Delay Tolerant Networking TCP Convergence Layer Protocol

This document describes a TCP based convergence layer protocol for Delay Tolerant Networking (TCPCL). Delay Tolerant Networking is an end-to-end architecture providing communications in and/or through highly stressed environments, including those with intermittent connectivity, long and/or variable delays, and high bit error rates. More detailed descriptions of the rationale and capabilities of these networks can be found in the Delay-Tolerant Network Architecture Internet Draft. An important goal of the DTN architecture is to accomodate a wide range of networking technologies and environments. The protocol used for DTN communications is the Bundling Protocol (BP), an application-layer protocol that is used to construct a store-and-forward overlay network. As described in the bundle protocol specification, BP requires the services of a "convergence layer adapter" (CLA) to send and receive bundles using an underlying internet protocol. This document describes one such convergence layer adapter that uses the well-known Transmission Control Protocol (TCP). The locations of the TCPCL and BP in the Internet model protocol stack are shown in Figure . In particular, both the BP and the TCPCL sit at the application layer, while TCP and IP sit at their normal layers.

+-------------------------+ | DTN Application | -\ +-------------------------| | | Bundle Protocol (BP) | -> Application Layer +-------------------------+ | | TCP Conv. Layer (TCPCL) | -/ +-------------------------+ | TCP | ---> Transport Layer +-------------------------+ | IP | ---> Network Layer +-------------------------+ | Link-Layer Protocol | ---> Link Layer +-------------------------+ | Physical Medium | ---> Physical Layer +-------------------------+ This document describes the format of the protocol data units passed between entities participating in TCPCL communications. This document does not address: The format of protocol data units of the bundling protocol, as those are defined elsewhere . Mechanisms for locating or identifying other bundle nodes within an internet. Operational logic or procedures used to implement this protocol. Note that this document describes version 3 of the protocol. Versions 0, 1, and 2 were never specified in any Internet Draft, RFC, or any other public document. These prior versions of the protocol were, however, implemented in the DTN reference implementation, in prior releases, hence the current version number reflects those prior versions.

The following set of definitions are abbreviated versions of those which appear in the Bundle Protocol Specification. To the extent in which terms appear in both documents, they are intended to have the same meaning. A bundle is a protocol data unit of the DTN bundle protocol. A bundle payload (or simply "payload") is the application data whose conveyance to the bundle's destination is the purpose for the transmission of a given bundle. A fragment is a bundle whose payload contains a range of bytes from another bundle's payload. A bundle node (or simply a "node") is any entity that can send and/or receive bundles. The particular instantiation of this entity is deliberately unconstrained, allowing for implementations in software libraries, long-running processes, or even hardware. One component of the bundle node is the implementation of a convergence layer adapter. A convergence layer adapter (CLA) sends and receives bundles utilizing the services of some 'native' internet protocol. This document describes the manner in which a CLA sends and receives bundles when using the TCP protocol for inter-node communication. A self describing numeric value (SDNV) is a variable length encoding for integer values, defined in the bundle protocol specification.

This section contains definitions that are interpreted to be specific to the operation of the TCPCL protocol, as described below. A TCP connection refers to a network connection using TCP as the as transport protocol. A TCPCL connection (as opposed to a TCP connection) is a TCPCL communication relationship between two bundle nodes. The lifetime of a TCPCL connection is one-to-one with the lifetime of an underlying TCP connection. Therefore a TCPCL connection is initated when a bundle node initiates a TCP connection to be initiated for the purposes of bundle communication. A TCPCL connection is terminated when the TCP connection ends, due either to one or both nodes actively terminating the TCP connection or due to network errors causing a failure of the TCP connection. For the remainder of this document, the term "connection" without the prefix "TCPCL" shall refer to a TCPCL connection. The connection parameters are a set of values used to affect the operation of the TCPCL for a given connection. The manner in which these parameters are conveyed to the bundle node and thereby to the TCPCL is implementation-dependent. However, the mechanism by which two bundle nodes exchange and negotiate the values to be used for a given session is described in Section . The connection initiator is the bundle node that causes the establishment of a new connection by creating a new TCP connection (for example, by using the connect() call in the BSD sockets API) and then following the procedures described in . The connection acceptor is the bundle node that establishes a connection in response to an active connection attempt by another bundle node (for example, by using the listen() and accept() calls of the BSD sockets API) and then following the procedures described in . Transmission refers to the procedures and mechanisms (described below) for conveyance of a bundle from one node to another.

This protocol provides bundle conveyance over a TCP connection. The general operation of the protocol is as follows: First one node establishes a connection to the other by initiating a TCP connection. At the beginning of the connection, an initial contact header is exchanged in both directions to set parameters of the connection and exchange a singleton endpoint identifier for each node, to denote the bundle-layer identity of each DTN node. Once the connection is established, bundles can be transmitted in either direction. Each bundle is transmitted in one or more logical segments of formatted bundle data. Each logical data segment consists of a DATA_SEGMENT message header, an SDNV containing the length of the segment, and finally the byte range of the bundle data. The choice of the length to use for segments is an implementation manner. The first segment for a bundle must set the 'start' flag and the last must set the 'end' flag in the DATA_SEGMENT message header. An optional feature of the protocol is for the receiving node to send acknowledgements as bundle data segments arrive (ACK_SEGMENT). The rationale behind these acknowledgements is to enable the sender node to determine how much of the bundle has been received, so that in case the connection is interrupted, it can perform reactive fragmentation to avoid re-sending the already transmitted part of the bundle. When acknowledgements are enabled, then for each data segment that is received, the receiving node sends an ACK_SEGMENT code followed by an SDNV containing the cumulative length of the bundle that has been received. Note that in the case of concurrent bidirectional transmission, then ack segments may be interleaved with data segments. Another optional feature is that a receiver may interrupt the transmission of a bundle at any point in time by replying with a negative acknowledgement (REFUSE_BUNDLE) which causes the sender to stop transmission of the current bundle, after completing transmission of all partially sent data segments. Note: This allows a receiver that detects it already has received a certain bundle to interrupt transmission as early as possible and thus save transmission capacity for other bundles. For connections that are idle, a KEEPALIVE code may optionally be sent at a negotiated interval. This is used to convey liveness information. Finally, before connections close, a SHUTDOWN code is sent on the channel. After sending a SHUTDOWN message, the sender of this message may send further acknowledgements (ACK_SEGMENT or REFUSE_BUNDLE) but no further data messages (DATA_SEGMENT). A SHUTDOWN message may also be used to refuse a connection setup by a peer.

The following figure visually depicts the protocol exchange for a simple session, showing the connection establishment, and the transmission of a single bundle split into three data segments (of lengths L1, L2, and L3) from Node A to Node B. Note that the sending node may transmit multiple DATA_SEGMENT messages without necessarily waiting for the corresponding ACK_SEGMENT responses. This enables pipelining of messages on a channel. Although this example only demonstrates a single bundle transmission, it is also possible to pipeline multiple DATA_SEGMENT messages for different bundles without necessarily waiting for ACK_SEGMENT messages to be returned for each one. No error or rejection conditions are shown in this example.

Node A Node B ====== ====== +-------------------------+ +-------------------------+ | Contact Header | -> <- | Contact Header | +-------------------------+ +-------------------------+ +-------------------------+ | DATA_SEGMENT (start) | -> | SDNV length [L1] | -> | Bundle Data 0..L1 | -> +-------------------------+ +-------------------------+ +-------------------------+ | DATA_SEGMENT | -> <- | ACK_SEGMENT | | SDNV length [L2] | -> <- | SDNV length [L1] | | Bundle Data L1..L2 | -> +-------------------------+ +-------------------------+ +-------------------------+ +-------------------------+ | DATA_SEGMENT (end) | -> <- | ACK_SEGMENT | | SDNV length [L3] | -> <- | SDNV length [L1+L2] | | Bundle Data L2..L3 | -> +-------------------------+ +-------------------------+ +-------------------------+ <- | ACK_SEGMENT | <- | SDNV length [L1+L2+L3] | +-------------------------+ +-------------------------+ +-------------------------+ | SHUTDOWN | -> <- | SHUTDOWN | +-------------------------+ +-------------------------+

For bundle transmissions to occur using the TCPCL, a connection must first be established between communicating nodes. The manner in which a bundle node makes the decision to establish such a connection is implementation dependant. For example, some connections may be opened proactively and maintained for as long as is possible given the network conditions, while other connections may be opened only when there is a bundle that is queued for transmission and the routing algorithm selects a certain next hop node. To establish a TCPCL connection, a node must first establish a TCP connection with the intended peer node, typically by using the services provided by the operating system. If the node is unable to establish a TCP connection for any reason, then it is an implementation manner to determine how to handle the failed connection. For example, a node may decide to re-attempt to establish the connection, perhaps after some delay or it may attempt to find an alternate route for bundle data. Once a TCP connection is established, the connection initiator MUST immediately transmit a contact header over the TCP connection. The connection acceptor MUST also immediately transmit a contact header over the TCP connection. The format of the contact header is described in ). Upon receipt of the contact header, both nodes perform the validation and negotiation procedures defined in After receiving the contact header from the other node, either node MAY also refuse the connection by sending a SHUTDOWN message. If connection setup is refused a reason MUST be included in the SHUTDOWN message.

Once a TCP connection is established, both parties exchange a contact header. This section describes the format of the contact header and the meaning of its fields.

The format for the Contact Header is as follows: 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 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 +---------------+---------------+---------------+---------------+ | magic='dtn!' | +---------------+---------------+---------------+---------------+ | version | flags | keepalive_interval | +---------------+---------------+---------------+---------------+ | local eid length (SDNV) | +---------------+---------------+---------------+---------------+ | local eid (variable) | +---------------+---------------+---------------+---------------+ The fields of the contact header are: A four byte field that always contains the byte sequence 0x64 0x76 0x6e 0x21, i.e. the ASCII string "dtn!". A one byte field value containing the current version of the protocol. A one byte field containing optional connection flags. The first five bits are unused and must be set to zero. The last three bits are interpreted as follows: A two byte integer field containing the number of seconds between exchanges of keepalive messages on the connection (see ). This value is in network byte order, as are all other multi-byte fields described in this protocol. A variable length SDNV field containing the length of the endpoint identifier (EID) for some singleton endpoint in which the sending node is a member. A four byte SDNV is depicted for clarity of the figure. An octet string containing the EID of some singleton endpoint in which the sending node is a member, in the canonical format of <scheme name>:<scheme-specific part>. A four byte EID is shown the clarity of the figure. Value Meaning 00000001 Request acknowledgement of bundle segments. 00000010 Request enabling of reactive fragmentation. 00000100 Indicate support for negative acknowledgements. This flags MUST NOT be used unless support for acknowledgements is also indicated. The manner in which values are configured and chosen for the various flags and parameters in the contact header is implementation dependent.

Upon reception of the contact header, both the connection initiator and the connection acceptor follow the following procedures for ensuring the validity of the connection and to negotiate values for the connection parameters. If the magic string is not present or is not valid, the connection MUST be terminated. The intent of the magic string is to provide a some protection against an inadvertent TCP connection by a different protocol than the one described in this document. To prevent a flood of repeated connections from a misconfigured application, a node MAY elect to hold an invalid connection open and idle for some time before closing it. If a node receives a contact header containing a version that is greater than the current version of the protocol that the node implements, then the node SHOULD interpret all fields and messages as it would normally. If a node receives a contact header with a version that is lower than the version of the protocol that the node implements, the node may either terminate the connection due to the version mismatch, or may adapt its operation to conform to the older version of the protocol. This decision is an implementation manner. A node calculates the parameters for a connection by negotiating the values from its own preferences (conveyed by the contact header it sent) with the preferences of the peer node (expressed in the contact header that it received). This negotiation should proceed in the following manner: The segment acknowledgements enabled parameter is set to true iff the corresponding flag is set in both contact headers. The reactive fragmentation enabled parameter is set to true iff the corresponding flag is set in both contact headers. Negative acknowledgements to interrupt transmission (actually: refuse reception) of a bundle may only be used iff both peers indicate support for negative acknowledements in their contact header. The keealive_interval parameter should be set to the minimum value from both contact headers. If one or both contact headers contains the value zero, then the keepalive feature (described in ) is disabled. Once this process of parameter negotiation is completed, the protocol defines no additional mechanism to change the parameters of an established connection; to effect such a change, the connection MUST be terminated and a new connection established.

This section describes the protocol operation for the duration of an established connection, including the mechanisms for transmitting bundles over the connection.

After the initial exchange of a contact header, all messages transmitted over the connection are denoted by a one octet header with the following structure: 0 1 2 3 4 5 6 7 +-+-+-+-+-+-+-+-+ | type | flags | +-+-+-+-+-+-+-+-+ type: Indicates the type of the message as per below flags: Optional flags defined on a per message type basis. The types and values for the message type code are as follows. Type Code Comment DATA_SEGMENT 0x1 Indicates the transmission of a segment of bundle data, described in . ACK_SEGMENT 0x2 Acknowledges reception of a data segment, described in REFUSE_BUNDLE 0x3 Indicates that the transmission of the current bundle shall be stopped, decscribed in . KEEPALIVE 0x4 Keepalive message for the connection, described in . SHUTDOWN 0x5 Indicates that one of the nodes participating in the connection wishes to cleanly terminate the connection, described in .

Each bundle is transmitted in one or more data segments. The format of a data segment message is as follows: 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 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 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | 0x1 |0|0|S|E| length ... | contents.... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ The type portion of the message header contains the value 0x1. The flags portion of the message header octet contains two optional values in the two low-order bits, denoted 'S' and 'E' above. The 'S' bit MUST be set to one iff it precedes the transmission of the first segment of a new bundle. The 'E' bit MUST be set to one when transmitting the last segment of a bundle. Determining the segment size is an implementation manner. In particular, a node may, based on local policy or configuration, only ever transmit bundle data in a single segment, in which case both the 'S' and 'E' bits MUST be set to one. However a node must be able to receive a bundle that has been tranmsmitted in any segment size. In the bundle protocol specification, a single bundle is comprised of a primary bundle block, a payload block, and zero or more additional bundle blocks. The relationship between the protocol blocks and the convergence layer segments is an implementation-specific decision. In particular, a segment may contain more than one protocol block; alternatively, a single protocol block (such as the payload) may be split into multiple segments. Once a transmission of a bundle has commenced, the node MUST only send segments containing sequential portions of that bundle until it sends a segment with the 'E' bit set. Following the message header, the length field is an SDNV containing the number of bytes of bundle data that are transmitted in this segment. Following this length is the actual data contents.

Although the TCP transport provides reliable transfer of data between hosts, the typical BSD sockets interface provides no means to inform a sending application of when the receiving application has processed some amount of transmitted data. Thus after transmitting some data, a bundle protocol agent needs an additional mechanism to determine whether the receiving agent has successfully received the segment. To this end, the TCPCL protocol offers an optional feature whereby a receiving node transmits acknowledgements of reception of data segments. This feature is enabled if and only if during the exchange of contact headers, both parties set the flag to indicate that segment acknowledgements are enabled (see ). If so, then the receiver must transmit a bundle acknowledgement header when it successfully receives each data segment. The format of a bundle acknowledgement is: 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 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 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | 0x2 |0|0|0|0| acknowledged length ... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ To transmit an acknowledgement, a node first transmits a message header with the ACK_SEGMENT type code and all flags set to zero, then transmits an SDNV containing the cumulative length of the received segment(s) of the current bundle. For example, suppose the sending node transmits four segments of bundle data with lengths 100, 200, 500, and 1000 respectively. After receiving the first segment, the node sends an acknowledgement of length 100. After the second segment is received, the node sends an acknowledgement of length 300. The third and fourth acknowledgements are of length 800 and 1800 respectively.

As bundles may be large, the TCPCL supports an optional mechanisms by which a receiving node may indicate to the sender that it does not want to receive the corresponding bundle. To do so, upon receiving a DATA_SEGMENT message, the node may transmit a REFUSE_BUNDLE message. As data segments and acknowledgements may cross on the wire, the data segment (and thus the bundle) that is being refused is implicitly identified by the sequence in which positive and negative acknowledgements are received. The receiver MUST have acknowledged (positively or negatively) all other received DATA_SEGMENTs before the one to be refused so that the sender can identify the bundles accepted and refused by means of a simple FIFO list of segments and acknowledgments. The bundle refusal MAY be sent before the entire data segment is received. If a sender receives a REFUSE_BUNDLE message, the sender MUST complete the transmission of any partially-sent DATA_SEGMENT message (so that the receiver stays in sync). The sender MUST NOT commence transmission of any further segments of the rejected bundles subsequently. Note, however, that this requirement does not ensure that a node will not receive another DATA_SEGMENT for the same bundle after transmitting a REFUSE_BUNDLE message since messages may cross on the wire.

The protocol includes a provision for transmission of keepalive messages over the TCP connection to help determine if the connection has been disrupted. As described in , one of the parameters in the contact header is the keepalive_interval. Both sides populate this field with their requested intervals (in seconds) between keepalive messages. The format of a keepalive message is a one byte message type code of KEEPALIVE (as described in , with no additional data. Both sides should send a keepalive message whenever the negotiated interval has elapsed with no transmission of any message (keepalive or other). If no message (keepalive or other) has been received for at least twice the keepalive interval, then either party may terminate the session by transmitting a one byte message type code of SHUTDOWN (as described in ) and closing the TCP connection.

This section describes the procedures for ending a TCPCL connection.

To cleanly shut down a connection, a SHUTDOWN message can be transmitted by either the sender or the receiver at any point following complete transmission of any other message. In case acknowledgements have been negotiated, it is advisable to acknowledge all received data segments first and then shut down the connection. The format of the shutdown message is as follows: 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 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 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | 0x3 |0|0|R|D| reason (opt) | reconnection delay (opt) | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ It is possible for a node to convey additional information regarding the reason for connection termination. To do so, the node sets the 'R' bit in the message header flags, and transmits a one-byte reason code immediately following the message header. The specified values of the reason code are: Code Meaning Comment 0x00 Idle timeout The connection is being closed due to idleness. 0x01 Version mismatch The node cannot conform to the specified TCPCL protocol version. 0x02 Busy The node is too busy to handle the current connection. It is also possible to convey a requested reconnection delay to indicate how long the other node must wait before attempting connection re-establishment. To do so, the node sets the 'D' bit in the message header flags, then transmits an SDNV specifying the requested delay, in seconds, following the message header (and optionally the shutdown reason code). The value 0 shall be interpreted as an infinite delay, i.e. that the node should not ever re-establish the connection. In contrast, if the node does not wish to request a delay, it should omit the delay field (and set the 'D' bit to zero). Note that in the figure above, a two octet SDNV is shown for convenience in representation. A connection shutdown MAY occur immediately after TCP connection establishment or reception of a contact header (and prior to any further data exchange). This may, for example, be used to notify a the initiator that the node is currenly not capable of or willing to communicate. Note, however, that a node MUST always send the contact header to its peer first. If either node terminates a connection prematurely in this manner, it SHOULD send a SHUTDOWN message and MUST indicate a reason code unless the incoming connection did not include the magic string. If a node does not want its peer to re-open the connection immediately, it SHOULD set the 'D' bit in the flags and include a reconnection delay to indicate when the peer is allowed to attempt another connection setup. If a connection is terminated before another protocol message has completed, then the node must not transmit the SHUTDOWN message but still should close the TCP connection. In particular, if the connection is interrupted while a node is in the process of transmitting a bundle data segment, then the node may identify that the connection should be terminated before it has completed the transmission of the data segment. Thus were the node to transmit the SHUTDOWN message, the receiving node might erroneously interpret the SHUTDOWN message to be part of the data segment.

The protocol includes a provision for clean shutdown of idle TCP connections. Determining the length of time to wait before closing idle connections, if they are to be closed at all, is an implementation and configuration matter. If there is a configured time to close idle links, then if no bundle data (other than keepalive messages) has been received for at least that amount of time, then either node may terminate the connection by transmitting a SHUTDOWN message indicating the reason code of 'idle timeout' (as described above). After receiving a SHUTDOWN message in response, both sides may close the TCP connection.

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 .

One security consideration for this protocol relates to the fact that nodes present their endpoint identifier as part of the connection header exchange. It would be possible for a node to falsify this value and present the identity of a singleton endpoint in which the node is not a member, essentially masquerading as another DTN node. If this identifier is used without further verification as a means to determine which bundles are transmitted over the connection, then the node that has falsified its identity may be able to obtain bundles that it should not have. These concerns may be mitigated through the use of the Bundle Security Protocols. In particular, the Bundle Authentication Header defines mechanism for secure exchange of bundles between DTN nodes. Thus an implementation could delay trusting the presented endpoint identifier until the node can securely validate that its peer is in fact the only member of the given singleton endpoint. Another consideration for this protocol relates to denial of service attacks. A node may send a large amount of data over a TCP connection, requiring the receiving node to either handle the data, attempt to stop the flood of data by sending a REFUSE_BUNDLE message, or forcibly terminate the connection. This burden could cause denial of service on other, well-behaving connections. There is also nothing to prevent a malicious node from continually establishing connections and repeatedly trying to send copious amounts of bundle data.

There should be a well-known TCP port assignment for this protocol.