RFC 5044






Network Working Group                                          P. Culley
Request for Comments: 5044                       Hewlett-Packard Company
Category: Standards Track                                       U. Elzur
                                                    Broadcom Corporation
                                                                R. Recio
                                                         IBM Corporation
                                                               S. Bailey
                                                   Sandburst Corporation
                                                              J. Carrier
                                                               Cray Inc.
                                                            October 2007


            Marker PDU Aligned Framing for TCP Specification

Status of This Memo



   This document specifies an Internet standards track protocol for the
   Internet community, and requests discussion and suggestions for
   improvements.  Please refer to the current edition of the "Internet
   Official Protocol Standards" (STD 1) for the standardization state
   and status of this protocol.  Distribution of this memo is unlimited.

Abstract



   Marker PDU Aligned Framing (MPA) is designed to work as an
   "adaptation layer" between TCP and the Direct Data Placement protocol
   (DDP) as described in RFC 5041.  It preserves the reliable, in-order
   delivery of TCP, while adding the preservation of higher-level
   protocol record boundaries that DDP requires.  MPA is fully compliant
   with applicable TCP RFCs and can be utilized with existing TCP
   implementations.  MPA also supports integrated implementations that
   combine TCP, MPA and DDP to reduce buffering requirements in the
   implementation and improve performance at the system level.

















Culley, et al.              Standards Track                     [Page 1]

RFC 5044                  MPA Framing for TCP               October 2007


Table of Contents



   1. Introduction ....................................................4
      1.1. Motivation .................................................4
      1.2. Protocol Overview ..........................................5
   2. Glossary ........................................................8
   3. MPA's Interactions with DDP ....................................11
   4. MPA Full Operation Phase .......................................13
      4.1. FPDU Format ...............................................13
      4.2. Marker Format .............................................14
      4.3. MPA Markers ...............................................14
      4.4. CRC Calculation ...........................................16
      4.5. FPDU Size Considerations ..................................21
   5. MPA's interactions with TCP ....................................22
      5.1. MPA transmitters with a standard layered TCP ..............22
      5.2. MPA receivers with a standard layered TCP .................23
   6. MPA Receiver FPDU Identification ...............................24
   7. Connection Semantics ...........................................24
      7.1. Connection Setup ..........................................24
           7.1.1. MPA Request and Reply Frame Format .................26
           7.1.2. Connection Startup Rules ...........................28
           7.1.3. Example Delayed Startup Sequence ...................30
           7.1.4. Use of Private Data ................................33
                  7.1.4.1. Motivation ................................33
                  7.1.4.2. Example Immediate Startup Using
                           Private Data ..............................35
           7.1.5. "Dual Stack" Implementations .......................37
      7.2. Normal Connection Teardown ................................38
   8. Error Semantics ................................................39
   9. Security Considerations ........................................40
      9.1. Protocol-Specific Security Considerations .................40
           9.1.1. Spoofing ...........................................40
                  9.1.1.1. Impersonation .............................41
                  9.1.1.2. Stream Hijacking ..........................41
                  9.1.1.3. Man-in-the-Middle Attack ..................41
           9.1.2. Eavesdropping ......................................42
      9.2. Introduction to Security Options ..........................42
      9.3. Using IPsec with MPA ......................................43
      9.4. Requirements for IPsec Encapsulation of MPA/DDP ...........43
   10. IANA Considerations ...........................................44
   Appendix A. Optimized MPA-Aware TCP Implementations ...............45
      A.1. Optimized MPA/TCP Transmitters ............................46
      A.2. Effects of Optimized MPA/TCP Segmentation .................46
      A.3. Optimized MPA/TCP Receivers ...............................48
      A.4. Re-segmenting Middleboxes and Non-Optimized MPA/TCP
           Senders ...................................................49
      A.5. Receiver Implementation ...................................50
           A.5.1. Network Layer Reassembly Buffers ...................51



Culley, et al.              Standards Track                     [Page 2]

RFC 5044                  MPA Framing for TCP               October 2007


           A.5.2. TCP Reassembly Buffers .............................52
   Appendix B. Analysis of MPA over TCP Operations ...................52
      B.1. Assumptions ...............................................53
           B.1.1. MPA Is Layered beneath DDP .........................53
           B.1.2. MPA Preserves DDP Message Framing ..................53
           B.1.3. The Size of the ULPDU Passed to MPA Is Less Than
                  EMSS Under Normal Conditions .......................53
           B.1.4. Out-of-Order Placement but NO Out-of-Order Delivery.54
     B.2.  The Value of FPDU Alignment ...............................54
           B.2.1. Impact of Lack of FPDU Alignment on the Receiver
                  Computational Load and Complexity ..................56
           B.2.2. FPDU Alignment Effects on TCP Wire Protocol ........60
   Appendix C. IETF Implementation Interoperability with RDMA
               Consortium Protocols ..................................62
     C.1. Negotiated Parameters ......................................63
     C.2. RDMAC RNIC and Non-Permissive IETF RNIC ....................64
          C.2.1. RDMAC RNIC Initiator ................................65
          C.2.2. Non-Permissive IETF RNIC Initiator ..................65
          C.2.3. RDMAC RNIC and Permissive IETF RNIC .................65
          C.2.4. RDMAC RNIC Initiator ................................66
          C.2.5. Permissive IETF RNIC Initiator ......................67
     C.3. Non-Permissive IETF RNIC and Permissive IETF RNIC ..........67
   Normative References ..............................................68
   Informative References ............................................68
   Contributors ......................................................70

Table of Figures

   Figure 1: ULP MPA TCP Layering .....................................5
   Figure 2: FPDU Format .............................................13
   Figure 3: Marker Format ...........................................14
   Figure 4: Example FPDU Format with Marker .........................16
   Figure 5: Annotated Hex Dump of an FPDU ...........................19
   Figure 6: Annotated Hex Dump of an FPDU with Marker ...............20
   Figure 7: Fully Layered Implementation ............................22
   Figure 8: MPA Request/Reply Frame .................................26
   Figure 9: Example Delayed Startup Negotiation .....................31
   Figure 10: Example Immediate Startup Negotiation ..................35
   Figure 11: Optimized MPA/TCP Implementation .......................45
   Figure 12: Non-Aligned FPDU Freely Placed in TCP Octet Stream .....56
   Figure 13: Aligned FPDU Placed Immediately after TCP Header .......58
   Figure 14: Connection Parameters for the RNIC Types ...............63
   Figure 15: MPA Negotiation between an RDMAC RNIC and a
              Non-Permissive IETF RNIC ...............................65
   Figure 16: MPA Negotiation between an RDMAC RNIC and a Permissive
              IETF RNIC ..............................................66
   Figure 17: MPA Negotiation between a Non-Permissive IETF RNIC and
              a Permissive IETF RNIC .................................67



Culley, et al.              Standards Track                     [Page 3]

RFC 5044                  MPA Framing for TCP               October 2007


1.  Introduction



   This section discusses the reason for creating MPA on TCP and a
   general overview of the protocol.

1.1.  Motivation



   The Direct Data Placement protocol [DDP], when used with TCP
   [RFC793], requires a mechanism to detect record boundaries.  The DDP
   records are referred to as Upper Layer Protocol Data Units by this
   document.  The ability to locate the Upper Layer Protocol Data Unit
   (ULPDU) boundary is useful to a hardware network adapter that uses
   DDP to directly place the data in the application buffer based on the
   control information carried in the ULPDU header.  This may be done
   without requiring that the packets arrive in order.  Potential
   benefits of this capability are the avoidance of the memory copy
   overhead and a smaller memory requirement for handling out-of-order
   or dropped packets.

   Many approaches have been proposed for a generalized framing
   mechanism.  Some are probabilistic in nature and others are
   deterministic.  An example probabilistic approach is characterized by
   a detectable value embedded in the octet stream, with no method of
   preventing that value elsewhere within user data.  It is
   probabilistic because under some conditions the receiver may
   incorrectly interpret application data as the detectable value.
   Under these conditions, the protocol may fail with unacceptable
   frequency.  One deterministic approach is characterized by embedded
   controls at known locations in the octet stream.  Because the
   receiver can guarantee it will only examine the data stream at
   locations that are known to contain the embedded control, the
   protocol can never misinterpret application data as being embedded
   control data.  For unambiguous handling of an out-of-order packet, a
   deterministic approach is preferred.

   The MPA protocol provides a framing mechanism for DDP running over
   TCP using the deterministic approach.  It allows the location of the
   ULPDU to be determined in the TCP stream even if the TCP segments
   arrive out of order.












Culley, et al.              Standards Track                     [Page 4]

RFC 5044                  MPA Framing for TCP               October 2007


1.2.  Protocol Overview



   The layering of PDUs with MPA is shown in Figure 1, below.

               +------------------+
               |     ULP client   |
               +------------------+  <- Consumer messages
               |        DDP       |
               +------------------+  <- ULPDUs
               |        MPA*      |
               +------------------+  <- FPDUs (containing ULPDUs)
               |        TCP*      |
               +------------------+  <- TCP Segments (containing FPDUs)
               |      IP etc.     |
               +------------------+
                * These may be fully layered or optimized together.

                       Figure 1: ULP MPA TCP Layering

   MPA is described as an extra layer above TCP and below DDP.  The
   operation sequence is:

   1.  A TCP connection is established by ULP action.  This is done
       using methods not described by this specification.  The ULP may
       exchange some amount of data in streaming mode prior to starting
       MPA, but is not required to do so.

   2.  The Consumer negotiates the use of DDP and MPA at both ends of a
       connection.  The mechanisms to do this are not described in this
       specification.  The negotiation may be done in streaming mode, or
       by some other mechanism (such as a pre-arranged port number).

   3.  The ULP activates MPA on each end in the Startup Phase, either as
       an Initiator or a Responder, as determined by the ULP.  This mode
       verifies the usage of MPA, specifies the use of CRC and Markers,
       and allows the ULP to communicate some additional data via a
       Private Data exchange.  See Section 7.1, Connection Setup, for
       more details on the startup process.

   4.  At the end of the Startup Phase, the ULP puts MPA (and DDP) into
       Full Operation and begins sending DDP data as further described
       below.  In this document, DDP data chunks are called ULPDUs.  For
       a description of the DDP data, see [DDP].








Culley, et al.              Standards Track                     [Page 5]

RFC 5044                  MPA Framing for TCP               October 2007


   Following is a description of data transfer when MPA is in Full
   Operation.

   1.  DDP determines the Maximum ULPDU (MULPDU) size by querying MPA
       for this value.  MPA derives this information from TCP or IP,
       when it is available, or chooses a reasonable value.

   2.  DDP creates ULPDUs of MULPDU size or smaller, and hands them to
       MPA at the sender.



   3.  MPA creates a Framed Protocol Data Unit (FPDU) by prepending a
       header, optionally inserting Markers, and appending a CRC field
       after the ULPDU and PAD (if any).  MPA delivers the FPDU to TCP.

   4.  The TCP sender puts the FPDUs into the TCP stream.  If the sender
       is optimized MPA/TCP, it segments the TCP stream in such a way
       that a TCP Segment boundary is also the boundary of an FPDU.  TCP
       then passes each segment to the IP layer for transmission.

   5.  The receiver may or may not be optimized.  If it is optimized
       MPA/TCP, it may separate passing the TCP payload to MPA from
       passing the TCP payload ordering information to MPA.  In either
       case, RFC-compliant TCP wire behavior is observed at both the
       sender and receiver.

   6.  The MPA receiver locates and assembles complete FPDUs within the
       stream, verifies their integrity, and removes MPA Markers (when
       present), ULPDU_Length, PAD, and the CRC field.

   7.  MPA then provides the complete ULPDUs to DDP.  MPA may also
       separate passing MPA payload to DDP from passing the MPA payload
       ordering information.

   A fully layered MPA on TCP is implemented as a data stream ULP for
   TCP and is therefore RFC compliant.

   An optimized DDP/MPA/TCP uses a TCP layer that potentially contains
   some additional behaviors as suggested in this document.  When
   DDP/MPA/TCP are cross-layer optimized, the behavior of TCP
   (especially sender segmentation) may change from that of the un-
   optimized implementation, but the changes are within the bounds
   permitted by the TCP RFC specifications, and will interoperate with
   an un-optimized TCP.  The additional behaviors are described in
   Appendix A and are not normative; they are described at a TCP
   interface layer as a convenience.  Implementations may achieve the
   described functionality using any method, including cross-layer
   optimizations between TCP, MPA, and DDP.




Culley, et al.              Standards Track                     [Page 6]

RFC 5044                  MPA Framing for TCP               October 2007


   An optimized DDP/MPA/TCP sender is able to segment the data stream
   such that TCP segments begin with FPDUs (FPDU Alignment).  This has
   significant advantages for receivers.  When segments arrive with
   aligned FPDUs, the receiver usually need not buffer any portion of
   the segment, allowing DDP to place it in its destination memory
   immediately, thus avoiding copies from intermediate buffers (DDP's
   reason for existence).

   An optimized DDP/MPA/TCP receiver allows a DDP on MPA implementation
   to locate the start of ULPDUs that may be received out of order.  It
   also allows the implementation to determine if the entire ULPDU has
   been received.  As a result, MPA can pass out-of-order ULPDUs to DDP
   for immediate use.  This enables a DDP on MPA implementation to save
   a significant amount of intermediate storage by placing the ULPDUs in
   the right locations in the application buffers when they arrive,
   rather than waiting until full ordering can be restored.

   The ability of a receiver to recover out-of-order ULPDUs is optional
   and declared to the transmitter during startup.  When the receiver
   declares that it does not support out-of-order recovery, the
   transmitter does not add the control information to the data stream
   needed for out-of-order recovery.

   If the receiver is fully layered, then MPA receives a strictly
   ordered stream of data and does not deal with out-of-order ULPDUs.
   In this case, MPA passes each ULPDU to DDP when the last bytes arrive
   from TCP, along with the indication that they are in order.

   MPA implementations that support recovery of out-of-order ULPDUs MUST
   support a mechanism to indicate the ordering of ULPDUs as the sender
   transmitted them and indicate when missing intermediate segments
   arrive.  These mechanisms allow DDP to reestablish record ordering
   and report Delivery of complete messages (groups of records).

   MPA also addresses enhanced data integrity.  Some users of TCP have
   noted that the TCP checksum is not as strong as could be desired (see
   [CRCTCP]).  Studies such as [CRCTCP] have shown that the TCP checksum
   indicates segments in error at a much higher rate than the underlying
   link characteristics would indicate.  With these higher error rates,
   the chance that an error will escape detection, when using only the
   TCP checksum for data integrity, becomes a concern.  A stronger
   integrity check can reduce the chance of data errors being missed.

   MPA includes a CRC check to increase the ULPDU data integrity to the
   level provided by other modern protocols, such as SCTP [RFC4960].  It
   is possible to disable this CRC check; however, CRCs MUST be enabled
   unless it is clear that the end-to-end connection through the network
   has data integrity at least as good as an MPA with CRC enabled (for



Culley, et al.              Standards Track                     [Page 7]

RFC 5044                  MPA Framing for TCP               October 2007


   example, when IPsec is implemented end to end).  DDP's ULP expects
   this level of data integrity and therefore the ULP does not have to
   provide its own duplicate data integrity and error recovery for lost
   data.

2.  Glossary

   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].

   Consumer - the ULPs or applications that lie above MPA and DDP.  The
       Consumer is responsible for making TCP connections, starting MPA
       and DDP connections, and generally controlling operations.

   CRC - Cyclic Redundancy Check.

   Delivery - (Delivered, Delivers) - For MPA, Delivery is defined as
       the process of informing DDP that a particular PDU is ordered for
       use.  A PDU is Delivered in the exact order that it was sent by
       the original sender; MPA uses TCP's byte stream ordering to
       determine when Delivery is possible.  This is specifically
       different from "passing the PDU to DDP", which may generally
       occur in any order, while the order of Delivery is strictly
       defined.

   EMSS - Effective Maximum Segment Size.  EMSS is the smaller of the
       TCP maximum segment size (MSS) as defined in RFC 793 [RFC793],
       and the current path Maximum Transmission Unit (MTU) [RFC1191].

   FPDU - Framed Protocol Data Unit.  The unit of data created by an MPA
       sender.

   FPDU Alignment - The property that an FPDU is Header Aligned with the
       TCP segment, and the TCP segment includes an integer number of
       FPDUs.  A TCP segment with an FPDU Alignment allows immediate
       processing of the contained FPDUs without waiting on other TCP
       segments to arrive or combining with prior segments.

   FPDU Pointer (FPDUPTR) - This field of the Marker is used to indicate
       the beginning of an FPDU.

   Full Operation (Full Operation Phase) - After the completion of the
       Startup Phase, MPA begins exchanging FPDUs.







Culley, et al.              Standards Track                     [Page 8]

RFC 5044                  MPA Framing for TCP               October 2007


   Header Alignment - The property that a TCP segment begins with an
       FPDU.  The FPDU is Header Aligned when the FPDU header is exactly
       at the start of the TCP segment (right behind the TCP headers on
       the wire).

   Initiator - The endpoint of a connection that sends the MPA Request
       Frame, i.e., the first to actually send data (which may not be
       the one that sends the TCP SYN).

   Marker - A four-octet field that is placed in the MPA data stream at
       fixed octet intervals (every 512 octets).

   MPA-aware TCP - A TCP implementation that is aware of the receiver
       efficiencies of MPA FPDU Alignment and is capable of sending TCP
       segments that begin with an FPDU.

   MPA-enabled - MPA is enabled if the MPA protocol is visible on the
       wire.  When the sender is MPA-enabled, it is inserting framing
       and Markers.  When the receiver is MPA-enabled, it is
       interpreting framing and Markers.

   MPA Request Frame - Data sent from the MPA Initiator to the MPA
       Responder during the Startup Phase.

   MPA Reply Frame - Data sent from the MPA Responder to the MPA
       Initiator during the Startup Phase.

   MPA - Marker-based ULP PDU Aligned Framing for TCP protocol.  This
       document defines the MPA protocol.

   MULPDU - Maximum ULPDU.  The current maximum size of the record that
       is acceptable for DDP to pass to MPA for transmission.

   Node - A computing device attached to one or more links of a network.
       A Node in this context does not refer to a specific application
       or protocol instantiation running on the computer.  A Node may
       consist of one or more MPA on TCP devices installed in a host
       computer.

   PAD - A 1-3 octet group of zeros used to fill an FPDU to an exact
       modulo 4 size.

   PDU - Protocol data unit

   Private Data - A block of data exchanged between MPA endpoints during
       initial connection setup.





Culley, et al.              Standards Track                     [Page 9]

RFC 5044                  MPA Framing for TCP               October 2007


   Protection Domain - An RDMA concept (see [VERBS-RDMA] and [RDMASEC])
       that ties use of various endpoint resources (memory access, etc.)
       to the specific RDMA/DDP/MPA connection.

   RDDP - A suite of protocols including MPA, [DDP], [RDMAP], an overall
       security document [RDMASEC], a problem statement [RFC4297], an
       architecture document [RFC4296], and an applicability document
       [APPL].

   RDMA - Remote Direct Memory Access; a protocol that uses DDP and MPA
       to enable applications to transfer data directly from memory
       buffers.  See [RDMAP].

   Remote Peer - The MPA protocol implementation on the opposite end of
       the connection.  Used to refer to the remote entity when
       describing protocol exchanges or other interactions between two
       Nodes.

   Responder - The connection endpoint that responds to an incoming MPA
       connection request (the MAP Request Frame).  This may not be the
       endpoint that awaited the TCP SYN.

   Startup Phase - The initial exchanges of an MPA connection that
       serves to more fully identify MPA endpoints to each other and
       pass connection specific setup information to each other.

   ULP - Upper Layer Protocol.  The protocol layer above the protocol
       layer currently being referenced.  The ULP for MPA is DDP [DDP].

   ULPDU - Upper Layer Protocol Data Unit.  The data record defined by
       the layer above MPA (DDP).  ULPDU corresponds to DDP's DDP
       segment.

   ULPDU_Length - A field in the FPDU describing the length of the
       included ULPDU.
















Culley, et al.              Standards Track                    [Page 10]

RFC 5044                  MPA Framing for TCP               October 2007


3.  MPA's Interactions with DDP



   DDP requires MPA to maintain DDP record boundaries from the sender to
   the receiver.  When using MPA on TCP to send data, DDP provides
   records (ULPDUs) to MPA.  MPA will use the reliable transmission
   abilities of TCP to transmit the data, and will insert appropriate
   additional information into the TCP stream to allow the MPA receiver
   to locate the record boundary information.

   As such, MPA accepts complete records (ULPDUs) from DDP at the sender
   and returns them to DDP at the receiver.

   MPA MUST encapsulate the ULPDU such that there is exactly one ULPDU
   contained in one FPDU.

   MPA over a standard TCP stack can usually provide FPDU Alignment with
   the TCP Header if the FPDU is equal to TCP's EMSS.  An optimized
   MPA/TCP stack can also maintain alignment as long as the FPDU is less
   than or equal to TCP's EMSS.  Since FPDU Alignment is generally
   desired by the receiver, DDP cooperates with MPA to ensure FPDUs'
   lengths do not exceed the EMSS under normal conditions.  This is done
   with the MULPDU mechanism.

   MPA MUST provide information to DDP on the current maximum size of
   the record that is acceptable to send (MULPDU).  DDP SHOULD limit
   each record size to MULPDU.  The range of MULPDU values MUST be
   between 128 octets and 64768 octets, inclusive.

   The sending DDP MUST NOT post a ULPDU larger than 64768 octets to
   MPA.  DDP MAY post a ULPDU of any size between one and 64768 octets;
   however, MPA is not REQUIRED to support a ULPDU Length that is
   greater than the current MULPDU.

   While the maximum theoretical length supported by the MPA header
   ULPDU_Length field is 65535, TCP over IP requires the IP datagram
   maximum length to be 65535 octets.  To enable MPA to support FPDU
   Alignment, the maximum size of the FPDU must fit within an IP
   datagram.  Thus, the ULPDU limit of 64768 octets was derived by
   taking the maximum IP datagram length, subtracting from it the
   maximum total length of the sum of the IPv4 header, TCP header, IPv4
   options, TCP options, and the worst-case MPA overhead, and then
   rounding the result down to a 128-octet boundary.

   Note that MULPDU will be significantly smaller than the theoretical
   maximum in most implementations for most circumstances, due to link
   MTUs, use of extra headers such as required for IPsec, etc.





Culley, et al.              Standards Track                    [Page 11]

RFC 5044                  MPA Framing for TCP               October 2007


   On receive, MPA MUST pass each ULPDU with its length to DDP when it
   has been validated.

   If an MPA implementation supports passing out-of-order ULPDUs to DDP,
   the MPA implementation SHOULD:

   *   Pass each ULPDU with its length to DDP as soon as it has been
       fully received and validated.

   *   Provide a mechanism to indicate the ordering of ULPDUs as the
       sender transmitted them.  One possible mechanism might be
       providing the TCP sequence number for each ULPDU.

   *   Provide a mechanism to indicate when a given ULPDU (and prior
       ULPDUs) are complete (Delivered to DDP).  One possible mechanism
       might be to allow DDP to see the current outgoing TCP ACK
       sequence number.

   *   Provide an indication to DDP that the TCP has closed or has begun
       to close the connection (e.g., received a FIN).

   MPA MUST provide the protocol version negotiated with its peer to
   DDP.  DDP will use this version to set the version in its header and
   to report the version to [RDMAP].



























Culley, et al.              Standards Track                    [Page 12]

RFC 5044                  MPA Framing for TCP               October 2007


4.  MPA Full Operation Phase



   The following sections describe the main semantics of the Full
   Operation Phase of MPA.

4.1.  FPDU Format



   MPA senders create FPDUs out of ULPDUs.  The format of an FPDU shown
   below MUST be used for all MPA FPDUs.  For purposes of clarity,
   Markers are not shown in Figure 2.

       0                   1                   2                   3
       0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |          ULPDU_Length         |                               |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+                               +
      |                                                               |
      ~                                                               ~
      ~                            ULPDU                              ~
      |                                                               |
      |                               +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                               |          PAD (0-3 octets)     |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                             CRC                               |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                           Figure 2: FPDU Format

   ULPDU_Length: 16 bits (unsigned integer).  This is the number of
   octets of the contained ULPDU.  It does not include the length of the
   FPDU header itself, the pad, the CRC, or of any Markers that fall
   within the ULPDU.  The 16-bit ULPDU Length field is large enough to
   support the largest IP datagrams for IPv4 or IPv6.

   PAD: The PAD field trails the ULPDU and contains between 0 and 3
   octets of data.  The pad data MUST be set to zero by the sender and
   ignored by the receiver (except for CRC checking).  The length of the
   pad is set so as to make the size of the FPDU an integral multiple of
   four.

   CRC: 32 bits.  When CRCs are enabled, this field contains a CRC32c
   check value, which is used to verify the entire contents of the FPDU,
   using CRC32c.  See Section 4.4, CRC Calculation.  When CRCs are not
   enabled, this field is still present, may contain any value, and MUST
   NOT
be checked.






Culley, et al.              Standards Track                    [Page 13]

RFC 5044                  MPA Framing for TCP               October 2007


   The FPDU adds a minimum of 6 octets to the length of the ULPDU.  In
   addition, the total length of the FPDU will include the length of any
   Markers and from 0 to 3 pad octets added to round-up the ULPDU size.

4.2.  Marker Format



   The format of a Marker MUST be as specified in Figure 3:

       0                   1                   2                   3
       0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |           RESERVED            |            FPDUPTR            |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                          Figure 3: Marker Format

   RESERVED: The Reserved field MUST be set to zero on transmit and
   ignored on receive (except for CRC calculation).

   FPDUPTR: The FPDU Pointer is a relative pointer, 16 bits long,
   interpreted as an unsigned integer that indicates the number of
   octets in the TCP stream from the beginning of the ULPDU Length field
   to the first octet of the entire Marker.  The least significant two
   bits MUST always be set to zero at the transmitter, and the receivers
   MUST always treat these as zero for calculations.

4.3.  MPA Markers



   MPA Markers are used to identify the start of FPDUs when packets are
   received out of order.  This is done by locating the Markers at fixed
   intervals in the data stream (which is correlated to the TCP sequence
   number) and using the Marker value to locate the preceding FPDU
   start.

   All MPA Markers are included in the containing FPDU CRC calculation
   (when both CRCs and Markers are in use).

   The MPA receiver's ability to locate out-of-order FPDUs and pass the
   ULPDUs to DDP is implementation dependent.  MPA/DDP allows those
   receivers that are able to deal with out-of-order FPDUs in this way
   to require the insertion of Markers in the data stream.  When the
   receiver cannot deal with out-of-order FPDUs in this way, it may
   disable the insertion of Markers at the sender.  All MPA senders MUST
   be able to generate Markers when their use is declared by the
   opposing receiver (see Section 7.1, Connection Setup).






Culley, et al.              Standards Track                    [Page 14]

RFC 5044                  MPA Framing for TCP               October 2007


   When Markers are enabled, MPA senders MUST insert a Marker into the
   data stream at a 512-octet periodic interval in the TCP Sequence
   Number Space.  The Marker contains a 16-bit unsigned integer referred
   to as the FPDUPTR (FPDU Pointer).

   If the FPDUPTR's value is non-zero, the FPDU Pointer is a 16-bit
   relative back-pointer.  FPDUPTR MUST contain the number of octets in
   the TCP stream from the beginning of the ULPDU Length field to the
   first octet of the Marker, unless the Marker falls between FPDUs.
   Thus, the location of the first octet of the previous FPDU header can
   be determined by subtracting the value of the given Marker from the
   current octet-stream sequence number (i.e., TCP sequence number) of
   the first octet of the Marker.  Note that this computation MUST take
   into account that the TCP sequence number could have wrapped between
   the Marker and the header.

   An FPDUPTR value of 0x0000 is a special case -- it is used when the
   Marker falls exactly between FPDUs (between the preceding FPDU CRC
   field and the next FPDU's ULPDU Length field).  In this case, the
   Marker is considered to be contained in the following FPDU; the
   Marker MUST be included in the CRC calculation of the FPDU following
   the Marker (if CRCs are being generated or checked).  Thus, an
   FPDUPTR value of 0x0000 means that immediately following the Marker
   is an FPDU header (the ULPDU Length field).

   Since all FPDUs are integral multiples of 4 octets, the bottom two
   bits of the FPDUPTR as calculated by the sender are zero.  MPA
   reserves these bits so they MUST be treated as zero for computation
   at the receiver.

   When Markers are enabled (see Section 7.1, Connection Setup), the MPA
   Markers MUST be inserted immediately preceding the first FPDU of Full
   Operation Phase, and at every 512th octet of the TCP octet stream
   thereafter.  As a result, the first Marker has an FPDUPTR value of
   0x0000.  If the first Marker begins at octet sequence number
   SeqStart, then Markers are inserted such that the first octet of the
   Marker is at octet sequence number SeqNum if the remainder of (SeqNum
   - SeqStart) mod 512 is zero.  Note that SeqNum can wrap.

   For example, if the TCP sequence number were used to calculate the
   insertion point of the Marker, the starting TCP sequence number is
   unlikely to be zero, and 512-octet multiples are unlikely to fall on
   a modulo 512 of zero.  If the MPA connection is started at TCP
   sequence number 11, then the 1st Marker will begin at 11, and
   subsequent Markers will begin at 523, 1035, etc.






Culley, et al.              Standards Track                    [Page 15]

RFC 5044                  MPA Framing for TCP               October 2007


   If an FPDU is large enough to contain multiple Markers, they MUST all
   point to the same point in the TCP stream: the first octet of the
   ULPDU Length field for the FPDU.

   If a Marker interval contains multiple FPDUs (the FPDUs are small),
   the Marker MUST point to the start of the ULPDU Length field for the
   FPDU containing the Marker unless the Marker falls between FPDUs, in
   which case the Marker MUST be zero.

   The following example shows an FPDU containing a Marker.

   0                   1                   2                   3
   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |       ULPDU Length (0x0010)   |                               |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+                               +
   |                                                               |
   +                                                               +
   |                         ULPDU (octets 0-9)                    |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |            (0x0000)           |        FPDU ptr (0x000C)      |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |                        ULPDU (octets 10-15)                   |
   |                               +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |                               |          PAD (2 octets:0,0)   |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |                              CRC                              |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

              Figure 4: Example FPDU Format with Marker

   MPA Receivers MUST preserve ULPDU boundaries when passing data to
   DDP.  MPA Receivers MUST pass the ULPDU data and the ULPDU Length to
   DDP and not the Markers, headers, and CRC.

4.4.  CRC Calculation



   An MPA implementation MUST implement CRC support and MUST either:

   (1)  always use CRCs; the MPA provider is not REQUIRED to support an
        administrator's request that CRCs not be used.

        or

   (2a) only indicate a preference not to use CRCs on the explicit
        request of the system administrator, via an interface not
        defined in this spec.  The default configuration for a
        connection MUST be to use CRCs.



Culley, et al.              Standards Track                    [Page 16]

RFC 5044                  MPA Framing for TCP               October 2007


   (2b) disable CRC checking (and possibly generation) if both the local
        and remote endpoints indicate preference not to use CRCs.

   An administrative decision to have a host request CRC suppression
   SHOULD NOT be made unless there is assurance that the TCP connection
   involved provides protection from undetected errors that is at least
   as strong as an end-to-end CRC32c.  End-to-end usage of an IPsec
   cryptographic integrity check is among the ways to provide such
   protection, and the use of channel bindings [NFSv4CHANNEL] by the ULP
   can provide a high level of assurance that the IPsec protection scope
   is end-to-end with respect to the ULP.

   The process MUST be invisible to the ULP.

   After receipt of an MPA startup declaration indicating that its peer
   requires CRCs, an MPA instance MUST continue generating and checking
   CRCs until the connection terminates.  If an MPA instance has
   declared that it does not require CRCs, it MUST turn off CRC checking
   immediately after receipt of an MPA mode declaration indicating that
   its peer also does not require CRCs.  It MAY continue generating
   CRCs.  See Section 7.1, Connection Setup, for details on the MPA
   startup.

   When sending an FPDU, the sender MUST include a CRC field.  When CRCs
   are enabled, the CRC field in the MPA FPDU MUST be computed using the
   CRC32c polynomial in the manner described in the iSCSI Protocol
   [iSCSI] document for Header and Data Digests.

   The fields which MUST be included in the CRC calculation when sending
   an FPDU are as follows:

   1)  If a Marker does not immediately precede the ULPDU Length field,
       the CRC-32c is calculated from the first octet of the ULPDU
       Length field, through all the ULPDU and Markers (if present), to
       the last octet of the PAD (if present), inclusive.  If there is a
       Marker immediately following the PAD, the Marker is included in
       the CRC calculation for this FPDU.

   2)  If a Marker immediately precedes the first octet of the ULPDU
       Length field of the FPDU, (i.e., the Marker fell between FPDUs,
       and thus is required to be included in the second FPDU), the
       CRC-32c is calculated from the first octet of the Marker, through
       the ULPDU Length header, through all the ULPDU and Markers (if
       present), to the last octet of the PAD (if present), inclusive.

   3)  After calculating the CRC-32c, the resultant value is placed into
       the CRC field at the end of the FPDU.




Culley, et al.              Standards Track                    [Page 17]

RFC 5044                  MPA Framing for TCP               October 2007


   When an FPDU is received, and CRC checking is enabled, the receiver
   MUST first perform the following:

   1)  Calculate the CRC of the incoming FPDU in the same fashion as
       defined above.

   2)  Verify that the calculated CRC-32c value is the same as the
       received CRC-32c value found in the FPDU CRC field.  If not, the
       receiver MUST treat the FPDU as an invalid FPDU.

   The procedure for handling invalid FPDUs is covered in Section 8,
   Error Semantics.

   The following is an annotated hex dump of an example FPDU sent as the
   first FPDU on the stream.  As such, it starts with a Marker.  The
   FPDU contains a 42 octet ULPDU (an example DDP segment) which in turn
   contains 24 octets of the contained ULPDU, which is a data load that
   is all zeros.  The CRC32c has been correctly calculated and can be
   used as a reference.  See the [DDP] and [RDMAP] specification for
   definitions of the DDP Control field, Queue, MSN, MO, and Send Data.































Culley, et al.              Standards Track                    [Page 18]

RFC 5044                  MPA Framing for TCP               October 2007


       Octet Contents  Annotation
       Count

       0000    00      Marker: Reserved
       0001    00
       0002    00      Marker: FPDUPTR
       0003    00
       0004    00      ULPDU Length
       0005    2a
       0006    41      DDP Control Field, Send with Last flag set
       0007    43
       0008    00      Reserved (DDP STag position with no STag)
       0009    00
       000a    00
       000b    00
       000c    00      DDP Queue = 0
       000d    00
       000e    00
       000f    00
       0010    00      DDP MSN = 1
       0011    00
       0012    00
       0013    01
       0014    00      DDP MO = 0
       0015    00
       0016    00
       0017    00
       0018    00      DDP Send Data (24 octets of zeros)
       ...
       002f    00
       0030    52      CRC32c
       0031    23
       0032    99
       0033    83

                  Figure 5: Annotated Hex Dump of an FPDU















Culley, et al.              Standards Track                    [Page 19]

RFC 5044                  MPA Framing for TCP               October 2007


      The following is an example sent as the second FPDU of the stream
      where the first FPDU (which is not shown here) had a length of 492
      octets and was also a Send to Queue 0 with Last Flag set.  This
      example contains a Marker.

       Octet Contents  Annotation
       Count

       01ec    00      Length
       01ed    2a
       01ee    41      DDP Control Field: Send with Last Flag set
       01ef    43
       01f0    00      Reserved (DDP STag position with no STag)
       01f1    00
       01f2    00
       01f3    00
       01f4    00      DDP Queue = 0
       01f5    00
       01f6    00
       01f7    00
       01f8    00      DDP MSN = 2
       01f9    00
       01fa    00
       01fb    02
       01fc    00      DDP MO = 0
       01fd    00
       01fe    00
       01ff    00
       0200    00      Marker: Reserved
       0201    00
       0202    00      Marker: FPDUPTR
       0203    14
       0204    00      DDP Send Data (24 octets of zeros)
       ...
       021b    00
       021c    84      CRC32c
       021d    92
       021e    58
       021f    98

            Figure 6: Annotated Hex Dump of an FPDU with Marker










Culley, et al.              Standards Track                    [Page 20]

RFC 5044                  MPA Framing for TCP               October 2007


4.5.  FPDU Size Considerations



   MPA defines the Maximum Upper Layer Protocol Data Unit (MULPDU) as
   the size of the largest ULPDU fitting in an FPDU.  For an empty TCP
   Segment, MULPDU is EMSS minus the FPDU overhead (6 octets) minus
   space for Markers and pad octets.

       The maximum ULPDU Length for a single ULPDU when Markers are
       present MUST be computed as:

       MULPDU = EMSS - (6 + 4 * Ceiling(EMSS / 512) + EMSS mod 4)

   The formula above accounts for the worst-case number of Markers.

       The maximum ULPDU Length for a single ULPDU when Markers are NOT
       present MUST be computed as:

       MULPDU = EMSS - (6 + EMSS mod 4)

   As a further optimization of the wire efficiency an MPA
   implementation MAY dynamically adjust the MULPDU (see Section 5 for
   latency and wire efficiency trade-offs).  When one or more FPDUs are
   already packed into a TCP Segment, MULPDU MAY be reduced accordingly.

   DDP SHOULD provide ULPDUs that are as large as possible, but less
   than or equal to MULPDU.

   If the TCP implementation needs to adjust EMSS to support MTU changes
   or changing TCP options, the MULPDU value is changed accordingly.

   In certain rare situations, the EMSS may shrink below 128 octets in
   size.  If this occurs, the MPA on TCP sender MUST NOT shrink the
   MULPDU below 128 octets and is not required to follow the
   segmentation rules in Section 5.1 and Appendix A.

   If one or more FPDUs are already packed into a TCP segment, such that
   the remaining room is less than 128 octets, MPA MUST NOT provide a
   MULPDU smaller than 128.  In this case, MPA would typically provide a
   MULPDU for the next full sized segment, but may still pack the next
   FPDU into the small remaining room, provide that the next FPDU is
   small enough to fit.

   The value 128 is chosen as to allow DDP designers room for the DDP
   Header and some user data.







Culley, et al.              Standards Track                    [Page 21]

RFC 5044                  MPA Framing for TCP               October 2007


5.  MPA's interactions with TCP



   The following sections describe MPA's interactions with TCP.  This
   section discusses using a standard layered TCP stack with MPA
   attached above a TCP socket.  Discussion of using an optimized MPA-
   aware TCP with an MPA implementation that takes advantage of the
   extra optimizations is done in Appendix A.

                   +-----------------------------------+
                   | +-----+       +-----------------+ |
                   | | MPA |       | Other Protocols | |
                   | +-----+       +-----------------+ |
                   |    ||                  ||         |
                   |  ----- socket API --------------  |
                   |            ||                     |
                   |         +-----+                   |
                   |         | TCP |                   |
                   |         +-----+                   |
                   |            ||                     |
                   |         +-----+                   |
                   |         | IP  |                   |
                   |         +-----+                   |
                   +-----------------------------------+

                   Figure 7: Fully Layered Implementation

   The Fully layered implementation is described for completeness;
   however, the user is cautioned that the reduced probability of FPDU
   alignment when transmitting with this implementation will tend to
   introduce a higher overhead at optimized receivers.  In addition, the
   lack of out-of-order receive processing will significantly reduce the
   value of DDP/MPA by imposing higher buffering and copying overhead in
   the local receiver.

5.1.  MPA transmitters with a standard layered TCP



   MPA transmitters SHOULD calculate a MULPDU as described in Section
   4.5.  If the TCP implementation allows EMSS to be determined by MPA,
   that value should be used.  If the transmit side TCP implementation
   is not able to report the EMSS, MPA SHOULD use the current MTU value
   to establish a likely FPDU size, taking into account the various
   expected header sizes.

   MPA transmitters SHOULD also use whatever facilities the TCP stack
   presents to cause the TCP transmitter to start TCP segments at FPDU
   boundaries.  Multiple FPDUs MAY be packed into a single TCP segment
   as determined by the EMSS calculation as long as they are entirely
   contained in the TCP segment.



Culley, et al.              Standards Track                    [Page 22]

RFC 5044                  MPA Framing for TCP               October 2007


   For example, passing FPDU buffers sized to the current EMSS to the
   TCP socket and using the TCP_NODELAY socket option to disable the
   Nagle [RFC896] algorithm will usually result in many of the segments
   starting with an FPDU.

   It is recognized that various effects can cause an FPDU Alignment to
   be lost.  Following are a few of the effects:

   *   ULPDUs that are smaller than the MULPDU.  If these are sent in a
       continuous stream, FPDU Alignment will be lost.  Note that
       careful use of a dynamic MULPDU can help in this case; the MULPDU
       for future FPDUs can be adjusted to re-establish alignment with
       the segments based on the current EMSS.

   *   Sending enough data that the TCP receive window limit is reached.
       TCP may send a smaller segment to exactly fill the receive
       window.

   *   Sending data when TCP is operating up against the congestion
       window.  If TCP is not tracking the congestion window in
       segments, it may transmit a smaller segment to exactly fill the
       receive window.

   *   Changes in EMSS due to varying TCP options, or changes in MTU.

   If FPDU Alignment with TCP segments is lost for any reason, the
   alignment is regained after a break in transmission where the TCP
   send buffers are emptied.  Many usage models for DDP/MPA will include
   such breaks.

   MPA receivers are REQUIRED to be able to operate correctly even if
   alignment is lost (see Section 6).

5.2.  MPA receivers with a standard layered TCP



   MPA receivers will get TCP data in the usual ordered stream.  The
   receivers MUST identify FPDU boundaries by using the ULPDU_LENGTH
   field, as described in Section 6.  Receivers MAY utilize markers to
   check for FPDU boundary consistency, but they are NOT required to
   examine the markers to determine the FPDU boundaries.











Culley, et al.              Standards Track                    [Page 23]

RFC 5044                  MPA Framing for TCP               October 2007


6.  MPA Receiver FPDU Identification



   An MPA receiver MUST first verify the FPDU before passing the ULPDU
   to DDP.  To do this, the receiver MUST:

   *   locate the start of the FPDU unambiguously,

   *   verify its CRC (if CRC checking is enabled).

   If the above conditions are true, the MPA receiver passes the ULPDU
   to DDP.

   To detect the start of the FPDU unambiguously one of the following
   MUST be used:

   1:  In an ordered TCP stream, the ULPDU Length field in the current
       FPDU when FPDU has a valid CRC, can be used to identify the
       beginning of the next FPDU.

   2:  For optimized MPA/TCP receivers that support out-of-order
       reception of FPDUs (see Section 4.3, MPA Markers) a Marker can
       always be used to locate the beginning of an FPDU (in FPDUs with
       valid CRCs).  Since the location of the Marker is known in the
       octet stream (sequence number space), the Marker can always be
       found.

   3:  Having found an FPDU by means of a Marker, an optimized MPA/TCP
       receiver can find following contiguous FPDUs by using the ULPDU
       Length fields (from FPDUs with valid CRCs) to establish the next
       FPDU boundary.

   The ULPDU Length field (see Section 4) MUST be used to determine if
   the entire FPDU is present before forwarding the ULPDU to DDP.

   CRC calculation is discussed in Section 4.4 above.

7.  Connection Semantics



7.1.  Connection Setup



   MPA requires that the Consumer MUST activate MPA, and any TCP
   enhancements for MPA, on a TCP half connection at the same location
   in the octet stream at both the sender and the receiver.  This is
   required in order for the Marker scheme to correctly locate the
   Markers (if enabled) and to correctly locate the first FPDU.

   MPA, and any TCP enhancements for MPA are enabled by the ULP in both
   directions at once at an endpoint.



Culley, et al.              Standards Track                    [Page 24]

RFC 5044                  MPA Framing for TCP               October 2007


   This can be accomplished several ways, and is left up to DDP's ULP:

   *   DDP's ULP MAY require DDP on MPA startup immediately after TCP
       connection setup.  This has the advantage that no streaming mode
       negotiation is needed.  An example of such a protocol is shown in
       Figure 10: Example Immediate Startup negotiation.

       This may be accomplished by using a well-known port, or a service
       locator protocol to locate an appropriate port on which DDP on
       MPA is expected to operate.

   *   DDP's ULP MAY negotiate the start of DDP on MPA sometime after a
       normal TCP startup, using TCP streaming data exchanges on the
       same connection.  The exchange establishes that DDP on MPA (as
       well as other ULPs) will be used, and exactly locates the point
       in the octet stream where MPA is to begin operation.  Note that
       such a negotiation protocol is outside the scope of this
       specification.  A simplified example of such a protocol is shown
       in Figure 9: Example Delayed Startup negotiation on page 33.

   An MPA endpoint operates in two distinct phases.

   The Startup Phase is used to verify correct MPA setup, exchange CRC
   and Marker configuration, and optionally pass Private Data between
   endpoints prior to completing a DDP connection.  During this phase,
   specifically formatted frames are exchanged as TCP byte streams
   without using CRCs or Markers.  During this phase a DDP endpoint need
   not be "bound" to the MPA connection.  In fact, the choice of DDP
   endpoint and its operating parameters may not be known until the
   Consumer supplied Private Data (if any) has been examined by the
   Consumer.

   The second distinct phase is Full Operation during which FPDUs are
   sent using all the rules that pertain (CRCs, Markers, MULPDU
   restrictions, etc.).  A DDP endpoint MUST be "bound" to the MPA
   connection at entry to this phase.

   When Private Data is passed between ULPs in the Startup Phase, the
   ULP is responsible for interpreting that data, and then placing MPA
   into Full Operation.

   Note: The following text differentiates the two endpoints by calling
       them Initiator and Responder.  This is quite arbitrary and is NOT
       related to the TCP startup (SYN, SYN/ACK sequence).  The
       Initiator is the side that sends first in the MPA startup
       sequence (the MPA Request Frame).





Culley, et al.              Standards Track                    [Page 25]

RFC 5044                  MPA Framing for TCP               October 2007


   Note: The possibility that both endpoints would be allowed to make a
       connection at the same time, sometimes called an active/active
       connection, was considered by the work group and rejected.  There
       were several motivations for this decision.  One was that
       applications needing this facility were few (none other than
       theoretical at the time of this document).  Another was that the
       facility created some implementation difficulties, particularly
       with the "dual stack" designs described later on.  A last issue
       was that dealing with rejected connections at startup would have
       required at least an additional frame type, and more recovery
       actions, complicating the protocol.  While none of these issues
       was overwhelming, the group and implementers were not motivated
       to do the work to resolve these issues.  The protocol includes a
       method of detecting these active/active startup attempts so that
       they can be rejected and an error reported.

   The ULP is responsible for determining which side is Initiator or
   Responder.  For client/server type ULPs, this is easy.  For peer-peer
   ULPs (which might utilize a TCP style active/active startup), some
   mechanism (not defined by this specification) must be established, or
   some streaming mode data exchanged prior to MPA startup to determine
   which side starts in Initiator and which starts in Responder MPA
   mode.

7.1.1  MPA Request and Reply Frame Format



       0                   1                   2                   3
       0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   0  |                                                               |
      +         Key (16 bytes containing "MPA ID Req Frame")          +
   4  |      (4D 50 41 20 49 44 20 52 65 71 20 46 72 61 6D 65)        |
      +         Or  (16 bytes containing "MPA ID Rep Frame")          +
   8  |      (4D 50 41 20 49 44 20 52 65 70 20 46 72 61 6D 65)        |
      +                                                               +
   12 |                                                               |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   16 |M|C|R| Res     |     Rev       |          PD_Length            |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                                                               |
      ~                                                               ~
      ~                   Private Data                                ~
      |                                                               |
      |                               +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                               |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                     Figure 8: MPA Request/Reply Frame



Culley, et al.              Standards Track                    [Page 26]

RFC 5044                  MPA Framing for TCP               October 2007


   Key: This field contains the "key" used to validate that the sender
       is an MPA sender.  Initiator mode senders MUST set this field to
       the fixed value "MPA ID Req Frame" or (in byte order) 4D 50 41 20
       49 44 20 52 65 71 20 46 72 61 6D 65 (in hexadecimal).  Responder
       mode receivers MUST check this field for the same value, and
       close the connection and report an error locally if any other
       value is detected.  Responder mode senders MUST set this field to
       the fixed value "MPA ID Rep Frame" or (in byte order) 4D 50 41 20
       49 44 20 52 65 70 20 46 72 61 6D 65 (in hexadecimal).  Initiator
       mode receivers MUST check this field for the same value, and
       close the connection and report an error locally if any other
       value is detected.

   M: This bit declares an endpoint's REQUIRED Marker usage.  When this
       bit is '1' in an MPA Request Frame, the Initiator declares that
       Markers are REQUIRED in FPDUs sent from the Responder.  When set
       to '1' in an MPA Reply Frame, this bit declares that Markers are
       REQUIRED in FPDUs sent from the Initiator.  When in a received
       MPA Request Frame or MPA Reply Frame and the value is '0',
       Markers MUST NOT be added to the data stream by that endpoint.
       When '1' Markers MUST be added as described in Section 4.3, MPA
       Markers.

   C: This bit declares an endpoint's preferred CRC usage.  When this
       field is '0' in the MPA Request Frame and the MPA Reply Frame,
       CRCs MUST not be checked and need not be generated by either
       endpoint.  When this bit is '1' in either the MPA Request Frame
       or MPA Reply Frame, CRCs MUST be generated and checked by both
       endpoints.  Note that even when not in use, the CRC field remains
       present in the FPDU.  When CRCs are not in use, the CRC field
       MUST be considered valid for FPDU checking regardless of its
       contents.

   R: This bit is set to zero, and not checked on reception in the MPA
       Request Frame.  In the MPA Reply Frame, this bit is the Rejected
       Connection bit, set by the Responders ULP to indicate acceptance
       '0', or rejection '1', of the connection parameters provided in
       the Private Data.

   Res: This field is reserved for future use.  It MUST be set to zero
       when sending, and not checked on reception.










Culley, et al.              Standards Track                    [Page 27]

RFC 5044                  MPA Framing for TCP               October 2007


   Rev: This field contains the revision of MPA.  For this version of
       the specification, senders MUST set this field to one.  MPA
       receivers compliant with this version of the specification MUST
       check this field.  If the MPA receiver cannot interoperate with
       the received version, then it MUST close the connection and
       report an error locally.  Otherwise, the MPA receiver should
       report the received version to the ULP.

   PD_Length: This field MUST contain the length in octets of the
       Private Data field.  A value of zero indicates that there is no
       Private Data field present at all.  If the receiver detects that
       the PD_Length field does not match the length of the Private Data
       field, or if the length of the Private Data field exceeds 512
       octets, the receiver MUST close the connection and report an
       error locally.  Otherwise, the MPA receiver should pass the
       PD_Length value and Private Data to the ULP.

   Private Data: This field may contain any value defined by ULPs or may
       not be present.  The Private Data field MUST be between 0 and 512
       octets in length.  ULPs define how to size, set, and validate
       this field within these limits.  Private Data usage is further
       discussed in Section 7.1.4.

7.1.2.  Connection Startup Rules



   The following rules apply to MPA connection Startup Phase:

   1.  When MPA is started in the Initiator mode, the MPA implementation
       MUST send a valid MPA Request Frame.  The MPA Request Frame MAY
       include ULP-supplied Private Data.

   2.  When MPA is started in the Responder mode, the MPA implementation
       MUST wait until an MPA Request Frame is received and validated
       before entering Full MPA/DDP Operation.

       If the MPA Request Frame is improperly formatted, the
       implementation MUST close the TCP connection and exit MPA.

       If the MPA Request Frame is properly formatted but the Private
       Data is not acceptable, the implementation SHOULD return an MPA
       Reply Frame with the Rejected Connection bit set to '1'; the MPA
       Reply Frame MAY include ULP-supplied Private Data; the
       implementation MUST exit MPA, leaving the TCP connection open.
       The ULP may close TCP or use the connection for other purposes.

       If the MPA Request Frame is properly formatted and the Private
       Data is acceptable, the implementation SHOULD return an MPA Reply
       Frame with the Rejected Connection bit set to '0'; the MPA Reply



Culley, et al.              Standards Track                    [Page 28]

RFC 5044                  MPA Framing for TCP               October 2007


       Frame MAY include ULP-supplied Private Data; and the Responder
       SHOULD prepare to interpret any data received as FPDUs and pass
       any received ULPDUs to DDP.

       Note: Since the receiver's ability to deal with Markers is
           unknown until the Request and Reply Frames have been
           received, sending FPDUs before this occurs is not possible.


       Note: The requirement to wait on a Request Frame before sending a
           Reply Frame is a design choice.  It makes for a well-ordered
           sequence of events at each end, and avoids having to specify
           how to deal with situations where both ends start at the same
           time.

   3.  MPA Initiator mode implementations MUST receive and validate an
       MPA Reply Frame.

       If the MPA Reply Frame is improperly formatted, the
       implementation MUST close the TCP connection and exit MPA.

       If the MPA Reply Frame is properly formatted but is the Private
       Data is not acceptable, or if the Rejected Connection bit is set
       to '1', the implementation MUST exit MPA, leaving the TCP
       connection open.  The ULP may close TCP or use the connection for
       other purposes.

       If the MPA Reply Frame is properly formatted and the Private Data
       is acceptable, and the Reject Connection bit is set to '0', the
       implementation SHOULD enter Full MPA/DDP Operation Phase;
       interpreting any received data as FPDUs and sending DDP ULPDUs as
       FPDUs.

   4.  MPA Responder mode implementations MUST receive and validate at
       least one FPDU before sending any FPDUs or Markers.

       Note: This requirement is present to allow the Initiator time to
           get its receiver into Full Operation before an FPDU arrives,
           avoiding potential race conditions at the Initiator.  This
           was also subject to some debate in the work group before
           rough consensus was reached.  Eliminating this requirement
           would allow faster startup in some types of applications.
           However, that would also make certain implementations
           (particularly "dual stack") much harder.

   5.  If a received "Key" does not match the expected value (see
       Section 7.1.1, MPA Request and Reply Frame Format) the TCP/DDP
       connection MUST be closed, and an error returned to the ULP.



Culley, et al.              Standards Track                    [Page 29]

RFC 5044                  MPA Framing for TCP               October 2007


   6.  The received Private Data fields may be used by Consumers at
       either end to further validate the connection and set up DDP or
       other ULP parameters.  The Initiator ULP MAY close the
       TCP/MPA/DDP connection as a result of validating the Private Data
       fields.  The Responder SHOULD return an MPA Reply Frame with the
       "Reject Connection" bit set to '1' if the validation of the
       Private Data is not acceptable to the ULP.

   7.  When the first FPDU is to be sent, then if Markers are enabled,
       the first octets sent are the special Marker 0x00000000, followed
       by the start of the FPDU (the FPDU's ULPDU Length field).  If
       Markers are not enabled, the first octets sent are the start of
       the FPDU (the FPDU's ULPDU Length field).

   8.  MPA implementations MUST use the difference between the MPA
       Request Frame and the MPA Reply Frame to check for incorrect
       "Initiator/Initiator" startups.  Implementations SHOULD put a
       timeout on waiting for the MPA Request Frame when started in
       Responder mode, to detect incorrect "Responder/Responder"
       startups.

   9.  MPA implementations MUST validate the PD_Length field.  The
       buffer that receives the Private Data field MUST be large enough
       to receive that data; the amount of Private Data MUST not exceed
       the PD_Length or the application buffer.  If any of the above
       fails, the startup frame MUST be considered improperly formatted.

   10. MPA implementations SHOULD implement a reasonable timeout while
       waiting for the entire set of startup frames; this prevents
       certain denial-of-service attacks.  ULPs SHOULD implement a
       reasonable timeout while waiting for FPDUs, ULPDUs, and
       application level messages to guard against application failures
       and certain denial-of-service attacks.

7.1.3.  Example Delayed Startup Sequence



   A variety of startup sequences are possible when using MPA on TCP.
   Following is an example of an MPA/DDP startup that occurs after TCP
   has been running for a while and has exchanged some amount of
   streaming data.  This example does not use any Private Data (an
   example that does is shown later in Section 7.1.4.2, Example
   Immediate Startup Using Private Data), although it is perfectly legal
   to include the Private Data.  Note that since the example does not
   use any Private Data, there are no ULP interactions shown between
   receiving "startup frames" and putting MPA into Full Operation.






Culley, et al.              Standards Track                    [Page 30]

RFC 5044                  MPA Framing for TCP               October 2007


         Initiator                                 Responder

  +---------------------------+
  |ULP streaming mode         |
  |  <Hello> request to       |
  |  transition to DDP/MPA    |           +---------------------------+
  |  mode (optional).         | --------> |ULP gets request;          |
  +---------------------------+           |  enables MPA Responder    |
                                          |  mode with last (optional)|
                                          |  streaming mode           |
                                          |  <Hello Ack> for MPA to   |
                                          |  send.                    |
  +---------------------------+           |MPA waits for incoming     |
  |ULP receives streaming     | <-------- |  <MPA Request Frame>.     |
  |  <Hello Ack>;             |           +---------------------------+
  |Enters MPA Initiator mode; |
  |MPA sends                  |
  |  <MPA Request Frame>;     |
  |MPA waits for incoming     |           +---------------------------+
  |  <MPA Reply Frame>.       | - - - - > |MPA receives               |
  +---------------------------+           |  <MPA Request Frame>.     |
                                          |Consumer binds DDP to MPA; |
                                          |MPA sends the              |
                                          |  <MPA Reply Frame>.       |
                                          |DDP/MPA enables FPDU       |
  +---------------------------+           |  decoding, but does not   |
  |MPA receives the           | < - - - - |  send any FPDUs.          |
  |  <MPA Reply Frame>        |           +---------------------------+
  |Consumer binds DDP to MPA; |
  |DDP/MPA begins Full        |
  |  Operation.               |
  |MPA sends first FPDU (as   |           +---------------------------+
  |  DDP ULPDUs become        | ========> |MPA receives first FPDU.   |
  |  available).              |           |MPA sends first FPDU (as   |
  +---------------------------+           |  DDP ULPDUs become        |
                                  <====== |  available).              |
                                          +---------------------------+

              Figure 9: Example Delayed Startup Negotiation












Culley, et al.              Standards Track                    [Page 31]

RFC 5044                  MPA Framing for TCP               October 2007


   An example Delayed Startup sequence is described below:

       *   Active and passive sides start up a TCP connection in the
           usual fashion, probably using sockets APIs.  They exchange
           some amount of streaming mode data.  At some point, one side
           (the MPA Initiator) sends streaming mode data that
           effectively says "Hello, let's go into MPA/DDP mode".

   *   When the remote side (the MPA Responder) gets this streaming mode
       message, the Consumer would send a last streaming mode message
       that effectively says "I acknowledge your Hello, and am now in
       MPA Responder mode".  The exchange of these messages establishes
       the exact point in the TCP stream where MPA is enabled.  The
       Responding Consumer enables MPA in the Responder mode and waits
       for the initial MPA startup message.

       *   The Initiating Consumer would enable MPA startup in the
           Initiator mode which then sends the MPA Request Frame.  It is
           assumed that no Private Data messages are needed for this
           example, although it is possible to do so.  The Initiating
           MPA (and Consumer) would also wait for the MPA connection to
           be accepted.

   *   The Responding MPA would receive the initial MPA Request Frame
       and would inform the Consumer that this message arrived.  The
       Consumer can then accept the MPA/DDP connection or close the TCP
       connection.

   *   To accept the connection request, the Responding Consumer would
       use an appropriate API to bind the TCP/MPA connections to a DDP
       endpoint, thus enabling MPA/DDP into Full Operation.  In the
       process of going to Full Operation, MPA sends the MPA Reply
       Frame.  MPA/DDP waits for the first incoming FPDU before sending
       any FPDUs.

   *   If the initial TCP data was not a properly formatted MPA Request
       Frame, MPA will close or reset the TCP connection immediately.

       *   The Initiating MPA would receive the MPA Reply Frame and
           would report this message to the Consumer.  The Consumer can
           then accept the MPA/DDP connection, or close or reset the TCP
           connection to abort the process.

       *   On determining that the connection is acceptable, the
           Initiating Consumer would use an appropriate API to bind the
           TCP/MPA connections to a DDP endpoint thus enabling MPA/DDP
           into Full Operation.  MPA/DDP would begin sending DDP
           messages as MPA FPDUs.



Culley, et al.              Standards Track                    [Page 32]

RFC 5044                  MPA Framing for TCP               October 2007


7.1.4.  Use of Private Data



   This section is advisory in nature, in that it suggests a method by
   which a ULP can deal with pre-DDP connection information exchange.

7.1.4.1.  Motivation



   Prior RDMA protocols have been developed that provide Private Data
   via out-of-band mechanisms.  As a result, many applications now
   expect some form of Private Data to be available for application use
   prior to setting up the DDP/RDMA connection.  Following are some
   examples of the use of Private Data.

   An RDMA endpoint (referred to as a Queue Pair, or QP, in InfiniBand
   and the [VERBS-RDMA]) must be associated with a Protection Domain.
   No receive operations may be posted to the endpoint before it is
   associated with a Protection Domain.  Indeed under both the
   InfiniBand and proposed RDMA/DDP verbs [VERBS-RDMA] an endpoint/QP is
   created within a Protection Domain.

   There are some applications where the choice of Protection Domain is
   dependent upon the identity of the remote ULP client.  For example,
   if a user session requires multiple connections, it is highly
   desirable for all of those connections to use a single Protection
   Domain.  Note: Use of Protection Domains is further discussed in
   [RDMASEC].

   InfiniBand, the DAT APIs [DAT-API], and the IT-API [IT-API] all
   provide for the active-side ULP to provide Private Data when
   requesting a connection.  This data is passed to the ULP to allow it
   to determine whether to accept the connection, and if so with which
   endpoint (and implicitly which Protection Domain).

   The Private Data can also be used to ensure that both ends of the
   connection have configured their RDMA endpoints compatibly on such
   matters as the RDMA Read capacity (see [RDMAP]).  Further ULP-
   specific uses are also presumed, such as establishing the identity of
   the client.

   Private Data is also allowed for when accepting the connection, to
   allow completion of any negotiation on RDMA resources and for other
   ULP reasons.

   There are several potential ways to exchange this Private Data.  For
   example, the InfiniBand specification includes a connection
   management protocol that allows a small amount of Private Data to be
   exchanged using datagrams before actually starting the RDMA
   connection.



Culley, et al.              Standards Track                    [Page 33]

RFC 5044                  MPA Framing for TCP               October 2007


   This document allows for small amounts of Private Data to be
   exchanged as part of the MPA startup sequence.  The actual Private
   Data fields are carried in the MPA Request Frame and the MPA Reply
   Frame.

   If larger amounts of Private Data or more negotiation is necessary,
   TCP streaming mode messages may be exchanged prior to enabling MPA.












































Culley, et al.              Standards Track                    [Page 34]

RFC 5044                  MPA Framing for TCP               October 2007


7.1.4.2.  Example Immediate Startup Using Private Data



          Initiator                                 Responder

   +---------------------------+
   |TCP SYN sent.              |           +--------------------------+
   +---------------------------+ --------> |TCP gets SYN packet;      |
   +---------------------------+           |  sends SYN-Ack.          |
   |TCP gets SYN-Ack           | <-------- +--------------------------+
   |  sends Ack.               |
   +---------------------------+ --------> +--------------------------+
   +---------------------------+           |Consumer enables MPA      |
   |Consumer enables MPA       |           |Responder mode, waits for |
   |Initiator mode with        |           |  <MPA Request frame>.    |
   |Private Data; MPA sends    |           +--------------------------+
   |  <MPA Request Frame>;     |
   |MPA waits for incoming     |           +--------------------------+
   |  <MPA Reply Frame>.       | - - - - > |MPA receives              |
   +---------------------------+           |  <MPA Request Frame>.    |
                                           |Consumer examines Private |
                                           |Data, provides MPA with   |
                                           |return Private Data,      |
                                           |binds DDP to MPA, and     |
                                           |enables MPA to send an    |
                                           |  <MPA Reply Frame>.      |
                                           |DDP/MPA enables FPDU      |
   +---------------------------+           |decoding, but does not    |
   |MPA receives the           | < - - - - |send any FPDUs.           |
   |  <MPA Reply Frame>.       |           +--------------------------+
   |Consumer examines Private  |
   |Data, binds DDP to MPA,    |
   |and enables DDP/MPA to     |
   |begin Full Operation.      |
   |MPA sends first FPDU (as   |           +--------------------------+
   |DDP ULPDUs become          | ========> |MPA receives first FPDU.  |
   |available).                |           |MPA sends first FPDU (as  |
   +---------------------------+           |DDP ULPDUs become         |
                                   <====== |available).               |
                                           +--------------------------+

             Figure 10: Example Immediate Startup Negotiation

   Note: The exact order of when MPA is started in the TCP connection
       sequence is implementation dependent; the above diagram shows one
       possible sequence.  Also, the Initiator "Ack" to the Responder's
       "SYN-Ack" may be combined into the same TCP segment containing
       the MPA Request Frame (as is allowed by TCP RFCs).




Culley, et al.              Standards Track                    [Page 35]

RFC 5044                  MPA Framing for TCP               October 2007


   The example immediate startup sequence is described below:

   *   The passive side (Responding Consumer) would listen on the TCP
       destination port, to indicate its readiness to accept a
       connection.

       *   The active side (Initiating Consumer) would request a
           connection from a TCP endpoint (that expected to upgrade to
           MPA/DDP/RDMA and expected the Private Data) to a destination
           address and port.

       *   The Initiating Consumer would initiate a TCP connection to
           the destination port.  Acceptance/rejection of the connection
           would proceed as per normal TCP connection establishment.

   *   The passive side (Responding Consumer) would receive the TCP
       connection request as usual allowing normal TCP gatekeepers, such
       as INETD and TCPserver, to exercise their normal
       safeguard/logging functions.  On acceptance of the TCP
       connection, the Responding Consumer would enable MPA in the
       Responder mode and wait for the initial MPA startup message.

       *   The Initiating Consumer would enable MPA startup in the
           Initiator mode to send an initial MPA Request Frame with its
           included Private Data message to send.  The Initiating MPA
           (and Consumer) would also wait for the MPA connection to be
           accepted, and any returned Private Data.

   *   The Responding MPA would receive the initial MPA Request Frame
       with the Private Data message and would pass the Private Data
       through to the Consumer.  The Consumer can then accept the
       MPA/DDP connection, close the TCP connection, or reject the MPA
       connection with a return message.

   *   To accept the connection request, the Responding Consumer would
       use an appropriate API to bind the TCP/MPA connections to a DDP
       endpoint, thus enabling MPA/DDP into Full Operation.  In the
       process of going to Full Operation, MPA sends the MPA Reply
       Frame, which includes the Consumer-supplied Private Data
       containing any appropriate Consumer response.  MPA/DDP waits for
       the first incoming FPDU before sending any FPDUs.

   *   If the initial TCP data was not a properly formatted MPA Request
       Frame, MPA will close or reset the TCP connection immediately.







Culley, et al.              Standards Track                    [Page 36]

RFC 5044                  MPA Framing for TCP               October 2007


   *   To reject the MPA connection request, the Responding Consumer
       would send an MPA Reply Frame with any ULP-supplied Private Data
       (with reason for rejection), with the "Rejected Connection" bit
       set to '1', and may close the TCP connection.

       *   The Initiating MPA would receive the MPA Reply Frame with the
           Private Data message and would report this message to the
           Consumer, including the supplied Private Data.

           If the "Rejected Connection" bit is set to a '1', MPA will
           close the TCP connection and exit.

           If the "Rejected Connection" bit is set to a '0', and on
           determining from the MPA Reply Frame Private Data that the
           connection is acceptable, the Initiating Consumer would use
           an appropriate API to bind the TCP/MPA connections to a DDP
           endpoint thus enabling MPA/DDP into Full Operation.  MPA/DDP
           would begin sending DDP messages as MPA FPDUs.

7.1.5.  "Dual Stack" Implementations



   MPA/DDP implementations are commonly expected to be implemented as
   part of a "dual stack" architecture.  One stack is the traditional
   TCP stack, usually with a sockets interface API (Application
   Programming Interface).  The second stack is the MPA/DDP stack with
   its own API, and potentially separate code or hardware to deal with
   the MPA/DDP data.  Of course, implementations may vary, so the
   following comments are of an advisory nature only.

   The use of the two stacks offers advantages:

       TCP connection setup is usually done with the TCP stack.  This
       allows use of the usual naming and addressing mechanisms.  It
       also means that any mechanisms used to "harden" the connection
       setup against security threats are also used when starting
       MPA/DDP.

       Some applications may have been originally designed for TCP, but
       are "enhanced" to utilize MPA/DDP after a negotiation reveals the
       capability to do so.  The negotiation process takes place in
       TCP's streaming mode, using the usual TCP APIs.

       Some new applications, designed for RDMA or DDP, still need to
       exchange some data prior to starting MPA/DDP.  This exchange can
       be of arbitrary length or complexity, but often consists of only
       a small amount of Private Data, perhaps only a single message.
       Using the TCP streaming mode for this exchange allows this to be
       done using well-understood methods.



Culley, et al.              Standards Track                    [Page 37]

RFC 5044                  MPA Framing for TCP               October 2007


   The main disadvantage of using two stacks is the conversion of an
   active TCP connection between them.  This process must be done with
   care to prevent loss of data.

   To avoid some of the problems when using a "dual stack" architecture,
   the following additional restrictions may be required by the
   implementation:

   1.  Enabling the DDP/MPA stack SHOULD be done only when no incoming
       stream data is expected.  This is typically managed by the ULP
       protocol.  When following the recommended startup sequence, the
       Responder side enters DDP/MPA mode, sends the last streaming mode
       data, and then waits for the MPA Request Frame.  No additional
       streaming mode data is expected.  The Initiator side ULP receives
       the last streaming mode data, and then enters DDP/MPA mode.
       Again, no additional streaming mode data is expected.

   2.  The DDP/MPA MAY provide the ability to send a "last streaming
       message" as part of its Responder DDP/MPA enable function.  This
       allows the DDP/MPA stack to more easily manage the conversion to
       DDP/MPA mode (and avoid problems with a very fast return of the
       MPA Request Frame from the Initiator side).

   Note: Regardless of the "stack" architecture used, TCP's rules MUST
       be followed.  For example, if network data is lost, re-segmented,
       or re-ordered, TCP MUST recover appropriately even when this
       occurs while switching stacks.

7.2.  Normal Connection Teardown



   Each half connection of MPA terminates when DDP closes the
   corresponding TCP half connection.

   A mechanism SHOULD be provided by MPA to DDP for DDP to be made aware
   that a graceful close of the TCP connection has been received by the
   TCP (e.g., FIN is received).















Culley, et al.              Standards Track                    [Page 38]

RFC 5044                  MPA Framing for TCP               October 2007


8.  Error Semantics



   The following errors MUST be detected by MPA and the codes SHOULD be
   provided to DDP or other Consumer:

   Code Error

   1   TCP connection closed, terminated, or lost.  This includes lost
       by timeout, too many retries, RST received, or FIN received.

   2   Received MPA CRC does not match the calculated value for the
       FPDU.

   3   In the event that the CRC is valid, received MPA Marker (if
       enabled) and ULPDU Length fields do not agree on the start of an
       FPDU.  If the FPDU start determined from previous ULPDU Length
       fields does not match with the MPA Marker position, MPA SHOULD
       deliver an error to DDP.  It may not be possible to make this
       check as a segment arrives, but the check SHOULD be made when a
       gap creating an out-of-order sequence is closed and any time a
       Marker points to an already identified FPDU.  It is OPTIONAL for
       a receiver to check each Marker, if multiple Markers are present
       in an FPDU, or if the segment is received in order.

   4   Invalid MPA Request Frame or MPA Response Frame received.  In
       this case, the TCP connection MUST be immediately closed.  DDP
       and other ULPs should treat this similar to code 1, above.

   When conditions 2 or 3 above are detected, an optimized MPA/TCP
   implementation MAY choose to silently drop the TCP segment rather
   than reporting the error to DDP.  In this case, the sending TCP will
   retry the segment, usually correcting the error, unless the problem
   was at the source.  In that case, the source will usually exceed the
   number of retries and terminate the connection.

   Once MPA delivers an error of any type, it MUST NOT pass or deliver
   any additional FPDUs on that half connection.

   For Error codes 2 and 3, MPA MUST NOT close the TCP connection
   following a reported error.  Closing the connection is the
   responsibility of DDP's ULP.

       Note that since MPA will not Deliver any FPDUs on a half
       connection following an error detected on the receive side of
       that connection, DDP's ULP is expected to tear down the
       connection.  This may not occur until after one or more last
       messages are transmitted on the opposite half connection.  This
       allows a diagnostic error message to be sent.



Culley, et al.              Standards Track                    [Page 39]

RFC 5044                  MPA Framing for TCP               October 2007


9.  Security Considerations



   This section discusses the security considerations for MPA.

9.1.  Protocol-Specific Security Considerations



   The vulnerabilities of MPA to third-party attacks are no greater than
   any other protocol running over TCP.  A third party, by sending
   packets into the network that are delivered to an MPA receiver, could
   launch a variety of attacks that take advantage of how MPA operates.
   For example, a third party could send random packets that are valid
   for TCP, but contain no FPDU headers.  An MPA receiver reports an
   error to DDP when any packet arrives that cannot be validated as an
   FPDU when properly located on an FPDU boundary.  A third party could
   also send packets that are valid for TCP, MPA, and DDP, but do not
   target valid buffers.  These types of attacks ultimately result in
   loss of connection and thus become a type of DOS (Denial Of Service)
   attack.  Communication security mechanisms such as IPsec [RFC2401,
   RFC4301] may be used to prevent such attacks.

   Independent of how MPA operates, a third party could use ICMP
   messages to reduce the path MTU to such a small size that performance
   would likewise be severely impacted.  Range checking on path MTU
   sizes in ICMP packets may be used to prevent such attacks.

   [RDMAP] and [DDP] are used to control, read, and write data buffers
   over IP networks.  Therefore, the control and the data packets of
   these protocols are vulnerable to the spoofing, tampering, and
   information disclosure attacks listed below.  In addition, connection
   to/from an unauthorized or unauthenticated endpoint is a potential
   problem with most applications using RDMA, DDP, and MPA.

9.1.1.  Spoofing



   Spoofing attacks can be launched by the Remote Peer or by a network
   based attacker.  A network-based spoofing attack applies to all
   Remote Peers.  Because the MPA Stream requires a TCP Stream in the
   ESTABLISHED state, certain types of traditional forms of wire attacks
   do not apply -- an end-to-end handshake must have occurred to
   establish the MPA Stream.  So, the only form of spoofing that applies
   is one when a remote node can both send and receive packets.  Yet
   even with this limitation the Stream is still exposed to the
   following spoofing attacks.








Culley, et al.              Standards Track                    [Page 40]

RFC 5044                  MPA Framing for TCP               October 2007


9.1.1.1.  Impersonation



   A network-based attacker can impersonate a legal MPA/DDP/RDMAP peer
   (by spoofing a legal IP address) and establish an MPA/DDP/RDMAP
   Stream with the victim.  End-to-end authentication (i.e., IPsec or
   ULP authentication) provides protection against this attack.

9.1.1.2.  Stream Hijacking



   Stream hijacking happens when a network-based attacker follows the
   Stream establishment phase, and waits until the authentication phase
   (if such a phase exists) is completed successfully.  He can then
   spoof the IP address and redirect the Stream from the victim to its
   own machine.  For example, an attacker can wait until an iSCSI
   authentication is completed successfully, and hijack the iSCSI
   Stream.

   The best protection against this form of attack is end-to-end
   integrity protection and authentication, such as IPsec, to prevent
   spoofing.  Another option is to provide physical security.
   Discussion of physical security is out of scope for this document.

9.1.1.3.  Man-in-the-Middle Attack



   If a network-based attacker has the ability to delete, inject,
   replay, or modify packets that will still be accepted by MPA (e.g.,
   TCP sequence number is correct, FPDU is valid, etc.), then the Stream
   can be exposed to a man-in-the-middle attack.  The attacker could
   potentially use the services of [DDP] and [RDMAP] to read the
   contents of the associated Data Buffer, to modify the contents of the
   associated Data Buffer, or to disable further access to the buffer.
   Other attacks on the connection setup sequence and even on TCP can be
   used to cause denial of service.  The only countermeasure for this
   form of attack is to either secure the MPA/DDP/RDMAP Stream (i.e.,
   integrity protect) or attempt to provide physical security to prevent
   man-in-the-middle type attacks.

   The best protection against this form of attack is end-to-end
   integrity protection and authentication, such as IPsec, to prevent
   spoofing or tampering.  If Stream or session level authentication and
   integrity protection are not used, then a man-in-the-middle attack
   can occur, enabling spoofing and tampering.

   Another approach is to restrict access to only the local subnet/link
   and provide some mechanism to limit access, such as physical security
   or 802.1.x.  This model is an extremely limited deployment scenario
   and will not be further examined here.




Culley, et al.              Standards Track                    [Page 41]

RFC 5044                  MPA Framing for TCP               October 2007


9.1.2.  Eavesdropping



   Generally speaking, Stream confidentiality protects against
   eavesdropping.  Stream and/or session authentication and integrity
   protection are a counter measurement against various spoofing and
   tampering attacks.  The effectiveness of authentication and integrity
   against a specific attack depend on whether the authentication is
   machine-level authentication (as the one provided by IPsec) or ULP
   authentication.

9.2.  Introduction to Security Options



   The following security services can be applied to an MPA/DDP/RDMAP
   Stream:

   1.  Session confidentiality - protects against eavesdropping.

   2.  Per-packet data source authentication - protects against the
       following spoofing attacks: network-based impersonation, Stream
       hijacking, and man in the middle.

   3.  Per-packet integrity - protects against tampering done by
       network-based modification of FPDUs (indirectly affecting buffer
       content through DDP services).

   4.  Packet sequencing - protects against replay attacks, which is a
       special case of the above tampering attack.

   If an MPA/DDP/RDMAP Stream may be subject to impersonation attacks,
   or Stream hijacking attacks, it is recommended that the Stream be
   authenticated, integrity protected, and protected from replay
   attacks.  It may use confidentiality protection to protect from
   eavesdropping (in case the MPA/DDP/RDMAP Stream traverses a public
   network).

   IPsec is capable of providing the above security services for IP and
   TCP traffic.

   ULP protocols may be able to provide part of the above security
   services.  See [NFSv4CHAN] for additional information on a promising
   approach called "channel binding".  From [NFSv4CHAN]:

       "The concept of channel bindings allows applications to prove
       that the end-points of two secure channels at different network
       layers are the same by binding authentication at one channel to
       the session protection at the other channel.  The use of channel





Culley, et al.              Standards Track                    [Page 42]

RFC 5044                  MPA Framing for TCP               October 2007


       bindings allows applications to delegate session protection to
       lower layers, which may significantly improve performance for
       some applications."

9.3.  Using IPsec with MPA



   IPsec can be used to protect against the packet injection attacks
   outlined above.  Because IPsec is designed to secure individual IP
   packets, MPA can run above IPsec without change.  IPsec packets are
   processed (e.g., integrity checked and decrypted) in the order they
   are received, and an MPA receiver will process the decrypted FPDUs
   contained in these packets in the same manner as FPDUs contained in
   unsecured IP packets.

   MPA implementations MUST implement IPsec as described in Section 9.4
   below.  The use of IPsec is up to ULPs and administrators.

9.4.  Requirements for IPsec Encapsulation of MPA/DDP



   The IP Storage working group has spent significant time and effort to
   define the normative IPsec requirements for IP storage [RFC3723].
   Portions of that specification are applicable to a wide variety of
   protocols, including the RDDP protocol suite.  In order not to
   replicate this effort, an MPA on TCP implementation MUST follow the
   requirements defined in RFC 3723, Sections 2.3 and 5, including the
   associated normative references for those sections.

   Additionally, since IPsec acceleration hardware may only be able to
   handle a limited number of active Internet Key Exchange Protocol
   (IKE) Phase 2 security associations (SAs), Phase 2 delete messages
   MAY be sent for idle SAs, as a means of keeping the number of active
   Phase 2 SAs to a minimum.  The receipt of an IKE Phase 2 delete
   message MUST NOT be interpreted as a reason for tearing down a
   DDP/RDMA Stream.  Rather, it is preferable to leave the Stream up,
   and if additional traffic is sent on it, to bring up another IKE
   Phase 2 SA to protect it.  This avoids the potential for continually
   bringing Streams up and down.

   The IPsec requirements for RDDP are based on the version of IPsec
   specified in RFC 2401 [RFC2401] and related RFCs, as profiled by RFC
   3723 [RFC3723], despite the existence of a newer version of IPsec
   specified in RFC 4301 [RFC4301] and related RFCs.  One of the
   important early applications of the RDDP protocols is their use with
   iSCSI [iSER]; RDDP's IPsec requirements follow those of IPsec in
   order to facilitate that usage by allowing a common profile of IPsec
   to be used with iSCSI and the RDDP protocols.  In the future, RFC





Culley, et al.              Standards Track                    [Page 43]

RFC 5044                  MPA Framing for TCP               October 2007


   3723 may be updated to the newer version of IPsec; the IPsec security
   requirements of any such update should apply uniformly to iSCSI and
   the RDDP protocols.



   Note that there are serious security issues if IPsec is not
   implemented end-to-end.  For example, if IPsec is implemented as a
   tunnel in the middle of the network, any hosts between the peer and
   the IPsec tunneling device can freely attack the unprotected Stream.

10.  IANA Considerations



   No IANA actions are required by this document.

   If a well-known port is chosen as the mechanism to identify a DDP on
   MPA on TCP, the well-known port must be registered with IANA.
   Because the use of the port is DDP specific, registration of the port
   with IANA is left to DDP.


































Culley, et al.              Standards Track                    [Page 44]

RFC 5044                  MPA Framing for TCP               October 2007


Appendix A.  Optimized MPA-Aware TCP Implementations



   This appendix is for information only and is NOT part of the
   standard.

   This appendix covers some Optimized MPA-aware TCP implementation
   guidance to implementers.  It is intended for those implementations
   that want to send/receive as much traffic as possible in an aligned
   and zero-copy fashion.

                   +-----------------------------------+
                   | +-----------+ +-----------------+ |
                   | | Optimized | | Other Protocols | |
                   | |  MPA/TCP  | +-----------------+ |
                   | +-----------+        ||           |
                   |         \\     --- socket API --- |
                   |          \\          ||           |
                   |           \\      +-----+         |
                   |            \\     | TCP |         |
                   |             \\    +-----+         |
                   |              \\    //             |
                   |             +-------+             |
                   |             |  IP   |             |
                   |             +-------+             |
                   +-----------------------------------+

                Figure 11: Optimized MPA/TCP Implementation

   The diagram above shows a block diagram of a potential
   implementation.  The network sub-system in the diagram can support
   traditional sockets-based connections using the normal API as shown
   on the right side of the diagram.  Connections for DDP/MPA/TCP are
   run using the facilities shown on the left side of the diagram.

   The DDP/MPA/TCP connections can be started using the facilities shown
   on the left side using some suitable API, or they can be initiated
   using the facilities shown on the right side and transitioned to the
   left side at the point in the connection setup where MPA goes to
   "Full MPA/DDP Operation Phase" as described in Section 7.1.2.

   The optimized MPA/TCP implementations (left side of diagram and
   described below) are only applicable to MPA.  All other TCP
   applications continue to use the standard TCP stacks and interfaces
   shown in the right side of the diagram.







Culley, et al.              Standards Track                    [Page 45]

RFC 5044                  MPA Framing for TCP               October 2007


A.1.  Optimized MPA/TCP Transmitters



   The various TCP RFCs allow considerable choice in segmenting a TCP
   stream.  In order to optimize FPDU recovery at the MPA receiver, an
   optimized MPA/TCP implementation uses additional segmentation rules.

   To provide optimum performance, an optimized MPA/TCP transmit side
   implementation should be enabled to:

   *   With an EMSS large enough to contain the FPDU(s), segment the
       outgoing TCP stream such that the first octet of every TCP
       segment begins with an FPDU.  Multiple FPDUs may be packed into a
       single TCP segment as long as they are entirely contained in the
       TCP segment.

   *   Report the current EMSS from the TCP to the MPA transmit layer.

   There are exceptions to the above rule.  Once an ULPDU is provided to
   MPA, the MPA/TCP sender transmits it or fails the connection; it
   cannot be repudiated.  As a result, during changes in MTU and EMSS,
   or when TCP's Receive Window size (RWIN) becomes too small, it may be
   necessary to send FPDUs that do not conform to the segmentation rule
   above.

   A possible, but less desirable, alternative is to use IP
   fragmentation on accepted FPDUs to deal with MTU reductions or
   extremely small EMSS.

   Even when alignment with TCP segments is lost, the sender still
   formats the FPDU according to FPDU format as shown in Figure 2.

   On a retransmission, TCP does not necessarily preserve original TCP
   segmentation boundaries.  This can lead to the loss of FPDU Alignment
   and containment within a TCP segment during TCP retransmissions.  An
   optimized MPA/TCP sender should try to preserve original TCP
   segmentation boundaries on a retransmission.

A.2.  Effects of Optimized MPA/TCP Segmentation



   Optimized MPA/TCP senders will fill TCP segments to the EMSS with a
   single FPDU when a DDP message is large enough.  Since the DDP
   message may not exactly fit into TCP segments, a "message tail" often
   occurs that results in an FPDU that is smaller than a single TCP
   segment.  Additionally, some DDP messages may be considerably shorter
   than the EMSS.  If a small FPDU is sent in a single TCP segment, the
   result is a "short" TCP segment.





Culley, et al.              Standards Track                    [Page 46]

RFC 5044                  MPA Framing for TCP               October 2007


   Applications expected to see strong advantages from Direct Data
   Placement include transaction-based applications and throughput
   applications.  Request/response protocols typically send one FPDU per
   TCP segment and then wait for a response.  Under these conditions,
   these "short" TCP segments are an appropriate and expected effect of
   the segmentation.

   Another possibility is that the application might be sending multiple
   messages (FPDUs) to the same endpoint before waiting for a response.
   In this case, the segmentation policy would tend to reduce the
   available connection bandwidth by under-filling the TCP segments.

   Standard TCP implementations often utilize the Nagle [RFC896]
   algorithm to ensure that segments are filled to the EMSS whenever the
   round-trip latency is large enough that the source stream can fully
   fill segments before ACKs arrive.  The algorithm does this by
   delaying the transmission of TCP segments until a ULP can fill a
   segment, or until an ACK arrives from the far side.  The algorithm
   thus allows for smaller segments when latencies are shorter to keep
   the ULP's end-to-end latency to reasonable levels.

   The Nagle algorithm is not mandatory to use [RFC1122].

   When used with optimized MPA/TCP stacks, Nagle and similar algorithms
   can result in the "packing" of multiple FPDUs into TCP segments.

   If a "message tail", small DDP messages, or the start of a larger DDP
   message are available, MPA may pack multiple FPDUs into TCP segments.
   When this is done, the TCP segments can be more fully utilized, but,
   due to the size constraints of FPDUs, segments may not be filled to
   the EMSS.  A dynamic MULPDU that informs DDP of the size of the
   remaining TCP segment space makes filling the TCP segment more
   effective.

       Note that MPA receivers do more processing of a TCP segment that
       contains multiple FPDUs; this may affect the performance of some
       receiver implementations.

   It is up to the ULP to decide if Nagle is useful with DDP/MPA.  Note
   that many of the applications expected to take advantage of MPA/DDP
   prefer to avoid the extra delays caused by Nagle.  In such scenarios,
   it is anticipated there will be minimal opportunity for packing at
   the transmitter and receivers may choose to optimize their
   performance for this anticipated behavior.







Culley, et al.              Standards Track                    [Page 47]

RFC 5044                  MPA Framing for TCP               October 2007


   Therefore, the application is expected to set TCP parameters such
   that it can trade off latency and wire efficiency.  Implementations
   should provide a connection option that disables Nagle for MPA/TCP
   similar to the way the TCP_NODELAY socket option is provided for a
   traditional sockets interface.

   When latency is not critical, application is expected to leave Nagle
   enabled.  In this case, the TCP implementation may pack any available
   FPDUs into TCP segments so that the segments are filled to the EMSS.
   If the amount of data available is not enough to fill the TCP segment
   when it is prepared for transmission, TCP can send the segment partly
   filled, or use the Nagle algorithm to wait for the ULP to post more
   data.

A.3.  Optimized MPA/TCP Receivers



   When an MPA receive implementation and the MPA-aware receive side TCP
   implementation support handling out-of-order ULPDUs, the TCP receive
   implementation performs the following functions:

   1)  The implementation passes incoming TCP segments to MPA as soon as
       they have been received and validated, even if not received in
       order.  The TCP layer commits to keeping each segment before it
       can be passed to the MPA.  This means that the segment must have
       passed the TCP, IP, and lower layer data integrity validation
       (i.e., checksum), must be in the receive window, must be part of
       the same epoch (if timestamps are used to verify this), and must
       have passed any other checks required by TCP RFCs.

       This is not to imply that the data must be completely ordered
       before use.  An implementation can accept out-of-order segments,
       SACK them [RFC2018], and pass them to MPA immediately, before the
       reception of the segments needed to fill in the gaps.  MPA
       expects to utilize these segments when they are complete FPDUs or
       can be combined into complete FPDUs to allow the passing of
       ULPDUs to DDP when they arrive, independent of ordering.  DDP
       uses the passed ULPDU to "place" the DDP segments (see [DDP] for
       more details).

       Since MPA performs a CRC calculation and other checks on received
       FPDUs, the MPA/TCP implementation ensures that any TCP segments
       that duplicate data already received and processed (as can happen
       during TCP retries) do not overwrite already received and
       processed FPDUs.  This avoids the possibility that duplicate data
       may corrupt already validated FPDUs.






Culley, et al.              Standards Track                    [Page 48]

RFC 5044                  MPA Framing for TCP               October 2007


   2)  The implementation provides a mechanism to indicate the ordering
       of TCP segments as the sender transmitted them.  One possible
       mechanism might be attaching the TCP sequence number to each
       segment.

   3)  The implementation also provides a mechanism to indicate when a
       given TCP segment (and the prior TCP stream) is complete.  One
       possible mechanism might be to utilize the leading (left) edge of
       the TCP Receive Window.

       MPA uses the ordering and completion indications to inform DDP
       when a ULPDU is complete; MPA Delivers the FPDU to DDP.  DDP uses
       the indications to "deliver" its messages to the DDP consumer
       (see [DDP] for more details).

       DDP on MPA utilizes the above two mechanisms to establish the
       Delivery semantics that DDP's consumers agree to.  These
       semantics are described fully in [DDP].  These include
       requirements on DDP's consumer to respect ownership of buffers
       prior to the time that DDP delivers them to the Consumer.

   The use of SACK [RFC2018] significantly improves network utilization
   and performance and is therefore recommended.  When combined with the
   out-of-order passing of segments to MPA and DDP, significant
   buffering and copying of received data can be avoided.

A.4.  Re-Segmenting Middleboxes and Non-Optimized MPA/TCP Senders



   Since MPA senders often start FPDUs on TCP segment boundaries, a
   receiving optimized MPA/TCP implementation may be able to optimize
   the reception of data in various ways.

   However, MPA receivers MUST NOT depend on FPDU Alignment on TCP
   segment boundaries.

   Some MPA senders may be unable to conform to the sender requirements
   because their implementation of TCP is not designed with MPA in mind.
   Even for optimized MPA/TCP senders, the network may contain
   "middleboxes" which modify the TCP stream by changing the
   segmentation.  This is generally interoperable with TCP and its users
   and MPA must be no exception.

   The presence of Markers in MPA (when enabled) allows an optimized
   MPA/TCP receiver to recover the FPDUs despite these obstacles,
   although it may be necessary to utilize additional buffering at the
   receiver to do so.





Culley, et al.              Standards Track                    [Page 49]

RFC 5044                  MPA Framing for TCP               October 2007


   Some of the cases that a receiver may have to contend with are listed
   below as a reminder to the implementer:

   *   A single aligned and complete FPDU, either in order or out of
       order:  This can be passed to DDP as soon as validated, and
       Delivered when ordering is established.

   *   Multiple FPDUs in a TCP segment, aligned and fully contained,
       either in order or out of order:  These can be passed to DDP as
       soon as validated, and Delivered when ordering is established.

   *   Incomplete FPDU: The receiver should buffer until the remainder
       of the FPDU arrives.  If the remainder of the FPDU is already
       available, this can be passed to DDP as soon as validated, and
       Delivered when ordering is established.

   *   Unaligned FPDU start: The partial FPDU must be combined with its
       preceding portion(s).  If the preceding parts are already
       available, and the whole FPDU is present, this can be passed to
       DDP as soon as validated, and Delivered when ordering is
       established.  If the whole FPDU is not available, the receiver
       should buffer until the remainder of the FPDU arrives.

   *   Combinations of unaligned or incomplete FPDUs (and potentially
       other complete FPDUs) in the same TCP segment:  If any FPDU is
       present in its entirety, or can be completed with portions
       already available, it can be passed to DDP as soon as validated,
       and Delivered when ordering is established.

A.5.  Receiver Implementation



   Transport & Network Layer Reassembly Buffers:

   The use of reassembly buffers (either TCP reassembly buffers or IP
   fragmentation reassembly buffers) is implementation dependent.  When
   MPA is enabled, reassembly buffers are needed if out-of-order packets
   arrive and Markers are not enabled.  Buffers are also needed if FPDU
   alignment is lost or if IP fragmentation occurs.  This is because the
   incoming out-of-order segment may not contain enough information for
   MPA to process all of the FPDU.  For cases where a re-segmenting
   middlebox is present, or where the TCP sender is not optimized, the
   presence of Markers significantly reduces the amount of buffering
   needed.

   Recovery from IP fragmentation is transparent to the MPA Consumers.






Culley, et al.              Standards Track                    [Page 50]

RFC 5044                  MPA Framing for TCP               October 2007


A.5.1  Network Layer Reassembly Buffers



   The MPA/TCP implementation should set the IP Don't Fragment bit at
   the IP layer.  Thus, upon a path MTU change, intermediate devices
   drop the IP datagram if it is too large and reply with an ICMP
   message that tells the source TCP that the path MTU has changed.
   This causes TCP to emit segments conformant with the new path MTU
   size.  Thus, IP fragments under most conditions should never occur at
   the receiver.  But it is possible.

   There are several options for implementation of network layer
   reassembly buffers:

   1.  drop any IP fragments, and reply with an ICMP message according
       to [RFC792] (fragmentation needed and DF set) to tell the Remote
       Peer to resize its TCP segment.

   2.  support an IP reassembly buffer, but have it of limited size
       (possibly the same size as the local link's MTU).  The end node
       would normally never Advertise a path MTU larger than the local
       link MTU.  It is recommended that a dropped IP fragment cause an
       ICMP message to be generated according to RFC 792.

   3.  multiple IP reassembly buffers, of effectively unlimited size.

   4.  support an IP reassembly buffer for the largest IP datagram (64
       KB).

   5.  support for a large IP reassembly buffer that could span multiple
       IP datagrams.

   An implementation should support at least 2 or 3 above, to avoid
   dropping packets that have traversed the entire fabric.

   There is no end-to-end ACK for IP reassembly buffers, so there is no
   flow control on the buffer.  The only end-to-end ACK is a TCP ACK,
   which can only occur when a complete IP datagram is delivered to TCP.
   Because of this, under worst case, pathological scenarios, the
   largest IP reassembly buffer is the TCP receive window (to buffer
   multiple IP datagrams that have all been fragmented).

   Note that if the Remote Peer does not implement re-segmentation of
   the data stream upon receiving the ICMP reply updating the path MTU,
   it is possible to halt forward progress because the opposite peer
   would continue to retransmit using a transport segment size that is
   too large.  This deadlock scenario is no different than if the fabric
   MTU (not last-hop MTU) was reduced after connection setup, and the
   remote node's behavior is not compliant with [RFC1122].



Culley, et al.              Standards Track                    [Page 51]

RFC 5044                  MPA Framing for TCP               October 2007


A.5.2  TCP Reassembly Buffers



   A TCP reassembly buffer is also needed.  TCP reassembly buffers are
   needed if FPDU Alignment is lost when using TCP with MPA or when the
   MPA FPDU spans multiple TCP segments.  Buffers are also needed if
   Markers are disabled and out-of-order packets arrive.

   Since lost FPDU Alignment often means that FPDUs are incomplete, an
   MPA on TCP implementation must have a reassembly buffer large enough
   to recover an FPDU that is less than or equal to the MTU of the
   locally attached link (this should be the largest possible Advertised
   TCP path MTU).  If the MTU is smaller than 140 octets, a buffer of at
   least 140 octets long is needed to support the minimum FPDU size.
   The 140 octets allow for the minimum MULPDU of 128, 2 octets of pad,
   2 of ULPDU_Length, 4 of CRC, and space for a possible Marker.  As
   usual, additional buffering is likely to provide better performance.

   Note that if the TCP segments were not stored, it would be possible
   to deadlock the MPA algorithm.  If the path MTU is reduced, FPDU
   Alignment requires the source TCP to re-segment the data stream to
   the new path MTU.  The source MPA will detect this condition and
   reduce the MPA segment size, but any FPDUs already posted to the
   source TCP will be re-segmented and lose FPDU Alignment.  If the
   destination does not support a TCP reassembly buffer, these segments
   can never be successfully transmitted and the protocol deadlocks.

   When a complete FPDU is received, processing continues normally.

Appendix B.  Analysis of MPA over TCP Operations



   This appendix is for information only and is NOT part of the
   standard.

   This appendix is an analysis of MPA on TCP and why it is useful to
   integrate MPA with TCP (with modifications to typical TCP
   implementations) to reduce overall system buffering and overhead.

   One of MPA's high-level goals is to provide enough information, when
   combined with the Direct Data Placement Protocol [DDP], to enable
   out-of-order placement of DDP payload into the final Upper Layer
   Protocol (ULP) Buffer.  Note that DDP separates the act of placing
   data into a ULP Buffer from that of notifying the ULP that the ULP
   Buffer is available for use.  In DDP terminology, the former is
   defined as "Placement", and the later is defined as "Delivery".  MPA
   supports in-order Delivery of the data to the ULP, including support
   for Direct Data Placement in the final ULP Buffer location when TCP
   segments arrive out of order.  Effectively, the goal is to use the




Culley, et al.              Standards Track                    [Page 52]

RFC 5044                  MPA Framing for TCP               October 2007


   pre-posted ULP Buffers as the TCP receive buffer, where the
   reassembly of the ULP Protocol Data Unit (PDU) by TCP (with MPA and
   DDP) is done in place, in the ULP Buffer, with no data copies.

   This appendix walks through the advantages and disadvantages of the
   TCP sender modifications proposed by MPA:

   1) that MPA prefers that the TCP sender to do Header Alignment, where
      a TCP segment should begin with an MPA Framing Protocol Data Unit
      (FPDU) (if there is payload present).

   2) that there be an integral number of FPDUs in a TCP segment (under
      conditions where the path MTU is not changing).

   This appendix concludes that the scaling advantages of FPDU Alignment
   are strong, based primarily on fairly drastic TCP receive buffer
   reduction requirements and simplified receive handling.  The analysis
   also shows that there is little effect to TCP wire behavior.

B.1.  Assumptions



B.1.1  MPA Is Layered beneath DDP



   MPA is an adaptation layer between DDP and TCP.  DDP requires
   preservation of DDP segment boundaries and a CRC32c digest covering
   the DDP header and data.  MPA adds these features to the TCP stream
   so that DDP over TCP has the same basic properties as DDP over SCTP.

B.1.2.  MPA Preserves DDP Message Framing



   MPA was designed as a framing layer specifically for DDP and was not
   intended as a general-purpose framing layer for any other ULP using
   TCP.

   A framing layer allows ULPs using it to receive indications from the
   transport layer only when complete ULPDUs are present.  As a framing
   layer, MPA is not aware of the content of the DDP PDU, only that it
   has received and, if necessary, reassembled a complete PDU for
   Delivery to the DDP.

B.1.3.  The Size of the ULPDU Passed to MPA Is Less Than EMSS under
        Normal Conditions



   To make reception of a complete DDP PDU on every received segment
   possible, DDP passes to MPA a PDU that is no larger than the EMSS of
   the underlying fabric.  Each FPDU that MPA creates contains
   sufficient information for the receiver to directly place the ULP
   payload in the correct location in the correct receive buffer.



Culley, et al.              Standards Track                    [Page 53]

RFC 5044                  MPA Framing for TCP               October 2007


   Edge cases when this condition does not occur are dealt with, but do
   not need to be on the fast path.

B.1.4.  Out-of-Order Placement but NO Out-of-Order Delivery



   DDP receives complete DDP PDUs from MPA.  Each DDP PDU contains the
   information necessary to place its ULP payload directly in the
   correct location in host memory.

   Because each DDP segment is self-describing, it is possible for DDP
   segments received out of order to have their ULP payload placed
   immediately in the ULP receive buffer.

   Data delivery to the ULP is guaranteed to be in the order the data
   was sent.  DDP only indicates data delivery to the ULP after TCP has
   acknowledged the complete byte stream.

B.2.  The Value of FPDU Alignment



   Significant receiver optimizations can be achieved when Header
   Alignment and complete FPDUs are the common case.  The optimizations
   allow utilizing significantly fewer buffers on the receiver and less
   computation per FPDU.  The net effect is the ability to build a
   "flow-through" receiver that enables TCP-based solutions to scale to
   10G and beyond in an economical way.  The optimizations are
   especially relevant to hardware implementations of receivers that
   process multiple protocol layers -- Data Link Layer (e.g., Ethernet),
   Network and Transport Layer (e.g., TCP/IP), and even some ULP on top
   of TCP (e.g., MPA/DDP).  As network speed increases, there is an
   increasing desire to use a hardware-based receiver in order to
   achieve an efficient high performance solution.

   A TCP receiver, under worst-case conditions, has to allocate buffers
   (BufferSizeTCP) whose capacities are a function of the bandwidth-
   delay product.  Thus:

       BufferSizeTCP = K * bandwidth [octets/second] * Delay [seconds].

   Where bandwidth is the end-to-end bandwidth of the connection, delay
   is the round-trip delay of the connection, and K is an
   implementation-dependent constant.

   Thus, BufferSizeTCP scales with the end-to-end bandwidth (10x more
   buffers for a 10x increase in end-to-end bandwidth).  As this
   buffering approach may scale poorly for hardware or software
   implementations alike, several approaches allow reduction in the
   amount of buffering required for high-speed TCP communication.




Culley, et al.              Standards Track                    [Page 54]

RFC 5044                  MPA Framing for TCP               October 2007


   The MPA/DDP approach is to enable the ULP's Buffer to be used as the
   TCP receive buffer.  If the application pre-posts a sufficient amount
   of buffering, and each TCP segment has sufficient information to
   place the payload into the right application buffer, when an out-of-
   order TCP segment arrives it could potentially be placed directly in
   the ULP Buffer.  However, placement can only be done when a complete
   FPDU with the placement information is available to the receiver, and
   the FPDU contents contain enough information to place the data into
   the correct ULP Buffer (e.g., there is a DDP header available).

   For the case when the FPDU is not aligned with the TCP segment, it
   may take, on average, 2 TCP segments to assemble one FPDU.
   Therefore, the receiver has to allocate BufferSizeNAF (Buffer Size,
   Non-Aligned FPDU) octets:

       BufferSizeNAF = K1* EMSS * number_of_connections + K2 * EMSS

   Where K1 and K2 are implementation-dependent constants and EMSS is
   the effective maximum segment size.

   For example, a 1 GB/sec link with 10,000 connections and an EMSS of
   1500 B would require 15 MB of memory.  Often the number of
   connections used scales with the network speed, aggravating the
   situation for higher speeds.

   FPDU Alignment would allow the receiver to allocate BufferSizeAF
   (Buffer Size, Aligned FPDU) octets:

       BufferSizeAF = K2 * EMSS

   for the same conditions.  An FPDU Aligned receiver may require memory
   in the range of ~100s of KB -- which is feasible for an on-chip
   memory and enables a "flow-through" design, in which the data flows
   through the network interface card (NIC) and is placed directly in
   the destination buffer.  Assuming most of the connections support
   FPDU Alignment, the receiver buffers no longer scale with number of
   connections.

   Additional optimizations can be achieved in a balanced I/O sub-system
   -- where the system interface of the network controller provides
   ample bandwidth as compared with the network bandwidth.  For almost
   twenty years this has been the case and the trend is expected to
   continue.  While Ethernet speeds have scaled by 1000 (from 10
   megabit/sec to 10 gigabit/sec), I/O bus bandwidth of volume CPU
   architectures has scaled from ~2 MB/sec to ~2 GB/sec (PC-XT bus to
   PCI-X DDR).  Under these conditions, the FPDU Alignment approach
   allows BufferSizeAF to be indifferent to network speed.  It is
   primarily a function of the local processing time for a given frame.



Culley, et al.              Standards Track                    [Page 55]

RFC 5044                  MPA Framing for TCP               October 2007


   Thus, when the FPDU Alignment approach is used, receive buffering is
   expected to scale gracefully (i.e., less than linear scaling) as
   network speed is increased.

B.2.1.  Impact of Lack of FPDU Alignment on the Receiver Computational
        Load and Complexity



   The receiver must perform IP and TCP processing, and then perform
   FPDU CRC checks, before it can trust the FPDU header placement
   information.  For simplicity of the description, the assumption is
   that an FPDU is carried in no more than 2 TCP segments.  In reality,
   with no FPDU Alignment, an FPDU can be carried by more than 2 TCP
   segments (e.g., if the path MTU was reduced).

   ----++-----------------------------++-----------------------++-----
   +---||---------------+    +--------||--------+   +----------||----+
   |   TCP Seg X-1      |    |     TCP Seg X    |   |  TCP Seg X+1   |
   +---||---------------+    +--------||--------+   +----------||----+
   ----++-----------------------------++-----------------------++-----
                   FPDU #N-1                  FPDU #N

     Figure 12: Non-Aligned FPDU Freely Placed in TCP Octet Stream

   The receiver algorithm for processing TCP segments (e.g., TCP segment
   #X in Figure 12) carrying non-aligned FPDUs (in order or out of
   order) includes:

   Data Link Layer processing (whole frame) -- typically including a CRC
   calculation.

       1.  Network Layer processing (assuming not an IP fragment, the
           whole Data Link Layer frame contains one IP datagram.  IP
           fragments should be reassembled in a local buffer.  This is
           not a performance optimization goal.)

       2.  Transport Layer processing -- TCP protocol processing, header
           and checksum checks.

           a.  Classify incoming TCP segment using the 5 tuple (IP SRC,
               IP DST, TCP SRC Port, TCP DST Port, protocol).











Culley, et al.              Standards Track                    [Page 56]

RFC 5044                  MPA Framing for TCP               October 2007


       3.  Find FPDU message boundaries.

           a.  Get MPA state information for the connection.

               If the TCP segment is in order, use the receiver-managed
               MPA state information to calculate where the previous
               FPDU message (#N-1) ends in the current TCP segment X.
               (previously, when the MPA receiver processed the first
               part of FPDU #N-1, it calculated the number of bytes
               remaining to complete FPDU #N-1 by using the MPA Length
               field).

                   Get the stored partial CRC for FPDU #N-1.

                   Complete CRC calculation for FPDU #N-1 data (first
                       portion of TCP segment #X).

                   Check CRC calculation for FPDU #N-1.

                   If no FPDU CRC errors, placement is allowed.

                   Locate the local buffer for the first portion of
                       FPDU#N-1, CopyData(local buffer of first portion
                       of FPDU #N-1, host buffer address, length).

                   Compute host buffer address for second portion of
                       FPDU #N-1.

                   CopyData (local buffer of second portion of FPDU #N-
                       1, host buffer address for second portion,
                       length).

                   Calculate the octet offset into the TCP segment for
                       the next FPDU #N.

                   Start calculation of CRC for available data for FPDU.
                       #N

                   Store partial CRC results for FPDU #N.

                   Store local buffer address of first portion of FPDU
                       #N.

                   No further action is possible on FPDU #N, before it
                       is completely received.






Culley, et al.              Standards Track                    [Page 57]

RFC 5044                  MPA Framing for TCP               October 2007


               If the TCP segment is out of order, the receiver must
               buffer the data until at least one complete FPDU is
               received.  Typically, buffering for more than one TCP
               segment per connection is required.  Use the MPA-based
               Markers to calculate where FPDU boundaries are.

                   When a complete FPDU is available, a similar
                   procedure to the in-order algorithm above is used.
                   There is additional complexity, though, because when
                   the missing segment arrives, this TCP segment must be
                   run through the CRC engine after the CRC is
                   calculated for the missing segment.

   If we assume FPDU Alignment, the following diagram and the algorithm
   below apply.  Note that when using MPA, the receiver is assumed to
   actively detect presence or loss of FPDU Alignment for every TCP
   segment received.

      +--------------------------+      +--------------------------+
   +--|--------------------------+   +--|--------------------------+
   |  |       TCP Seg X          |   |  |         TCP Seg X+1      |
   +--|--------------------------+   +--|--------------------------+
      +--------------------------+      +--------------------------+
                FPDU #N                          FPDU #N+1

      Figure 13: Aligned FPDU Placed Immediately after TCP Header

























Culley, et al.              Standards Track                    [Page 58]

RFC 5044                  MPA Framing for TCP               October 2007


   The receiver algorithm for FPDU Aligned frames (in order or out of
   order) includes:

       1)  Data Link Layer processing (whole frame) -- typically
           including a CRC calculation.

       2)  Network Layer processing (assuming not an IP fragment, the
           whole Data Link Layer frame contains one IP datagram.  IP
           fragments should be reassembled in a local buffer.  This is
           not a performance optimization goal.)

       3)  Transport Layer processing -- TCP protocol processing, header
           and checksum checks.

           a.  Classify incoming TCP segment using the 5 tuple (IP SRC,
               IP DST, TCP SRC Port, TCP DST Port, protocol).

       4)  Check for Header Alignment (described in detail in Section
           6).  Assuming Header Alignment for the rest of the algorithm
           below.

           a.  If the header is not aligned, see the algorithm defined
               in the prior section.

       5)  If TCP segment is in order or out of order, the MPA header is
           at the beginning of the current TCP payload.  Get the FPDU
           length from the FPDU header.

       6)  Calculate CRC over FPDU.

       7)  Check CRC calculation for FPDU #N.

       8)  If no FPDU CRC errors, placement is allowed.

       9)  CopyData(TCP segment #X, host buffer address, length).

       10) Loop to #5 until all the FPDUs in the TCP segment are
           consumed in order to handle FPDU packing.

   Implementation note: In both cases, the receiver has to classify the
   incoming TCP segment and associate it with one of the flows it
   maintains.  In the case of no FPDU Alignment, the receiver is forced
   to classify incoming traffic before it can calculate the FPDU CRC.
   In the case of FPDU Alignment, the operations order is left to the
   implementer.






Culley, et al.              Standards Track                    [Page 59]

RFC 5044                  MPA Framing for TCP               October 2007


   The FPDU Aligned receiver algorithm is significantly simpler.  There
   is no need to locally buffer portions of FPDUs.  Accessing state
   information is also substantially simplified -- the normal case does
   not require retrieving information to find out where an FPDU starts
   and ends or retrieval of a partial CRC before the CRC calculation can
   commence.  This avoids adding internal latencies, having multiple
   data passes through the CRC machine, or scheduling multiple commands
   for moving the data to the host buffer.

   The aligned FPDU approach is useful for in-order and out-of-order
   reception.  The receiver can use the same mechanisms for data storage
   in both cases, and only needs to account for when all the TCP
   segments have arrived to enable Delivery.  The Header Alignment,
   along with the high probability that at least one complete FPDU is
   found with every TCP segment, allows the receiver to perform data
   placement for out-of-order TCP segments with no need for intermediate
   buffering.  Essentially, the TCP receive buffer has been eliminated
   and TCP reassembly is done in place within the ULP Buffer.

   In case FPDU Alignment is not found, the receiver should follow the
   algorithm for non-aligned FPDU reception, which may be slower and
   less efficient.

B.2.2.  FPDU Alignment Effects on TCP Wire Protocol



   In an optimized MPA/TCP implementation, TCP exposes its EMSS to MPA.
   MPA uses the EMSS to calculate its MULPDU, which it then exposes to
   DDP, its ULP.  DDP uses the MULPDU to segment its payload so that
   each FPDU sent by MPA fits completely into one TCP segment.  This has
   no impact on wire protocol, and exposing this information is already
   supported on many TCP implementations, including all modern flavors
   of BSD networking, through the TCP_MAXSEG socket option.

   In the common case, the ULP (i.e., DDP over MPA) messages provided to
   the TCP layer are segmented to MULPDU size.  It is assumed that the
   ULP message size is bounded by MULPDU, such that a single ULP message
   can be encapsulated in a single TCP segment.  Therefore, in the
   common case, there is no increase in the number of TCP segments
   emitted.  For smaller ULP messages, the sender can also apply
   packing, i.e., the sender packs as many complete FPDUs as possible
   into one TCP segment.  The requirement to always have a complete FPDU
   may increase the number of TCP segments emitted.  Typically, a ULP
   message size varies from a few bytes to multiple EMSSs (e.g., 64
   Kbytes).  In some cases, the ULP may post more than one message at a
   time for transmission, giving the sender an opportunity for packing.
   In the case where more than one FPDU is available for transmission
   and the FPDUs are encapsulated into a TCP segment and there is no
   room in the TCP segment to include the next complete FPDU, another



Culley, et al.              Standards Track                    [Page 60]

RFC 5044                  MPA Framing for TCP               October 2007


   TCP segment is sent.  In this corner case, some of the TCP segments
   are not full size.  In the worst-case scenario, the ULP may choose an
   FPDU size that is EMSS/2 +1 and has multiple messages available for
   transmission.  For this poor choice of FPDU size, the average TCP
   segment size is therefore about 1/2 of the EMSS and the number of TCP
   segments emitted is approaching 2x of what is possible without the
   requirement to encapsulate an integer number of complete FPDUs in
   every TCP segment.  This is a dynamic situation that only lasts for
   the duration where the sender ULP has multiple non-optimal messages
   for transmission and this causes a minor impact on the wire
   utilization.

   However, it is not expected that requiring FPDU Alignment will have a
   measurable impact on wire behavior of most applications.  Throughput
   applications with large I/Os are expected to take full advantage of
   the EMSS.  Another class of applications with many small outstanding
   buffers (as compared to EMSS) is expected to use packing when
   applicable.  Transaction-oriented applications are also optimal.

   TCP retransmission is another area that can affect sender behavior.
   TCP supports retransmission of the exact, originally transmitted
   segment (see [RFC793], Sections 2.6 and 3.7 (under "Managing the
   Window") and [RFC1122], Section 4.2.2.15).  In the unlikely event
   that part of the original segment has been received and acknowledged
   by the Remote Peer (e.g., a re-segmenting middlebox, as documented in
   A.4">Appendix A.4, Re-Segmenting Middleboxes and Non-Optimized MPA/TCP
   Senders), a better available bandwidth utilization may be possible by
   retransmitting only the missing octets.  If an optimized MPA/TCP
   retransmits complete FPDUs, there may be some marginal bandwidth
   loss.

   Another area where a change in the TCP segment number may have impact
   is that of slow start and congestion avoidance.  Slow-start
   exponential increase is measured in segments per second, as the
   algorithm focuses on the overhead per segment at the source for
   congestion that eventually results in dropped segments.  Slow-start
   exponential bandwidth growth for optimized MPA/TCP is similar to any
   TCP implementation.  Congestion avoidance allows for a linear growth
   in available bandwidth when recovering after a packet drop.  Similar
   to the analysis for slow start, optimized MPA/TCP doesn't change the
   behavior of the algorithm.  Therefore, the average size of the
   segment versus EMSS is not a major factor in the assessment of the
   bandwidth growth for a sender.  Both slow start and congestion
   avoidance for an optimized MPA/TCP will behave similarly to any TCP
   sender and allow an optimized MPA/TCP to enjoy the theoretical
   performance limits of the algorithms.





Culley, et al.              Standards Track                    [Page 61]

RFC 5044                  MPA Framing for TCP               October 2007


   In summary, the ULP messages generated at the sender (e.g., the
   amount of messages grouped for every transmission request) and
   message size distribution has the most significant impact over the
   number of TCP segments emitted.  The worst-case effect for certain
   ULPs (with average message size of EMSS/2+1 to EMSS) is bounded by an
   increase of up to 2x in the number of TCP segments and acknowledges.
   In reality, the effect is expected to be marginal.

Appendix C.  IETF Implementation Interoperability with RDMA Consortium
             Protocols



   This appendix is for information only and is NOT part of the
   standard.

   This appendix covers methods of making MPA implementations
   interoperate with both IETF and RDMA Consortium versions of the
   protocols.

   The RDMA Consortium created early specifications of the MPA/DDP/RDMA
   protocols, and some manufacturers created implementations of those
   protocols before the IETF versions were finalized.  These protocols
   are very similar to the IETF versions making it possible for
   implementations to be created or modified to support either set of
   specifications.

   For those interested, the RDMA Consortium protocol documents (draft-
   culley-iwarp-mpa-v1.0.pdf [RDMA-MPA], draft-shah-iwarp-ddp-v1.0.pdf
   [RDMA-DDP], and draft-recio-iwarp-rdmac-v1.0.pdf [RDMA-RDMAC]) can be
   obtained at http://www.rdmaconsortium.org/home.

   In this section, implementations of MPA/DDP/RDMA that conform to the
   RDMAC specifications are called RDMAC RNICs.  Implementations of
   MPA/DDP/RDMA that conform to the IETF RFCs are called IETF RNICs.

   Without the exchange of MPA Request/Reply Frames, there is no
   standard mechanism for enabling RDMAC RNICs to interoperate with IETF
   RNICs.  Even if a ULP uses a well-known port to start an IETF RNIC
   immediately in RDMA mode (i.e., without exchanging the MPA
   Request/Reply messages), there is no reason to believe an IETF RNIC
   will interoperate with an RDMAC RNIC because of the differences in
   the version number in the DDP and RDMAP headers on the wire.

   Therefore, the ULP or other supporting entity at the RDMAC RNIC must
   implement MPA Request/Reply Frames on behalf of the RNIC in order to
   negotiate the connection parameters.  The following section describes
   the results following the exchange of the MPA Request/Reply Frames
   before the conversion from streaming to RDMA mode.




Culley, et al.              Standards Track                    [Page 62]

RFC 5044                  MPA Framing for TCP               October 2007


C.1.  Negotiated Parameters



   Three types of RNICs are considered:

   Upgraded RDMAC RNIC - an RNIC implementing the RDMAC protocols that
   has a ULP or other supporting entity that exchanges the MPA
   Request/Reply Frames in streaming mode before the conversion to RDMA
   mode.

   Non-permissive IETF RNIC - an RNIC implementing the IETF protocols
   that is not capable of implementing the RDMAC protocols.  Such an
   RNIC can only interoperate with other IETF RNICs.

   Permissive IETF RNIC - an RNIC implementing the IETF protocols that
   is capable of implementing the RDMAC protocols on a per-connection
   basis.

   The Permissive IETF RNIC is recommended for those implementers that
   want maximum interoperability with other RNIC implementations.

   The values used by these three RNIC types for the MPA, DDP, and RDMAP
   versions as well as MPA Markers and CRC are summarized in Figure 14.

    +----------------++-----------+-----------+-----------+-----------+
    | RNIC TYPE      || DDP/RDMAP |    MPA    |    MPA    |    MPA    |
    |                ||  Version  | Revision  |  Markers  |    CRC    |
    +----------------++-----------+-----------+-----------+-----------+
    +----------------++-----------+-----------+-----------+-----------+
    | RDMAC          ||     0     |     0     |     1     |     1     |
    |                ||           |           |           |           |
    +----------------++-----------+-----------+-----------+-----------+
    | IETF           ||     1     |     1     |  0 or 1   |  0 or 1   |
    | Non-permissive ||           |           |           |           |
    +----------------++-----------+-----------+-----------+-----------+
    | IETF           ||  1 or 0   |  1 or 0   |  0 or 1   |  0 or 1   |
    | permissive     ||           |           |           |           |
    +----------------++-----------+-----------+-----------+-----------+

           Figure 14: Connection Parameters for the RNIC Types
            for MPA Markers and MPA CRC, enabled=1, disabled=0.

   It is assumed there is no mixing of versions allowed between MPA,
   DDP, and RDMAP.  The RNIC either generates the RDMAC protocols on the
   wire (version is zero) or uses the IETF protocols (version is one).







Culley, et al.              Standards Track                    [Page 63]

RFC 5044                  MPA Framing for TCP               October 2007


   During the exchange of the MPA Request/Reply Frames, each peer
   provides its MPA Revision, Marker preference (M: 0=disabled,
   1=enabled), and CRC preference.  The MPA Revision provided in the MPA
   Request Frame and the MPA Reply Frame may differ.

   From the information in the MPA Request/Reply Frames, each side sets
   the Version field (V: 0=RDMAC, 1=IETF) of the DDP/RDMAP protocols as
   well as the state of the Markers for each half connection.  Between
   DDP and RDMAP, no mixing of versions is allowed.  Moreover, the DDP
   and RDMAP version MUST be identical in the two directions.  The RNIC
   either generates the RDMAC protocols on the wire (version is zero) or
   uses the IETF protocols (version is one).

   In the following sections, the figures do not discuss CRC negotiation
   because there is no interoperability issue for CRCs.  Since the RDMAC
   RNIC will always request CRC use, then, according to the IETF MPA
   specification, both peers MUST generate and check CRCs.

C.2.  RDMAC RNIC and Non-Permissive IETF RNIC



   Figure 15 shows that a Non-permissive IETF RNIC cannot interoperate
   with an RDMAC RNIC, despite the fact that both peers exchange MPA
   Request/Reply Frames.  For a Non-permissive IETF RNIC, the MPA
   negotiation has no effect on the DDP/RDMAP version and it is unable
   to interoperate with the RDMAC RNIC.

   The rows in the figure show the state of the Marker field in the MPA
   Request Frame sent by the MPA Initiator.  The columns show the state
   of the Marker field in the MPA Reply Frame sent by the MPA Responder.
   Each type of RNIC is shown as an Initiator and a Responder.  The
   connection results are shown in the lower right corner, at the
   intersection of the different RNIC types, where V=0 is the RDMAC
   DDP/RDMAP version, V=1 is the IETF DDP/RDMAC version, M=0 means MPA
   Markers are disabled, and M=1 means MPA Markers are enabled.  The
   negotiated Marker state is shown as X/Y, for the receive direction of
   the Initiator/Responder.















Culley, et al.              Standards Track                    [Page 64]

RFC 5044                  MPA Framing for TCP               October 2007


          +---------------------------++-----------------------+
          |   MPA                     ||          MPA          |
          | CONNECT                   ||       Responder       |
          |   MODE  +-----------------++-------+---------------+
          |         |   RNIC          || RDMAC |     IETF      |
          |         |   TYPE          ||       | Non-permissive|
          |         |          +------++-------+-------+-------+
          |         |          |MARKER|| M=1   | M=0   |  M=1  |
          +---------+----------+------++-------+-------+-------+
          +---------+----------+------++-------+-------+-------+
          |         |   RDMAC  | M=1  || V=0   | close | close |
          |         |          |      || M=1/1 |       |       |
          |         +----------+------++-------+-------+-------+
          |   MPA   |          | M=0  || close | V=1   | V=1   |
          |Initiator|   IETF   |      ||       | M=0/0 | M=0/1 |
          |         |Non-perms.+------++-------+-------+-------+
          |         |          | M=1  || close | V=1   | V=1   |
          |         |          |      ||       | M=1/0 | M=1/1 |
          +---------+----------+------++-------+-------+-------+

           Figure 15: MPA Negotiation between an RDMAC RNIC and
                      a Non-Permissive IETF RNIC

C.2.1.  RDMAC RNIC Initiator



   If the RDMAC RNIC is the MPA Initiator, its ULP sends an MPA Request
   Frame with Rev field set to zero and the M and C bits set to one.
   Because the Non-permissive IETF RNIC cannot dynamically downgrade the
   version number it uses for DDP and RDMAP, it would send an MPA Reply
   Frame with the Rev field equal to one and then gracefully close the
   connection.

C.2.2.  Non-Permissive IETF RNIC Initiator



   If the Non-permissive IETF RNIC is the MPA Initiator, it sends an MPA
   Request Frame with Rev field equal to one.  The ULP or supporting
   entity for the RDMAC RNIC responds with an MPA Reply Frame that has
   the Rev field equal to zero and the M bit set to one.  The Non-
   permissive IETF RNIC will gracefully close the connection after it
   reads the incompatible Rev field in the MPA Reply Frame.

C.2.3.  RDMAC RNIC and Permissive IETF RNIC



   Figure 16 shows that a Permissive IETF RNIC can interoperate with an
   RDMAC RNIC regardless of its Marker preference.  The figure uses the
   same format as shown with the Non-permissive IETF RNIC.





Culley, et al.              Standards Track                    [Page 65]

RFC 5044                  MPA Framing for TCP               October 2007


          +---------------------------++-----------------------+
          |   MPA                     ||          MPA          |
          | CONNECT                   ||       Responder       |
          |   MODE  +-----------------++-------+---------------+
          |         |   RNIC          || RDMAC |     IETF      |
          |         |   TYPE          ||       |  Permissive   |
          |         |          +------++-------+-------+-------+
          |         |          |MARKER|| M=1   | M=0   | M=1   |
          +---------+----------+------++-------+-------+-------+
          +---------+----------+------++-------+-------+-------+
          |         |   RDMAC  | M=1  || V=0   | N/A   | V=0   |
          |         |          |      || M=1/1 |       | M=1/1 |
          |         +----------+------++-------+-------+-------+
          |   MPA   |          | M=0  || V=0   | V=1   | V=1   |
          |Initiator|   IETF   |      || M=1/1 | M=0/0 | M=0/1 |
          |         |Permissive+------++-------+-------+-------+
          |         |          | M=1  || V=0   | V=1   | V=1   |
          |         |          |      || M=1/1 | M=1/0 | M=1/1 |
          +---------+----------+------++-------+-------+-------+

           Figure 16: MPA Negotiation between an RDMAC RNIC and
                         a Permissive IETF RNIC

   A truly Permissive IETF RNIC will recognize an RDMAC RNIC from the
   Rev field of the MPA Req/Rep Frames and then adjust its receive
   Marker state and DDP/RDMAP version to accommodate the RDMAC RNIC.  As
   a result, as an MPA Responder, the Permissive IETF RNIC will never
   return an MPA Reply Frame with the M bit set to zero.  This case is
   shown as a not applicable (N/A) in Figure 16.

C.2.4.  RDMAC RNIC Initiator



   When the RDMAC RNIC is the MPA Initiator, its ULP or other supporting
   entity prepares an MPA Request message and sets the revision to zero
   and the M bit and C bit to one.

   The Permissive IETF Responder receives the MPA Request message and
   checks the revision field.  Since it is capable of generating RDMAC
   DDP/RDMAP headers, it sends an MPA Reply message with revision set to
   zero and the M and C bits set to one.  The Responder must inform its
   ULP that it is generating version zero DDP/RDMAP messages.










Culley, et al.              Standards Track                    [Page 66]

RFC 5044                  MPA Framing for TCP               October 2007


C.2.5  Permissive IETF RNIC Initiator



   If the Permissive IETF RNIC is the MPA Initiator, it prepares the MPA
   Request Frame setting the Rev field to one.  Regardless of the value
   of the M bit in the MPA Request Frame, the ULP or other supporting
   entity for the RDMAC RNIC will create an MPA Reply Frame with Rev
   equal to zero and the M bit set to one.

   When the Initiator reads the Rev field of the MPA Reply Frame and
   finds that its peer is an RDMAC RNIC, it must inform its ULP that it
   should generate version zero DDP/RDMAP messages and enable MPA
   Markers and CRC.

C.3.  Non-Permissive IETF RNIC and Permissive IETF RNIC



   For completeness, Figure 17 below shows the results of MPA
   negotiation between a Non-permissive IETF RNIC and a Permissive IETF
   RNIC.  The important point from this figure is that an IETF RNIC
   cannot detect whether its peer is a Permissive or Non-permissive
   RNIC.

      +---------------------------++-------------------------------+
      |   MPA                     ||              MPA              |
      | CONNECT                   ||            Responder          |
      |   MODE  +-----------------++---------------+---------------+
      |         |   RNIC          ||     IETF      |     IETF      |
      |         |   TYPE          || Non-permissive|  Permissive   |
      |         |          +------++-------+-------+-------+-------+
      |         |          |MARKER|| M=0   | M=1   | M=0   | M=1   |
      +---------+----------+------++-------+-------+-------+-------+
      +---------+----------+------++-------+-------+-------+-------+
      |         |          | M=0  || V=1   | V=1   | V=1   | V=1   |
      |         |   IETF   |      || M=0/0 | M=0/1 | M=0/0 | M=0/1 |
      |         |Non-perms.+------++-------+-------+-------+-------+
      |         |          | M=1  || V=1   | V=1   | V=1   | V=1   |
      |         |          |      || M=1/0 | M=1/1 | M=1/0 | M=1/1 |
      |   MPA   +----------+------++-------+-------+-------+-------+
      |Initiator|          | M=0  || V=1   | V=1   | V=1   | V=1   |
      |         |   IETF   |      || M=0/0 | M=0/1 | M=0/0 | M=0/1 |
      |         |Permissive+------++-------+-------+-------+-------+
      |         |          | M=1  || V=1   | V=1   | V=1   | V=1   |
      |         |          |      || M=1/0 | M=1/1 | M=1/0 | M=1/1 |
      +---------+----------+------++-------+-------+-------+-------+

    Figure 17: MPA negotiation between a Non-permissive IETF RNIC and a
                           Permissive IETF RNIC.





Culley, et al.              Standards Track                    [Page 67]

RFC 5044                  MPA Framing for TCP               October 2007


Normative References

   [iSCSI]      Satran, J., Meth, K., Sapuntzakis, C., Chadalapaka, M.,
                and E. Zeidner, "Internet Small Computer Systems
                Interface (iSCSI)", RFC 3720, April 2004.

   [RFC1191]    Mogul, J. and S. Deering, "Path MTU discovery", RFC
                1191, November 1990.

   [RFC2018]    Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, "TCP
                Selective Acknowledgment Options", RFC 2018, October
                1996.

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

   [RFC2401]    Kent, S. and R. Atkinson, "Security Architecture for the
                Internet Protocol", RFC 2401, November 1998.

   [RFC3723]    Aboba, B., Tseng, J., Walker, J., Rangan, V., and F.
                Travostino, "Securing Block Storage Protocols over IP",
                RFC 3723, April 2004.

   [RFC793]     Postel, J., "Transmission Control Protocol", STD 7, RFC
                793, September 1981.

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

Informative References

   [APPL]       Bestler, C. and L. Coene, "Applicability of Remote
                Direct Memory Access Protocol (RDMA) and Direct Data
                Placement (DDP)", RFC 5045, October 2007.

   [CRCTCP]     Stone J., Partridge, C., "When the CRC and TCP checksum
                disagree", ACM Sigcomm, Sept. 2000.

   [DAT-API]    DAT Collaborative, "kDAPL (Kernel Direct Access
                Programming Library) and uDAPL (User Direct Access
                Programming Library)", Http://www.datcollaborative.org.

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





Culley, et al.              Standards Track                    [Page 68]

RFC 5044                  MPA Framing for TCP               October 2007


   [iSER]       Ko, M., Chadalapaka, M., Hufferd, J., Elzur, U., Shah,
                H., and P. Thaler, "Internet Small Computer System
                Interface (iSCSI) Extensions for Remote Direct Memory
                Access (RDMA)" RFC 5046, October 2007.

   [IT-API]     The Open Group, "Interconnect Transport API (IT-API)"
                Version 2.1, http://www.opengroup.org.

   [NFSv4CHAN]  Williams, N., "On the Use of Channel Bindings to Secure
                Channels", Work in Progress, June 2006.

   [RDMA-DDP]   "Direct Data Placement over Reliable Transports (Version
                1.0)", RDMA Consortium, October 2002,
                <http://www.rdmaconsortium.org/home/draft-shah-iwarp-
                ddp-v1.0.pdf>.

   [RDMA-MPA]   "Marker PDU Aligned Framing for TCP Specification
                (Version 1.0)", RDMA Consortium, October 2002,
                <http://www.rdmaconsortium.org/home/draft-culley-iwarp-
                mpa-v1.0.pdf>.

   [RDMA-RDMAC] "An RDMA Protocol Specification (Version 1.0)", RDMA
                Consortium, October 2002,
                <http://www.rdmaconsortium.org/home/draft-recio-iwarp-
                rdmac-v1.0.pdf>.

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

   [RFC792]     Postel, J., "Internet Control Message Protocol", STD 5,
                RFC 792, September 1981.

   [RFC896]     Nagle, J., "Congestion control in IP/TCP internetworks",
                RFC 896, January 1984.

   [RFC1122]    Braden, R., "Requirements for Internet Hosts -
                Communication Layers", STD 3, RFC 1122, October 1989.

   [RFC4960]    Stewart, R., Ed., "Stream Control Transmission
                Protocol", RFC 4960, September 2007.

   [RFC4296]    Bailey, S. and T. Talpey, "The Architecture of Direct
                Data Placement (DDP) and Remote Direct Memory Access
                (RDMA) on Internet Protocols", RFC 4296, December 2005.






Culley, et al.              Standards Track                    [Page 69]

RFC 5044                  MPA Framing for TCP               October 2007


   [RFC4297]    Romanow, A., Mogul, J., Talpey, T., and S. Bailey,
                "Remote Direct Memory Access (RDMA) over IP Problem
                Statement", RFC 4297, December 2005.

   [RFC4301]    Kent, S. and K. Seo, "Security Architecture for the
                Internet Protocol", RFC 4301, December 2005.

   [VERBS-RMDA] "RDMA Protocol Verbs Specification", RDMA Consortium
                standard, April 2003, <http://www.rdmaconsortium.org/
                home/draft-hilland-iwarp-verbs-v1.0-RDMAC.pdf>.

Contributors

   Dwight Barron
   Hewlett-Packard Company
   20555 SH 249
   Houston, TX 77070-2698 USA
   Phone: 281-514-2769
   EMail: dwight.barron@hp.com

   Jeff Chase
   Department of Computer Science
   Duke University
   Durham, NC 27708-0129 USA
   Phone: +1 919 660 6559
   EMail: chase@cs.duke.edu

   Ted Compton
   EMC Corporation
   Research Triangle Park, NC 27709 USA
   Phone: 919-248-6075
   EMail: compton_ted@emc.com

   Dave Garcia
   24100 Hutchinson Rd.
   Los Gatos, CA  95033
   Phone: 831 247 4464
   EMail: Dave.Garcia@StanfordAlumni.org

   Hari Ghadia
   Gen10 Technology, Inc.
   1501 W Shady Grove Road
   Grand Prairie, TX 75050
   Phone: (972) 301 3630
   EMail: hghadia@gen10technology.com






Culley, et al.              Standards Track                    [Page 70]

RFC 5044                  MPA Framing for TCP               October 2007


   Howard C. Herbert
   Intel Corporation
   MS CH7-404
   5000 West Chandler Blvd.
   Chandler, AZ 85226
   Phone: 480-554-3116
   EMail: howard.c.herbert@intel.com

   Jeff Hilland
   Hewlett-Packard Company
   20555 SH 249
   Houston, TX 77070-2698 USA
   Phone: 281-514-9489
   EMail: jeff.hilland@hp.com

   Mike Ko
   IBM
   650 Harry Rd.
   San Jose, CA 95120
   Phone: (408) 927-2085
   EMail: mako@us.ibm.com

   Mike Krause
   Hewlett-Packard Corporation, 43LN
   19410 Homestead Road
   Cupertino, CA 95014 USA
   Phone: +1 (408) 447-3191
   EMail: krause@cup.hp.com

   Dave Minturn
   Intel Corporation
   MS JF1-210
   5200 North East Elam Young Parkway
   Hillsboro, Oregon  97124
   Phone: 503-712-4106
   EMail: dave.b.minturn@intel.com

   Jim Pinkerton
   Microsoft, Inc.
   One Microsoft Way
   Redmond, WA 98052 USA
   EMail: jpink@microsoft.com









Culley, et al.              Standards Track                    [Page 71]

RFC 5044                  MPA Framing for TCP               October 2007


   Hemal Shah
   Broadcom Corporation
   5300 California Avenue
   Irvine, CA 92617 USA
   Phone: +1 (949) 926-6941
   EMail: hemal@broadcom.com

   Allyn Romanow
   Cisco Systems
   170 W Tasman Drive
   San Jose, CA 95134 USA
   Phone: +1 408 525 8836
   EMail: allyn@cisco.com

   Tom Talpey
   Network Appliance
   1601 Trapelo Road #16
   Waltham, MA  02451 USA
   Phone: +1 (781) 768-5329
   EMail: thomas.talpey@netapp.com

   Patricia Thaler
   Broadcom
   16215 Alton Parkway
   Irvine, CA 92618
   Phone: 916 570 2707
   EMail: pthaler@broadcom.com

   Jim Wendt
   Hewlett Packard Corporation
   8000 Foothills Boulevard MS 5668
   Roseville, CA 95747-5668 USA
   Phone: +1 916 785 5198
   EMail: jim_wendt@hp.com

   Jim Williams
   Emulex Corporation
   580 Main Street
   Bolton, MA 01740 USA
   Phone: +1 978 779 7224
   EMail: jim.williams@emulex.com










Culley, et al.              Standards Track                    [Page 72]

RFC 5044                  MPA Framing for TCP               October 2007


Authors' Addresses



   Paul R. Culley
   Hewlett-Packard Company
   20555 SH 249
   Houston, TX 77070-2698 USA
   Phone: 281-514-5543
   EMail: paul.culley@hp.com

   Uri Elzur
   5300 California Avenue
   Irvine, CA 92617, USA
   Phone: 949.926.6432
   EMail: uri@broadcom.com

   Renato J Recio
   IBM
   Internal Zip 9043
   11400 Burnett Road
   Austin, Texas 78759
   Phone: 512-838-3685
   EMail: recio@us.ibm.com

   Stephen Bailey
   Sandburst Corporation
   600 Federal Street
   Andover, MA 01810 USA
   Phone: +1 978 689 1614
   EMail: steph@sandburst.com

   John Carrier
   Cray Inc.
   411 First Avenue S, Suite 600
   Seattle, WA 98104-2860
   Phone: 206-701-2090
   EMail: carrier@cray.com















Culley, et al.              Standards Track                    [Page 73]

RFC 5044                  MPA Framing for TCP               October 2007


Full Copyright Statement



   Copyright (C) The IETF Trust (2007).

   This document is subject to the rights, licenses and restrictions
   contained in BCP 78, and except as set forth therein, the authors
   retain all their rights.

   This document and the information contained herein are provided on an
   "AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS
   OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY, THE IETF TRUST AND
   THE INTERNET ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS
   OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF
   THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED
   WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.

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 documents 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
   specification can be obtained from the IETF on-line IPR repository at
   http://www.ietf.org/ipr.

   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-ipr@ietf.org.












Culley, et al.              Standards Track                    [Page 74]