data at the client will not be required.  These chunks undergo a
   similar registration and advertisement via "write chunk lists" built
   as a part of XDR encoding.

   Some RPC client implementations are not able to determine where an
   RPC call's results reside during the "encode" phase.  This makes it
   difficult or impossible for the RPC client layer to encode the write
   chunk list at the time of building the request.  In this case, it is
   difficult for the RPC implementation to provide transparency to the
   RPC consumer, which may require recoding to provide result
   information at this earlier stage.

   Therefore, if the RPC client does not make a write chunk list
   available to receive the result, then the RPC server MAY return data
   inline in the reply, or if the upper-layer specification permits, it
   MAY be returned via a read chunk list.  It is NOT RECOMMENDED that
   upper-layer RPC client protocol specifications omit write chunk lists
   for eligible replies, due to the lower performance of the additional
   handshaking to perform data transfer, and the requirement that the
   RPC server must expose (and preserve) the reply data for a period of




Talpey & Callaghan           Standards Track                    [Page 9]

RFC 5666                 RDMA Transport for RPC             January 2010


   time.  In the absence of a server-provided read chunk list in the
   reply, if the encoded reply overflows the posted receive buffer, the
   RPC will fail with an RDMA transport error.

   When any data within a message is provided via either read or write
   chunks, the chunk itself refers only to the data portion of the XDR
   stream element.  In particular, for counted fields (e.g., a "<>"
   encoding) the byte count that is encoded as part of the field remains
   in the XDR stream, and is also encoded in the chunk list.  The data
   portion is however elided from the encoded XDR stream, and is
   transferred as part of chunk list processing.  It is important to
   maintain upper-layer implementation compatibility -- both the count
   and the data must be transferred as part of the logical XDR stream.
   While the chunk list processing results in the data being available
   to the upper-layer peer for XDR decoding, the length present in the
   chunk list entries is not.  Any byte count in the XDR stream MUST
   match the sum of the byte counts present in the corresponding read or
   write chunk list.  If they do not agree, an RPC protocol encoding
   error results.

   The following items are contained in a chunk list entry.

   Handle
        Steering tag or handle obtained when the chunk memory is
        registered for RDMA.

   Length
        The length of the chunk in bytes.

   Offset
        The offset or beginning memory address of the chunk.  In order
        to support the widest array of RDMA implementations, as well as
        the most general steering tag scheme, this field is
        unconditionally included in each chunk list entry.

        While zero-based offset schemes are available in many RDMA
        implementations, their use by RPC requires individual
        registration of each read or write chunk.  On many such
        implementations, this can be a significant overhead.  By
        providing an offset in each chunk, many pre-registration or
        region-based registrations can be readily supported, and by
        using a single, universal chunk representation, the RPC RDMA
        protocol implementation is simplified to its most general form.

   Position
        For data that is to be encoded, the position in the XDR stream
        where the chunk would normally reside.  Note that the chunk
        therefore inserts its data into the XDR stream at this position,



Talpey & Callaghan           Standards Track                   [Page 10]

RFC 5666                 RDMA Transport for RPC             January 2010


        but its transfer is no longer "inline".  Also note therefore
        that all chunks belonging to a single RPC argument or result
        will have the same position.  For data that is to be decoded, no
        position is used.

   When XDR marshaling is complete, the chunk list is XDR encoded, then
   sent to the receiver prepended to the RPC message.  Any source data
   for a read chunk, or the destination of a write chunk, remain behind
   in the sender's registered memory, and their actual payload is not
   marshaled into the request or reply.

      +----------------+----------------+-------------
      | RPC-over-RDMA  |                |
      |    header w/   |   RPC Header   | Non-chunk args/results
      |     chunks     |                |
      +----------------+----------------+-------------

   Read chunk lists and write chunk lists are structured somewhat
   differently.  This is due to the different usage -- read chunks are
   decoded and indexed by their argument's or result's position in the
   XDR data stream;  their size is always known.  Write chunks, on the
   other hand, are used only for results, and have neither a preassigned
   offset in the XDR stream nor a size until the results are produced,
   since the buffers may be only partially filled, or may not be used
   for results at all.  Their presence in the XDR stream is therefore
   not known until the reply is processed.  The mapping of write chunks
   onto designated NFS procedures and their results is described in
   [RFC5667].

   Therefore, read chunks are encoded into a read chunk list as a single
   array, with each entry tagged by its (known) size and its argument's
   or result's position in the XDR stream.  Write chunks are encoded as
   a list of arrays of RDMA buffers, with each list element (an array)
   providing buffers for a separate result.  Individual write chunk list
   elements MAY thereby result in being partially or fully filled, or in
   fact not being filled at all.  Unused write chunks, or unused bytes
   in write chunk buffer lists, are not returned as results, and their
   memory is returned to the upper layer as part of RPC completion.
   However, the RPC layer MUST NOT assume that the buffers have not been
   modified.

3.5.  XDR Decoding with Read Chunks



   The XDR decode process moves data from an XDR stream into a data
   structure provided by the RPC client or server application.  Where
   elements of the destination data structure are buffers or strings,
   the RPC application can either pre-allocate storage to receive the




Talpey & Callaghan           Standards Track                   [Page 11]

RFC 5666                 RDMA Transport for RPC             January 2010


   data or leave the string or buffer fields null and allow the XDR
   decode stage of RPC processing to automatically allocate storage of
   sufficient size.

   When decoding a message from an RDMA transport, the receiver first
   XDR decodes the chunk lists from the RPC-over-RDMA header, then
   proceeds to decode the body of the RPC message (arguments or
   results).  Whenever the XDR offset in the decode stream matches that
   of a chunk in the read chunk list, the XDR routine initiates an RDMA
   Read to bring over the chunk data into locally registered memory for
   the destination buffer.

   When processing an RPC request, the RPC receiver (RPC server)
   acknowledges its completion of use of the source buffers by simply
   replying to the RPC sender (client), and the peer may then free all
   source buffers advertised by the request.

   When processing an RPC reply, after completing such a transfer, the
   RPC receiver (client) MUST issue an RDMA_DONE message (described in
   Section 3.8) to notify the peer (server) that the source buffers can
   be freed.

   The read chunk list is constructed and used entirely within the
   RPC/XDR layer.  Other than specifying the minimum chunk size, the
   management of the read chunk list is automatic and transparent to an
   RPC application.

3.6.  XDR Decoding with Write Chunks



   When a write chunk list is provided for the results of the RPC call,
   the RPC server MUST provide any corresponding data via RDMA Write to
   the memory referenced in the chunk list entries.  The RPC reply
   conveys this by returning the write chunk list to the client with the
   lengths rewritten to match the actual transfer.  The XDR decode of
   the reply therefore performs no local data transfer but merely
   returns the length obtained from the reply.

   Each decoded result consumes one entry in the write chunk list, which
   in turn consists of an array of RDMA segments.  The length is
   therefore the sum of all returned lengths in all segments comprising
   the corresponding list entry.  As each list entry is decoded, the
   entire entry is consumed.

   The write chunk list is constructed and used by the RPC application.
   The RPC/XDR layer simply conveys the list between client and server
   and initiates the RDMA Writes back to the client.  The mapping of





Talpey & Callaghan           Standards Track                   [Page 12]

RFC 5666                 RDMA Transport for RPC             January 2010


   write chunk list entries to procedure arguments MUST be determined
   for each protocol.  An example of a mapping is described in
   [RFC5667].

3.7.  XDR Roundup and Chunks



   The XDR protocol requires 4-byte alignment of each new encoded
   element in any XDR stream.  This requirement is for efficiency and
   ease of decode/unmarshaling at the receiver -- if the XDR stream
   buffer begins on a native machine boundary, then the XDR elements
   will lie on similarly predictable offsets in memory.

   Within XDR, when non-4-byte encodes (such as an odd-length string or
   bulk data) are marshaled, their length is encoded literally, while
   their data is padded to begin the next element at a 4-byte boundary
   in the XDR stream.  For TCP or RDMA inline encoding, this minimal
   overhead is required because the transport-specific framing relies on
   the fact that the relative offset of the elements in the XDR stream
   from the start of the message determines the XDR position during
   decode.

   On the other hand, RPC/RDMA Read chunks carry the XDR position of
   each chunked element and length of the Chunk segment, and can be
   placed by the receiver exactly where they belong in the receiver's
   memory without regard to the alignment of their position in the XDR
   stream.  Since any rounded-up data is not actually part of the upper
   layer's message, the receiver will not reference it, and there is no
   reason to set it to any particular value in the receiver's memory.

   When roundup is present at the end of a sequence of chunks, the
   length of the sequence will terminate it at a non-4-byte XDR
   position.  When the receiver proceeds to decode the remaining part of
   the XDR stream, it inspects the XDR position indicated by the next
   chunk.  Because this position will not match (else roundup would not
   have occurred), the receiver decoding will fall back to inspecting
   the remaining inline portion.  If in turn, no data remains to be
   decoded from the inline portion, then the receiver MUST conclude that
   roundup is present, and therefore it advances the XDR decode position
   to that indicated by the next chunk (if any).  In this way, roundup
   is passed without ever actually transferring additional XDR bytes.

   Some protocol operations over RPC/RDMA, for instance NFS writes of
   data encountered at the end of a file or in direct I/O situations,
   commonly yield these roundups within RDMA Read Chunks.  Because any
   roundup bytes are not actually present in the data buffers being
   written, memory for these bytes would come from noncontiguous
   buffers, either as an additional memory registration segment or as an
   additional Chunk.  The overhead of these operations can be



Talpey & Callaghan           Standards Track                   [Page 13]

RFC 5666                 RDMA Transport for RPC             January 2010


   significant to both the sender to marshal them and even higher to the
   receiver to which to transfer them.  Senders SHOULD therefore avoid
   encoding individual RDMA Read Chunks for roundup whenever possible.
   It is acceptable, but not necessary, to include roundup data in an
   existing RDMA Read Chunk, but only if it is already present in the
   XDR stream to carry upper-layer data.

   Note that there is no exposure of additional data at the sender due
   to eliding roundup data from the XDR stream, since any additional
   sender buffers are never exposed to the peer.  The data is literally
   not there to be transferred.

   For RDMA Write Chunks, a simpler encoding method applies.  Again,
   roundup bytes are not transferred, instead the chunk length sent to
   the receiver in the reply is simply increased to include any roundup.
   Because of the requirement that the RDMA Write Chunks are filled
   sequentially without gaps, this situation can only occur on the final
   chunk receiving data.  Therefore, there is no opportunity for roundup
   data to insert misalignment or positional gaps into the XDR stream.

3.8.  RPC Call and Reply



   The RDMA transport for RPC provides three methods of moving data
   between RPC client and server:

   Inline
        Data is moved between RPC client and server within an RDMA Send.

   RDMA Read
        Data is moved between RPC client and server via an RDMA Read
        operation via steering tag; address and offset obtained from a
        read chunk list.

   RDMA Write
        Result data is moved from RPC server to client via an RDMA Write
        operation via steering tag; address and offset obtained from a
        write chunk list or reply chunk in the client's RPC call
        message.

   These methods of data movement may occur in combinations within a
   single RPC.  For instance, an RPC call may contain some inline data
   along with some large chunks to be transferred via RDMA Read to the
   server.  The reply to that call may have some result chunks that the
   server RDMA Writes back to the client.  The following protocol
   interactions illustrate RPC calls that use these methods to move RPC
   message data:





Talpey & Callaghan           Standards Track                   [Page 14]

RFC 5666                 RDMA Transport for RPC             January 2010


   An RPC with write chunks in the call message:

       RPC Client                           RPC Server
           |     RPC Call + Write Chunk list     |
      Send |   ------------------------------>   |
           |                                     |
           |               Chunk 1               |
           |   <------------------------------   | Write
           |                  :                  |
           |               Chunk n               |
           |   <------------------------------   | Write
           |                                     |
           |               RPC Reply             |
           |   <------------------------------   | Send

   In the presence of write chunks, RDMA ordering provides the guarantee
   that all data in the RDMA Write operations has been placed in memory
   prior to the client's RPC reply processing.

   An RPC with read chunks in the call message:

       RPC Client                           RPC Server
           |     RPC Call + Read Chunk list      |
      Send |   ------------------------------>   |
           |                                     |
           |               Chunk 1               |
           |   +------------------------------   | Read
           |   v----------------------------->   |
           |                  :                  |
           |               Chunk n               |
           |   +------------------------------   | Read
           |   v----------------------------->   |
           |                                     |
           |               RPC Reply             |
           |   <------------------------------   | Send
















Talpey & Callaghan           Standards Track                   [Page 15]

RFC 5666                 RDMA Transport for RPC             January 2010


   An RPC with read chunks in the reply message:

       RPC Client                           RPC Server
           |               RPC Call              |
      Send |   ------------------------------>   |
           |                                     |
           |     RPC Reply + Read Chunk list     |
           |   <------------------------------   | Send
           |                                     |
           |               Chunk 1               |
      Read |   ------------------------------+   |
           |   <-----------------------------v   |
           |                  :                  |
           |               Chunk n               |
      Read |   ------------------------------+   |
           |   <-----------------------------v   |
           |                                     |
           |                 Done                |
      Send |   ------------------------------>   |

   The final Done message allows the RPC client to signal the server
   that it has received the chunks, so the server can de-register and
   free the memory holding the chunks.  A Done completion is not
   necessary for an RPC call, since the RPC reply Send is itself a
   receive completion notification.  In the event that the client fails
   to return the Done message within some timeout period, the server MAY
   conclude that a protocol violation has occurred and close the RPC
   connection, or it MAY proceed with a de-register and free its chunk
   buffers.  This may result in a fatal RDMA error if the client later
   attempts to perform an RDMA Read operation, which amounts to the same
   thing.

   The use of read chunks in RPC reply messages is much less efficient
   than providing write chunks in the originating RPC calls, due to the
   additional message exchanges, the need for the RPC server to
   advertise buffers to the peer, the necessity of the server
   maintaining a timer for the purpose of recovery from misbehaving
   clients, and the need for additional memory registration.  Their use
   is NOT RECOMMENDED by upper layers where efficiency is a primary
   concern [RFC5667].  However, they MAY be employed by upper-layer
   protocol bindings that are primarily concerned with transparency,
   since they can frequently be implemented completely within the RPC
   lower layers.

   It is important to note that the Done message consumes a credit at
   the RPC server.  The RPC server SHOULD provide sufficient credits to
   the client to allow the Done message to be sent without deadlock
   (driving the outstanding credit count to zero).  The RPC client MUST



Talpey & Callaghan           Standards Track                   [Page 16]

RFC 5666                 RDMA Transport for RPC             January 2010


   account for its required Done messages to the server in its
   accounting of available credits, and the server SHOULD replenish any
   credit consumed by its use of such exchanges at its earliest
   opportunity.

   Finally, it is possible to conceive of RPC exchanges that involve any
   or all combinations of write chunks in the RPC call, read chunks in
   the RPC call, and read chunks in the RPC reply.  Support for such
   exchanges is straightforward from a protocol perspective, but in
   practice such exchanges would be quite rare, limited to upper-layer
   protocol exchanges that transferred bulk data in both the call and
   corresponding reply.

3.9.  Padding



   Alignment of specific opaque data enables certain scatter/gather
   optimizations.  Padding leverages the useful property that RDMA
   transfers preserve alignment of data, even when they are placed into
   pre-posted receive buffers by Sends.

   Many servers can make good use of such padding.  Padding allows the
   chaining of RDMA receive buffers such that any data transferred by
   RDMA on behalf of RPC requests will be placed into appropriately
   aligned buffers on the system that receives the transfer.  In this
   way, the need for servers to perform RDMA Read to satisfy all but the
   largest client writes is obviated.

   The effect of padding is demonstrated below showing prior bytes on an
   XDR stream ("XXX" in the figure below) followed by an opaque field
   consisting of four length bytes ("LLLL") followed by data bytes
   ("DDD").  The receiver of the RDMA Send has posted two chained
   receive buffers.  Without padding, the opaque data is split across
   the two buffers.  With the addition of padding bytes ("ppp") prior to
   the first data byte, the data can be forced to align correctly in the
   second buffer.

                                            Buffer 1       Buffer 2
      Unpadded                           --------------  --------------


       XXXXXXXLLLLDDDDDDDDDDDDDD    ---> XXXXXXXLLLLDDD  DDDDDDDDDDD


      Padded


       XXXXXXXLLLLpppDDDDDDDDDDDDDD ---> XXXXXXXLLLLppp  DDDDDDDDDDDDDD




Talpey & Callaghan           Standards Track                   [Page 17]

RFC 5666                 RDMA Transport for RPC             January 2010


   Padding is implemented completely within the RDMA transport encoding,
   flagged with a specific message type.  Where padding is applied, two
   values are passed to the peer:  an "rdma_align", which is the padding
   value used, and "rdma_thresh", which is the opaque data size at or
   above which padding is applied.  For instance, if the server is using
   chained 4 KB receive buffers, then up to (4 KB - 1) padding bytes
   could be used to achieve alignment of the data.  The XDR routine at
   the peer MUST consult these values when decoding opaque values.
   Where the decoded length exceeds the rdma_thresh, the XDR decode MUST
   skip over the appropriate padding as indicated by rdma_align and the
   current XDR stream position.

4.  RPC RDMA Message Layout



   RPC call and reply messages are conveyed across an RDMA transport
   with a prepended RPC-over-RDMA header.  The RPC-over-RDMA header
   includes data for RDMA flow control credits, padding parameters, and
   lists of addresses that provide direct data placement via RDMA Read
   and Write operations.  The layout of the RPC message itself is
   unchanged from that described in [RFC5531] except for the possible
   exclusion of large data chunks that will be moved by RDMA Read or
   Write operations.  If the RPC message (along with the RPC-over-RDMA
   header) is too long for the posted receive buffer (even after any
   large chunks are removed), then the entire RPC message MAY be moved
   separately as a chunk, leaving just the RPC-over-RDMA header in the
   RDMA Send.

4.1.  RPC-over-RDMA Header



   The RPC-over-RDMA header begins with four 32-bit fields that are
   always present and that control the RDMA interaction including RDMA-
   specific flow control.  These are then followed by a number of items
   such as chunk lists and padding that MAY or MUST NOT be present
   depending on the type of transmission.  The four fields that are
   always present are:

   1. Transaction ID (XID).
      The XID generated for the RPC call and reply.  Having the XID at
      the beginning of the message makes it easy to establish the
      message context.  This XID MUST be the same as the XID in the RPC
      header.  The receiver MAY perform its processing based solely on
      the XID in the RPC-over-RDMA header, and thereby ignore the XID in
      the RPC header, if it so chooses.

   2. Version number.
      This version of the RPC RDMA message protocol is 1.  The version
      number MUST be increased by 1 whenever the format of the RPC RDMA
      messages is changed.



Talpey & Callaghan           Standards Track                   [Page 18]

RFC 5666                 RDMA Transport for RPC             January 2010


   3. Flow control credit value.
      When sent in an RPC call message, the requested value is provided.
      When sent in an RPC reply message, the granted value is returned.
      RPC calls SHOULD NOT be sent in excess of the currently granted
      limit.

   4. Message type.

      o  RDMA_MSG = 0 indicates that chunk lists and RPC message follow.

      o  RDMA_NOMSG = 1 indicates that after the chunk lists there is no
         RPC message.  In this case, the chunk lists provide information
         to allow the message proper to be transferred using RDMA Read
         or Write and thus is not appended to the RPC-over-RDMA header.

      o  RDMA_MSGP = 2 indicates that a chunk list and RPC message with
         some padding follow.

      o  RDMA_DONE = 3 indicates that the message signals the completion
         of a chunk transfer via RDMA Read.

      o  RDMA_ERROR = 4 is used to signal any detected error(s) in the
         RPC RDMA chunk encoding.

   Because the version number is encoded as part of this header, and the
   RDMA_ERROR message type is used to indicate errors, these first four
   fields and the start of the following message body MUST always remain
   aligned at these fixed offsets for all versions of the RPC-over-RDMA
   header.

   For a message of type RDMA_MSG or RDMA_NOMSG, the Read and Write
   chunk lists follow.  If the Read chunk list is null (a 32-bit word of
   zeros), then there are no chunks to be transferred separately and the
   RPC message follows in its entirety.  If non-null, then it's the
   beginning of an XDR encoded sequence of Read chunk list entries.  If
   the Write chunk list is non-null, then an XDR encoded sequence of
   Write chunk entries follows.

   If the message type is RDMA_MSGP, then two additional fields that
   specify the padding alignment and threshold are inserted prior to the
   Read and Write chunk lists.

   A header of message type RDMA_MSG or RDMA_MSGP MUST be followed by
   the RPC call or RPC reply message body, beginning with the XID.  The
   XID in the RDMA_MSG or RDMA_MSGP header MUST match this.






Talpey & Callaghan           Standards Track                   [Page 19]

RFC 5666                 RDMA Transport for RPC             January 2010


   +--------+---------+---------+-----------+-------------+----------
   |        |         |         | Message   |   NULLs     | RPC Call
   |  XID   | Version | Credits |  Type     |    or       |    or
   |        |         |         |           | Chunk Lists | Reply Msg
   +--------+---------+---------+-----------+-------------+----------

   Note that in the case of RDMA_DONE and RDMA_ERROR, no chunk list or
   RPC message follows.  As an implementation hint: a gather operation
   on the Send of the RDMA RPC message can be used to marshal the
   initial header, the chunk list, and the RPC message itself.

4.2.  RPC-over-RDMA Header Errors



   When a peer receives an RPC RDMA message, it MUST perform the
   following basic validity checks on the header and chunk contents.  If
   such errors are detected in the request, an RDMA_ERROR reply MUST be
   generated.

   Two types of errors are defined, version mismatch and invalid chunk
   format.  When the peer detects an RPC-over-RDMA header version that
   it does not support (currently this document defines only version 1),
   it replies with an error code of ERR_VERS, and provides the low and
   high inclusive version numbers it does, in fact, support.  The
   version number in this reply MUST be any value otherwise valid at the
   receiver.  When other decoding errors are detected in the header or
   chunks, either an RPC decode error MAY be returned or the RPC/RDMA
   error code ERR_CHUNK MUST be returned.

4.3.  XDR Language Description



   Here is the message layout in XDR language.

      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_read_chunk {
         uint32 position;        /* Position in XDR stream */
         struct xdr_rdma_segment target;
      };

      struct xdr_read_list {
         struct xdr_read_chunk entry;
         struct xdr_read_list  *next;
      };




Talpey & Callaghan           Standards Track                   [Page 20]

RFC 5666                 RDMA Transport for RPC             January 2010


      struct xdr_write_chunk {
         struct xdr_rdma_segment target<>;
      };

      struct xdr_write_list {
         struct xdr_write_chunk entry;
         struct xdr_write_list  *next;
      };

      struct rdma_msg {
         uint32    rdma_xid;     /* Mirrors the RPC header xid */
         uint32    rdma_vers;    /* Version of this protocol */
         uint32    rdma_credit;  /* Buffers requested/granted */
         rdma_body rdma_body;
      };

      enum rdma_proc {
         RDMA_MSG=0,   /* An RPC call or reply msg */
         RDMA_NOMSG=1, /* An RPC call or reply msg - separate body */
         RDMA_MSGP=2,  /* An RPC call or reply msg with padding */
         RDMA_DONE=3,  /* Client signals reply completion */
         RDMA_ERROR=4  /* An RPC RDMA encoding error */
      };

      union rdma_body switch (rdma_proc proc) {
         case RDMA_MSG:
           rpc_rdma_header rdma_msg;
         case RDMA_NOMSG:
           rpc_rdma_header_nomsg rdma_nomsg;
         case RDMA_MSGP:
           rpc_rdma_header_padded rdma_msgp;
         case RDMA_DONE:
           void;
         case RDMA_ERROR:
           rpc_rdma_error rdma_error;
      };

      struct rpc_rdma_header {
         struct xdr_read_list   *rdma_reads;
         struct xdr_write_list  *rdma_writes;
         struct xdr_write_chunk *rdma_reply;
         /* rpc body follows */
      };

      struct rpc_rdma_header_nomsg {
         struct xdr_read_list   *rdma_reads;
         struct xdr_write_list  *rdma_writes;
         struct xdr_write_chunk *rdma_reply;



Talpey & Callaghan           Standards Track                   [Page 21]

RFC 5666                 RDMA Transport for RPC             January 2010


      };

      struct rpc_rdma_header_padded {
         uint32                 rdma_align;   /* Padding alignment */
         uint32                 rdma_thresh;  /* Padding threshold */
         struct xdr_read_list   *rdma_reads;
         struct xdr_write_list  *rdma_writes;
         struct xdr_write_chunk *rdma_reply;
         /* rpc body follows */
      };

      enum rpc_rdma_errcode {
         ERR_VERS = 1,
         ERR_CHUNK = 2
      };

      union rpc_rdma_error switch (rpc_rdma_errcode err) {
         case ERR_VERS:
           uint32               rdma_vers_low;
           uint32               rdma_vers_high;
         case ERR_CHUNK:
           void;
         default:
           uint32               rdma_extra[8];
      };

5.  Long Messages



   The receiver of RDMA Send messages is required by RDMA to have
   previously posted one or more adequately sized buffers.  The RPC
   client can inform the server of the maximum size of its RDMA Send
   messages via the Connection Configuration Protocol described later in
   this document.

   Since RPC messages are frequently small, memory savings can be
   achieved by posting small buffers.  Even large messages like NFS READ
   or WRITE will be quite small once the chunks are removed from the
   message.  However, there may be large messages that would demand a
   very large buffer be posted, where the contents of the buffer may not
   be a chunkable XDR element.  A good example is an NFS READDIR reply,
   which may contain a large number of small filename strings.  Also,
   the NFS version 4 protocol [RFC3530] features COMPOUND request and
   reply messages of unbounded length.

   Ideally, each upper layer will negotiate these limits.  However, it
   is frequently necessary to provide a transparent solution.





Talpey & Callaghan           Standards Track                   [Page 22]

RFC 5666                 RDMA Transport for RPC             January 2010


5.1.  Message as an RDMA Read Chunk



   One relatively simple method is to have the client identify any RPC
   message that exceeds the RPC server's posted buffer size and move it
   separately as a chunk, i.e., reference it as the first entry in the
   read chunk list with an XDR position of zero.

   Normal Message

   +--------+---------+---------+------------+-------------+----------
   |        |         |         |            |             | RPC Call
   |  XID   | Version | Credits |  RDMA_MSG  | Chunk Lists |    or
   |        |         |         |            |             | Reply Msg
   +--------+---------+---------+------------+-------------+----------

   Long Message

   +--------+---------+---------+------------+-------------+
   |        |         |         |            |             |
   |  XID   | Version | Credits | RDMA_NOMSG | Chunk Lists |
   |        |         |         |            |             |
   +--------+---------+---------+------------+-------------+
                                                |
                                                |  +----------
                                                |  | Long RPC Call
                                                +->|    or
                                                   | Reply Message
                                                   +----------

   If the receiver gets an RPC-over-RDMA header with a message type of
   RDMA_NOMSG and finds an initial read chunk list entry with a zero XDR
   position, it allocates a registered buffer and issues an RDMA Read of
   the long RPC message into it.  The receiver then proceeds to XDR
   decode the RPC message as if it had received it inline with the Send
   data.  Further decoding may issue additional RDMA Reads to bring over
   additional chunks.

   Although the handling of long messages requires one extra network
   turnaround, in practice these messages will be rare if the posted
   receive buffers are correctly sized, and of course they will be
   non-existent for RDMA-aware upper layers.










Talpey & Callaghan           Standards Track                   [Page 23]

RFC 5666                 RDMA Transport for RPC             January 2010


   A long call RPC with request supplied via RDMA Read

       RPC Client                           RPC Server
           |        RDMA-over-RPC Header         |
      Send |   ------------------------------>   |
           |                                     |
           |          Long RPC Call Msg          |
           |   +------------------------------   | Read
           |   v----------------------------->   |
           |                                     |
           |         RDMA-over-RPC Reply         |
           |   <------------------------------   | Send

   An RPC with long reply returned via RDMA Read

       RPC Client                           RPC Server
           |             RPC Call                |
      Send |   ------------------------------>   |
           |                                     |
           |         RDMA-over-RPC Header        |
           |   <------------------------------   | Send
           |                                     |
           |          Long RPC Reply Msg         |
      Read |   ------------------------------+   |
           |   <-----------------------------v   |
           |                                     |
           |                Done                 |
      Send |   ------------------------------>   |

   It is possible for a single RPC procedure to employ both a long call
   for its arguments and a long reply for its results.  However, such an
   operation is atypical, as few upper layers define such exchanges.

5.2.  RDMA Write of Long Replies (Reply Chunks)



   A superior method of handling long RPC replies is to have the RPC
   client post a large buffer into which the server can write a large
   RPC reply.  This has the advantage that an RDMA Write may be slightly
   faster in network latency than an RDMA Read, and does not require the
   server to wait for the completion as it must for RDMA Read.
   Additionally, for a reply it removes the need for an RDMA_DONE
   message if the large reply is returned as a Read chunk.

   This protocol supports direct return of a large reply via the
   inclusion of an OPTIONAL rdma_reply write chunk after the read chunk
   list and the write chunk list.  The client allocates a buffer sized
   to receive a large reply and enters its steering tag, address and
   length in the rdma_reply write chunk.  If the reply message is too



Talpey & Callaghan           Standards Track                   [Page 24]

RFC 5666                 RDMA Transport for RPC             January 2010


   long to return inline with an RDMA Send (exceeds the size of the
   client's posted receive buffer), even with read chunks removed, then
   the RPC server performs an RDMA Write of the RPC reply message into
   the buffer indicated by the rdma_reply chunk.  If the client doesn't
   provide an rdma_reply chunk, or if it's too small, then if the upper-
   layer specification permits, the message MAY be returned as a Read
   chunk.

   An RPC with long reply returned via RDMA Write


    RPC Client                           RPC Server
        |      RPC Call with rdma_reply       |
   Send |   ------------------------------>   |
        |                                     |
        |          Long RPC Reply Msg         |
        |   <------------------------------   | Write
        |                                     |
        |         RDMA-over-RPC Header        |
        |   <------------------------------   | Send

   The use of RDMA Write to return long replies requires that the client
   applications anticipate a long reply and have some knowledge of its
   size so that an adequately sized buffer can be allocated.  This is
   certainly true of NFS READDIR replies; where the client already
   provides an upper bound on the size of the encoded directory fragment
   to be returned by the server.

   The use of these "reply chunks" is highly efficient and convenient
   for both RPC client and server.  Their use is encouraged for eligible
   RPC operations such as NFS READDIR, which would otherwise require
   extensive chunk management within the results or use of RDMA Read and
   a Done message [RFC5667].

6.  Connection Configuration Protocol



   RDMA Send operations require the receiver to post one or more buffers
   at the RDMA connection endpoint, each large enough to receive the
   largest Send message.  Buffers are consumed as Send messages are
   received.  If a buffer is too small, or if there are no buffers
   posted, the RDMA transport MAY return an error and break the RDMA
   connection.  The receiver MUST post sufficient, adequately buffers to
   avoid buffer overrun or capacity errors.

   The protocol described above includes only a mechanism for managing
   the number of such receive buffers and no explicit features to allow
   the RPC client and server to provision or control buffer sizing, nor
   any other session parameters.



Talpey & Callaghan           Standards Track                   [Page 25]

RFC 5666                 RDMA Transport for RPC             January 2010


   In the past, this type of connection management has not been
   necessary for RPC.  RPC over UDP or TCP does not have a protocol to
   negotiate the link.  The server can get a rough idea of the maximum
   size of messages from the server protocol code.  However, a protocol
   to negotiate transport features on a more dynamic basis is desirable.

   The Connection Configuration Protocol allows the client to pass its
   connection requirements to the server, and allows the server to
   inform the client of its connection limits.

   Use of the Connection Configuration Protocol by an upper layer is
   OPTIONAL.

6.1.  Initial Connection State



   This protocol MAY be used for connection setup prior to the use of
   another RPC protocol that uses the RDMA transport.  It operates
   in-band, i.e., it uses the connection itself to negotiate the
   connection parameters.  To provide a basis for connection
   negotiation, the connection is assumed to provide a basic level of
   interoperability: the ability to exchange at least one RPC message at
   a time that is at least 1 KB in size.  The server MAY exceed this
   basic level of configuration, but the client MUST NOT assume more
   than one, and MUST receive a valid reply from the server carrying the
   actual number of available receive messages, prior to sending its
   next request.

6.2.  Protocol Description



   Version 1 of the Connection Configuration Protocol consists of a
   single procedure that allows the client to inform the server of its
   connection requirements and the server to return connection
   information to the client.

   The maxcall_sendsize argument is the maximum size of an RPC call
   message that the client MAY send inline in an RDMA Send message to
   the server.  The server MAY return a maxcall_sendsize value that is
   smaller or larger than the client's request.  The client MUST NOT
   send an inline call message larger than what the server will accept.
   The maxcall_sendsize limits only the size of inline RPC calls.  It
   does not limit the size of long RPC messages transferred as an
   initial chunk in the Read chunk list.

   The maxreply_sendsize is the maximum size of an inline RPC message
   that the client will accept from the server.






Talpey & Callaghan           Standards Track                   [Page 26]

RFC 5666                 RDMA Transport for RPC             January 2010


   The maxrdmaread is the maximum number of RDMA Reads that may be
   active at the peer.  This number correlates to the RDMA incoming RDMA
   Read count ("IRD") configured into each originating endpoint by the
   client or server.  If more than this number of RDMA Read operations
   by the connected peer are issued simultaneously, connection loss or
   suboptimal flow control may result; therefore, the value SHOULD be
   observed at all times.  The peers' values need not be equal.  If
   zero, the peer MUST NOT issue requests that require RDMA Read to
   satisfy, as no transfer will be possible.

   The align value is the value recommended by the server for opaque
   data values such as strings and counted byte arrays.  The client MAY
   use this value to compute the number of prepended pad bytes when XDR
   encoding opaque values in the RPC call message.

      typedef unsigned int uint32;

      struct config_rdma_req {
           uint32  maxcall_sendsize;
                       /* max size of inline RPC call */
           uint32  maxreply_sendsize;
                       /* max size of inline RPC reply */
           uint32  maxrdmaread;
                       /* max active RDMA Reads at client */
      };

      struct config_rdma_reply {
           uint32  maxcall_sendsize;
                       /* max call size accepted by server */
           uint32  align;
                       /* server's receive buffer alignment */
           uint32  maxrdmaread;
                       /* max active RDMA Reads at server */
      };

      program CONFIG_RDMA_PROG {
         version VERS1 {
            /*
             * Config call/reply
             */
            config_rdma_reply CONF_RDMA(config_rdma_req) = 1;
         } = 1;
      } = 100417;








Talpey & Callaghan           Standards Track                   [Page 27]

RFC 5666                 RDMA Transport for RPC             January 2010


7.  Memory Registration Overhead



   RDMA requires that all data be transferred between registered memory
   regions at the source and destination.  All protocol headers as well
   as separately transferred data chunks use registered memory.  Since
   the cost of registering and de-registering memory can be a large
   proportion of the RDMA transaction cost, it is important to minimize
   registration activity.  This is easily achieved within RPC controlled
   memory by allocating chunk list data and RPC headers in a reusable
   way from pre-registered pools.

   The data chunks transferred via RDMA MAY occupy memory that persists
   outside the bounds of the RPC transaction.  Hence, the default
   behavior of an RPC-over-RDMA transport is to register and de-register
   these chunks on every transaction.  However, this is not a limitation
   of the protocol -- only of the existing local RPC API.  The API is
   easily extended through such functions as rpc_control(3) to change
   the default behavior so that the application can assume
   responsibility for controlling memory registration through an RPC-
   provided registered memory allocator.

8.  Errors and Error Recovery



   RPC RDMA protocol errors are described in Section 4.  RPC errors and
   RPC error recovery are not affected by the protocol, and proceed as
   for any RPC error condition.  RDMA transport error reporting and
   recovery are outside the scope of this protocol.

   It is assumed that the link itself will provide some degree of error
   detection and retransmission.  iWARP's Marker PDU Aligned (MPA) layer
   (when used over TCP), Stream Control Transmission Protocol (SCTP), as
   well as the InfiniBand link layer all provide Cyclic Redundancy Check
   (CRC) protection of the RDMA payload, and CRC-class protection is a
   general attribute of such transports.  Additionally, the RPC layer
   itself can accept errors from the link level and recover via
   retransmission.  RPC recovery can handle complete loss and
   re-establishment of the link.

   See Section 11 for further discussion of the use of RPC-level
   integrity schemes to detect errors and related efficiency issues.

9.  Node Addressing



   In setting up a new RDMA connection, the first action by an RPC
   client will be to obtain a transport address for the server.  The
   mechanism used to obtain this address, and to open an RDMA connection
   is dependent on the type of RDMA transport, and is the responsibility
   of each RPC protocol binding and its local implementation.



Talpey & Callaghan           Standards Track                   [Page 28]

RFC 5666                 RDMA Transport for RPC             January 2010


10.  RPC Binding



   RPC services normally register with a portmap or rpcbind [RFC1833]
   service, which associates an RPC program number with a service
   address.  (In the case of UDP or TCP, the service address for NFS is
   normally port 2049.)  This policy is no different with RDMA
   interconnects, although it may require the allocation of port numbers
   appropriate to each upper-layer binding that uses the RPC framing
   defined here.

   When mapped atop the iWARP [RFC5040, RFC5041] transport, which uses
   IP port addressing due to its layering on TCP and/or SCTP, port
   mapping is trivial and consists merely of issuing the port in the
   connection process.  The NFS/RDMA protocol service address has been
   assigned port 20049 by IANA, for both iWARP/TCP and iWARP/SCTP.

   When mapped atop InfiniBand [IB], which uses a Group Identifier
   (GID)-based service endpoint naming scheme, a translation MUST be
   employed.  One such translation is defined in the InfiniBand Port
   Addressing Annex [IBPORT], which is appropriate for translating IP
   port addressing to the InfiniBand network.  Therefore, in this case,
   IP port addressing may be readily employed by the upper layer.

   When a mapping standard or convention exists for IP ports on an RDMA
   interconnect, there are several possibilities for each upper layer to
   consider:

      One possibility is to have an upper-layer server register its
      mapped IP port with the rpcbind service, under the netid (or
      netid's) defined here.  An RPC/RDMA-aware client can then resolve
      its desired service to a mappable port, and proceed to connect.
      This is the most flexible and compatible approach, for those upper
      layers that are defined to use the rpcbind service.

      A second possibility is to have the server's portmapper register
      itself on the RDMA interconnect at a "well known" service address.
      (On UDP or TCP, this corresponds to port 111.)  A client could
      connect to this service address and use the portmap protocol to
      obtain a service address in response to a program number, e.g., an
      iWARP port number, or an InfiniBand GID.

      Alternatively, the client could simply connect to the mapped well-
      known port for the service itself, if it is appropriately defined.
      By convention, the NFS/RDMA service, when operating atop such an
      InfiniBand fabric, will use the same 20049 assignment as for
      iWARP.





Talpey & Callaghan           Standards Track                   [Page 29]

RFC 5666                 RDMA Transport for RPC             January 2010


   Historically, different RPC protocols have taken different approaches
   to their port assignment; therefore, the specific method is left to
   each RPC/RDMA-enabled upper-layer binding, and not addressed here.

   In Section 12, "IANA Considerations", this specification defines two
   new "netid" values, to be used for registration of upper layers atop
   iWARP [RFC5040, RFC5041] and (when a suitable port translation
   service is available) InfiniBand [IB].  Additional RDMA-capable
   networks MAY define their own netids, or if they provide a port
   translation, MAY share the one defined here.

11.  Security Considerations



   RPC provides its own security via the RPCSEC_GSS framework [RFC2203].
   RPCSEC_GSS can provide message authentication, integrity checking,
   and privacy.  This security mechanism will be unaffected by the RDMA
   transport.  The data integrity and privacy features alter the body of
   the message, presenting it as a single chunk.  For large messages the
   chunk may be large enough to qualify for RDMA Read transfer.
   However, there is much data movement associated with computation and
   verification of integrity, or encryption/decryption, so certain
   performance advantages may be lost.

   For efficiency, a more appropriate security mechanism for RDMA links
   may be link-level protection, such as certain configurations of
   IPsec, which may be co-located in the RDMA hardware.  The use of
   link-level protection MAY be negotiated through the use of the new
   RPCSEC_GSS mechanism defined in [RFC5403] in conjunction with the
   Channel Binding mechanism [RFC5056] and IPsec Channel Connection
   Latching [RFC5660].  Use of such mechanisms is REQUIRED where
   integrity and/or privacy is desired, and where efficiency is
   required.

   An additional consideration is the protection of the integrity and
   privacy of local memory by the RDMA transport itself.  The use of
   RDMA by RPC MUST NOT introduce any vulnerabilities to system memory
   contents, or to memory owned by user processes.  These protections
   are provided by the RDMA layer specifications, and specifically their
   security models.  It is REQUIRED that any RDMA provider used for RPC
   transport be conformant to the requirements of [RFC5042] in order to
   satisfy these protections.

   Once delivered securely by the RDMA provider, any RDMA-exposed
   addresses will contain only RPC payloads in the chunk lists,
   transferred under the protection of RPCSEC_GSS integrity and privacy.
   By these means, the data will be protected end-to-end, as required by
   the RPC layer security model.




Talpey & Callaghan           Standards Track                   [Page 30]

RFC 5666                 RDMA Transport for RPC             January 2010


   Where upper-layer protocols choose to supply results to the requester
   via read chunks, a server resource deficit can arise if the client
   does not promptly acknowledge their status via the RDMA_DONE message.
   This can potentially lead to a denial-of-service situation, with a
   single client unfairly (and unnecessarily) consuming server RDMA
   resources.  Servers for such upper-layer protocols MUST protect
   against this situation, originating from one or many clients.  For
   example, a time-based window of buffer availability may be offered,
   if the client fails to obtain the data within the window, it will
   simply retry using ordinary RPC retry semantics.  Or, a more severe
   method would be for the server to simply close the client's RDMA
   connection, freeing the RDMA resources and allowing the server to
   reclaim them.

   A fairer and more useful method is provided by the protocol itself.
   The server MAY use the rdma_credit value to limit the number of
   outstanding requests for each client.  By including the number of
   outstanding RDMA_DONE completions in the computation of available
   client credits, the server can limit its exposure to each client, and
   therefore provide uninterrupted service as its resources permit.

   However, the server must ensure that it does not decrease the credit
   count to zero with this method, since the RDMA_DONE message is not
   acknowledged.  If the credit count were to drop to zero solely due to
   outstanding RDMA_DONE messages, the client would deadlock since it
   would never obtain a new credit with which to continue.  Therefore,
   if the server adjusts credits to zero for outstanding RDMA_DONE, it
   MUST withhold its reply to at least one message in order to provide
   the next credit.  The time-based window (or any other appropriate
   method) SHOULD be used by the server to recover resources in the
   event that the client never returns.

   The Connection Configuration Protocol, when used, MUST be protected
   by an appropriate RPC security flavor, to ensure it is not attacked
   in the process of initiating an RPC/RDMA connection.

12.  IANA Considerations



   Three new assignments are specified by this document:

   - A new set of RPC "netids" for resolving RPC/RDMA services

   - Optional service port assignments for upper-layer bindings

   - An RPC program number assignment for the configuration protocol

   These assignments have been established, as below.




Talpey & Callaghan           Standards Track                   [Page 31]

RFC 5666                 RDMA Transport for RPC             January 2010


   The new RPC transport has been assigned an RPC "netid", which is an
   rpcbind [RFC1833] string used to describe the underlying protocol in
   order for RPC to select the appropriate transport framing, as well as
   the format of the service addresses and ports.

   The following "Netid" registry strings are defined for this purpose:

      NC_RDMA "rdma"
      NC_RDMA6 "rdma6"

   These netids MAY be used for any RDMA network satisfying the
   requirements of Section 2, and able to identify service endpoints
   using IP port addressing, possibly through use of a translation
   service as described above in Section 10, "RPC Binding".  The "rdma"
   netid is to be used when IPv4 addressing is employed by the
   underlying transport, and "rdma6" for IPv6 addressing.

   The netid assignment policy and registry are defined in [RFC5665].

   As a new RPC transport, this protocol has no effect on RPC program
   numbers or existing registered port numbers.  However, new port
   numbers MAY be registered for use by RPC/RDMA-enabled services, as
   appropriate to the new networks over which the services will operate.

   For example, the NFS/RDMA service defined in [RFC5667] has been
   assigned the port 20049, in the IANA registry:

      nfsrdma 20049/tcp Network File System (NFS) over RDMA
      nfsrdma 20049/udp Network File System (NFS) over RDMA
      nfsrdma 20049/sctp Network File System (NFS) over RDMA

   The OPTIONAL Connection Configuration Protocol described herein
   requires an RPC program number assignment.  The value "100417" has
   been assigned:

      rdmaconfig 100417 rpc.rdmaconfig

   The RPC program number assignment policy and registry are defined in
   [RFC5531].

13.  Acknowledgments



   The authors wish to thank Rob Thurlow, John Howard, Chet Juszczak,
   Alex Chiu, Peter Staubach, Dave Noveck, Brian Pawlowski, Steve
   Kleiman, Mike Eisler, Mark Wittle, Shantanu Mehendale, David
   Robinson, and Mallikarjun Chadalapaka for their contributions to this
   document.




Talpey & Callaghan           Standards Track                   [Page 32]

RFC 5666                 RDMA Transport for RPC             January 2010


14.  References



14.1.  Normative References



   [RFC1833]  Srinivasan, R., "Binding Protocols for ONC RPC Version 2",
              RFC 1833, August 1995.

   [RFC2203]  Eisler, M., Chiu, A., and L. Ling, "RPCSEC_GSS Protocol
              Specification", RFC 2203, September 1997.

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119, March 1997.

   [RFC4506]  Eisler, M., Ed., "XDR: External Data Representation
              Standard", STD 67, RFC 4506, May 2006.

   [RFC5042]  Pinkerton, J. and E. Deleganes, "Direct Data Placement
              Protocol (DDP) / Remote Direct Memory Access Protocol
              (RDMAP) Security", RFC 5042, October 2007.

   [RFC5056]  Williams, N., "On the Use of Channel Bindings to Secure
              Channels", RFC 5056, November 2007.

   [RFC5403]  Eisler, M., "RPCSEC_GSS Version 2", RFC 5403, February
              2009.

   [RFC5531]  Thurlow, R., "RPC: Remote Procedure Call Protocol
              Specification Version 2", RFC 5531, May 2009.

   [RFC5660]  Williams, N., "IPsec Channels: Connection Latching", RFC
              5660, October 2009.

   [RFC5665]  Eisler, M., "IANA Considerations for Remote Procedure Call
              (RPC) Network Identifiers and Universal Address Formats",
              RFC 5665, January 2010.

14.2.  Informative References



   [RFC1094]  Sun Microsystems, "NFS: Network File System Protocol
              specification", RFC 1094, March 1989.

   [RFC1813]  Callaghan, B., Pawlowski, B., and P. Staubach, "NFS
              Version 3 Protocol Specification", RFC 1813, June 1995.

   [RFC3530]  Shepler, S., Callaghan, B., Robinson, D., Thurlow, R.,
              Beame, C., Eisler, M., and D. Noveck, "Network File System
              (NFS) version 4 Protocol", RFC 3530, April 2003.




Talpey & Callaghan           Standards Track                   [Page 33]

RFC 5666                 RDMA Transport for RPC             January 2010


   [RFC5040]  Recio, R., Metzler, B., Culley, P., Hilland, J., and D.
              Garcia, "A Remote Direct Memory Access Protocol
              Specification", RFC 5040, October 2007.

   [RFC5041]  Shah, H., Pinkerton, J., Recio, R., and P. Culley, "Direct
              Data Placement over Reliable Transports", RFC 5041,
              October 2007.

   [RFC5532]  Talpey, T. and C. Juszczak, "Network File System (NFS)
              Remote Direct Memory Access (RDMA) Problem Statement", RFC
              5532, May 2009.

   [RFC5661]  Shepler, S., Ed., Eisler, M., Ed., and D. Noveck, Ed.,
              "Network File System Version 4 Minor Version 1 Protocol",
              RFC 5661, January 2010.

   [RFC5667]  Talpey, T. and B. Callaghan, "Network File System (NFS)
              Direct Data Placement", RFC 5667, January 2010.

   [IB]       InfiniBand Trade Association, InfiniBand Architecture
              Specifications, available from
              http://www.infinibandta.org.

   [IBPORT]   InfiniBand Trade Association, "IP Addressing Annex",
              available from http://www.infinibandta.org.

Authors' Addresses



   Tom Talpey
   170 Whitman St.
   Stow, MA 01775 USA

   EMail: tmtalpey@gmail.com


   Brent Callaghan
   Apple Computer, Inc.
   MS: 302-4K
   2 Infinite Loop
   Cupertino, CA 95014 USA

   EMail: brentc@apple.com









Talpey & Callaghan           Standards Track                   [Page 34]