QUIC Working Group                                         Q. De Coninck
Internet-Draft                                            O. Bonaventure
Intended status: Standards Track                               UCLouvain
Expires: 21 February 2021                                 20 August 2020

                Multipath Extensions for QUIC (MP-QUIC)


   This document specifies extensions to the QUIC protocol to enable the
   simultaneous usage of multiple paths for a single connection.

   These extensions are compliant with the single-path QUIC design and
   preserve QUIC privacy features.

Status of This Memo

   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 21 February 2021.

Copyright Notice

   Copyright (c) 2020 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 Simplified BSD License text
   as described in Section 4.e of the Trust Legal Provisions and are
   provided without warranty as described in the Simplified BSD License.

De Coninck & BonaventureExpires 21 February 2021                [Page 1]

Internet-Draft                   MP-QUIC                     August 2020

Table of Contents

   1.  Introduction  . . . . . . . . . . . . . . . . . . . . . . . .   3
   2.  Conventions and Definitions . . . . . . . . . . . . . . . . .   4
     2.1.  Notation Conventions  . . . . . . . . . . . . . . . . . .   5
   3.  Overview  . . . . . . . . . . . . . . . . . . . . . . . . . .   5
     3.1.  Moving from Bidirectional Paths to Uniflows . . . . . . .   5
     3.2.  Beyond Connection Migration . . . . . . . . . . . . . . .   7
     3.3.  Establishment of a Multipath QUIC Connection  . . . . . .   8
     3.4.  Architecture of Multipath QUIC  . . . . . . . . . . . . .   8
     3.5.  Uniflow Establishment . . . . . . . . . . . . . . . . . .  11
     3.6.  Exchanging Data over Multiple Uniflows  . . . . . . . . .  12
     3.7.  Exchanging Addresses  . . . . . . . . . . . . . . . . . .  14
     3.8.  Coping with Address Removals  . . . . . . . . . . . . . .  15
     3.9.  Uniflow Migration . . . . . . . . . . . . . . . . . . . .  15
     3.10. Congestion Control  . . . . . . . . . . . . . . . . . . .  15
   4.  Mapping Uniflow IDs to Connection IDs . . . . . . . . . . . .  15
   5.  Using Multiple Uniflows . . . . . . . . . . . . . . . . . . .  16
     5.1.  Multipath Negotiation . . . . . . . . . . . . . . . . . .  16
       5.1.1.  Transport Parameter Definition  . . . . . . . . . . .  17
     5.2.  Coping with Additional Remote Addresses . . . . . . . . .  17
     5.3.  Receiving Uniflow State . . . . . . . . . . . . . . . . .  17
     5.4.  Sending Uniflow State . . . . . . . . . . . . . . . . . .  18
     5.5.  Losing Addresses  . . . . . . . . . . . . . . . . . . . .  20
   6.  New Frames  . . . . . . . . . . . . . . . . . . . . . . . . .  20
     6.1.  MP_NEW_CONNECTION_ID Frame  . . . . . . . . . . . . . . .  20
     6.2.  MP_RETIRE_CONNECTION_ID Frame . . . . . . . . . . . . . .  21
     6.3.  MP_ACK Frame  . . . . . . . . . . . . . . . . . . . . . .  22
     6.4.  ADD_ADDRESS Frame . . . . . . . . . . . . . . . . . . . .  23
     6.5.  REMOVE_ADDRESS Frame  . . . . . . . . . . . . . . . . . .  24
     6.6.  UNIFLOWS Frame  . . . . . . . . . . . . . . . . . . . . .  25
   7.  Extension of the Meaning of Existing QUIC Frames  . . . . . .  26
   8.  Security Considerations . . . . . . . . . . . . . . . . . . .  27
     8.1.  Nonce Computation . . . . . . . . . . . . . . . . . . . .  27
     8.2.  Validation of Exchanged Addresses . . . . . . . . . . . .  27
   9.  IANA Considerations . . . . . . . . . . . . . . . . . . . . .  27
     9.1.  QUIC Transport Parameter Registry . . . . . . . . . . . .  27
   10. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . .  28
   11. References  . . . . . . . . . . . . . . . . . . . . . . . . .  28
     11.1.  Normative References . . . . . . . . . . . . . . . . . .  28
     11.2.  Informative References . . . . . . . . . . . . . . . . .  29
   Appendix A.  Comparison with Multipath TCP  . . . . . . . . . . .  30
     A.1.  Multipath TCP Bidirectional Paths vs. QUIC Uniflows . . .  31
     A.2.  Uniflow Establishment . . . . . . . . . . . . . . . . . .  31
     A.3.  Exchanging Data over Multiple Uniflows  . . . . . . . . .  31
     A.4.  Congestion Control  . . . . . . . . . . . . . . . . . . .  31
     A.5.  ACK Frame . . . . . . . . . . . . . . . . . . . . . . . .  31
   Appendix B.  To move in companion drafts  . . . . . . . . . . . .  32

De Coninck & BonaventureExpires 21 February 2021                [Page 2]

Internet-Draft                   MP-QUIC                     August 2020

     B.1.  Uniflow Establishment . . . . . . . . . . . . . . . . . .  32
     B.2.  Exchanging Addresses  . . . . . . . . . . . . . . . . . .  32
     B.3.  Uniflow Migration . . . . . . . . . . . . . . . . . . . .  32
     B.4.  Scheduling Strategies . . . . . . . . . . . . . . . . . .  34
   Appendix C.  Change Log . . . . . . . . . . . . . . . . . . . . .  34
     C.1.  Since draft-deconinck-quic-multipath-04 . . . . . . . . .  34
     C.2.  Since draft-deconinck-quic-multipath-03 . . . . . . . . .  35
     C.3.  Since draft-deconinck-quic-multipath-02 . . . . . . . . .  35
     C.4.  Since draft-deconinck-quic-multipath-01 . . . . . . . . .  35
     C.5.  Since draft-deconinck-quic-multipath-00 . . . . . . . . .  35
     C.6.  Since draft-deconinck-multipath-quic-00 . . . . . . . . .  36
   Authors' Addresses  . . . . . . . . . . . . . . . . . . . . . . .  36

1.  Introduction

   Today's endhosts are equipped with several network interfaces.  Users
   expect to be able to seamlessly switch from one interface to another
   one or use them simultaneously to aggregate bandwidth whenever
   needed.  During the last years, several multipath extensions to
   transport protocols have been proposed (e.g., [RFC6824], [MPRTP], or
   [SCTPCMT]).  Multipath TCP [RFC6824] is the most mature one.  It is
   already deployed on popular smartphones, but also for other use cases
   [RFC8041] [IETFJ].

   With regular TCP and UDP, all the packets that belong to a given flow
   share the same 4-tuple {source IP address, source port number,
   destination IP address, destination port number} that acts as an
   identifier for this flow.  This prevents these flows from using
   multiple paths.

   QUIC [I-D.ietf-quic-transport] does not use the 4-tuple as an
   implicit connection identifier.  Instead, a QUIC flow is identified
   by a Connection ID.  This enables QUIC to cope with events affecting
   the 4-tuple, such as NAT rebinding or IP address changes.  The QUIC
   connection migration feature, specified in Section 9 of
   [I-D.ietf-quic-transport], enables migrating a flow from one 4-tuple
   to another one, sustaining in particular a connection over different
   network paths.

De Coninck & BonaventureExpires 21 February 2021                [Page 3]

Internet-Draft                   MP-QUIC                     August 2020

   Still, there is a void to specify simultaneous usage of QUIC over
   available network paths for a single connection.  Use cases such as
   bandwidth aggregation or seamless network handovers would be
   applicable to QUIC, as they are now with Multipath TCP
   [RFC8041][IETFJ].  Experience with Multipath TCP on smartphones shows
   that the ability to simultaneously use WLAN and cellular during
   handovers improves the user perceived quality of experience.  A
   performance evaluation of an early solution for such use cases and a
   comparison between Multipath QUIC and Multipath TCP may be found in

   In this document, we leverage many of the lessons learned from the
   design of Multipath TCP and the comments received on the first
   versions of this document to propose extensions to the current QUIC
   design to enable it to simultaneously use several network paths.
   This document focuses mainly on network paths that are
   distinguishable by the endpoints.

   This document is organized as follows.  It first provides in
   Section 3 an overview of the operation of Multipath QUIC.  It then
   states the required changes in the current QUIC design
   [I-D.ietf-quic-transport] and specifies in Section 5 the usage of
   multiple paths.  Finally, it discusses some security considerations.

2.  Conventions and Definitions

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "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.

   We assume that the reader is familiar with the terminology used in
   [I-D.ietf-quic-transport].  In addition, we define the following

   *  Uniflow: A unidirectional flow of packets between a QUIC host and
      its peer.  This flow is identified by an internal identifier,
      called Uniflow ID.  Packets sent over a uniflow use a Destination
      Connection ID that may change during the lifetime of the
      connection.  When being in use, an uniflow is temporarily bound to
      a 4-tuple (Source IP Address, Source Port Number, Destination IP
      Address, Destination Port Number).

De Coninck & BonaventureExpires 21 February 2021                [Page 4]

Internet-Draft                   MP-QUIC                     August 2020

   *  Initial Uniflows: The two uniflows used by peers for the
      establishment of a QUIC connection.  One is the uniflow from the
      client to the server and the other is the uniflow from the server
      to the client.  The cryptographic handshake is done on these
      uniflows.  These are identified by Uniflow ID 0.

2.1.  Notation Conventions

   Packet and frame diagrams use the format described in Section 12 of

3.  Overview

   The design of QUIC [I-D.ietf-quic-transport] provides reliable
   transport with multiplexing, confidentiality, integrity, and
   authenticity of data flows.  A wide range of devices on today's
   Internet are multihomed.  Examples include smartphones equipped with
   both WLAN and cellular interfaces, but also regular dual-stack hosts
   that use both IPv4 and IPv6.

   The current design of QUIC does not enable multihomed devices to
   efficiently use different paths simultaneously.  This document
   proposes multipath extensions with the following design goals:

   *  Each host keeps control on the number of uniflows being used over
      the connection.

   *  The simultaneous usage of multiple uniflows should not introduce
      new privacy concerns.

   *  A host must ensure that all the paths it uses actually reach its
      peer to avoid packet flooding towards a victim (see
      Section 21.12.3 of [I-D.ietf-quic-transport])

   *  The multipath extensions should handle the asymmetrical nature of
      paths between two peers.

   We first explain why a multipath extension would be beneficial to
   QUIC and then describe it at a high level.

3.1.  Moving from Bidirectional Paths to Uniflows

   To understand the overall architecture of the multipath extensions,
   let us first refine the notion of "path".  A path may be denoted by a
   4-tuple (Source IP Address, Source Port Number, Destination IP
   Address, Destination Port Number).  In QUIC, this is namely a UDP
   path from the local host to the remote one.  Considering a smartphone
   interacting with a single-homed server, the smartphone might want to

De Coninck & BonaventureExpires 21 February 2021                [Page 5]

Internet-Draft                   MP-QUIC                     August 2020

   use one path over the WLAN network and another over the cellular one.
   Those paths are not necessarily disjoint.  For example, when
   interacting with a dual-stack server, a smartphone may create two
   paths over WLAN: one over IPv4 and the other one over IPv6.

   A regular QUIC connection is composed of two independent active
   packet flows.  The first flow gathers the packets from the client to
   the server and the other the packets from the server to the client.
   To illustrate this, let us consider the example depicted in Figure 1.
   In this example, the client has two IP addresses: IPc1 and IPc2.  The
   server has one single address: IPs1.

                     Probed flow IPc2 to IPs1
    |                                                         |
    |   From IPc1 to IPs1                                     v
    |   +--------+          Client to Server Flow         +--------+
    |   |        | =====================================> |        |
    +-- | Client |                                        | Server |
        |        | <===================================== |        |
        +--------+          Server to Client Flow         +--------+
                                                 From IPc1* to IPs1*

             Figure 1: Identifying Unidirectional Flows in QUIC

   The client initiates the QUIC connection by sending packets towards
   the server.  The server then replies to the client if it accepts the
   connection.  If the handshake succeeds, the connection is
   established.  Still, this "path" actually consists in two independent
   UDP flows.  Each host has its own view of (i) the 4-tuple used to
   send packets and (ii) the 4-tuple on which it receives packets.
   While the 4-tuple used by the client to send packets may be the same
   as the one seen and used by the server, this is not always the case
   since middleboxes (e.g., NATs) may alter the 4-tuple of packets.

   To further emphasize on this flow asymmetry, QUIC embeds a path
   validation mechanism [I-D.ietf-quic-transport] assessing whether a
   host can reach its peer through a given 4-tuple.  This process is
   unidirectional, i.e., the sender checks that it can reach the
   receiver, but not the reverse.  A host receiving a PATH_CHALLENGE
   frame on a new 4-tuple may in turn initiate a path validation, but
   this is up to the peer.

   A QUIC connection is a collection of unidirectional flows (called,
   uniflows).  A plain QUIC connection is composed of a main uniflow
   from client to server and another main uniflow from server to client.
   These uniflows have their own Connection IDs.  They are host-
   specific, i.e., the uniflow(s) from A to B are different from the

De Coninck & BonaventureExpires 21 February 2021                [Page 6]

Internet-Draft                   MP-QUIC                     August 2020

   ones from B to A.  This potentially enables the use of unidirectional
   links such as non-broadcast satellite links [RFC3077], which cannot
   be used with TCP.

3.2.  Beyond Connection Migration

   Unlike TCP [RFC0793], QUIC is not bound to a particular 4-tuple
   during the lifetime of a connection.  A QUIC connection is identified
   by a (set of) Connection ID(s), placed in the public header of each
   QUIC packet.  This enables hosts to continue the connection even if
   the 4-tuple changes due to, e.g., NAT rebinding.  This ability to
   shift a connection from one 4-tuple to another is called: Connection
   Migration.  One of its use cases is fail-over when the IP address in
   use fails but another one is available.  A smartphone losing the WLAN
   connectivity can then continue the connection over its cellular
   interface, for instance.

   A QUIC connection can thus start on a given set of uniflows, denoted
   as the initial uniflows, and end on another ones.  However, the
   current QUIC design [I-D.ietf-quic-transport] assumes that only one
   pair of uniflows is in use for a given connection.  The current
   transport specification [I-D.ietf-quic-transport] does not provide
   means to distinguish path migration from simultaneous usage of
   available uniflows for a given connection.

   This document fills that void.  It first proposes mechanisms to
   communicate endhost addresses to the peer.  It then leverages the
   Address Validation procedure with PATH_CHALLENGE and PATH_RESPONSE
   frames described in Section 8 of [I-D.ietf-quic-transport] to verify
   whether the additional addresses advertised by the host are
   reachable.  In this case, those addresses can be used to initiate new
   uniflows to spread packets over several network paths following a
   packet scheduling policy that is out of scope of this document.

   TODO: Add a companion document discussing the packet scheduling and
   path management considerations.

   The example of Figure 2 illustrates a data exchange between a dual-
   homed client sending a request spanning two packets and a single-
   homed server.  Uniflow IDs are independently chosen by each host.  In
   the presented example, the client sends packets over WLAN on Uniflow
   0 and over LTE on Uniflow 1, while the packets sent by the server
   over WLAN are on Uniflow 2 and those over LTE are on Uniflow 1.

De Coninck & BonaventureExpires 21 February 2021                [Page 7]

Internet-Draft                   MP-QUIC                     August 2020

   Server                        Phone                        Server
   via WLAN                                                  via LTE
   -------                      -------                        -----
     | Pkt(DCID=A,PN=5,frames=[    |                             |
     |  STREAM("Request (1/2)")])  | Pkt(DCID=B,PN=1,frames=[    |
     |<----------------------------|  STREAM("Request (2/2)")])  |
     | Pkt(DCID=E,PN=1,frames=[    |--------                     |
     |  ACK(LargestAcked=5)])      |       |----------           |
     |---------------------------->|                 |--------   |
     | Pkt(DCID=E,PN=2,frames=[    |                         |-->|
     |  STREAM("Response 1")])     | Pkt(DCID=D,PN=1,frames=[    |
     |---------------------------->|  MPACK(UID=1,LargestAck=1), |
     |                             |  STREAM("Response 2")])  ---|
     | Pkt(DCID=A,PN=6,frames=[    |                 ---------|  |
     |  MPACK(UID=2,LargestAck=2), |       ----------|           |
     |  MPACK(UID=1,LargestAck=1)])|<------|                     |
     |<----------------------------|                             |
     | Pkt(DCID=E,PN=3,frames=[    | Pkt(DCID=D,PN=2,frames=[    |
     |  STREAM("Response 3")])     |  STREAM("Response 4")])     |
     |---------------------------->|                         ----|
     |                             |                   ------|   |
     |            ...              |    ...  <---------|         |

                  Figure 2: Data flow with Multipath QUIC

   The remaining of this section presents a high-level overview of the
   multipath operations in QUIC.

3.3.  Establishment of a Multipath QUIC Connection

   A Multipath QUIC connection starts as a regular QUIC connection
   [I-D.ietf-quic-transport] [I-D.ietf-quic-tls].  The multipath
   extensions defined in this document are negotiated using the
   "max_sending_uniflow_id" transport parameter.  Any value for this
   transport parameter advertises the support of the multipath

   Notice that a host advertising a value of 0 for the
   "max_sending_uniflow_id" transport parameter indicates that it does
   not want additional uniflows to send packets, but it still supports
   the multipath extensions.  Such situation might be useful when the
   host does not require multiple uniflows for packet sending but still
   wants to let the peer use multiple uniflows to reach it.

3.4.  Architecture of Multipath QUIC

   To illustrate the architecture of a Multipath QUIC connection,
   consider Figure 3.

De Coninck & BonaventureExpires 21 February 2021                [Page 8]

Internet-Draft                   MP-QUIC                     August 2020

   +--------+          CID A - Uniflow ID 1          +--------+
   |        | =====================================> |        |
   |        |          CID B - Uniflow ID 0          |        |
   |        | =====================================> |        |
   |        |                                        |        |
   | Client |          CID C - Uniflow ID 0          | Server |
   |        | <===================================== |        |
   |        |          CID D - Uniflow ID 1          |        |
   |        | <===================================== |        |
   |        |          CID E - Uniflow ID 2          |        |
   |        | <===================================== |        |
   +--------+                                        +--------+

       Figure 3: An Example of Uniflow Distribution over a Multipath
                              QUIC Connection

   Once established, a Multipath QUIC connection consists in one or more
   uniflows from the client to the server and one or more uniflows from
   the server to the client.  The number of uniflows in one direction
   can be different from the one in the other direction.  The example in
   Figure 3 shows two uniflows from the client to the server and three
   uniflows from the server to the client.  From the end-hosts'
   viewpoint, they observe two kinds of uniflows:

   *  Sending uniflows: uniflows over which the host can send packets

   *  Receiving uniflows: uniflows over which the host can receive

   Reconsidering the example in Figure 3, the client has two sending
   uniflows and three receiving uniflows.  The server has three sending
   uniflows and two receiving uniflows.  There is thus a one-to-one
   mapping between the sending uniflows of a host and the receiving
   uniflows of its peer.  A uniflow is seen as a sending uniflow from
   the sender's perspective and as a receiving uniflow from the
   receiver's viewpoint.

   Each uniflow is associated with a specific four-tuple and identified
   by a Uniflow ID, as shown in Figure 4.

De Coninck & BonaventureExpires 21 February 2021                [Page 9]

Internet-Draft                   MP-QUIC                     August 2020

   Client state
       |                      Connection                     |
       |   +-----------+ +-----------+ ... +-------------+   |
       |   |  Sending  | |  Sending  | ... |   Sending   |   |
       |   | Uniflow 0 | | Uniflow 1 |     | Uniflow N-1 |   |
       |   | UCID set A| | UCID set B| ... |  UCID set C |   |
       |   |  Tuple A  | |  Tuple B  | ... |   Tuple C   |   |
       |   |   PNS A   | |   PNS B   |     |    PNS C    |   |
       |   +-----------+ +-----------+ ... +-------------+   |
       |   +-----------+ +-----------+ ... +-------------+   |
       |   | Receiving | | Receiving | ... |  Receiving  |   |
       |   | Uniflow 0 | | Uniflow 1 |     | Uniflow M-1 |   |
       |   | UCID set X| | UCID set Y| ... |  UCID set Z |   |
       |   |  Tuple X  | |  Tuple Y  | ... |   Tuple Z   |   |
       |   |   PNS X   | |   PNS Y   |     |    PNS Z    |   |
       |   +-----------+ +-----------+ ... +-------------+   |

   Server state
       |                      Connection                     |
       |   +-----------+ +-----------+ ... +-------------+   |
       |   |  Sending  | |  Sending  | ... |   Sending   |   |
       |   | Uniflow 0 | | Uniflow 1 |     | Uniflow M-1 |   |
       |   | UCID set X| | UCID set Y| ... |  UCID set Z |   |
       |   |  Tuple X* | |  Tuple Y* | ... |   Tuple Z*  |   |
       |   |   PNS X   | |   PNS Y   |     |    PNS Z    |   |
       |   +-----------+ +-----------+ ... +-------------+   |
       |   +-----------+ +-----------+ ... +-------------+   |
       |   | Receiving | | Receiving | ... |  Receiving  |   |
       |   | Uniflow 0 | | Uniflow 1 |     | Uniflow N-1 |   |
       |   | UCID set A| | UCID set B| ... |  UCID set C |   |
       |   |  Tuple A* | |  Tuple B* | ... |   Tuple C*  |   |
       |   |   PNS A   | |   PNS B   |     |    PNS C    |   |
       |   +-----------+ +-----------+ ... +-------------+   |

      Figure 4: Architectural view of Multipath QUIC for a host having
                N sending uniflows and M receiving uniflows

   A Multipath QUIC connection starts using two Initial Uniflows,
   identified by Uniflow ID 0 on each peer.  The packets can then be
   spread over several uniflows.  Each uniflow has its (set of) Uniflow
   Connection ID(s) (UCID) packets that are used to explicitly mark
   where they belong to.  Depending on the direction of the uniflow, the
   host keeps either the Uniflow Source Connection ID (USCID, for the
   receiving uniflows) or the Uniflow Destination Connection ID (USCID,

De Coninck & BonaventureExpires 21 February 2021               [Page 10]

Internet-Draft                   MP-QUIC                     August 2020

   for the sending uniflows).  Notice that the (set of) UDCID(s) of a
   sending uniflow of a host is the same as the (set of) USCID(s) of the
   corresponding receive uniflow of the remote peer.

   Preventing the linkability of different uniflows is an important
   requirement for the multipath extensions
   [I-D.huitema-quic-mpath-req].  We address it by using UCIDs as
   implicit uniflow identifiers.  This makes the linkability harder than
   having explicit signaling as in earlier version of this draft.
   Furthermore, it does not require any public header change and thus
   preserves the QUIC invariants [I-D.ietf-quic-invariants].

   When a uniflow is in use, each endhost associates it with a network
   path.  In practice, this consists in a particular 4-tuple over which
   packets are sent (or received) on a sending (or receiving) uniflow.
   Each endhost has a specific vision of the 4-tuple, which might differ
   between endhosts.  For instance, a client located behind a NAT sends
   data from a private IP address and the server will receive packets
   coming from the NAT's public IP address.  Notice that while uniflows
   may share a common network path, this is not mandatory.

   Each uniflow is an independent flow of packets over a given network
   path.  Uniflows can experience very different network conditions
   (latency, bandwidth, ...).  To handle this, each uniflow has its own
   packet sequence number space.

   In addition to the UCIDs, 4-tuple and packet number space, some
   additional information is maintained for each uniflow.  The Uniflow
   ID identifies the uniflow at the frame level and ensures uniqueness
   of the nonce (see Section 8.1 for details) while limiting the number
   of concurrently used uniflows.

3.5.  Uniflow Establishment

   The "max_sending_uniflow_id" transport parameter exchanged during the
   cryptographic handshake fixes an upper bound on the number of sending
   uniflows a host wants to support.  Then, hosts provide to their peer
   Uniflow Connection IDs to use on uniflows.  Both hosts dynamically
   control how many sending uniflows can currently be in use by the
   peer, i.e., the number of different Uniflow IDs proposed to the peer.
   While the sender determines the upper bound of sending paths it can
   have, it is the receiver that initializes uniflows, as the sender
   needs a UCID communicated by the receiver before using a uniflow.

   Notice that the peers might advertise different values for the
   "max_sending_uniflow_id" transport parameters, setting different
   upper bounds to the sending and receiving uniflows of each host.

De Coninck & BonaventureExpires 21 February 2021               [Page 11]

Internet-Draft                   MP-QUIC                     August 2020

   Hosts initiate the creation of their receiving uniflows by sending
   MP_NEW_CONNECTION_ID frames (see Section 6.1) which are an extended
   version of the NEW_CONNECTION_ID frame.  This frame associates a UCID
   to a uniflow.  Upon reception of the MP_NEW_CONNECTION_ID frame, a
   host can start using the proposed sending uniflow having the
   referenced Uniflow ID by marking sent packets with the provided UCID.
   Therefore, once a host sends a MP_NEW_CONNECTION_ID frame, it
   announces that it is ready to receive packets from that Uniflow ID
   with the proposed UCID.  As frames are encrypted, adding new uniflows
   over a QUIC connection does not leak cleartext identifiers

   A server might provide several Uniflow Connection IDs for the same
   Uniflow ID with multiple MP_NEW_CONNECTION_ID frames.  This can be
   useful to cope with migration cases, as described in Section 3.9.
   Multipath QUIC is thus asymmetrical.

3.6.  Exchanging Data over Multiple Uniflows

   A QUIC packet acts as a container for one or more frames.  Multipath
   QUIC uses the same STREAM frames as QUIC to carry data.  A byte
   offset is associated with the data payload.  One of the key design of
   (Multipath) QUIC is that frames are independent of the packets
   carrying them.  This implies that a frame transmitted over one
   uniflow could be retransmitted later on another uniflow without any
   change.  Furthermore, all current QUIC frames are idempotent and
   could be optimistically duplicated over several uniflows.

   The uniflow on which data is sent is a packet-level information.
   This means that a frame can be sent regardless of the uniflow of the
   packet carrying it.  Other flow control considerations like the
   stream receive window advertised by the MAX_STREAM_DATA frame remain
   unchanged when there are multiple sending uniflows.

   As previously described, Multipath QUIC might face reordering at
   packet-level when using uniflows having different latencies.  The
   presence of different Uniflow Connection IDs ensures that the packets
   sent over a given uniflow will contain monotonically increasing
   packet numbers.  To ensure more flexibility and potentially to reduce
   the ACK block section of the (MP_)ACK frame when aggregating
   bandwidth of uniflows exhibiting different network characteristics,
   each uniflow keeps its own monotonically increasing Packet Number
   space.  This potentially allows sending up to 2 * 2^64 packets on a
   QUIC connection since each uniflow has its own packet number space
   (see Section 8.1 for the detail of this limit).

De Coninck & BonaventureExpires 21 February 2021               [Page 12]

Internet-Draft                   MP-QUIC                     August 2020

   With the introduction of multiple uniflows, there is a need to
   acknowledge packets sent on different uniflows separately.  The
   packets sent on Initial Uniflows (with Uniflow ID 0) are still
   acknowledged with regular ACK frames, such that no modification is
   introduced in a core frame.  For the other uniflows, the multipath
   extensions introduce a MP_ACK frame which prefixes the ACK frame with
   a Uniflow ID field indicating from which receiving uniflow the host
   acknowledges packets.  To better explain this, let us consider the
   situation illustrated in Figure 5.

   Sending uniflow 0 - CID A      |    Receiving uniflow 0 - CID A
   Sending uniflow 1 - CID B      |    Receiving uniflow 1 - CID B
   Receiving uniflow 0 - CID C    |      Sending uniflow 0 - CID C
   Receiving uniflow 1 - CID D    |      Sending uniflow 1 - CID D
   Receiving uniflow 2 - CID E    |      Sending uniflow 2 - CID E

   Client                                                   Server
   ------                                                   ------
      |                                                        |
      |       Pkt(DCID=B,PN=42,frames=[STREAM("Request")])     |
      |---------------------------                             |
      |                          |---------------------------->|
      |                                                        |
      |  Pkt(DCID=E,PN=58,frames=[                             |
      |   STREAM("Response"), MP_ACK(UID=1,LargestAcked=42)])  |
      |                          ------------------------------|
      |<-------------------------|                             |
      |                                                        |

              Figure 5: Acknowledging Packets Sent on Uniflows

   Here five uniflows are in use, two in the client to server direction
   and three in the reverse one.  The client first sends a packet on its
   sending Uniflow 1 (linked to CID B).  The server receives the packet
   on its receiving Uniflow 1.  Therefore, it generates a MP_ACK frame
   for Uniflow ID 1 and transmits it to the client.  The server can
   choose any of its sending uniflows to transmit this frame.  In the
   provided situation, the server sends its packets on Uniflow 2.  The
   client thus receives this packet on its receiving Uniflow 2.

   Similarly, packets sent over a given uniflow might be acknowledged by
   (MP_)ACK frames sent on another uniflow that does not share the same
   network path.  Looking at Figure 2 again, "Response 2" packet on
   server's sending uniflow 1 with DCID D using the LTE network is
   acknowledged by a MP_ACK frame received on a uniflow using the WLAN

De Coninck & BonaventureExpires 21 February 2021               [Page 13]

Internet-Draft                   MP-QUIC                     August 2020

3.7.  Exchanging Addresses

   When a multi-homed device connects to a dual-stacked server using its
   IPv4 address, it is aware of its local addresses (e.g., the WLAN and
   the cellular ones) and the IPv4 remote address used to establish the
   QUIC connection.  If the client wants to create new uniflows and use
   them over the IPv6 network, it needs to learn the other addresses of
   the remote peer.

   This is possible with the ADD_ADDRESS frames that are sent by a
   Multipath QUIC host to advertise its current addresses.  Each
   advertised address is identified by an Address ID.  The addresses
   attached to a host can vary during the lifetime of a Multipath QUIC
   connection.  A new ADD_ADDRESS frame is transmitted when a host has a
   new address.  This ADD_ADDRESS frame is protected as other QUIC
   control frames, which implies that it cannot be spoofed by attackers.
   The communicated address MUST first be validated by the receiving
   host before it starts using it as described in Section 8 of
   [I-D.ietf-quic-transport].  This process ensures that the advertised
   address actually belongs to the peer and that the peer can receive
   packets sent by the host on the provided address.  It also prevents
   hosts from launching amplification attacks to a victim address.

   If the client is behind a NAT, it could announce a private address in
   an ADD_ADDRESS frame.  In such situations, the server would not be
   able to validate the communicated address.  The client might still
   use its NATed addresses to start using its sending uniflows.  To
   enable the server to make the link between the private and the public
   addresses and hence conciliate the different 4-tuple views, Multipath
   QUIC provides the UNIFLOWS frame that lists the current active
   sending Uniflow IDs along with their associated local Address ID.
   Notice that a host might also discover the public addresses of its
   peer by observing its remote IP addresses associated to the

   A receiving uniflow is active as soon as the host has sent the
   MP_NEW_CONNECTION_ID frames proposing the corresponding Uniflow
   Connection IDs to its peer.  A sending uniflow is active when it has
   received its Uniflow Connection IDs and is bound to a validated
   4-tuple.  The UNIFLOWS frame indicates the local Address IDs that the
   uniflow uses from the sender's perspective.  With this information,
   the remote host can validate the public address and associate the
   advertised one with the perceived addresses.

De Coninck & BonaventureExpires 21 February 2021               [Page 14]

Internet-Draft                   MP-QUIC                     August 2020

3.8.  Coping with Address Removals

   During the lifetime of a QUIC connection, a host might lose some of
   its addresses.  A concrete example is a smartphone going out of reach
   of a WLAN network or shutting off one of its network interfaces.
   Such address removals are advertised using REMOVE_ADDRESS frames.
   The REMOVE_ADDRESS frame contains the Address ID of the lost address
   previously communicated through ADD_ADDRESS.  Notice that because a
   given Address ID might encounter several events that need to be
   ordered (e.g., ADD_ADDRESS, REMOVE_ADDRESS and ADD_ADDRESS again),
   both ADD_ADDRESS and REMOVE_ADDRESS frames include an Address ID
   related Sequence Number.

3.9.  Uniflow Migration

   At a given time, a Multipath QUIC endpoint gathers a set of active
   sending and receiving uniflows, each associated to a 4-tuple.  This
   association is mutable.  Hosts can change the 4-tuple used by their
   sending uniflows at any time, enabling QUIC to migrate uniflows from
   one network path to another.  Yet, to address privacy issues due to
   the linkability of addresses, hosts should avoid reusing the same
   Connection ID used by a sending uniflow when the 4-tuple changes, as
   described in Section 9.5 of [I-D.ietf-quic-transport].

3.10.  Congestion Control

   The QUIC congestion control scheme is defined in
   [I-D.ietf-quic-recovery].  This congestion control scheme is not
   suitable when several sending uniflows are active.  Using the
   congestion control scheme defined in [I-D.ietf-quic-recovery] with
   Multipath QUIC would result in unfairness.  Each sending uniflow of a
   Multipath QUIC connection MUST have its own congestion control state.
   As for Multipath TCP, the windows of the different sending uniflows
   MUST be coupled together [RFC6356].

4.  Mapping Uniflow IDs to Connection IDs

   As described in the overview section, hosts need to identify on which
   uniflows packets are sent.  The Uniflow ID must then be inferred from
   the public header.  This is done by using the Destination Connection
   ID field of Short Header packets.

De Coninck & BonaventureExpires 21 February 2021               [Page 15]

Internet-Draft                   MP-QUIC                     August 2020

   The Initial Uniflow Connection IDs are determined during the
   cryptographic handshake and actually correspond to both Connection
   IDs in the current single-path QUIC design [I-D.ietf-quic-transport].
   Additional Uniflow Connection IDs for the Initial Uniflows can be
   provided with the regular NEW_CONNECTION_ID frames.  The Uniflow
   Connection IDs of the other uniflows are determined when the
   MP_NEW_CONNECTION_ID frames are exchanged.

   Hosts MUST accept packets coming from their peer using the UCIDs they
   proposed in the (MP_)NEW_CONNECTION_ID frames they sent and associate
   them with the corresponding receiving Uniflow ID.  Upon reception of
   a (MP_)NEW_CONNECTION_ID frame, hosts MUST acknowledge it and MUST
   store the advertised Uniflow Destination Connection ID and the
   Uniflow ID of the proposed sending uniflow.

   Hosts MUST ensure that all advertised Uniflow Connection IDs are
   available for the whole connection lifetime, unless they have been
   retired by their peer in the meantime by the reception of a

   A host MUST NOT send MP_NEW_CONNECTION_ID frames with a Uniflow ID
   greater than the value of "max_sending_uniflow_id" advertised by its

5.  Using Multiple Uniflows

   This section describes in details the Multipath QUIC operations.

5.1.  Multipath Negotiation

   The Multipath negotiation takes place during the cryptographic
   handshake with the "max_sending_uniflow_id" transport parameter.  A
   QUIC connection is initially single-path in QUIC.  During this
   handshake, hosts advertise their support for multipath operations.
   When a host advertises a value for the "max_sending_uniflow_id"
   transport parameter, it indicates that it supports the multipath
   extensions, i.e., the extensions defined in this document (not to be
   mixed with the availability of local multiple network paths).  If any
   host does not advertise the "max_sending_uniflow_id" transport
   parameter, multipath extensions are disabled.

   The usage of multiple uniflows relies on the ability to use several
   Connection IDs over a same QUIC connection.  Therefore, zero-length
   Connection IDs MUST NOT be used if the peer advertises a value
   different from 0 for the "max_sending_uniflow_id" transport

De Coninck & BonaventureExpires 21 February 2021               [Page 16]

Internet-Draft                   MP-QUIC                     August 2020

5.1.1.  Transport Parameter Definition

   A host MAY use the following transport parameter:

   max_sending_uniflow_id (0x40):  Indicates the support of the
      multipath extension presented in this document, regardless of the
      carried value.  Its integer value puts an upper bound on the
      number of sending uniflows the host advertising the value is ready
      to support.  If absent, this means that the host does not agree to
      use the multipath extension over the connection.

5.2.  Coping with Additional Remote Addresses

   Hosts can learn remote addresses either by receiving ADD_ADDRESS
   frames or observing the 4-tuple of incoming packets.  Hosts MUST
   first validate the newly learned remote IP addresses before starting
   sending packets to those addresses.  This requirement is explained in
   Section 8.2.  Hosts MUST initiate Address Validation Procedure as
   specified in [I-D.ietf-quic-transport].

   A host MAY cache a validated address for a limited amount of time.

5.3.  Receiving Uniflow State

   When proposing uniflows to their peer, hosts need to maintain some
   state for their receiving uniflows.  This state is created upon the
   sending of a first MP_NEW_CONNECTION_ID frame proposing the
   corresponding Uniflow ID.  As long as there is still one active
   Uniflow Connection ID for this receiving uniflow (i.e., one UCID
   which was not retired yet using a MP_RETIRE_CONNECTION_ID), the host
   MUST accept packets over the receiving uniflow.  Once created, hosts
   MUST keep the following receiving uniflow information:

   Uniflow ID:  An integer that uniquely identifies the receiving
      uniflow in the connection.  This value is immutable.

   Uniflow Connection IDs:  Possible values for the Connection ID field
      of packets belonging to this receiving uniflow.  This value
      contains the sequence of active UCIDs that were advertised in
      previously sent MP_NEW_CONNECTION_ID frames.  Notice that this
      sequence might be empty, e.g., when all advertised UCIDs have been
      retired by the peer.

   Packet Number Space:  Packet number space dedicated to this receiving
      uniflow.  Packet number considerations described in Section 12.3
      of [I-D.ietf-quic-transport] apply within a given receiving

De Coninck & BonaventureExpires 21 February 2021               [Page 17]

Internet-Draft                   MP-QUIC                     August 2020

   Associated 4-tuple:  The tuple (sIP, dIP, sport, dport) currently
      observed to receive packets over this uniflow.  This value is
      mutable, because a host might receive a packet with a different
      (possibly) validated remote address and/or port than the one
      previously recorded.  If a host observes a change in the 4-tuple
      of the receiving uniflow, it follows the considerations of
      Section 9.5 of [I-D.ietf-quic-transport].

   Associated local Address ID:  The Address ID advertised in
      ADD_ADDRESS frames sent by the peer corresponding to the local
      address used to receive packets.  This helps to generate UNIFLOWS
      frames advertising the mapping between uniflows and addresses.
      The addresses on which the connection was established have Address
      ID 0.

   Hosts can also collect network measurements on a per-uniflow basis,
   like the number of packets received.

5.4.  Sending Uniflow State

   During a Multipath QUIC connection, hosts maintain some state for
   sending uniflows.  The state of the sending uniflow determines
   information that hosts are required to store.  The possible sending
   uniflow states are depicted in Figure 6.

   TODO: intermediate state for address validation?  Yes

         | receive a first MP_NEW_CONNECTION_ID
         |    with the associated Uniflow ID
         v       path usage over a validated 4-tuple
   +----------+ ------------------------------------> +----------+
   |  UNUSED  |                                       |  ACTIVE  |
   +----------+ <------------------------------------ +----------+
                    address change or retired UCID

      Figure 6: Finite-State Machine describing the possible states of
                             a sending uniflow

   Once a sending uniflow has been proposed by the peer in a received
   MP_NEW_CONNECTION_ID frame, it is in the UNUSED state.  In this
   situation, hosts MUST keep the following sending uniflow information:

   Uniflow ID:  An integer that uniquely identifies the sending uniflow
      in the connection.  This value is immutable.

De Coninck & BonaventureExpires 21 February 2021               [Page 18]

Internet-Draft                   MP-QUIC                     August 2020

   Uniflow Connection IDs:  Possible values for the Connection ID field
      of packets belonging to this sending uniflow.  This value contains
      the sequence of active UCIDs that were advertised in previously
      received MP_NEW_CONNECTION_ID frames.  Notice that this sequence
      might be empty, e.g., when all advertised UCIDs have been retired.

   Sending Uniflow State:  The current state of the sending uniflow,
      being one of the values presented in Figure 6.

   Packet Number Space:  Packet number space dedicated to this sending
      uniflow.  Packet number considerations described in Section 12.3
      of [I-D.ietf-quic-transport] apply within a given sending uniflow.

   When the host wants to start using the sending uniflow over a
   validated address, the sending uniflow goes to the ACTIVE state.
   This is the state where a sending uniflow can be used to send
   packets.  Having an uniflow in ACTIVE state only guarantees that it
   can be used, but the host is not forced to.  In addition to the
   fields required in the UNUSED state, the following elements MUST be

   Associated 4-tuple:  The tuple (sIP, dIP, sport, dport) currently
      used to packets over this uniflow.  This value is mutable, as the
      host might decide to change its local (or remote) address and/or
      port.  A host that changes the 4-tuple of a sending uniflow SHOULD
      migrate it.

   Associated (local Address ID, remote Address ID) tuple:  Those
      identifiers come from the ADD_ADDRESS sent (local) and received
      (remote).  This enables a host to temporarily stop using a sending
      uniflow when, e.g., the remote Address ID is declared as lost in a
      REMOVE_ADDRESS.  The addresses on which the connection was
      established have Address ID 0.  The reception of UNIFLOWS frames
      helps hosts associate the remote Address ID used by the sending

   Congestion controller:  A congestion window limiting the transmission
      rate of the sending uniflow.

   Performance metrics:  Basic statistics such as one-way delay or the
      number of packets sent.  This information can be useful when a
      host needs to perform packet scheduling decisions and flow

   It might happen that a sending path is temporarily unavailable,
   because one of the endpoint's addresses is no more available or
   because the host retired all the UCIDs of the sending uniflow.  In
   such cases, the path goes back to the UNUSED state.  When performing

De Coninck & BonaventureExpires 21 February 2021               [Page 19]

Internet-Draft                   MP-QUIC                     August 2020

   a transition back to the UNUSED state, hosts MUST reset the
   additional state added by the ACTIVE state.  In the UNUSED state, the
   host MUST NOT send non-probing packets on it.  At this state, the
   host might want to restart using the uniflow over another validated
   4-tuple, switching the uniflow state back to the ACTIVE state.
   However, its congestion controller state MUST be restarted and its
   performance metrics SHOULD be reset.

5.5.  Losing Addresses

   During the lifetime of a connection, a host might lose addresses,
   e.g., a network interface that was shut down.  All the ACTIVE sending
   uniflows that were using that local address MUST stop sending packets
   from that address.  To advertise the loss of an address to the peer,
   the host MUST send a REMOVE_ADDRESS frame indicating which local
   Address IDs has been lost.  The host MUST also send an UNIFLOWS frame
   indicating the status of the remaining ACTIVE uniflows.

   Upon reception of the REMOVE_ADDRESS, the receiving host MUST stop
   using the ACTIVE sending uniflows affected by the address removal.

   Hosts MAY reuse one of these sending uniflows by changing the
   assigned 4-tuple.  In this case, it MUST send an UNIFLOWS frame
   describing that change.

6.  New Frames

   To support the multipath operations, new frames have been defined to
   coordinate hosts.  All frames defined in this document MUST be
   exchanged in 1-RTT packets.  A host receiving one of the following
   frames in other encryption context MUST close the connection with a


   The MP_NEW_CONNECTION_ID frame (type=0x40) is an extension of the
   NEW_CONNECTION_ID frame defined by [I-D.ietf-quic-transport].  It
   provides the peer with alternative Connection IDs and associates them
   to a particular uniflow using the Uniflow ID.

   The format of the MP_NEW_CONNECTION_ID frame is as follows.

De Coninck & BonaventureExpires 21 February 2021               [Page 20]

Internet-Draft                   MP-QUIC                     August 2020

    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
   |                         Uniflow ID (i)                      ...
   |                      Sequence Number (i)                    ...
   |                      Retire Prior To (i)                    ...
   |   Length (8)  |                                               |
   +-+-+-+-+-+-+-+-+       Connection ID (8..160)                  +
   |                                                             ...
   |                                                               |
   +                                                               +
   |                                                               |
   +                   Stateless Reset Token (128)                 +
   |                                                               |
   +                                                               +
   |                                                               |

                    Figure 7: MP_NEW_CONNECTION_ID frame

   Compared to the frame specified in [I-D.ietf-quic-transport], an
   Uniflow ID varint field of is prefixed to associate the Connection ID
   with a uniflow.  This frame can be sent by both hosts.  Upon
   reception of the frame with a specified Uniflow ID, the peer MUST
   update the related sending uniflow state and store the communicated
   Connection ID.

   To limit the delay of the multipath usage upon handshake completion,
   hosts SHOULD send MP_NEW_CONNECTION_ID frames for receive uniflows
   they allow using as soon the connection establishment completes.

   The generation of Connection ID MUST follow the same considerations
   as presented in Section 5.1 of [I-D.ietf-quic-transport].


   The MP_RETIRE_CONNECTION_ID frame (type=0x41) is an extension of the
   RETIRE_CONNECTION_ID frame defined by [I-D.ietf-quic-transport].  It
   indicates that the end-host will no longer use a Connection ID
   related to a given uniflow that was issued by its peer.

   The format of the MP_RETIRE_CONNECTION_ID frame is shown below.

De Coninck & BonaventureExpires 21 February 2021               [Page 21]

Internet-Draft                   MP-QUIC                     August 2020

    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
   |                         Uniflow ID (i)                      ...
   |                      Sequence Number (i)                    ...

                  Figure 8: MP_RETIRE_CONNECTION_ID frame

   The frame is handled as described in [I-D.ietf-quic-transport] on an
   uniflow basis.

6.3.  MP_ACK Frame

   The MP_ACK frame (types 0x42 and 0x43) is an extension of the ACK
   frame defined by [I-D.ietf-quic-transport].  It allows hosts to
   acknowledge packets that were sent on non-initial uniflows.

   The format of the MP_ACK frame is shown below.

    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
   |                         Uniflow ID (i)                      ...
   |                    Largest Acknowledged (i)                 ...
   |                         ACK Delay (i)                       ...
   |                      ACK Range Count (i)                    ...
   |                      First ACK Range (i)                    ...
   |                         ACK Ranges (*)                      ...
   |                         [ECN Counts]                        ...

                  Figure 9: ACK frame adapted to Multipath

   Compared to the ACK frame, the MP_ACK frame is prefixed by a varint
   Uniflow ID field indicating to which uniflow the acknowledged packet
   sequence numbers relate.

De Coninck & BonaventureExpires 21 February 2021               [Page 22]

Internet-Draft                   MP-QUIC                     August 2020

6.4.  ADD_ADDRESS Frame

   The ADD_ADDRESS frame (type=0x44) is used by a host to advertise its
   currently reachable addresses.

   The format of the ADD_ADDRESS frame is shown below.

    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|0|0|P|IPVers.|Address ID (8) |
   |                       Sequence Number (i)                   ...
   |Interface T.(8)|
   |                       IP Address (32/128)                   ...
   |          [Port (16)]          |

                        Figure 10: ADD_ADDRESS Frame

   The ADD_ADDRESS frame contains the following fields.

   Reserved bits:  The three most-significant bits of the first byte are
      set to 0, and are reserved for future use.

   P bit:  The fourth most-significant bit of the first byte indicates,
      if set, the presence of the Port field.

   IPVers.:  The remaining four least-significant bits of the first byte
      contain the version of the IP address contained in the IP Address

   Address ID:  An unique identifier for the advertised address for
      tracking and removal purposes.  This is needed when, e.g., a NAT
      changes the IP address such that both hosts see different IP
      addresses for a same network path.

   Sequence Number:  An Address ID related sequence number of the event.
      The sequence number space is shared with REMOVE_ADDRESS frames
      mentioning the same Address ID.

   Interface Type:  Used to provide an indication about the interface
      type to which this address is bound.  The following values are

De Coninck & BonaventureExpires 21 February 2021               [Page 23]

Internet-Draft                   MP-QUIC                     August 2020

      *  0: fixed.  Used as default value.

      *  1: WLAN

      *  2: cellular

   IP Address:  The advertised IP address, in network order.

   Port:  This optional field indicates the port number related to the
      advertised IP address.  When this field is present, it indicates
      that an uniflow can use the 2-tuple (IP addr, port).

   Upon reception of an ADD_ADDRESS frame, the receiver SHOULD store the
   communicated address for future use.

   The receiver MUST NOT send packets others than validation ones to the
   communicated address without having validated it as specified in
   Section 8 of [I-D.ietf-quic-transport].  ADD_ADDRESS frames SHOULD
   contain globally reachable addresses.  Link-local and possibly
   private addresses SHOULD NOT be exchanged.


   The REMOVE_ADDRESS frame (type=0x45) is used by a host to signal that
   a previously announced address was lost.

   The format of the REMOVE_ADDRESS frame is shown below.

    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
   |Address ID (8) |
   |                       Sequence Number (i)                   ...

                      Figure 11: REMOVE_ADDRESS Frame

   The REMOVE_ADDRESS frame contains the following fields.

   Address ID:  The identifier of the address to remove.

De Coninck & BonaventureExpires 21 February 2021               [Page 24]

Internet-Draft                   MP-QUIC                     August 2020

   Sequence Number:  An Address ID related sequence number of the event.
      The sequence number space is shared with ADD_ADDRESS frames
      mentioning the same Address ID.  This help the receiver figure out
      that a REMOVE_ADDRESS might have been sent before an ADD_ADDRESS
      frame implying the same Address ID, even if for some reason the
      REMOVE_ADDRESS reaches the receiver after the newer ADD_ADDRESS

   A host SHOULD stop using sending uniflows using the removed address
   and set them in the UNUSED state.  If the REMOVE_ADDRESS contains an
   Address ID that was not previously announced, the receiver MUST
   silently ignore the frame.

6.6.  UNIFLOWS Frame

   The UNIFLOWS frame (type=0x46) communicates the uniflows' state of
   the sending host to the peer.  It allows the sender to communicate
   its active uniflows to the peer in order to detect potential
   connectivity issue over uniflows.

   The format of the UNIFLOWS frame is shown below.

    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
   |                          Sequence (i)                       ...
   |                      ReceivingUniflows (i)                  ...
   |                   ActiveSendingUniflows (i)                 ...
   |               Receiving Uniflow Info Section (*)            ...
   |                Sending Uniflow Info Section (*)             ...

                         Figure 12: UNIFLOWS Frame

   The UNIFLOWS frame contains the following fields.

   Sequence:  A variable-length integer.  This value starts at 0 and
      increases by 1 for each UNIFLOWS frame sent by the host.  It
      allows identifying the most recent UNIFLOWS frame.

   ReceivingUniflows:  The current number of receiving uniflows
      considered as being usable from the sender's point of view.

De Coninck & BonaventureExpires 21 February 2021               [Page 25]

Internet-Draft                   MP-QUIC                     August 2020

   ActiveSendingUniflows:  The current number of sending uniflows in the
      ACTIVE state from the sender's point of view.

   Receiving Uniflow Info Section:  Contains information about the
      receiving uniflows (there are ReceivingUniflows entries).

   Sending Uniflow Info Section:  Contains information about the sending
      uniflows in ACTIVE state (there are ActiveSendingUniflows

   Both Receiving Uniflow Info and Sending Uniflow Info Sections share
   the same format, which is shown below.

    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
   |                       Uniflow ID 0 (i)                      ...
   |LocAddrID 0 (8)|
   |                       Uniflow ID N (i)                      ...
   |LocAddrID N (8)|

                      Figure 13: Uniflow Info Section

   The fields in the Uniflow Info Section are the following.

   Uniflow ID:  The Uniflow ID of the uniflow the sending host provides
      information about.

   LocAddrID:  The local Address ID of the address currently used by the

   The Uniflow Info section only contains the local Address ID so far,
   but this section can be extended later with other potentially useful

7.  Extension of the Meaning of Existing QUIC Frames

   The multipath extensions do not modify the wire format of existing
   QUIC frames.  However, they extend the meaning of a few of them while
   keeping this addition transparent and consistent with the single-path
   QUIC design.  The concerned frames (and their extended meaning) are
   the following.

De Coninck & BonaventureExpires 21 February 2021               [Page 26]

Internet-Draft                   MP-QUIC                     August 2020

   NEW_CONNECTION_ID:  Equivalent to a MP_NEW_CONNECTION_ID frame with
      Uniflow ID set to 0.

      with Uniflow ID set to 0.

   ACK:  Equivalent to a MP_ACK frame with Uniflow ID set to 0.

8.  Security Considerations

8.1.  Nonce Computation

   With Multipath QUIC, each uniflow has its own packet number space.
   With the current nonce computation [I-D.ietf-quic-tls], using twice
   the same packet number over two different uniflows on the same
   direction leads to the same cryptographic nonce.  Using twice the
   same nonce MUST NOT happen, hence MP-QUIC has a different nonce
   computation than [I-D.ietf-quic-tls]

   The left most bits of nonce MUST be the Uniflow ID that identifies
   the current uniflow up to max_sending_uniflow_id.  The remaining bits
   of the nonce is formed by an exclusive OR of the least significant
   bits of the packet protection IV with the padded packet number (left-
   padded with 0s).  The nonce MUST be left-padded with a 0 if
   max_sending_uniflow_id <= 2, and the max_sending_uniflow_id MUST NOT
   be higher than 2^61.  If a uniflow has sent
   2^62-max_sending_uniflow_id packets, another uniflow MUST be used to
   avoid re-using the same nonce.

8.2.  Validation of Exchanged Addresses

   To use addresses communicated by the peer through ADD_ADDRESS frames,
   hosts are required to validate them before using uniflows to these
   addresses as described in Section 8 of [I-D.ietf-quic-transport].
   Section 21.12.3 of [I-D.ietf-quic-transport] provides additional
   motivation for this process.  In addition, hosts MUST send ADD
   ADDRESS frames in 1-RTT frames to prevent off-path attacks.

9.  IANA Considerations

9.1.  QUIC Transport Parameter Registry

   This document defines a new transport parameter for the negotiation
   of multiple paths.  The following entry in Table 1 should be added to
   the "QUIC Transport Parameters" registry under the "QUIC Protocol"

De Coninck & BonaventureExpires 21 February 2021               [Page 27]

Internet-Draft                   MP-QUIC                     August 2020

            | Value | Parameter Name         | Specification |
            | 0x40  | max_sending_uniflow_id | Section 5.1.1 |

              Table 1: Addition to QUIC Transport Parameters

10.  Acknowledgments

   We would like to thank Masahiro Kozuka and Kazuho Oku for their
   numerous comments on the first version of this draft.  We also thank
   Philipp Tiesel for his early comments that led to the current design,
   and Ian Swett for later feedback.  We also want to thank Christian
   Huitema for his draft about multipath requirements to identify
   critical elements about the multipath feature.  Mohamed Boucadair
   provided lot of useful inputs on the second version of this document.
   Maxime Piraux and Florentin Rochet helped us to improve the last
   versions of this draft.  This project was partially supported by the
   MQUIC project funded by the Walloon Government.

11.  References

11.1.  Normative References

              Thomson, M., "Version-Independent Properties of QUIC",
              Work in Progress, Internet-Draft, draft-ietf-quic-
              invariants-09, 9 June 2020, <http://www.ietf.org/internet-

              Iyengar, J. and I. Swett, "QUIC Loss Detection and
              Congestion Control", Work in Progress, Internet-Draft,
              draft-ietf-quic-recovery-29, 9 June 2020,

              Thomson, M. and S. Turner, "Using TLS to Secure QUIC",
              Work in Progress, Internet-Draft, draft-ietf-quic-tls-29,
              9 June 2020, <http://www.ietf.org/internet-drafts/draft-

              Iyengar, J. and M. Thomson, "QUIC: A UDP-Based Multiplexed
              and Secure Transport", Work in Progress, Internet-Draft,

De Coninck & BonaventureExpires 21 February 2021               [Page 28]

Internet-Draft                   MP-QUIC                     August 2020

              draft-ietf-quic-transport-29, 9 June 2020,

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
              May 2017, <https://www.rfc-editor.org/info/rfc8174>.

11.2.  Informative References

              Huitema, C., "Quic ACK Timestamps For Measuring One-Way
              Delays", Work in Progress, Internet-Draft, draft-huitema-
              quic-1wd-00, 3 January 2020, <http://www.ietf.org/

              Huitema, C., "QUIC Multipath Requirements", Work in
              Progress, Internet-Draft, draft-huitema-quic-mpath-req-01,
              7 January 2018, <http://www.ietf.org/internet-drafts/

   [IETFJ]    Bonaventure, O. and S. Seo, "Multipath TCP Deployments",
              IETF Journal , November 2016.

   [MPQUIC]   De Coninck, Q. and O. Bonaventure, "Multipath QUIC: Design
              and Evaluation", 13th International Conference on emerging
              Networking EXperiments and Technologies (CoNEXT 2017).
              http://multipath-quic.org , December 2017.

   [MPRTP]    Singh, V., Ahsan, S., and J. Ott, "MPRTP: Multipath
              considerations for real-time media", Proceedings of the
              4th ACM Multimedia Systems Conference , 2013.

   [OLIA]     Khalili, R., Gast, N., Popovic, M., Upadhyay, U., and J.-
              Y. Le Boudec, "MPTCP is not pareto-optimal: performance
              issues and a possible solution", Proceedings of the 8th
              international conference on Emerging networking
              experiments and technologies, ACM , 2012.

   [RFC0793]  Postel, J., "Transmission Control Protocol", STD 7,
              RFC 793, DOI 10.17487/RFC0793, September 1981,

De Coninck & BonaventureExpires 21 February 2021               [Page 29]

Internet-Draft                   MP-QUIC                     August 2020

   [RFC3077]  Duros, E., Dabbous, W., Izumiyama, H., Fujii, N., and Y.
              Zhang, "A Link-Layer Tunneling Mechanism for
              Unidirectional Links", RFC 3077, DOI 10.17487/RFC3077,
              March 2001, <https://www.rfc-editor.org/info/rfc3077>.

   [RFC6356]  Raiciu, C., Handley, M., and D. Wischik, "Coupled
              Congestion Control for Multipath Transport Protocols",
              RFC 6356, DOI 10.17487/RFC6356, October 2011,

   [RFC6824]  Ford, A., Raiciu, C., Handley, M., and O. Bonaventure,
              "TCP Extensions for Multipath Operation with Multiple
              Addresses", RFC 6824, DOI 10.17487/RFC6824, January 2013,

   [RFC7050]  Savolainen, T., Korhonen, J., and D. Wing, "Discovery of
              the IPv6 Prefix Used for IPv6 Address Synthesis",
              RFC 7050, DOI 10.17487/RFC7050, November 2013,

   [RFC7225]  Boucadair, M., "Discovering NAT64 IPv6 Prefixes Using the
              Port Control Protocol (PCP)", RFC 7225,
              DOI 10.17487/RFC7225, May 2014,

   [RFC8041]  Bonaventure, O., Paasch, C., and G. Detal, "Use Cases and
              Operational Experience with Multipath TCP", RFC 8041,
              DOI 10.17487/RFC8041, January 2017,

   [SCTPCMT]  Iyengar, J., Amer, P., and R. Stewart, "Concurrent
              multipath transfer using SCTP multihoming over independent
              end-to-end paths", IEEE/ACM Transactions on networking,
              Vol. 14, no 5 , 2006.

Appendix A.  Comparison with Multipath TCP

   Multipath TCP [RFC6824] is currently the most widely deployed
   multipath transport protocol on the Internet.  While its design
   impacted the initial versions of the Multipath extensions for the
   QUIC protocol, there are now major differences between both protocols
   that we now highlight.

De Coninck & BonaventureExpires 21 February 2021               [Page 30]

Internet-Draft                   MP-QUIC                     August 2020

A.1.  Multipath TCP Bidirectional Paths vs. QUIC Uniflows

   The notion of bidirectional paths, i.e., paths where packets flow in
   both directions, became a de facto standard with TCP.  The Multipath
   TCP extension [RFC6824] combines several TCP connections to spread a
   single data stream over them.  Hence, all the paths of a Multipath
   TCP connection must be bidirectional.  However, networking
   experiences showed that packets following a direction do not always
   share the exact same road as the packets in the opposite direction.
   Furthermore, QUIC does not require a network path to be bidirectional
   in order to be used.

A.2.  Uniflow Establishment

   Unlike Multipath TCP [RFC6824], both hosts dynamically control how
   many sending uniflows can currently be in use by the peer.

A.3.  Exchanging Data over Multiple Uniflows

   The uniflow on which data is sent is a packet-level information.
   This means that a frame can be sent regardless of the uniflow of the
   packet carrying it.  Furthermore, because the data offset is a frame-
   level information, there is no need to define additional sequence
   numbers to cope with reordering across uniflows, unlike Multipath TCP
   [RFC6824] that uses a Data Sequence Number at the Multipath TCP

   Decoupling the network paths of data with their acknowledgment can be
   useful to limit the latency due to (MP_)ACK transmissions on high-
   latency network paths and to enable the usage of unidirectional
   networks.  Such scheduling decision would not have been possible in
   Multipath TCP [RFC6824] which must acknowledge data on the
   (bidirectional) path it was received on.

A.4.  Congestion Control

   Multipath TCP uses the LIA congestion control scheme specified in
   [RFC6356].  This scheme can immediately be adapted to Multipath QUIC.
   Other coupled congestion control schemes have been proposed for
   Multipath TCP such as [OLIA].

A.5.  ACK Frame

   Since frames are independent of packets, and the uniflow notion
   relates to the packets, the (MP_)ACK frames can be sent on any
   uniflow, unlike Multipath TCP [RFC6824] which is constrained to send
   ACKs on the same path.

De Coninck & BonaventureExpires 21 February 2021               [Page 31]

Internet-Draft                   MP-QUIC                     August 2020

Appendix B.  To move in companion drafts

B.1.  Uniflow Establishment

   Sending useful data on a fresh new sending uniflow might lead to poor
   performance as the network path used by the QUIC uniflow might not be
   usable.  A typical case is when a server wants to initiate a new
   sending uniflow to a client behind a NAT.  The client would possibly
   never receive this packet, leading to connectivity issues on that
   uniflow.  In addition, an opportunistic usage of network paths might
   also lead to possible attacks, such as a client advertising the IP
   address of a victim hoping that the server will flood it.  To avoid
   these issues, a remote address MUST have been validated as described
   in [I-D.ietf-quic-transport] before associating it on a sending

   Because attaching to new networks may be volatile and an endpoint
   does not have full visibility on multiple paths that may be available
   (e.g., hosts connected to a CPE), a Multipath QUIC capable endhost
   SHOULD advertise a "max_sending_uniflow_id" value of at least 4 and
   SHOULD propose at least 4 receiving uniflows to its peer.

B.2.  Exchanging Addresses

   Likewise, the client may be located behind a NAT64.  As such it may
   announce an IPv6 address in an ADD_ADDRESS frame, that will be
   received over IPv4 by an IPv4-only server.  The server should not
   discard that address, even if it is not IPv6-capable.

   An IPv6-only client may also receive from the server an ADD_ADDRESS
   frame which may contain an IPv4 address.  The client should rely on
   means, such as [RFC7050] or [RFC7225], to learn the IPv6 prefix to
   build an IPv4-converted IPv6 address.

   Hosts that are connected behind an address sharing mechanism may
   collect the external IP address and port numbers assigned to the
   hosts and then use their addresses in the ADD_ADDRESS.  Means to
   gather such information include, but not limited to, UPnP IGD, PCP,
   or STUN.

B.3.  Uniflow Migration

   At a given time, a Multipath QUIC endpoint gathers a set of active
   sending and receiving uniflows, each associated to a 4-tuple.  To
   address privacy issues due to the linkability of addresses with
   Connection IDs, hosts should avoid changing the 4-tuple used by a
   sending uniflow.  There still remain situations where this change is
   unavoidable.  These can be categorized into two groups: host-aware

De Coninck & BonaventureExpires 21 February 2021               [Page 32]

Internet-Draft                   MP-QUIC                     August 2020

   changes (e.g., network handover from Wi-Fi to cellular) and host-
   unaware changes (e.g., NAT rebinding).

   For the host-aware case, let us consider the case of a Multipath QUIC
   connection where the client is a smartphone with both Wi-Fi and
   cellular.  It advertised both addresses and the server currently
   enables only one client's sending uniflow, the initial one.  The
   Initial Uniflow uses the Wi-Fi address.  Then, for some reason, the
   Wi-Fi address becomes unusable.  To preserve connectivity, the client
   might then decide to use the cellular address for its Initial sending
   uniflow.  It thus sends a REMOVE_ADDRESS announcing the loss of the
   Wi-Fi address and an UNIFLOWS frame to inform that its Initial
   sending uniflow is now using the cellular address.  If the cellular
   address validation succeeds (which could have been done as soon as
   the cellular address was advertised), the server can continue
   exchanging data through the cellular address.

   TODO: I don't think we have to change the Uniflow ID, we can just
   not sure the PATH_UPDATE frame is required, but will depend on the
   use cases

   TODO: update following paragraph to remove PATH_UPDATE

   However, both server and client might want to change the uniflow used
   on the cellular address for privacy concerns.  If the server provides
   an additional uniflow (e.g., with Uniflow ID 1) through
   MP_NEW_CONNECTION_ID frame at the beginning of the connection, the
   client can perform the network path change directly and avoid using
   the Initial Uniflow Connection ID on the cellular network.  This can
   be done using the PATH_UPDATE frame.  It can indicate that the host
   stopped to use the Initial sending uniflow to use the one with
   Uniflow ID 1 instead.  This frame is placed in the first packet sent
   to the new sending uniflow with its corresponding UCID.  The client
   can then send the REMOVE_ADDRESS and UNIFLOWS frames on this new
   uniflow.  Compared to the previous case, it is harder to link the
   uniflows with the IP addresses to observe that they belong to the
   same Multipath QUIC connection.

   For the host-unaware case, the situation is similar.  In case of NAT
   rebinding, the server will observe a change in the 2-tuple (source
   IP, source port) of the receiving uniflow of the packet.  The server
   first validates that the 2-tuple actually belongs to the client
   [I-D.ietf-quic-transport].  If it is the case, the server can send a
   PATH_UPDATE frame on a previously communicated but unused Uniflow ID.
   The client might have sent some packets with a given UCID on a
   different 4-tuple, but the server did not use the given UCID on that
   4-tuple.  Because some on-path devices may rewrite the source IP

De Coninck & BonaventureExpires 21 February 2021               [Page 33]

Internet-Draft                   MP-QUIC                     August 2020

   address to forward packets via the available network attachments
   (e.g., a host located behind a multi-homed CPE), the server may
   inadvertently conclude that an uniflow is not anymore valid leading
   thus to frequently sending PATH_UPDATE frames as a function of the
   traffic distribution scheme enforced by the on-path device.  To
   prevent such behavior, the server SHOULD wait for at least X seconds
   to ensure this is about a connection migration and not a side effect
   of an on-path multi-interfaced device.

B.4.  Scheduling Strategies

   The current QUIC design [I-D.ietf-quic-transport] offers a single
   scheduling space, i.e., which frames will be packed inside a given
   packet.  With the simultaneous use of several sending uniflows, a
   second dimension is added, i.e., the sending uniflow on which the
   packet will be sent.  This dimension can have a non negligible impact
   on the operations of Multipath QUIC, especially if the available
   sending uniflows exhibit very different network characteristics.

   The progression of the data flow depends on the reception of the
   MAX_DATA and MAX_STREAM_DATA frames.  Those frames SHOULD be
   duplicated on several or all ACTIVE sending uniflows.  This helps to
   limit the head-of-line blocking issue due to the transmission of the
   frames over a slow or lossy network path.

   The sending path on which (MP_)ACK frames are sent impacts the peer.
   The (MP_)ACK frame is notably used to determine the latency of a
   combination of uniflows.  In particular, the peer can compute the
   round-trip-time of the combination of its sending uniflow with its
   receive one.  The peer would compute the latency as the sum of the
   forward delay of the acknowledged uniflow and the return delay of the
   uniflow used to send the (MP_)ACK frame.  Choosing between
   acknowledging packets symmetrically (on uniflow B to A if packet was
   sent on A to B) or not is up to the implementation, if only possible.
   However, hosts SHOULD keep a consistent acknowledgment strategy.
   Selecting a random uniflow to acknowledge packets may affect the
   performance of the connection.  Notice that the inclusion of a
   timestamp field in the (MP_)ACK frame, as proposed by
   [I-D.huitema-quic-1wd], may help hosts estimate more precisely the
   one-way delay of each uniflow, therefore leading to improved
   scheduling decisions.  Unlike MAX_DATA and MAX_STREAM_DATA, (MP_)ACK
   frames SHOULD NOT be systematically duplicated on several sending
   uniflows as they can induce a large network overhead.

Appendix C.  Change Log

C.1.  Since draft-deconinck-quic-multipath-04

De Coninck & BonaventureExpires 21 February 2021               [Page 34]

Internet-Draft                   MP-QUIC                     August 2020

   *  Mostly editorial and reference fixes

C.2.  Since draft-deconinck-quic-multipath-03

   *  Clarify the notion of asymmetric paths by introducing uniflows

   *  Remove the PATH_UPDATE frame

   *  Rename PATHS frame to UNIFLOWS frame and adapt its content

   *  Add a sequence number to frames involving Address ID events (#4)

   *  Disallow Zero-length connection ID (#2)

   *  Correctly handle nonce computation (thanks to Florentin Rochet)

   *  Focus on the core concepts of multipath and delegate algorithms to
      companion drafts

   *  Updated text to match draft-ietf-quic-transport-27

C.3.  Since draft-deconinck-quic-multipath-02

   *  Consider asymmetric paths

C.4.  Since draft-deconinck-quic-multipath-01

   *  Include path policies considerations

   *  Add practical considerations thanks to Mohamed Boucadair inputs

   *  Adapt the RETIRE_CONNECTION_ID frame

   *  Updated text to match draft-ietf-quic-transport-18

C.5.  Since draft-deconinck-quic-multipath-00

   *  Comply with asymmetric Connection IDs

   *  Add max_paths transport parameter to negotiate initial number of
      active paths

   *  Path ID as a regular varint

   *  Remove max_path_id transport parameter

   *  Updated text to match draft-ietf-quic-transport-14

De Coninck & BonaventureExpires 21 February 2021               [Page 35]

Internet-Draft                   MP-QUIC                     August 2020

C.6.  Since draft-deconinck-multipath-quic-00

   *  Added PATH_UPDATE frame

   *  Added MAX_PATHS frame

   *  No more packet header change

   *  Implicit Path ID notification using Connection ID and
      NEW_CONNECTION_ID frames

   *  Variable-length encoding for Path ID

   *  Updated text to match draft-ietf-quic-transport-10

   *  Fixed various typos

Authors' Addresses

   Quentin De Coninck

   Email: quentin.deconinck@uclouvain.be

   Olivier Bonaventure

   Email: olivier.bonaventure@uclouvain.be

De Coninck & BonaventureExpires 21 February 2021               [Page 36]