| Internet-Draft | TCPLS | July 2022 |
| Piraux, et al. | Expires 12 January 2023 | [Page] |
This document specifies a protocol leveraging TCP and TLS to provide modern transport services such as multiplexing, connection migration and multipath in a secure manner.¶
This note is to be removed before publishing as an RFC.¶
Source for this draft and an issue tracker can be found at https://github.com/mpiraux/draft-piraux-tcpls.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 12 January 2023.¶
Copyright (c) 2022 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
The TCP/IP protocol stack continuously evolves. In the early days, most applications were interacting with the transport layer (mainly TCP, but also UDP) using the socket API. This is illustrated in Figure 1.¶
+------------------------------+ | Application | +------------------------------+ | TCP/UDP | +------------------------------+ | IPv4 | +------------------------------+
The TCP/IP stack has slowly evolved and the figure above does not anymore describe current Internet applications. IPv6 is now widely deployed next to IPv4 in the network layer. In the transport layer, protocols such as SCTP [RFC4960] or DCCP [RFC6335] and TCP extensions including Multipath TCP [RFC8684] or tcpcrypt [RFC8548] have been specified. The security aspects of the TCP/IP protocol suite are much more important today than in the past [RFC7258]. Many applications rely on TLS [RFC8446] and their stack is similar to the one shown in Figure 2.¶
+------------------------------+ | Application | +------------------------------+ | TLS | +------------------------------+ | TCP | +------------------------------+ | IPv4/IPv6 | +------------------------------+
Recently, the IETF went one step further in improving the transport layer with the QUIC protocol [RFC9000]. QUIC is a new secure transport protocol primarily designed for HTTP/3. It includes the reliability and congestion control features that are part of TCP and integrates the security features of TLS 1.3 [RFC8446]. This close integration between the reliability and security features brings a lot of benefits in QUIC. QUIC runs above UDP to be able to pass through most middleboxes and to be implementable in user space. While QUIC reuses TLS, it does not strictly layer TLS on top of UDP as DTLS [I-D.ietf-tls-dtls13]. This organization, illustrated in Figure 3 provides much more flexibility than simply layering TLS above UDP. For example, the QUIC migration capabilities enable an application to migrate an existing QUIC session from an IPv4 path to an IPv6 one.¶
+------------------------------+ | Application | +------------------------------+ |.......... | | TLS | QUIC ..........| |.......... | UDP | +------------------------------+ | IPv4/IPv6 | +------------------------------+
In this document, we revisit how TCP and TLS 1.3 can be used to provide modern transport services to applications. We apply a similar principle and combine TCP and TLS 1.3 in a protocol that we call TCPLS. TCPLS leverages the security features of TLS 1.3 like QUIC, but without begin simply layered above a single TCP connection. In addition, TCPLS reuses the existing TCP stacks and TCP's wider support in current networks. A preliminary version of the TCPLS protocol is described in [CONEXT21].¶
+------------------------------+ | Application | +------------------------------+ |.......... | | TLS | TCPLS ..........| |.......... | TCP | +------------------------------+ | IPv4/IPv6 | +------------------------------+
In this document, we use the term TLS/TCP to refer to the TLS 1.3 protocol running over one TCP connection. We reserve the word TCPLS for the protocol proposed in this document.¶
This document is organized as follows. First, Section 3 summarizes the different types of services that modern transports expose to application. Section 4 gives an overview of TCPLS and how it supports these services. Finally, Section 5 describes the TCPLS in more details and the TLS Extensions introduced in this document.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
This document uses the same conventions as defined in Section 1.3 of [RFC9000].¶
This document uses network byte order (that is, big endian) values. Fields are placed starting from the high-order bits of each byte.¶
Application requirements and the devices they run on evolve over time. In the early days, most applications involved single-file transfer and ran on single-homed computers with a fixed-line network. Today, web-based applications require exchanging multiple objects, with different priorities, on devices that can move from one access network to another and that often have multiple access networks available. Security is also a key requirement of applications that evolved from only guaranteeing the confidentiality and integrity of application messages to also preventing pervasive monitoring.¶
With TCP and TLS/TCP, applications use a single connection that supports a single bytestream in each direction. Some TCP applications such as HTTP/2 [RFC7540] use multiple streams, but these are mapped to a single TCP connection which leads to Head-of-Line (HoL) blocking when packet losses occur. SCTP [RFC4960] supports multiple truly-concurrent streams and QUIC adopted a similar approach to prevent HoL blocking.¶
Modern transport services also changed the utilization of the underlying network. With TCP, when a host creates a connection, it is bound to the IP addresses used by the client and the server during the handshake. When the client moves and receives a different IP address, it has to reestablish all TCP connections bound to the previous address. When the client and the server are dual-stack, they cannot easily switch from one address family to another. Happy Eyeballs [RFC8305] provides a partial answer to this problem for web applications with heuristics that clients can use to probe TCP connections with different address families. With Multipath TCP, the client and the server can learn other addresses of the remote host and combine several TCP connections within a single Multipath TCP connection that is exposed to the application. This supports various use cases [RFC8041]. QUIC [RFC9000] enables applications to migrate from one network path to another, but not to simultaneously use different paths.¶
In order for TCPLS to be widely compatible with middleboxes that inspect TCP segments and TLS records, TCPLS does not modify the TCP connection establishment and only adds a TLS extension to the TLS handshake. Figure 5 illustrates the opening of a TCPLS session which starts with the TCP 3-way handshake, followed by the TLS handshake. In the Extensions of the ClientHello and in the server EncryptedExtensions, the tcpls TLS Extension is introduced to announce the support of TCPLS.¶
Client Server | SYN | |------------------------------------------>| | SYN+ACK | |<------------------------------------------| | ACK, TLS ClientHello + tcpls | |------------------------------------------>| | TLS ServerHello, TLS EncryptedExtensions | | + tcpls, ... | |<------------------------------------------| | TLS Finished | |------------------------------------------>| | |
TCP/TLS offers a single encrypted bytestream service to the application. To achieve this, TLS records are used to encrypt and secure chunks of the application bytestream and are then sent through the TCP bytestream. TCPLS leverages TLS records differently. TCPLS defines its own framing that allows encoding application data and control information. A TCPLS frame is the basic unit of information for TCPLS. A TCPLS frame always fits in a single record. One or more TCPLS frames can be encoded in a TLS record. This TLS record is then reliably transported by a TCP connection. Figure 6 illustrates the relationship between TCPLS frames and TLS records.¶
TCPLS Data TCP Control TCPLS Data abcdef 0010010 mnopq... <---------> <-----------> <------------> / / / / | / | / | / | / | / | / +----------------+ +-----------------+ | TLS record n | | TLS record n+1 | .... +----------------+ +-----------------+
TCPLS extends the service provided by TCP with streams. Streams are independent bidirectional bytestreams that can be used by applications to concurrently convey several objects over a TCPLS session. Streams can be opened by the client and by the server.¶
Streams are identified by a 32-bit unsigned integer. The parity of this number indicates the initiator of the stream. The client opens even-numbered streams while the server opens odd-numbered streams. Streams are opened in sequence, e.g. a client that has opened stream 0 will use stream 2 as the next one.¶
Data is exchanged using Stream frames whose format is described in Section 5.2.3. Each Stream frame carries a chunk of data of a given stream. Applications can mark the end of a stream to close it.¶
TCPLS enables the receiver to decrypt and process TLS records in zero copy similarly to TLS 1.3 under circumstances discussed in Section 4.5.¶
Similarly to HTTP/2 [RFC7540], conveying several streams on a single TCP connection introduces Head-of-Line (HoL) blocking between the streams. To alleviate this, TCPLS provides means to the application to choose the degree of HoL blocking resilience it needs for its application objects by spreading streams among different underlying TCP connections.¶
TCPLS is not restricted to using a single TCP connection to exchange frames. A TCPLS session starts with the TCP connection that was used to transport the TLS handshake. After this handshake, other TCP connections can be added to a TCPLS session, either to spread the load or for failover. TCPLS manages the utilization of the underlying TCP connections within a TCPLS session.¶
Multipath TCP enables both the client and the server to establish additional TCP connections. However, experience has shown that additional subflows are only established by the clients. TCPLS focuses on this deployment and only allows clients to create additional TCP connections.¶
Using Multipath TCP, a client can try establishing a new TCP connection at any time. If a server wishes to restrict the number of TCP connections that correspond to one Multipath TCP connection, it has to respond with RST to the in excess connection attempts.¶
TCPLS takes another approach. To control the number of connections that a client can establish, a TCPLS server supplies unique tokens. A client includes one of the server supplied tokens when it attaches a new TCP connection to a TCPLS session. Each token can only be used once, hence limiting the amount of additional TCP connections.¶
TCPLS endpoints can advertise their local addresses, allowing new TCP connections for a given TCPLS session to be established between new pairs of addresses. When an endpoint is no more willing new TCP connections to use one of its advertised addresses, it can remove this address from the TCPLS session.¶
The TCPLS server provides tokens to the client in order to join new TCP connections to the TCPLS session. Figure 7 illustrates a client and server first establishing a new TCPLS session as described in Section 4. Then the server sends a token over this connection using the New Token frame. Each token has a sequence number (e.g. 1) and a value (e.g. "abc"). The client uses this token to open a new TCP connection and initiates the TCPLS handshake. It adds the token inside the TCPLS Join TLS extension in the ClientHello.¶
<-1.TCPLS Handshake->
.---------------------------------.
| <-2.New Token(1,abc) |
v v
+--------+ +--------+
| Client | | Server |
+--------+ +--------+
^ ^
| 3.TCPLS Handshake + Join(abc)-> | Legend:
.---------------------------------. --- TCP connection
When receiving a TCPLS Join Extension, the server validates the token and associates the TCP connection to the TCPLS session.¶
Each TCP connection that is part of a TCPLS session is identified by a 32-bit unsigned integer called its Connection ID. The first TCP connection of a session corresponds to Connection ID 0. When joining a new connection, the sequence number of the token, i.e. 1 in our example, becomes the Connection ID of the connection. The Connection ID enables the Client and the Server to identify a specific TCP connection within a given TCPLS session.¶
TCPLS supports two types of failover. In make-before-break, the client creates a TCP connection using the procedure described in Section 4.2.1 but only uses it once the initial connection fails.¶
In break-before-make, the client creates the initial TCP connection and uses it for the TCPLS handshake and the data. The server advertises one or more tokens over this connection. Upon failure of the initial TCP connection, the client initiates a second TCP connection using the server-provided token.¶
In both cases, some records sent by the client or the server might be in transit when the failure occurs. Some of these records could have been partially received but not yet delivered to the TCPLS layer when the underlying TCP connection fails. Other records could have already been received, decrypted and data of their frames could have been delivered to the application. To prevent data losses and duplication, TCPLS includes its own acknowledgments.¶
A TCPLS receiver acknowledges the received records using the ACK frame. Records are acknowledged after the record protection has been successfully removed. This enables the sender to know which records have been received. TCPLS enables the endpoint to send acknowledgments for a TCP connection over any connections, e.g. not only the receiving connection.¶
To migrate from a given TCP connection, an endpoint stops transmitting over this TCP connection and sends the following frames on other TCP connections. It leverages the acknowledgments to retransmit the frames of TLS records that have not been yet acknowledged.¶
When an endpoint abortfully closes a TCP connection, its peer leverages the acknowlegments to retransmit the TLS records that were not acknowlegded.¶
TCPLS also supports the utilization of different TCP connections, over different paths or interfaces, to improve throughput or spread stream frames over different TCP connections. When the endpoints have opened several TCP connections, they can send frames over the connections. TCPLS can send all the stream frames belonging to a given stream over one or more underlying TCP connections. The latter enables bandwidth aggregation by using TCP connections established over different network paths.¶
When adding new TCP connections to a TCPLS session, an endpoint does not complete the TLS handshake. TCPLS provides a nonce construction for TLS record protection that is used for all connections of a session. This reduces the cryptographic cost of adding connections. The endpoints SHOULD send TLS messages to form an apparent complete TLS handshake to middleboxes.¶
In order to use the TLS session over multiple connections, TCPLS adds a record sequence number space per connection that is maintained independently at both sides. Each record sent over a TCPLS session is identified by the Connection ID of its connection and its record sequence number. Each record nonce is constructed as defined in Figure 8.¶
N N-32 64 0
+---------------------------------------------------------------+
| client/server_write_iv |
+---------------------------------------------------------------+
XOR XOR
+-------------------+ +--------------------+
| Connection ID | | Conn. record sequ. |
+-------------------+ +--------------------+
This construction guarantees that every TLS record sent over the TLS session is protected with a unique nonce. As in TLS 1.3, the per-connection record sequence is implicit.¶
Endpoints notify their peers that they do not intend to send more data over a given TCPLS session by sending a TLS Alert "close_notify". The alert can be sent over one or more TCP connections of the session. The alert MUST be sent before closing the last TCP connection of the TCPLS session. The endpoint MAY close its side of the TCP connections after sending the alert.¶
When all TCP connections of a session are closed and the TLS Alert "close_notify" was exchanged in both directions, the TCPLS session is considered as closed.¶
We leave defining an abortful and idle session closure mechanisms for future versions of this document.¶
TCPLS enables the receiver to process TLS records in a zero-copy manner under several conditions. When they are met, the application data carried in TCPLS frames can be decrypted at the right place in the application buffers.¶
First, zero-copy can be achieved when Stream frames of a given stream arrive in order. When using several TCP connections, out-of-order Stream frames cannot be processed in zero copy. A Multipath scheduling algorithm may target the minimization of out-of-order packets.¶
Second, the composition of TCPLS frames in a TLS record is impactful. The sender SHOULD encode a single Stream Data frame as the first frame of the record, followed by control-related frames if needed. When the sender encodes several Stream frames, the frame at the start of the record SHOULD be the largest, in order to maximise the use of zero copy. When several Stream frames are included in a record, they SHOULD belong to different streams.¶
This document specifies two TLS extensions used by TCPLS. The first, "tcpls", is used to announce the support of TCPLS. The second, "tcpls_join", is used to join a TCP connection to a TCPLS session. Their types are defined as follows.¶
enum {
tcpls(TBD1),
tcpls_join(TBD2),
(65535)
} ExtensionType;
¶
The table below indicates the TLS messages where these extensions can appear. "CH" indicates ClientHello while "EE" indicates EncryptedExtensions.¶
| Extension | Allowed TLS messages |
|---|---|
| tcpls | CH, EE |
| tcpls_join | CH |
The "tcpls" extension is used by the client and the server to announce their support of TCPLS. The extension contains no value. When it is present in both the ClientHello and the EncryptedExtensions, the endpoints MUST use TCPLS after completing the TLS handshake.¶
struct {
opaque token<32>;
} Join;
¶
The "tcpls_join" extension is used by the client to join the TCP connection on which it is sent to a TCPLS session. The extension contains a Token provided by the server. The client MUST NOT send more than one "tcpls_join" extension in its ClientHello. When receiving a ClientHello with this extension, the server checks that the token is valid and joins the TCP connection to the corresponding TCPLS session. When the token is not valid, the server MUST abort the handshake with an illegal_parameter alert.¶
By controlling the amount of tokens given to the client, the server can control the number of active TCP connections of a TCPLS session. The server SHOULD replenish the tokens when TCP connections are removed from the TCPLS session.¶
TCPLS uses TLS Application Data records to exchange TCPLS frames. After decryption, the record payload consists of a sequence of TCPLS frames. Figure 9 illustrates the manner in which TCPLS frames are parsed from a decrypted TLS record. The receiver processes the frames starting from the last one to the first one. The fields of each frames are also parsed from the end towards the beginning of the TLS Application Data content. The parsing of a frame starts with the last byte indicating the frame type and then with type-specific fields preceeding it, forming a Type-Value unit. Such ordering enables a zero-copy processing of the type-specific fields as explained in Section 4.5.¶
First frame start
|
0 First frame end | n
+---------------------|------------|-+
| v v |
| *------------* *------------* |
| .... | Value |Type| | Value |Type| |
| *------------* *------------* |
+------------------------------------+
Decrypted TLS record
Table 2 lists the frames specified in this document.¶
| Type value | Frame name | Rules | Definition |
|---|---|---|---|
| 0x00 | Padding | N | Section 5.2.1 |
| 0x01 | Ping | Section 5.2.2 | |
| 0x02-0x03 | Stream | Section 5.2.3 | |
| 0x04 | ACK | N | Section 5.2.4 |
| 0x05 | New Token | S | Section 5.2.5 |
| 0x06 | Connection Reset | Section 5.2.6 | |
| 0x07 | New Address | Section 5.2.7 | |
| 0x08 | Remove Address | Section 5.2.8 | |
| 0x09 | Stream Change | Cs | Section 5.2.9 |
The "Rules" column in Table 2 indicates special requirements regarding certain frames.¶
Non-ack-eliciting. Receiving this frame does not elicit the sending of a TCPLS acknowledgment.¶
Server only. This frame MUST NOT be sent by the client.¶
Connection-specific semantics. This frame is not idempotent and has specific semantics based on the TCP connection over which it is exchanged.¶
This frame has no semantic value. It can be used to mitigate traffic analysis on the TLS records of a TCPLS session. The Padding frame has no content.¶
Padding frame {
Type (8) = 0x00,
}
This frame is used to elicit an acknowledgment from its peer. It has no content. When an endpoint receives a Ping frame, it acknowledges the TLS record that contains this frame. This frame can be used by an endpoint to check that its peer can receive TLS records over a particular TCP connection.¶
Ping frame {
Type (8) = 0x01,
}
This frame is used to carry chunks of data of a given stream.¶
Stream frame {
Stream Data (...),
Length (16),
Offset (64),
Stream ID (32),
FIN (1),
Type (7) = 0x01,
}
The last bit of the frame type bit indicates that this Stream frame ends the stream when its value is 1. The last byte of the stream is at the sum of the Offset and Length fields of this frame.¶
A 32-bit unsigned integer indicating the ID of the stream this frame relates to.¶
A 64-bit unsigned integer indicating the offset in bytes of the carried data in the stream.¶
A 16-bit unsigned integer indicating the length of the Stream Data field.¶
This frame is sent by the receiver to acknowledge the receipt of TLS records on a particular TCP connection of the TCPLS session. Although the reliability of the data exchange on a connection is handled by TCP, there are situations such as the failure of a TCP connection where a sender does not know whether the TLS frames that it sent have been correctly received by the peer. The ACK frame allows a TCPLS receiver to indicate the highest TLS record sequence number received on a specific connection. The ACK frame can be sent over any TCP connection of a TCPLS session.¶
ACK frame {
Highest Record Sequence Received (64),
Connection ID (32),
Type (8) = 0x04,
}
This frame is used by the server to provide tokens to the client. Each token can be used to join a new TCP connection to the TCPLS session, as described in Section 4.2.1. Clients MUST NOT send New Token frames.¶
New Token frame {
Token (256),
Sequence (8),
Type (8) = 0x05,
}
This frame is used by the receiver to inform the sender that a TCP connection has been reset.¶
Connection Reset frame {
Connection ID (32)
Type (8) = 0x06,
}
A 32-bit unsigned integer indicating the ID of the connection that failed.¶
This frame is used by an endpoint to add a new local address to the TCPLS session. This address can then be used to establish new TCP connections. The server advertises addresses that the client can use as destination when adding TCP connections. The client advertises address that it can use as source when adding TCP connections.¶
New Address frame {
Port (16),
Address (..),
Address Version (8),
Address ID (8),
Type (8) = 0x07,
}
A 8-bit identifier for this address. For a given Address ID, an endpoint receiving a frame with a content that differs from previously received frames MUST ignore the frame. An endpoint receiving a frame for an Address ID that was previously removed MUST ignore the frame.¶
A 8-bit value identifying the Internet address version of this address. The number 4 indicates IPv4 while 6 indicates IPv6.¶
The address value. Its size depends on its version. IPv4 addresses are 32-bit long while IPv6 addresses are 128-bit long.¶
A 16-bit value indicating the TCP port used with this address.¶
This frame is used by an endpoint to announce that it is not willing to use a given address to establish new TCP connections. After receiving this frame, a client MUST NOT establish new TCP connections to the given address. After receiving this frame, an endpoint MUST close all TCP connections using the given address.¶
Remove Address frame {
Address ID (8),
Type (8) = 0x08,
}
A 8-bit identifier for the address to remove. An endpoint receiving a frame for an address that was nonexistent or already removed MUST ignore the frame.¶
This frame is used by a sender to announce the Stream ID and Offset of the next record over a given TCP connection. It can be used to make explicit a change in stream scheduling over a connection to the receiver, enabling a zero-copy receive path as explained in Section 4.5. The hint contained in this frame relates to the connection over which it was exchanged.¶
Stream Change frame {
Type (8) = 0x09,
Next Offset (64),
Next Record Stream ID (32),
}
When issuing tokens to the client as presented in Section 4.2.1, the server SHOULD ensure that their values appear as random to observers and cannot be correlated together for a given TCPLS session.¶
The security considerations for TLS apply to TCPLS. The next versions of this document will elaborate on other security considerations following the guidelines of [RFC3552].¶
IANA is requested to create a new "TCPLS" heading for the new registry described in Section 5.2. New registrations in TCPLS registries follow the "Specification Required" policy of [RFC8126].¶
IANA is requested to add the following entries to the existing "TLS ExtensionType Values" registry.¶
| Value | Extension Name | TLS 1.3 | Recommended | Reference |
|---|---|---|---|---|
| TBD1 | tcpls | CH, EE | N | This document |
| TBD2 | tcpls_join | CH | N | This document |
Note that "Recommended" is set to N as these extensions are intended for uses as described in this document.¶
IANA is requested to create a new registry "TCPLS Frames Types" under the "TCPLS" heading.¶
The registry governs an 8-bit space. Entries in this registry must include a "Frame name" field containing a short mnemonic for the frame type. The initial content of the registry is present in Table 2, without the "Rules" column.¶
In this section, we discuss alternatives to mechanisms defined in this document.¶
When adding a new TCP connection to a TCPLS session, endpoints derive a Initial Vector based on the technique presented in Section 4.3. An alternative is to use New Session Tickets to derive separate cryptographic protection materials based on a pre-shared key sent by the server. In this mode of operation, the server provides New Session Tickets to control the amount of additional TCP connections that can be opened by the client. The server could encode the Connection ID in the ticket value.¶
Using this mechanism differs from our proposal in several ways. While it enables to use an existing TLS mechanism for this purpose, it has a number of differences. First, it requires the server to compute pre-shared keys which could be more costly than computing the TCPLS tokens defined in Section 4.2.1. Second, when used to establish a TLS session, additional TLS messages must be computed and exchanged to complete the handshake, which the current mechanism does not require. Third, TLS New Session Tickets have a lifetime that is separated from the session they are exchanged over. This is unneeded in the context of TCPLS and may require additional protocol specification and guidance to implementers.¶
This work has been partially supported by the ``Programme de recherche d'interet general WALINNOV - MQUIC project (convention number 1810018)'' and European Union through the NGI Pointer programme for the TCPLS project (Horizon 2020 Framework Programme, Grant agreement number 871528). The authors thank Quentin De Coninck and Louis Navarre for their comments on the first version of this draft.¶