QUIC-LB: Generating Routable QUIC Connection IDs

1.1. Terminology

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119].¶

In this document, these words will appear with that interpretation only when in ALL CAPS. Lower case uses of these words are not to be interpreted as carrying significance described in RFC 2119.¶

In this document, "client" and "server" refer to the endpoints of a QUIC connection unless otherwise indicated. A "load balancer" is an intermediary for that connection that does not possess QUIC connection keys, but it may rewrite IP addresses or conduct other IP or UDP processing. A "configuration agent" is the entity that determines the QUIC-LB configuration parameters for the network and leverages some system to distribute that configuration.¶

Note that stateful load balancers that act as proxies, by terminating a QUIC connection with the client and then retrieving data from the server using QUIC or another protocol, are treated as a server with respect to this specification.¶

For brevity, "Connection ID" will often be abbreviated as "CID".¶

1.2. Notation

All wire formats will be depicted using the notation defined in Section 1.3 of [RFC9000]. There is one addition: the function len() refers to the length of a field which can serve as a limit on a different field, so that the lengths of two fields can be concisely defined as limited to a sum, for example:¶

x(A..B) y(C..B-len(x))¶

indicates that x can be of any length between A and B, and y can be of any length between C and B provided that (len(x) + len(y)) does not exceed B.¶

The example below illustrates the basic framework:¶

Example Structure {
  One-bit Field (1),
  7-bit Field with Fixed Value (7) = 61,
  Field with Variable-Length Integer (i),
  Arbitrary-Length Field (..),
  Variable-Length Field (8..24),
  Variable-Length Field with Dynamic Limit (8..24-len(Variable-Length Field)),
  Field With Minimum Length (16..),
  Field With Maximum Length (..128),
  [Optional Field (64)],
  Repeated Field (8) ...,
}

2.1. Simplicity

QUIC is intended to provide unlinkability across connection migration, but servers are not required to provide additional connection IDs that effectively prevent linkability. If the coordination scheme is too difficult to implement, servers behind load balancers using connection IDs for routing will use trivially linkable connection IDs. Clients will therefore be forced to choose between terminating the connection during migration or remaining linkable, subverting a design objective of QUIC.¶

The solution should be both simple to implement and require little additional infrastructure for cryptographic keys, etc.¶

2.2. Security

In the limit where there are very few connections to a pool of servers, no scheme can prevent the linking of two connection IDs with high probability. In the opposite limit, where all servers have many connections that start and end frequently, it will be difficult to associate two connection IDs even if they are known to map to the same server.¶

QUIC-LB is relevant in the region between these extremes: when the information that two connection IDs map to the same server is helpful to linking two connection IDs. Obviously, any scheme that transparently communicates this mapping to outside observers compromises QUIC's defenses against linkability.¶

Though not an explicit goal of the QUIC-LB design, concealing the server mapping also complicates attempts to focus attacks on a specific server in the pool.¶

3.1. Config Rotation

The first two bits of any connection ID MUST encode an identifier for the configuration that the connection ID uses. This enables incremental deployment of new QUIC-LB settings (e.g., keys).¶

When new configuration is distributed to servers, there will be a transition period when connection IDs reflecting old and new configuration coexist in the network. The rotation bits allow load balancers to apply the correct routing algorithm and parameters to incoming packets.¶

Configuration Agents SHOULD deliver new configurations to load balancers before doing so to servers, so that load balancers are ready to process CIDs using the new parameters when they arrive.¶

A Configuration Agent SHOULD NOT use a codepoint to represent a new configuration until it takes precautions to make sure that all connections using CIDs with an old configuration at that codepoint have closed or transitioned.¶

Servers MUST NOT generate new connection IDs using an old configuration after receiving a new one from the configuration agent. Servers MUST send NEW_CONNECTION_ID frames that provide CIDs using the new configuration, and retire CIDs using the old configuration using the "Retire Prior To" field of that frame.¶

It also possible to use these bits for more long-lived distinction of different configurations, but this has privacy implications (see Section 11.3).¶

3.2. Configuration Failover

If a server has not received a valid QUIC-LB configuration, and believes that low-state, Connection-ID aware load balancers are in the path, it SHOULD generate connection IDs with the config rotation bits set to '11' and SHOULD use the "disable_active_migration" transport parameter in all new QUIC connections. It SHOULD NOT send NEW_CONNECTION_ID frames with new values.¶

A load balancer that sees a connection ID with config rotation bits set to '11' MUST revert to 5-tuple routing. These connection IDs may be of any length; however, see Section 11.6 for limits on this length.¶

3.3. Length Self-Description

Local hardware cryptographic offload devices may accelerate QUIC servers by receiving keys from the QUIC implementation indexed to the connection ID. However, on physical devices operating multiple QUIC servers, it is impractical to efficiently lookup these keys if the connection ID does not self-encode its own length.¶

Note that this is a function of particular server devices and is irrelevant to load balancers. As such, load balancers MAY omit this from their configuration. However, the remaining 6 bits in the first octet of the Connection ID are reserved to express the length of the following connection ID, not including the first octet.¶

A server not using this functionality SHOULD make the six bits appear to be random.¶

3.4. Format

First Octet {
  Config Rotation (2),
  CID Len or Random Bits (6),
}

The first octet has the following fields:¶

Config Rotation: Indicates the configuration used to interpret the CID.¶

CID Len or Random Bits: Length Self-Description (if applicable), or random bits otherwise. Encodes the length of the Connection ID following the First Octet.¶

4.1. Unroutable Connection IDs

QUIC-LB servers will generate Connection IDs that are decodable to extract a server ID in accordance with a specified algorithm and parameters. However, QUIC often uses client-generated Connection IDs prior to receiving a packet from the server.¶

These client-generated CIDs might not conform to the expectations of the routing algorithm and therefore not be routable by the load balancer. Those that are not routable are "unroutable DCIDs" and receive similar treatment regardless of why they're unroutable:¶

• The config rotation bits (Section 3.1) may not correspond to an active configuration. Note: a packet with a DCID that indicates 5-tuple routing (see Section 3.2) is always routable.¶
• The DCID might not be long enough for the decoder to process.¶
• The extracted server mapping might not correspond to an active server.¶

All other DCIDs are routable.¶

Load balancers MUST forward packets with routable DCIDs to a server in accordance with the chosen routing algorithm. Exception: if the load balancer can parse the QUIC packet and makes a routing decision depending on the contents (e.g., the SNI in a TLS client hello), it MAY route in accordance with this instead. However, load balancers MUST always route long header packets it cannot parse in accordance with the DCID (see Section 10).¶

Load balancers SHOULD drop short header packets with unroutable DCIDs.¶

When forwarding a packet with a long header and unroutable DCID, load balancers MUST use a fallback algorithm as specified in Section 4.2.¶

Load balancers MAY drop packets with long headers and unroutable DCIDs if and only if it knows that the encoded QUIC version does not allow an unroutable DCID in a packet with that signature. For example, a load balancer can safely drop a QUIC version 1 Handshake packet with an unroutable DCID, as a version 1 Handshake packet sent to a QUIC-LB routable server will always have a server-generated routable CID. The prohibition against dropping packets with long headers remains for unknown QUIC versions.¶

Furthermore, while the load balancer function MUST NOT drop packets, the device might implement other security policies, outside the scope of this specification, that might force a drop.¶

Servers that receive packets with unroutable CIDs MUST use the available mechanisms to induce the client to use a routable CID in future packets. In QUIC version 1, this requires using a routable CID in the Source CID field of server-generated long headers.¶

4.2. Fallback Algorithms

There are conditions described below where a load balancer routes a packet using a "fallback algorithm." It can choose any algorithm, without coordination with the servers, but the algorithm SHOULD be deterministic over short time scales so that related packets go to the same server. The design of this algorithm SHOULD consider the version-invariant properties of QUIC described in [RFC8999] to maximize its robustness to future versions of QUIC.¶

A fallback algorithm MUST NOT make the routing behavior dependent on any bits in the first octet of the QUIC packet header, except the first bit, which indicates a long header. All other bits are QUIC version-dependent and intermediaries SHOULD NOT base their design on version-specific templates.¶

For example, one fallback algorithm might convert a unroutable DCID to an integer and divided by the number of servers, with the modulus used to forward the packet. The number of servers is usually consistent on the time scale of a QUIC connection handshake. Another might simply hash the address/port 4-tuple. See also Section 10.¶

4.3. Server ID Allocation

Load Balancer configurations include a mapping of server IDs to forwarding addresses. The corresponding server configurations contain one or more unique server IDs.¶

The configuration agent chooses a server ID length for each configuration that MUST be at least one octet.¶

A QUIC-LB configuration MAY significantly over-provision the server ID space (i.e., provide far more codepoints than there are servers) to increase the probability that a randomly generated Destination Connection ID is unroutable.¶

The configuration agent SHOULD provide a means for servers to express the number of server IDs it can usefully employ, because a single routing address actually corresponds to multiple server entities (see Section 9.1).¶

Conceptually, each configuration has its own set of server ID allocations, though two static configurations with identical server ID lengths MAY use a common allocation between them.¶

A server encodes one of its assigned server IDs in any CID it generates using the relevant configuration.¶

5.1. CID format

All connection IDs use the following format:¶

QUIC-LB Connection ID {
    First Octet (8),
    Server ID (8..152-len(Nonce)),
    Nonce (32..152-len(Server ID)),
}

5.2. Configuration Agent Actions

The configuration agent assigns a server ID to every server in its pool in accordance with Section 4.3, and determines a server ID length (in octets) sufficiently large to encode all server IDs, including potential future servers.¶

Each configuration specifies the length of the Server ID and Nonce fields, with limits defined for each algorithm.¶

Optionally, it also defines a 16-octet key. Note that failure to define a key means that observers can determine the assigned server of any connection, significantly increasing the linkability of QUIC address migration.¶

The nonce length MUST be at least 4 octets. The server ID length MUST be at least 1 octet.¶

As QUIC version 1 limits connection IDs to 20 octets, the server ID and nonce lengths MUST sum to 19 octets or less.¶

5.3. Server Actions

The server writes the first octet and its server ID into their respective fields.¶

If there is no key in the configuration, the server MUST fill the Nonce field with bytes that appear to be random. If there is a key, the server fills the nonce field with a nonce of its choosing. See Section 11.6 for details.¶

The server MAY append additional bytes to the connection ID, up to the limit specified in that version of QUIC, for its own use. These bytes MUST NOT provide observers with any information that could link two connection IDs to the same connection, client, or server. In particular, all servers using a configuration MUST consistently add the same length to each connection ID, to preserve the linkability objectives of QUIC-LB. Any additional bytes SHOULD appear random unless individual servers are not distinguishable (e.g. any server using that configuration appends identical bytes to every connection ID).¶

If there is no key in the configuration, the Connection ID is complete. Otherwise, there are further steps, as described in the two following subsections.¶

Encryption below uses the AES-128-ECB cipher. Future standards could add new algorithms that use other ciphers to provide cryptographic agility in accordance with [RFC7696]. QUIC-LB implementations SHOULD be extensible to support new algorithms.¶

expand_left(0xaaba3c, 0x13) = 0xaaba3c00000000000000000000000013

expand_right(0xaaba3c, 0x13) = 0x13000000000000000000000000aaba3c

truncate_left(0x2094842ca49256198c2deaa0ba53caa0, 28) = 0x2094842
truncate_right(0x2094842ca49256198c2deaa0ba53caa0, 28) = 0xa53caa0

server_id = 0x3144a
nonce = 0x9c69c275
key = 0xfdf726a9893ec05c0632d3956680baf0

// step 1
plaintext_CID = 0x31441a9c69c275

// step 2
left_0 = 0x31441a9
right_0 = 0xc69c275

// step 3
aes_input = 0x31441a90000000000000000000000001
ciphertext = 0x4d140de42d0b85bdf554ba35c1d5c653

// step 4
right_1 = 0xc69c275 ^ 0x1d5c653 = 0xdbc0426

// step 5
aes_input = 0x0200000000000000000000000dbc0426
aes_output = 0x7e99160f3cf5b89c70584ccd2c2cd24b
left_1 = 0x31441a9 ^ 0x7e99160 = 0x4fdd0c9

// step 6
AES input = 0x4fdd0c90000000000000000000000003
AES output = 0x26c1d5a3a5e31ff8e3ca505da6061ac6
right_2 = 0xdbc0426 ^ 0x6061ac6 = 0xbba1ee0

// step 7
AES input = 0x0400000000000000000000000bba1ee0
AES output = 0xade1b8b25b436a94007d80cf3704377b
left_2 = 0x4fdd0c9 ^ 0xade1b8b = 0xe23cb42

// step 8
cid = first_octet || left_2 || right_2 = 0x07e23cb42bba1ee0

5.4. Load Balancer Actions

On each incoming packet, the load balancer extracts consecutive octets, beginning with the second octet. If there is no key, the first octets correspond to the server ID.¶

If there is a key, the load balancer takes one of two actions:¶

left_1 = left_2 ^ truncate_left(AES_ECB(key, expand_right(right_2), 0x04))
right_1 = right_2 ^ truncate_right(AES_ECB(key, expand_left(left_1, 0x03))
left_0 = left_1 ^ truncate_left(AES_ECB(key, expand_right(right_1), 0x02))

7.1. Common Requirements

Regardless of mechanism, a retry service has an active mode, where it is generating Retry packets, and an inactive mode, where it is not, based on its assessment of server load and the likelihood an attack is underway. The choice of mode MAY be made on a per-packet or per-connection basis, through a stochastic process or based on client address.¶

A configuration agent MUST distribute a list of QUIC versions the Retry Service supports. It MAY also distribute either an "Allow-List" or a "Deny-List" of other QUIC versions. It MUST NOT distribute both an Allow-List and a Deny-List.¶

The Allow-List or Deny-List MUST NOT include any versions included for Retry Service Support.¶

The Configuration Agent MUST provide a means for the entity that controls the Retry Service to report its supported version(s) to the configuration Agent. If the entity has not reported this information, it MUST NOT activate the Retry Service and the configuration agent MUST NOT distribute configuration that activates it.¶

The configuration agent MAY delete versions from the final supported version list if policy does not require the Retry Service to operate on those versions.¶

The configuration Agent MUST provide a means for the entities that control servers behind the Retry Service to report either an Allow-List or a Deny-List.¶

If all entities supply Allow-Lists, the consolidated list MUST be the union of these sets. If all entities supply Deny-Lists, the consolidated list MUST be the intersection of these sets.¶

If entities provide a mixture of Allow-Lists and Deny-Lists, the consolidated list MUST be a Deny-List that is the intersection of all provided Deny-Lists and the inverses of all Allow-Lists.¶

If no entities that control servers have reported Allow-Lists or Deny-Lists, the default is a Deny-List with the null set (i.e., all unsupported versions will be admitted). This preserves the future extensibilty of QUIC.¶

A retry service MUST forward all packets for a QUIC version it does not support that are not on a Deny-List or absent from an Allow-List. Note that if servers support versions the retry service does not, this may increase load on the servers.¶

Note that future versions of QUIC might not have Retry packets, require different information in Retry, or use different packet type indicators.¶

7.2. No-Shared-State Retry Service

The no-shared-state retry service requires no coordination, except that the server must be configured to accept this service and know which QUIC versions the retry service supports. The scheme uses the first bit of the token to distinguish between tokens from Retry packets (codepoint '0') and tokens from NEW_TOKEN frames (codepoint '1').¶

Non-Shared-State Retry Service Token {
  Token Type (1) = 0,
  ODCIL (7) = 8..20,
  Original Destination Connection ID (64..160),
  Opaque Data (..),
}

7.3. Shared-State Retry Service

A shared-state retry service uses a shared key, so that the server can decode the service's retry tokens. It does not require that all traffic pass through the Retry service, so servers MAY send Retry packets in response to Initial packets that don't include a valid token.¶

Both server and service must have time synchronized with respect to one another to prevent tokens being incorrectly marked as expired, though tight synchronization is unnecessary.¶

The tokens are protected using AES128-GCM AEAD, as explained in Section 7.3.1. All tokens, generated by either the server or retry service, MUST use the following format, which includes:¶

• A 1 bit token type identifier.¶
• A 7 bit token key identifier.¶
• A 96 bit unique token number transmitted in clear text, but protected as part of the AEAD associated data.¶
• A token body, encoding the Original Destination Connection ID and the Timestamp, optionally followed by server specific Opaque Data.¶

The token protection uses an 128 bit representation of the source IP address from the triggering Initial packet. The client IP address is 16 octets. If an IPv4 address, the last 12 octets are zeroes. It also uses the Source Connection ID of the Retry packet, which will cause an authentication failure if it differs from the Destination Connection ID of the packet bearing the token.¶

If there is a Network Address Translator (NAT) in the server infrastructure that changes the client IP, the Retry Service MUST either be positioned behind the NAT, or the NAT must have the token key to rewrite the Retry token accordingly. Note also that a host that obtains a token through a NAT and then attempts to connect over a path that does not have an identically configured NAT will fail address validation.¶

The 96 bit unique token number is set to a random value using a cryptography-grade random number generator.¶

The token key identifier and the corresponding AEAD key and AEAD IV are provisioned by the configuration agent.¶

The token body is encoded as follows:¶

Shared-State Retry Service Token Body {
   Timestamp (64),
   [ODCIL (8) = 8..20],
   [Original Destination Connection ID (64..160)],
   [Port (16)],
   Opaque Data (..),
}

The token body has the following fields:¶

Timestamp: The Timestamp is a 64-bit integer, in network order, that expresses the expiration time of the token as a number of seconds in POSIX time (see Sec. 4.16 of [TIME_T]).¶

ODCIL: The original destination connection ID length. Tokens in NEW_TOKEN frames do not have this field.¶

Original Destination Connection ID: The server or Retry Service copies this from the field in the client Initial packet. Tokens in NEW_TOKEN frames do not have this field.¶

Port: The Source Port of the UDP datagram that triggered the Retry packet. This field MUST be present if and only if the ODCIL is greater than zero. This field is therefore always absent in tokens in NEW_TOKEN frames.¶

Opaque Data: The server may use this field to encode additional information, such as congestion window, RTT, or MTU. The Retry Service MUST have zero-length opaque data.¶

Some implementations of QUIC encode in the token the Initial Packet Number used by the client, in order to verify that the client sends the retried Initial with a PN larger that the triggering Initial. Such implementations will encode the Initial Packet Number as part of the opaque data. As tokens may be generated by the Service, servers MUST NOT reject tokens because they lack opaque data and therefore the packet number.¶

Shared-state Retry Services use the AES-128-ECB cipher. Future standards could add new algorithms that use other ciphers to provide cryptographic agility in accordance with [RFC7696]. Retry Service and server implementations SHOULD be extensible to support new algorithms.¶

Shared-State Retry Service Token {
  Token Type (1),
  Key Sequence (7),
  Unique Token Number (96),
  Encrypted Shared-State Retry Service Token Body (64..),
  AEAD Integrity Check Value (128),
}

Shared-State Retry Service Token Pseudoheader {
  IP Address (128),
  Token Type (1),
  Key Sequence (7),
  Unique Token Number (96),
  [RSCIL (8)],
  [Retry Source Connection ID (0..20)],
}

9.1. Load balancer chains

Some network architectures may have multiple tiers of low-state load balancers, where a first tier of devices makes a routing decision to the next tier, and so on, until packets reach the server. Although QUIC-LB is not explicitly designed for this use case, it is possible to support it.¶

If each load balancer is assigned a range of server IDs that is a subset of the range of IDs assigned to devices that are closer to the client, then the first devices to process an incoming packet can extract the server ID and then map it to the correct forwarding address. Note that this solution is extensible to arbitrarily large numbers of load-balancing tiers, as the maximum server ID space is quite large.¶

9.2. Moving connections between servers

Some deployments may transparently move a connection from one server to another. The means of transferring connection state between servers is out of scope of this document.¶

To support a handover, a server involved in the transition could issue CIDs that map to the new server via a NEW_CONNECTION_ID frame, and retire CIDs associated with the new server using the "Retire Prior To" field in that frame.¶

Alternately, if the old server is going offline, the load balancer could simply map its server ID to the new server's address.¶

11.1. Attackers not between the load balancer and server

Any attacker might open a connection to the server infrastructure and aggressively simulate migration to obtain a large sample of IDs that map to the same server. It could then apply analytical techniques to try to obtain the server encoding.¶

An encrypted encoding provides robust protection against this. An unencrypted one provides none.¶

Were this analysis to obtain the server encoding, then on-path observers might apply this analysis to correlating different client IP addresses.¶

11.2. Attackers between the load balancer and server

Attackers in this privileged position are intrinsically able to map two connection IDs to the same server. The QUIC-LB algorithms do prevent the linkage of two connection IDs to the same individual connection if servers make reasonable selections when generating new IDs for that connection.¶

11.3. Multiple Configuration IDs

During the period in which there are multiple deployed configuration IDs (see Section 3.1), there is a slight increase in linkability. The server space is effectively divided into segments with CIDs that have different config rotation bits. Entities that manage servers SHOULD strive to minimize these periods by quickly deploying new configurations across the server pool.¶

11.4. Limited configuration scope

A simple deployment of QUIC-LB in a cloud provider might use the same global QUIC-LB configuration across all its load balancers that route to customer servers. An attacker could then simply become a customer, obtain the configuration, and then extract server IDs of other customers' connections at will.¶

To avoid this, the configuration agent SHOULD issue QUIC-LB configurations to mutually distrustful servers that have different keys for encryption algorithms. In many cases, the load balancers can distinguish these configurations by external IP address.¶

However, assigning multiple entities to an IP address is complimentary with concealing DNS requests (e.g., DoH [RFC8484]) and the TLS Server Name Indicator (SNI) ([I-D.ietf-tls-esni]) to obscure the ultimate destination of traffic. While the load balancer's fallback algorithm (Section 4.2) can use the SNI to make a routing decision on the first packet, there are three ways to route subsequent packets:¶

• all co-tenants can use the same QUIC-LB configuration, leaking the server mapping to each other as described above;¶
• co-tenants can be issued one of up to three configurations distinguished by the config rotation bits (Section 3.1), exposing information about the target domain to the entire network; or¶
• tenants can use 4-tuple routing in their CIDs (in which case they SHOULD disable migration in their connections), which neutralizes the value of QUIC-LB but preserves privacy.¶

When configuring QUIC-LB, administrators must evaluate the privacy tradeoff considering the relative value of each of these properties, given the trust model between tenants, the presence of methods to obscure the domain name, and value of address migration in the tenant use cases.¶

As the plaintext algorithm makes no attempt to conceal the server mapping, these deployments SHOULD simply use a common configuration.¶

11.5. Stateless Reset Oracle

Section 21.9 of [RFC9000] discusses the Stateless Reset Oracle attack. For a server deployment to be vulnerable, an attacking client must be able to cause two packets with the same Destination CID to arrive at two different servers that share the same cryptographic context for Stateless Reset tokens. As QUIC-LB requires deterministic routing of DCIDs over the life of a connection, it is a sufficient means of avoiding an Oracle without additional measures.¶

Note also that when a server starts using a new QUIC-LB config rotation codepoint, new CIDs might not be unique with respect to previous configurations that occupied that codepoint, and therefore different clients may have observed the same CID and stateless reset token. A straightforward method of managing stateless reset keys is to maintain a separate key for each config rotation codepoint, and replace each key when the configuration for that codepoint changes. Thus, a server transitions from one config to another, it will be able to generate correct tokens for connections using either type of CID.¶

11.6. Connection ID Entropy

If a server ever reuses a nonce in generating a CID for a given configuration, it risks exposing sensitive information. Given the same server ID, the CID will be identical (aside from a possible difference in the first octet). This can risk exposure of the QUIC-LB key. If two clients receive the same connection ID, they also have each other's stateless reset token unless that key has changed in the interim.¶

The encrypte mode needs to generate different cipher text for each generated Connection ID instance to protect the Server ID. To do so, at least four octets of the CID are reserved for a nonce that, if used only once, will result in unique cipher text for each Connection ID.¶

If servers simply increment the nonce by one with each generated connection ID, then it is safe to use the existing keys until any server's nonce counter exhausts the allocated space and rolls over. To maximize entropy, servers SHOULD start with a random nonce value, in which case the configuration is usable until the nonce value wraps around to zero and then reaches the initial value again.¶

Whether or not it implements the counter method, the server MUST NOT reuse a nonce until it switches to a configuration with new keys.¶

If the nonce is sent in plaintext, servers MUST generate nonces so that they appear to be random. Observable correlations between plaintext nonces would provide trivial linkability between individual connections, rather than just to a common server.¶

For any algorithm, configuration agents SHOULD implement an out-of-band method to discover when servers are in danger of exhausting their nonce space, and SHOULD respond by issuing a new configuration. A server that has exhausted its nonces MUST either switch to a different configuration, or if none exists, use the 4-tuple routing config rotation codepoint.¶

When sizing a nonce that is to be randomly generated, the configuration agent SHOULD consider that a server generating a N-bit nonce will create a duplicate about every 2^(N/2) attempts, and therefore compare the expected rate at which servers will generate CIDs with the lifetime of a configuration.¶

11.7. Shared-State Retry Keys

The Shared-State Retry Service defined in Section 7.3 describes the format of retry tokens or new tokens protected and encrypted using AES128-GCM. Each token includes a 96 bit randomly generated unique token number, and an 8 bit identifier used to get the AES-GCM encryption context. The AES-GCM encryption context contains a 128 bit key and an AEAD IV. There are three important security considerations for these tokens:¶

• An attacker that obtains a copy of the encryption key will be able to decrypt and forge tokens.¶
• Attackers may be able to retrieve the key if they capture a sufficently large number of retry tokens encrypted with a given key.¶
• Confidentiality of the token data will fail if separate tokens reuse the same 96 bit unique token number and the same key.¶

To protect against disclosure of keys to attackers, service and servers MUST ensure that the keys are stored securely. To limit the consequences of potential exposures, the time to live of any given key should be limited.¶

Section 6.6 of [RFC9001] states that "Endpoints MUST count the number of encrypted packets for each set of keys. If the total number of encrypted packets with the same key exceeds the confidentiality limit for the selected AEAD, the endpoint MUST stop using those keys." It goes on with the specific limit: "For AEAD_AES_128_GCM and AEAD_AES_256_GCM, the confidentiality limit is 2^23 encrypted packets; see Appendix B.1." It is prudent to adopt the same limit here, and configure the service in such a way that no more than 2^23 tokens are generated with the same key.¶

In order to protect against collisions, the 96 bit unique token numbers should be generated using a cryptographically secure pseudorandom number generator (CSPRNG), as specified in Appendix C.1 of the TLS 1.3 specification [RFC8446]. With proper random numbers, if fewer than 2^40 tokens are generated with a single key, the risk of collisions is lower than 0.001%.¶

A.1. Tree Diagram

This summary of the YANG models uses the notation in [RFC8340].¶

module: ietf-quic-lb-server
  +--rw quic-lb!
     +--rw config-id                         uint8
     +--rw first-octet-encodes-cid-length?   boolean
     +--rw server-id-length                  uint8
     +--rw nonce-length                      uint8
     +--rw cid-key?                          quic-lb-key
     +--rw server-id                         yang:hex-string
     +--rw retry-service-config
        +--rw supported-versions*            uint32
        +--rw unsupported-version-default?   enumeration
        +--rw version-exceptions*            uint32
        +--rw token-keys* [key-sequence-number]
           +--rw key-sequence-number    uint8
           +--rw token-key              quic-lb-key
           +--rw token-iv               yang:hex-string

module: ietf-quic-lb-middlebox
  +--rw quic-lb!
     +--rw cid-configs* [config-rotation-bits]
     |  +--rw config-rotation-bits    uint8
     |  +--rw server-id-length        uint8
     |  +--rw cid-key?                quic-lb-key
     |  +--rw nonce-length            uint8
     |  +--rw server-id-mappings* [server-id]
     |     +--rw server-id         yang:hex-string
     |     +--rw server-address    inet:ip-address
     +--rw retry-service-config
        +--rw supported-versions*            uint32
        +--rw unsupported-version-default?   enumeration
        +--rw version-exceptions*            uint32
        +--rw token-keys* [key-sequence-number]
           +--rw key-sequence-number    uint8
           +--rw token-key              quic-lb-key
           +--rw token-iv               yang:hex-string

B.1. Unencrypted CIDs

cr_bits sid nonce cid
0 c4605e 4504cc4f 07c4605e4504cc4f
1 350d28b420 3487d970b 40a350d28b4203487d970b

B.2. Encrypted CIDs

The key for all of these examples is 8f95f09245765f80256934e50c66207f. The test vectors include an example that uses the 16-octet single-pass special case, as well as an instance where the server ID length exceeds the nonce length, requiring a fourth decryption pass.¶

cr_bits sid nonce cid
0 ed793a ee080dbf 07b07d879e7f63c4
1 ed793a51d49b8f5fab65 ee080dbf48 4fcf549a705a6525af3f14c92d2fb6c6
2 ed793a51d49b8f5f ee080dbf48c0d1e5 904d4c1b1c024d2f59dd16c955d1b2a4ee
0 ed793a51d49b8f5fab ee080dbf48c0d1e55d 128e0ccb2ecf5cd08998fca8ac9bfe5abae55d

B.3. Shared State Retry Tokens

In this case, the shared-state retry token is issued by retry service, so the opaque data of shared-state retry token body would be null (Section 7.3).¶

LB configuration:
key_seq 0x00
encrypt_key 0x30313233343536373839303132333435
AEAD_IV 0x313233343536373839303132

Shared-State Retry Service Token Body:
ODCIL 0x12
RSCIL 0x10
port 0x1a0a
original_destination_connection_id 0x0c3817b544ca1c94313bba41757547eec937
retry_source_connection_id 0x0301e770d24b3b13070dd5c2a9264307
timestamp 0x0000000060c7bf4d

Shared-State Retry Service Token:
unique_token_number 0x59ef316b70575e793e1a8782
key_sequence 0x00
encrypted_shared_state_retry_service_token_body
0x7d38b274aa4427c7a1557c3fa666945931defc65da387a83855196a7cb73caac1e28e5346fd76868de94f8b62294
AEAD_ICV 0xf91174fdd711543a32d5e959867f9c22

AEAD related parameters:
client_ip_addr 127.0.0.1
client_port 6666
AEAD_nonce 0x68dd025f45616941072ab6b0
AEAD_associated_data 0x7f00000100000000000000000000000059ef316b70575e793e1a878200

C.1. DTLS 1.0 and 1.2

DTLS 1.0 [RFC4347] and 1.2 [RFC6347] use packet formats that a QUIC-LB router will interpret as short header packets with CIDs that request 4-tuple routing. As such, they will route such packets consistently as long as the 4-tuple does not change. Note that DTLS 1.0 has been deprecated by the IETF.¶

The first octet of every DTLS 1.0 or 1.2 datagram contains the content type. A QUIC-LB load balancer will interpret any content type less than 128 as a short header packet, meaning that the subsequent octets should contain a connection ID.¶

Existing TLS content types comfortably fit in the range below 128. Assignment of codepoints greater than 64 would require coordination in accordance with [RFC7983], and anyway would likely create problems demultiplexing DTLS and version 1 of QUIC. Therefore, this document believes it is extremely unlikely that TLS content types of 128 or greater will be assigned. Nevertheless, such an assignment would cause a QUIC-LB load balancer to interpret the packet as a QUIC long header with an essentially random connection ID, which is likely to be routed irregularly.¶

The second octet of every DTLS 1.0 or 1.2 datagram is the bitwise complement of the DTLS Major version (i.e. version 1.x = 0xfe). A QUIC-LB load balancer will interpret this as a connection ID that requires 4-tuple based load balancing, meaning that the routing will be consistent as long as the 4-tuple remains the same.¶

[I-D.ietf-tls-dtls-connection-id] defines an extension to add connection IDs to DTLS 1.2. Unfortunately, a QUIC-LB load balancer will not correctly parse the connection ID and will continue 4-tuple routing. An modified QUIC-LB load balancer that correctly identifies DTLS and parses a DTLS 1.2 datagram for the connection ID is outside the scope of this document.¶

C.2. DTLS 1.3

DTLS 1.3 [I-D.draft-ietf-tls-dtls13] changes the structure of datagram headers in relevant ways.¶

Handshake packets continue to have a TLS content type in the first octet and 0xfe in the second octet, so they will be 4-tuple routed, which should not present problems for likely NAT rebinding or address change events.¶

Non-handshake packets always have zero in their most significant bit and will therefore always be treated as QUIC short headers. If the connection ID is present, it follows in the succeeding octets. Therefore, a DTLS 1.3 association where the server utilizes Connection IDs and the encodings in this document will be routed correctly in the presence of client address and port changes.¶

However, if the client does not include the connection_id extension in its ClientHello, the server is unable to use connection IDs. In this case, non- handshake packets will appear to contain random connection IDs and be routed randomly. Thus, unmodified QUIC-LB load balancers will not work with DTLS 1.3 if the client does not advertise support for connection IDs, or the server does not request the use of a compliant connection ID.¶

A QUIC-LB load balancer might be modified to identify DTLS 1.3 packets and correctly parse the fields to identify when there is no connection ID and revert to 4-tuple routing, removing the server requirement above. However, such a modification is outside the scope of this document, and classifying some packets as DTLS might be incompatible with future versions of QUIC.¶

C.3. Future Versions of DTLS

As DTLS does not have an IETF consensus document that defines what parts of DTLS will be invariant in future versions, it is difficult to speculate about the applicability of this section to future versions of DTLS.¶

E.1. since draft-ietf-quic-load-balancers-10

• Refactored algorithm descriptions; made the 4-pass algorithm easier to implement¶
• Revised test vectors¶
• Split YANG model into a server and middlebox version¶

E.2. since draft-ietf-quic-load-balancers-09

• Renamed "Stream Cipher" and "Block Cipher" to "Encrypted Short" and "Encrypted Long"¶
• Added section on per-connection state¶
• Changed "Encrypted Short" to a 4-pass algorithm.¶
• Recommended a random initial nonce when incrementing.¶
• Clarified what SNI LBs should do with unknown QUIC versions.¶

E.3. since draft-ietf-quic-load-balancers-08

• Eliminate Dynamic SID allocation¶
• Eliminated server use bytes¶

E.4. since draft-ietf-quic-load-balancers-07

• Shortened SSCID nonce minimum length to 4 bytes¶
• Removed RSCID from Retry token body¶
• Simplified CID formats¶
• Shrunk size of SID table¶

E.5. since draft-ietf-quic-load-balancers-06

• Added interoperability with DTLS¶
• Changed "non-compliant" to "unroutable"¶
• Changed "arbitrary" algorithm to "fallback"¶
• Revised security considerations for mistrustful tenants¶
• Added retry service considerations for non-Initial packets¶

E.6. since draft-ietf-quic-load-balancers-05

• Added low-config CID for further discussion¶
• Complete revision of shared-state Retry Token¶
• Added YANG model¶
• Updated configuration limits to ensure CID entropy¶
• Switched to notation from quic-transport¶

E.7. since draft-ietf-quic-load-balancers-04

• Rearranged the shared-state retry token to simplify token processing¶
• More compact timestamp in shared-state retry token¶
• Revised server requirements for shared-state retries¶
• Eliminated zero padding from the test vectors¶
• Added server use bytes to the test vectors¶
• Additional compliant DCID criteria¶

E.8. since-draft-ietf-quic-load-balancers-03

• Improved Config Rotation text¶
• Added stream cipher test vectors¶
• Deleted the Obfuscated CID algorithm¶

E.9. since-draft-ietf-quic-load-balancers-02

• Replaced stream cipher algorithm with three-pass version¶
• Updated Retry format to encode info for required TPs¶
• Added discussion of version invariance¶
• Cleaned up text about config rotation¶
• Added Reset Oracle and limited configuration considerations¶
• Allow dropped long-header packets for known QUIC versions¶

E.10. since-draft-ietf-quic-load-balancers-01

• Test vectors for load balancer decoding¶
• Deleted remnants of in-band protocol¶
• Light edit of Retry Services section¶
• Discussed load balancer chains¶

E.11. since-draft-ietf-quic-load-balancers-00

• Removed in-band protocol from the document¶

E.12. Since draft-duke-quic-load-balancers-06

• Switch to IETF WG draft.¶

E.13. Since draft-duke-quic-load-balancers-05

• Editorial changes¶
• Made load balancer behavior independent of QUIC version¶
• Got rid of token in stream cipher encoding, because server might not have it¶
• Defined "non-compliant DCID" and specified rules for handling them.¶
• Added psuedocode for config schema¶

E.14. Since draft-duke-quic-load-balancers-04

• Added standard for retry services¶

E.15. Since draft-duke-quic-load-balancers-03

• Renamed Plaintext CID algorithm as Obfuscated CID¶
• Added new Plaintext CID algorithm¶
• Updated to allow 20B CIDs¶
• Added self-encoding of CID length¶

E.16. Since draft-duke-quic-load-balancers-02

• Added Config Rotation¶
• Added failover mode¶
• Tweaks to existing CID algorithms¶
• Added Block Cipher CID algorithm¶
• Reformatted QUIC-LB packets¶

E.17. Since draft-duke-quic-load-balancers-01

• Complete rewrite¶
• Supports multiple security levels¶
• Lightweight messages¶

E.18. Since draft-duke-quic-load-balancers-00

• Converted to markdown¶
• Added variable length connection IDs¶