QUIC: A UDP-Based Multiplexed and Secure Transport

[x]

Indicates that x is optional

x (A)

Indicates that x is A bits long

x (A/B/C) …

Indicates that x is one of A, B, or C bits long

x (i) …

Indicates that x uses the variable-length encoding in Section 8.1

x (*) …

Indicates that x is variable-length

QUIC relies on a combined cryptographic and transport handshake for setting up a secure transport connection. QUIC connections are expected to commonly use 0-RTT handshakes, meaning that for most QUIC connections, data can be sent immediately following the client handshake packet, without waiting for a reply from the server. QUIC provides a dedicated stream (Stream ID 0) to be used for performing the cryptographic handshake and QUIC options negotiation. The format of the QUIC options and parameters used during negotiation are described in this document, but the handshake protocol that runs on Stream ID 0 is described in the accompanying cryptographic handshake draft [QUIC-TLS].¶

When application messages are transported over TCP, independent application messages can suffer from head-of-line blocking. When an application multiplexes many streams atop TCP’s single-bytestream abstraction, a loss of a TCP segment results in blocking of all subsequent segments until a retransmission arrives, irrespective of the application streams that are encapsulated in subsequent segments. QUIC ensures that lost packets carrying data for an individual stream only impact that specific stream. Data received on other streams can continue to be reassembled and delivered to the application.¶

QUIC’s packet framing and acknowledgments carry rich information that help both congestion control and loss recovery in fundamental ways. Each QUIC packet carries a new packet number, including those carrying retransmitted data. This obviates the need for a separate mechanism to distinguish acknowledgments for retransmissions from those for original transmissions, avoiding TCP’s retransmission ambiguity problem. QUIC acknowledgments also explicitly encode the delay between the receipt of a packet and its acknowledgment being sent, and together with the monotonically-increasing packet numbers, this allows for precise network roundtrip-time (RTT) calculation. QUIC’s ACK frames support multiple ACK blocks, so QUIC is more resilient to reordering than TCP with SACK support, as well as able to keep more bytes on the wire when there is reordering or loss.¶

QUIC implements stream- and connection-level flow control. At a high level, a QUIC receiver advertises the maximum amount of data that it is willing to receive on each stream. As data is sent, received, and delivered on a particular stream, the receiver sends MAX_STREAM_DATA frames that increase the advertised limit for that stream, allowing the peer to send more data on that stream.¶

In addition to this stream-level flow control, QUIC implements connection-level flow control to limit the aggregate buffer that a QUIC receiver is willing to allocate to all streams on a connection. Connection-level flow control works in the same way as stream-level flow control, but the bytes delivered and the limits are aggregated across all streams.¶

TCP headers appear in plaintext on the wire and are not authenticated, causing a plethora of injection and header manipulation issues for TCP, such as receive-window manipulation and sequence-number overwriting. While some of these are mechanisms used by middleboxes to improve TCP performance, others are active attacks. Even “performance-enhancing” middleboxes that routinely interpose on the transport state machine end up limiting the evolvability of the transport protocol, as has been observed in the design of MPTCP [RFC6824] and in its subsequent deployability issues.¶

Generally, QUIC packets are always authenticated and the payload is typically fully encrypted. The parts of the packet header which are not encrypted are still authenticated by the receiver, so as to thwart any packet injection or manipulation by third parties. Some early handshake packets, such as the Version Negotiation packet, are not encrypted, but information sent in these unencrypted handshake packets is later verified as part of cryptographic processing.¶

QUIC connections are identified by a Connection ID, a 64-bit unsigned number randomly generated by the server. QUIC’s consistent connection ID allows connections to survive changes to the client’s IP and port, such as those caused by NAT rebindings or by the client changing network connectivity to a new address. QUIC provides automatic cryptographic verification of a rebound client, since the client continues to use the same session key for encrypting and decrypting packets. The consistent connection ID can be used to allow migration of the connection to a new server IP address as well, since the Connection ID remains consistent across changes in the client’s and the server’s network addresses.¶

QUIC version negotiation allows for multiple versions of the protocol to be deployed and used concurrently. Version negotiation is described in Section 7.2.¶

Header Form:

The most significant bit (0x80) of octet 0 (the first octet) is set to 1 for long headers.

Long Packet Type:

The remaining seven bits of octet 0 contain the packet type. This field can indicate one of 128 packet types. The types specified for this version are listed in Table 1.

Connection ID:

Octets 1 through 8 contain the connection ID. Section 5.6 describes the use of this field in more detail.

Version:

Octets 9 to 12 contain the selected protocol version. This field indicates which version of QUIC is in use and determines how the rest of the protocol fields are interpreted.

Packet Number:

Octets 13 to 16 contain the packet number. Section 5.7 describes the use of packet numbers.

Payload:

Octets from 17 onwards (the rest of QUIC packet) are the payload of the packet.

The header form, packet type, connection ID, packet number and version fields of a long header packet are version-independent. The types of packets defined in Table 1 are version-specific. See Section 5.8 for details on how packets from different versions of QUIC are interpreted.¶

The interpretation of the fields and the payload are specific to a version and packet type. Type-specific semantics for this version are described in the following sections.¶

Header Form:

The most significant bit (0x80) of octet 0 is set to 0 for the short header.

Omit Connection ID Flag:

The second bit (0x40) of octet 0 indicates whether the Connection ID field is omitted. If set to 0, then the Connection ID field is present; if set to 1, the Connection ID field is omitted. The Connection ID field can only be omitted if the omit_connection_id transport parameter (Section 7.4.1) is specified by the intended recipient of the packet.

Key Phase Bit:

The third bit (0x20) of octet 0 indicates the key phase, which allows a recipient of a packet to identify the packet protection keys that are used to protect the packet. See [QUIC-TLS] for details.

Short Packet Type:

The remaining 5 bits of octet 0 include one of 32 packet types. Table 2 lists the types that are defined for short packets.

Connection ID:

If the Omit Connection ID Flag is not set, a connection ID occupies octets 1 through 8 of the packet. See Section 5.6 for more details.

Packet Number:

The length of the packet number field depends on the packet type. This field can be 1, 2 or 4 octets long depending on the short packet type.

Protected Payload:

Packets with a short header always include a 1-RTT protected payload.

The packet type in a short header currently determines only the size of the packet number field. Additional types can be used to signal the presence of other fields.¶

The header form, omit connection ID flag, and connection ID of a short header packet are version-independent. The remaining fields are specific to the selected QUIC version. See Section 5.8 for details on how packets from different versions of QUIC are interpreted.¶

A Version Negotiation packet is inherently not version-specific, and does not use the packet headers defined above. Upon receipt by a client, it will appear to be a packet using the long header, but will be identified as a Version Negotiation packet based on the Version field.¶

The Version Negotiation packet is a response to a client packet that contains a version that is not supported by the server, and is only sent by servers.¶

The value in the Unused field is selected randomly by the server. The Connection ID field echoes the corresponding value from the triggering client packet. This allows clients some assurance that the server received the packet and that the Version Negotiation packet is in fact from the server. The Version field MUST be set to 0x00000000. The remainder of the Version Negotiation packet is a list of 32-bit versions which the server supports.¶

A Version Negotiation packet cannot be explicitly acknowledged in an ACK frame by a client. Receiving another Initial packet implicitly acknowledges a Version Negotiation packet.¶

See Section 7.2 for a description of the version negotiation process.¶

Once version negotiation is complete, the cryptographic handshake is used to agree on cryptographic keys. The cryptographic handshake is carried in Initial (Section 5.4.1), Retry (Section 5.4.2) and Handshake (Section 5.4.3) packets.¶

All these packets use the long header and contain the current QUIC version in the version field.¶

In order to prevent tampering by version-unaware middleboxes, handshake packets are protected with a connection- and version-specific key, as described in [QUIC-TLS]. This protection does not provide confidentiality or integrity against on-path attackers, but provides some level of protection against off-path attackers.¶

Packets that are protected with 0-RTT keys are sent with long headers; all packets protected with 1-RTT keys are sent with short headers. The different packet types explicitly indicate the encryption level and therefore the keys that are used to remove packet protection.¶

Packets protected with 0-RTT keys use a type value of 0x7C. The connection ID field for a 0-RTT packet is selected by the client.¶

The client can send 0-RTT packets after receiving a Handshake packet (Section 5.4.3), if that packet does not complete the handshake. Even if the client receives a different connection ID in the Handshake packet, it MUST continue to use the connection ID selected by the client for 0-RTT packets, see Section 5.6.¶

The version field for protected packets is the current QUIC version.¶

The packet number field contains a packet number, which increases with each packet sent, see Section 5.7 for details.¶

The payload is protected using authenticated encryption. [QUIC-TLS] describes packet protection in detail. After decryption, the plaintext consists of a sequence of frames, as described in Section 6.¶

QUIC connections are identified by their 64-bit Connection ID. All long headers contain a Connection ID. Short headers indicate the presence of a Connection ID using the Omit Connection ID flag. When present, the Connection ID is in the same location in all packet headers, making it straightforward for middleboxes, such as load balancers, to locate and use it.¶

The client MUST choose a random connection ID and use it in Initial packets (Section 5.4.1) and 0-RTT packets (Section 5.5).¶

When the server receives a Initial packet and decides to proceed with the handshake, it chooses a new value for the connection ID and sends that in a Retry (Section 5.4.2) or Handshake (Section 5.4.3) packet. The server MAY choose to use the value that the client initially selects.¶

Once the client receives the connection ID that the server has chosen, it MUST use it for all subsequent Handshake (Section 5.4.3) and 1-RTT (Section 5.5) packets but not for 0-RTT packets (Section 5.5).¶

Server’s Version Negotiation (Section 5.3) and Retry (Section 5.4.2) packets MUST use connection ID selected by the client.¶

The packet number is an integer in the range 0 to 2^62-1. The value is used in determining the cryptographic nonce for packet encryption. Each endpoint maintains a separate packet number for sending and receiving. The packet number for sending MUST increase by at least one after sending any packet, unless otherwise specified (see Section 5.7.1).¶

A QUIC endpoint MUST NOT reuse a packet number within the same connection (that is, under the same cryptographic keys). If the packet number for sending reaches 2^62 - 1, the sender MUST close the connection without sending a CONNECTION_CLOSE frame or any further packets; a server MAY send a Stateless Reset (Section 7.9.4) in response to further packets that it receives.¶

For the packet header, the number of bits required to represent the packet number are reduced by including only the least significant bits of the packet number. The actual packet number for each packet is reconstructed at the receiver based on the largest packet number received on a successfully authenticated packet.¶

A packet number is decoded by finding the packet number value that is closest to the next expected packet. The next expected packet is the highest received packet number plus one. For example, if the highest successfully authenticated packet had a packet number of 0xaa82f30e, then a packet containing a 16-bit value of 0x1f94 will be decoded as 0xaa831f94.¶

The sender MUST use a packet number size able to represent more than twice as large a range than the difference between the largest acknowledged packet and packet number being sent. A peer receiving the packet will then correctly decode the packet number, unless the packet is delayed in transit such that it arrives after many higher-numbered packets have been received. An endpoint SHOULD use a large enough packet number encoding to allow the packet number to be recovered even if the packet arrives after packets that are sent afterwards.¶

As a result, the size of the packet number encoding is at least one more than the base 2 logarithm of the number of contiguous unacknowledged packet numbers, including the new packet.¶

For example, if an endpoint has received an acknowledgment for packet 0x6afa2f, sending a packet with a number of 0x6b4264 requires a 16-bit or larger packet number encoding; whereas a 32-bit packet number is needed to send a packet with a number of 0x6bc107.¶

Version Negotiation (Section 5.3) and Retry (Section 5.4.2) packets have special rules for populating the packet number field.¶

• the location of the header form flag,
• the location of the Omit Connection ID flag in short headers,
• the location and size of the Connection ID field in both header forms,
• the location and size of the Version field in long headers,
• the format and semantics of the Version Negotiation packet.

Implementations MUST assume that an unsupported version uses an unknown packet format. All other fields MUST be ignored when processing a packet that contains an unsupported version.¶

Incoming packets are classified on receipt. Packets can either be associated with an existing connection, be discarded, or - for servers - potentially create a new connection.¶

Packets that can be associated with an existing connection are handled according to the current state of that connection. Packets are associated with existing connections using connection ID if it is present; this might include connection IDs that were advertised using NEW_CONNECTION_ID (Section 8.13). Packets without connection IDs and long-form packets for connections that have incomplete cryptographic handshakes are associated with an existing connection using the tuple of source and destination IP addresses and ports.¶

A packet that uses the short header could be associated with an existing connection with an incomplete cryptographic handshake. Such a packet could be a valid packet that has been reordered with respect to the long-form packets that will complete the cryptographic handshake. This might happen after the final set of cryptographic handshake messages from either peer. These packets are expected to be correlated with a connection using the tuple of IP addresses and ports. Packets that might be reordered in this fashion SHOULD be buffered in anticipation of the handshake completing.¶

0-RTT packets might be received prior to a Client Initial packet at a server. If the version of these packets is acceptable to the server, it MAY buffer these packets in anticipation of receiving a reordered Client Initial packet.¶

Buffering ensures that data is not lost, which improves performance; conversely, discarding these packets could create false loss signals for the congestion controllers. However, limiting the number and size of buffered packets might be needed to prevent exposure to denial of service.¶

For clients, any packet that cannot be associated with an existing connection SHOULD be discarded if it is not buffered. Discarded packets MAY be logged for diagnostic or security purposes.¶

For servers, packets that aren’t associated with a connection potentially create a new connection. However, only packets that use the long packet header and that are at least the minimum size defined for the protocol version can be initial packets. A server MAY discard packets with a short header or packets that are smaller than the smallest minimum size for any version that the server supports. A server that discards a packet that cannot be associated with a connection MAY also generate a stateless reset (Section 7.9.4).¶

This version of QUIC defines a minimum size for initial packets of 1200 octets (see Section 9). Versions of QUIC that define smaller minimum initial packet sizes need to be aware that initial packets will be discarded without action by servers that only support versions with larger minimums. Clients that support multiple QUIC versions can avoid this problem by ensuring that they increase the size of their initial packets to the largest minimum size across all of the QUIC versions they support. Servers need to recognize initial packets that are the minimum size of all QUIC versions they support.¶

QUIC’s connection establishment begins with version negotiation, since all communication between the endpoints, including packet and frame formats, relies on the two endpoints agreeing on a version.¶

A QUIC connection begins with a client sending an Initial packet (Section 5.4.1). The details of the handshake mechanisms are described in Section 7.3, but any Initial packet sent from the client to the server MUST use the long header format - which includes the version of the protocol being used - and they MUST be padded to at least 1200 octets.¶

The server receives this packet and determines whether it potentially creates a new connection (see Section 7.1). If the packet might generate a new connection, the server then checks whether it understands the version that the client has selected.¶

If the packet contains a version that is acceptable to the server, the server proceeds with the handshake (Section 7.3). This commits the server to the version that the client selected.¶

QUIC relies on a combined cryptographic and transport handshake to minimize connection establishment latency. QUIC allocates stream 0 for the cryptographic handshake. Version 0x00000001 of QUIC uses TLS 1.3 as described in [QUIC-TLS]; a different QUIC version number could indicate that a different cryptographic handshake protocol is in use.¶

• authenticated key exchange, where
  • a server is always authenticated,
  • a client is optionally authenticated,
  • every connection produces distinct and unrelated keys,
  • keying material is usable for packet protection for both 0-RTT and 1-RTT packets, and
  • 1-RTT keys have forward secrecy
• authenticated values for the transport parameters of the peer (see Section 7.4)
• authenticated confirmation of version negotiation (see Section 7.4.4)
• authenticated negotiation of an application protocol (TLS uses ALPN [RFC7301] for this purpose)
• for the server, the ability to carry data that provides assurance that the client can receive packets that are addressed with the transport address that is claimed by the client (see Section 7.6)

The initial cryptographic handshake message MUST be sent in a single packet. Any second attempt that is triggered by address validation MUST also be sent within a single packet. This avoids having to reassemble a message from multiple packets. Reassembling messages requires that a server maintain state prior to establishing a connection, exposing the server to a denial of service risk.¶

The first client packet of the cryptographic handshake protocol MUST fit within a 1232 octet QUIC packet payload. This includes overheads that reduce the space available to the cryptographic handshake protocol.¶

Details of how TLS is integrated with QUIC is provided in more detail in [QUIC-TLS].¶

During connection establishment, both endpoints make authenticated declarations of their transport parameters. These declarations are made unilaterally by each endpoint. Endpoints are required to comply with the restrictions implied by these parameters; the description of each parameter includes rules for its handling.¶

The format of the transport parameters is the TransportParameters struct from Figure 6. This is described using the presentation language from Section 3 of [I-D.ietf-tls-tls13].¶

The extension_data field of the quic_transport_parameters extension defined in [QUIC-TLS] contains a TransportParameters value. TLS encoding rules are therefore used to encode the transport parameters.¶

QUIC encodes transport parameters into a sequence of octets, which are then included in the cryptographic handshake. Once the handshake completes, the transport parameters declared by the peer are available. Each endpoint validates the value provided by its peer. In particular, version negotiation MUST be validated (see Section 7.4.4) before the connection establishment is considered properly complete.¶

Definitions for each of the defined transport parameters are included in Section 7.4.1. Any given parameter MUST appear at most once in a given transport parameters extension. An endpoint MUST treat receipt of duplicate transport parameters as a connection error of type TRANSPORT_PARAMETER_ERROR.¶

A server can process an initial cryptographic handshake messages from a client without committing any state. This allows a server to perform address validation (Section 7.6, or to defer connection establishment costs.¶

A server that generates a response to an initial packet without retaining connection state MUST use the Retry packet (Section 5.4.2). This packet causes a client to reset its transport state and to continue the connection attempt with new connection state while maintaining the state of the cryptographic handshake.¶

A server MUST NOT send multiple Retry packets in response to a client handshake packet. Thus, any cryptographic handshake message that is sent MUST fit within a single packet.¶

In TLS, the Retry packet type is used to carry the HelloRetryRequest message.¶

Transport protocols commonly spend a round trip checking that a client owns the transport address (IP and port) that it claims. Verifying that a client can receive packets sent to its claimed transport address protects against spoofing of this information by malicious clients.¶

This technique is used primarily to avoid QUIC from being used for traffic amplification attack. In such an attack, a packet is sent to a server with spoofed source address information that identifies a victim. If a server generates more or larger packets in response to that packet, the attacker can use the server to send more data toward the victim than it would be able to send on its own.¶

Several methods are used in QUIC to mitigate this attack. Firstly, the initial handshake packet is padded to at least 1200 octets. This allows a server to send a similar amount of data without risking causing an amplification attack toward an unproven remote address.¶

A server eventually confirms that a client has received its messages when the cryptographic handshake successfully completes. This might be insufficient, either because the server wishes to avoid the computational cost of completing the handshake, or it might be that the size of the packets that are sent during the handshake is too large. This is especially important for 0-RTT, where the server might wish to provide application data traffic - such as a response to a request - in response to the data carried in the early data from the client.¶

To send additional data prior to completing the cryptographic handshake, the server then needs to validate that the client owns the address that it claims.¶

Source address validation is therefore performed during the establishment of a connection. TLS provides the tools that support the feature, but basic validation is performed by the core transport protocol.¶

A different type of source address validation is performed after a connection migration, see Section 7.7.2.¶

QUIC connections are identified by their 64-bit Connection ID. QUIC’s consistent connection ID allows connections to survive changes to the client’s IP and/or port, such as those caused by client or server migrating to a new network. Connection migration allows a client to retain any shared state with a connection when they move networks. This includes state that can be hard to recover such as outstanding requests, which might otherwise be lost with no easy way to retry them.¶

An endpoint that receives packets that contain a source IP address and port that has not yet been used can start sending new packets with those as a destination IP address and port. Packets exchanged between endpoints can then follow the new path.¶

Due to variations in path latency or packet reordering, packets from different source addresses might be reordered. The packet with the highest packet number MUST be used to determine which path to use. Endpoints also need to be prepared to receive packets from an older source address.¶

An endpoint MUST validate that its peer can receive packets at the new address before sending any significant quantity of data to that address, or it risks being used for denial of service. See Section 7.7.2 for details.¶

A connection migration could be triggered by an attacker that is able to capture and forward a packet such that it arrives before the legitimate copy of that packet. Such a packet will appear to be a legitimate connection migration and the legitimate copy will be dropped as a duplicate.¶

After a spurious migration, validation of the source address will fail because the entity at the source address does not have the necessary cryptographic keys to read or respond to the PING frame that is sent to it, even if it wanted to. Such a spurious connection migration could result in the connection being dropped when the source address validation fails. This grants an attacker the ability to terminate the connection.¶

Receipt of packets with higher packet numbers from the legitimate address will trigger another connection migration. This will cause the validation of the address of the spurious migration to be abandoned.¶

To ensure that a peer sends packets from the legitimate address before the validation of the new address can fail, an endpoint SHOULD attempt to validate the old remote address before attempting to validate the new address. If the connection migration is spurious, then the legitimate address will be used to respond and the connection will migrate back to the old address.¶

As with any address validation, packets containing retransmissions of the PING frame validating an address MUST be sent to the address being validated. Consequently, during a migration of a peer, an endpoint could be sending to multiple remote addresses.¶

An endpoint MAY abandon address validation for an address that it considers to be already valid. That is, if successive connection migrations occur in quick succession with the final remote address being identical to the initial remote address, the endpoint MAY abandon address validation for that address.¶

• idle timeout (Section 7.9.2)
• immediate close (Section 7.9.3)
• stateless reset (Section 7.9.4)

QUIC frames use a common variable-length encoding for all non-negative integer values. This encoding ensures that smaller integer values need fewer octets to encode.¶

The QUIC variable-length integer encoding reserves the two most significant bits of the first octet to encode the base 2 logarithm of the integer encoding length in octets. The integer value is encoded on the remaining bits, in network byte order.¶

This means that integers are encoded on 1, 2, 4, or 8 octets and can encode 6, 14, 30, or 62 bit values respectively. Table 4 summarizes the encoding properties.¶

For example, the eight octet sequence c2 19 7c 5e ff 14 e8 8c (in hexadecimal) decodes to the decimal value 151288809941952652; the four octet sequence 9d 7f 3e 7d decodes to 494878333; the two octet sequence 7b bd decodes to 15293; and the single octet 25 decodes to 37 (as does the two octet sequence 40 25).¶

Error codes (Section 12.3) are described using integers, but do not use this encoding.¶

The PADDING frame (type=0x00) has no semantic value. PADDING frames can be used to increase the size of a packet. Padding can be used to increase an initial client packet to the minimum required size, or to provide protection against traffic analysis for protected packets.¶

A PADDING frame has no content. That is, a PADDING frame consists of the single octet that identifies the frame as a PADDING frame.¶

An endpoint may use a RST_STREAM frame (type=0x01) to abruptly terminate a stream.¶

After sending a RST_STREAM, an endpoint ceases transmission and retransmission of STREAM frames on the identified stream. A receiver of RST_STREAM can discard any data that it already received on that stream.¶

Stream ID:

A variable-length integer encoding of the Stream ID of the stream being terminated.

Application Protocol Error Code:

A 16-bit application protocol error code (see Section 12.4) which indicates why the stream is being closed.

Final Offset:

A variable-length integer indicating the absolute byte offset of the end of data written on this stream by the RST_STREAM sender.

An endpoint sends a CONNECTION_CLOSE frame (type=0x02) to notify its peer that the connection is being closed. CONNECTION_CLOSE is used to signal errors at the QUIC layer, or the absence of errors (with the NO_ERROR code).¶

If there are open streams that haven’t been explicitly closed, they are implicitly closed when the connection is closed.¶

Error Code:

A 16-bit error code which indicates the reason for closing this connection. CONNECTION_CLOSE uses codes from the space defined in Section 12.3 (APPLICATION_CLOSE uses codes from the application protocol error code space, see Section 12.4).

Reason Phrase Length:

A variable-length integer specifying the length of the reason phrase in bytes. Note that a CONNECTION_CLOSE frame cannot be split between packets, so in practice any limits on packet size will also limit the space available for a reason phrase.

Reason Phrase:

A human-readable explanation for why the connection was closed. This can be zero length if the sender chooses to not give details beyond the Error Code. This SHOULD be a UTF-8 encoded string [RFC3629].

An APPLICATION_CLOSE frame (type=0x03) uses the same format as the CONNECTION_CLOSE frame (Section 8.4), except that it uses error codes from the application protocol error code space (Section 12.4) instead of the transport error code space.¶

Other than the error code space, the format and semantics of the APPLICATION_CLOSE frame are identical to the CONNECTION_CLOSE frame.¶

The MAX_DATA frame (type=0x04) is used in flow control to inform the peer of the maximum amount of data that can be sent on the connection as a whole.¶

Maximum Data:

A variable-length integer indicating the maximum amount of data that can be sent on the entire connection, in units of octets.

All data sent in STREAM frames counts toward this limit, with the exception of data on stream 0. The sum of the largest received offsets on all streams - including streams in terminal states, but excluding stream 0 - MUST NOT exceed the value advertised by a receiver. An endpoint MUST terminate a connection with a QUIC_FLOW_CONTROL_RECEIVED_TOO_MUCH_DATA error if it receives more data than the maximum data value that it has sent, unless this is a result of a change in the initial limits (see Section 7.4.2).¶

The MAX_STREAM_DATA frame (type=0x05) is used in flow control to inform a peer of the maximum amount of data that can be sent on a stream.¶

Stream ID:

The stream ID of the stream that is affected encoded as a variable-length integer.

Maximum Stream Data:

A variable-length integer indicating the maximum amount of data that can be sent on the identified stream, in units of octets.

When counting data toward this limit, an endpoint accounts for the largest received offset of data that is sent or received on the stream. Loss or reordering can mean that the largest received offset on a stream can be greater than the total size of data received on that stream. Receiving STREAM frames might not increase the largest received offset.¶

The data sent on a stream MUST NOT exceed the largest maximum stream data value advertised by the receiver. An endpoint MUST terminate a connection with a FLOW_CONTROL_ERROR error if it receives more data than the largest maximum stream data that it has sent for the affected stream, unless this is a result of a change in the initial limits (see Section 7.4.2).¶

The MAX_STREAM_ID frame (type=0x06) informs the peer of the maximum stream ID that they are permitted to open.¶

Maximum Stream ID:

ID of the maximum unidirectional or bidirectional peer-initiated stream ID for the connection encoded as a variable-length integer. The limit applies to unidirectional steams if the second least signification bit of the stream ID is 1, and applies to bidirectional streams if it is 0.

Loss or reordering can mean that a MAX_STREAM_ID frame can be received which states a lower stream limit than the client has previously received. MAX_STREAM_ID frames which do not increase the maximum stream ID MUST be ignored.¶

A peer MUST NOT initiate a stream with a higher stream ID than the greatest maximum stream ID it has received. An endpoint MUST terminate a connection with a STREAM_ID_ERROR error if a peer initiates a stream with a higher stream ID than it has sent, unless this is a result of a change in the initial limits (see Section 7.4.2).¶

Endpoints can use PING frames (type=0x07) to verify that their peers are still alive or to check reachability to the peer.¶

The PING frame contains a variable-length payload.¶

Length:

This 8-bit value describes the length of the Data field.

Data:

This variable-length field contains arbitrary data.

A PING frame with an empty Data field causes the packet containing it to be acknowledged as normal. No other action is required of the recipient.¶

An empty PING frame can be used to keep a connection alive when an application or application protocol wishes to prevent the connection from timing out. An application protocol SHOULD provide guidance about the conditions under which generating a PING is recommended. This guidance SHOULD indicate whether it is the client or the server that is expected to send the PING. Having both endpoints send PING frames without coordination can produce an excessive number of packets and poor performance.¶

If the Data field is not empty, the recipient of this frame MUST generate a PONG frame (Section 8.15) containing the same Data. A PING frame with data is not appropriate for use in keeping a connection alive, because the PONG frame elicits an acknowledgement, causing the sender of the original PING to send two packets.¶

A connection will time out if no packets are sent or received for a period longer than the time specified in the idle_timeout transport parameter (see Section 7.9). However, state in middleboxes might time out earlier than that. Though REQ-5 in [RFC4787] recommends a 2 minute timeout interval, experience shows that sending packets every 15 to 30 seconds is necessary to prevent the majority of middleboxes from losing state for UDP flows.¶

A sender SHOULD send a BLOCKED frame (type=0x08) when it wishes to send data, but is unable to due to connection-level flow control (see Section 11.2.1). BLOCKED frames can be used as input to tuning of flow control algorithms (see Section 11.1.2).¶

The BLOCKED frame contains a single field.¶

Offset:

A variable-length integer indicating the connection-level offset at which the blocking occurred.

A sender SHOULD send a STREAM_BLOCKED frame (type=0x09) when it wishes to send data, but is unable to due to stream-level flow control. This frame is analogous to BLOCKED (Section 8.10).¶

Stream ID:

A variable-length integer indicating the stream which is flow control blocked.

Offset:

A variable-length integer indicating the offset of the stream at which the blocking occurred.

A sender MAY send a STREAM_ID_BLOCKED frame (type=0x0a) when it wishes to open a stream, but is unable to due to the maximum stream ID limit set by its peer (see Section 8.8). This does not open the stream, but informs the peer that a new stream was needed, but the stream limit prevented the creation of the stream.¶

The STREAM_ID_BLOCKED frame contains a single field.¶

Stream ID:

A variable-length integer indicating the highest stream ID that the sender was permitted to open.

A server sends a NEW_CONNECTION_ID frame (type=0x0b) to provide the client with alternative connection IDs that can be used to break linkability when migrating connections (see Section 7.7.1).¶

Sequence:

A variable-length integer. This value starts at 0 and increases by 1 for each connection ID that is provided by the server. The connection ID that is assigned during the handshake is assumed to have a sequence of -1. That is, the value selected during the handshake comes immediately before the first value that a server can send.

Connection ID:

A 64-bit connection ID.

Stateless Reset Token:

A 128-bit value that will be used to for a stateless reset when the associated connection ID is used (see Section 7.9.4).

An endpoint may use a STOP_SENDING frame (type=0x0c) to communicate that incoming data is being discarded on receipt at application request. This signals a peer to abruptly terminate transmission on a stream.¶

Stream ID:

A variable-length integer carrying the Stream ID of the stream being ignored.

Application Error Code:

A 16-bit, application-specified reason the sender is ignoring the stream (see Section 12.4).

The PONG frame (type=0x0d) is sent in response to a PING frame that contains data. Its format is identical to the PING frame (Section 8.9).¶

An endpoint that receives an unsolicited PONG frame - that is, a PONG frame containing a payload that is empty MUST generate a connection error of type FRAME_ERROR, indicating the PONG frame (that is, 0x10d). If the content of a PONG frame does not match the content of a PING frame previously sent by the endpoint, the endpoint MAY generate a connection error of type UNSOLICITED_PONG.¶

Receivers send ACK frames (type=0xe) to inform senders which packets they have received and processed. A sent packet that has never been acknowledged is missing. The ACK frame contains any number of ACK blocks. ACK blocks are ranges of acknowledged packets.¶

Unlike TCP SACKs, QUIC acknowledgements are irrevocable. Once a packet has been acknowledged, even if it does not appear in a future ACK frame, it remains acknowledged.¶

A client MUST NOT acknowledge Version Negotiation or Retry packets. These packet types contain packet numbers selected by the client, not the server.¶

A sender MAY intentionally skip packet numbers to introduce entropy into the connection, to avoid opportunistic acknowledgement attacks. The sender SHOULD close the connection if an unsent packet number is acknowledged. The format of the ACK frame is efficient at expressing even long blocks of missing packets, allowing for large, unpredictable gaps.¶

An ACK frame is shown below.¶

Largest Acknowledged:

A variable-length integer representing the largest packet number the peer is acknowledging; this is usually the largest packet number that the peer has received prior to generating the ACK frame. Unlike the packet number in the QUIC long or short header, the value in an ACK frame is not truncated.

ACK Delay:

A variable-length integer including the time in microseconds that the largest acknowledged packet, as indicated in the Largest Acknowledged field, was received by this peer to when this ACK was sent. The value of the ACK Delay field is scaled by multiplying the encoded value by the 2 to the power of the value of the ack_delay_exponent transport parameter set by the sender of the ACK frame. The ack_delay_exponent defaults to 3, or a multiplier of 8 (see Section 7.4.1). Scaling in this fashion allows for a larger range of values with a shorter encoding at the cost of lower resolution.

ACK Block Count:

The number of Additional ACK Block (and Gap) fields after the First ACK Block.

ACK Blocks:

Contains one or more blocks of packet numbers which have been successfully received, see Section 8.16.1.

STREAM frames implicitly create a stream and carry stream data. The STREAM frame takes the form 0b00010XXX (or the set of values from 0x10 to 0x17). The value of the three low-order bits of the frame type determine the fields that are present in the frame.¶

• The OFF bit (0x04) in the frame type is set to indicate that there is an Offset field present. When set to 1, the Offset field is present; when set to 0, the Offset field is absent and the Stream Data starts at an offset of 0 (that is, the frame contains the first octets of the stream, or the end of a stream that includes no data).
• The LEN bit (0x02) in the frame type is set to indicate that there is a Length field present. If this bit is set to 0, the Length field is absent and the Stream Data field extends to the end of the packet. If this bit is set to 1, the Length field is present.
• The FIN bit (0x01) of the frame type is set only on frames that contain the final offset of the stream. Setting this bit indicates that the frame marks the end of the stream.

A STREAM frame is shown below.¶

Stream ID:

A variable-length integer indicating the stream ID of the stream (see Section 10.1).

Offset:

A variable-length integer specifying the byte offset in the stream for the data in this STREAM frame. This field is present when the OFF bit is set to 1. When the Offset field is absent, the offset is 0.

Length:

A variable-length integer specifying the length of the Stream Data field in this STREAM frame. This field is present when the LEN bit is set to 1. When the LEN bit is set to 0, the Stream Data field consumes all the remaining octets in the packet.

Stream Data:

The bytes from the designated stream to be delivered.

A stream frame’s Stream Data MUST NOT be empty, unless the FIN bit is set. When the FIN flag is sent on an empty STREAM frame, the offset in the STREAM frame is the offset of the next byte that would be sent.¶

The first byte in the stream has an offset of 0. The largest offset delivered on a stream - the sum of the re-constructed offset and data length - MUST be less than 2^62.¶

Stream multiplexing is achieved by interleaving STREAM frames from multiple streams into one or more QUIC packets. A single QUIC packet can include multiple STREAM frames from one or more streams.¶

Implementation note: One of the benefits of QUIC is avoidance of head-of-line blocking across multiple streams. When a packet loss occurs, only streams with data in that packet are blocked waiting for a retransmission to be received, while other streams can continue making progress. Note that when data from multiple streams is bundled into a single QUIC packet, loss of that packet blocks all those streams from making progress. An implementation is therefore advised to bundle as few streams as necessary in outgoing packets without losing transmission efficiency to underfilled packets.¶

The QUIC packet size includes the QUIC header and integrity check, but not the UDP or IP header.¶

Clients MUST pad any Initial packet it sends to have a QUIC packet size of at least 1200 octets. Sending an Initial packet of this size ensures that the network path supports a reasonably sized packet, and helps reduce the amplitude of amplification attacks caused by server responses toward an unverified client address.¶

An Initial packet MAY exceed 1200 octets if the client knows that the Path Maximum Transmission Unit (PMTU) supports the size that it chooses.¶

A server MAY send a CONNECTION_CLOSE frame with error code PROTOCOL_VIOLATION in response to an Initial packet smaller than 1200 octets. It MUST NOT send any other frame type in response, or otherwise behave as if any part of the offending packet was processed as valid.¶

The Path Maximum Transmission Unit (PMTU) is the maximum size of the entire IP header, UDP header, and UDP payload. The UDP payload includes the QUIC packet header, protected payload, and any authentication fields.¶

All QUIC packets SHOULD be sized to fit within the estimated PMTU to avoid IP fragmentation or packet drops. To optimize bandwidth efficiency, endpoints SHOULD use Packetization Layer PMTU Discovery ([PLPMTUD]). Endpoints MAY use PMTU Discovery ([PMTUDv4], [PMTUDv6]) for detecting the PMTU, setting the PMTU appropriately, and storing the result of previous PMTU determinations.¶

In the absence of these mechanisms, QUIC endpoints SHOULD NOT send IP packets larger than 1280 octets. Assuming the minimum IP header size, this results in a QUIC packet size of 1232 octets for IPv6 and 1252 octets for IPv4. Some QUIC implementations MAY wish to be more conservative in computing allowed QUIC packet size given unknown tunneling overheads or IP header options.¶

QUIC endpoints that implement any kind of PMTU discovery SHOULD maintain an estimate for each combination of local and remote IP addresses. Each pairing of local and remote addresses could have a different maximum MTU in the path.¶

QUIC depends on the network path supporting a MTU of at least 1280 octets. This is the IPv6 minimum MTU and therefore also supported by most modern IPv4 networks. An endpoint MUST NOT reduce its MTU below this number, even if it receives signals that indicate a smaller limit might exist.¶

If a QUIC endpoint determines that the PMTU between any pair of local and remote IP addresses has fallen below 1280 octets, it MUST immediately cease sending QUIC packets on the affected path. This could result in termination of the connection if an alternative path cannot be found.¶

Streams are identified by an unsigned 62-bit integer, referred to as the Stream ID. The least significant two bits of the Stream ID are used to identify the type of stream (unidirectional or bidirectional) and the initiator of the stream.¶

The least significant bit (0x1) of the Stream ID identifies the initiator of the stream. Clients initiate even-numbered streams (those with the least significant bit set to 0); servers initiate odd-numbered streams (with the bit set to 1). Separation of the stream identifiers ensures that client and server are able to open streams without the latency imposed by negotiating for an identifier.¶

If an endpoint receives a frame for a stream that it expects to initiate (i.e., odd-numbered for the client or even-numbered for the server), but which it has not yet opened, it MUST close the connection with error code STREAM_STATE_ERROR.¶

The second least significant bit (0x2) of the Stream ID differentiates between unidirectional streams and bidirectional streams. Unidirectional streams always have this bit set to 1 and bidirectional streams have this bit set to 0.¶

The two type bits from a Stream ID therefore identify streams as summarized in Table 5.¶

Stream ID 0 (0x0) is a client-initiated, bidirectional stream that is used for the cryptographic handshake. Stream 0 MUST NOT be used for application data.¶

A QUIC endpoint MUST NOT reuse a Stream ID. Open streams can be used in any order. Streams that are used out of order result in opening all lower-numbered streams of the same type in the same direction.¶

Stream IDs are encoded as a variable-length integer (see Section 8.1).¶

This section describes the two types of QUIC stream in terms of the states of their send or receive components. Two state machines are described: one for streams on which an endpoint transmits data (Section 10.2.1); another for streams from which an endpoint receives data (Section 10.2.2).¶

Unidirectional streams use the applicable state machine directly. Bidirectional streams use both state machines. For the most part, the use of these state machines is the same whether the stream is unidirectional or bidirectional. The conditions for opening a stream are slightly more complex for a bidirectional stream because the opening of either send or receive causes the stream to open in both directions.¶

Opening a stream causes all lower-numbered streams of the same type to implicitly open. This includes both send and receive streams if the stream is bidirectional. For bidirectional streams, an endpoint can send data on an implicitly opened stream. On both unidirectional and bidirectional streams, an endpoint MAY send MAX_STREAM_DATA or STOP_SENDING on implicitly opened streams. An endpoint SHOULD NOT implicitly open streams that it initiates, instead opening streams in order.¶

Note:

These states are largely informative. This document uses stream states to describe rules for when and how different types of frames can be sent and the reactions that are expected when different types of frames are received. Though these state machines are intended to be useful in implementing QUIC, these states aren’t intended to constrain implementations. An implementation can define a different state machine as long as its behavior is consistent with an implementation that implements these states.

If an endpoint is no longer interested in the data it is receiving on a stream, it MAY send a STOP_SENDING frame identifying that stream to prompt closure of the stream in the opposite direction. This typically indicates that the receiving application is no longer reading data it receives from the stream, but is not a guarantee that incoming data will be ignored.¶

STREAM frames received after sending STOP_SENDING are still counted toward the connection and stream flow-control windows, even though these frames will be discarded upon receipt. This avoids potential ambiguity about which STREAM frames count toward flow control.¶

A STOP_SENDING frame requests that the receiving endpoint send a RST_STREAM frame. An endpoint that receives a STOP_SENDING frame MUST send a RST_STREAM frame for that stream, and can use an error code of STOPPING. If the STOP_SENDING frame is received on a send stream that is already in the “Data Sent” state, a RST_STREAM frame MAY still be sent in order to cancel retransmission of previously-sent STREAM frames.¶

STOP_SENDING SHOULD only be sent for a receive stream that has not been reset. STOP_SENDING is most useful for streams in the “Recv” or “Size Known” states.¶

An endpoint is expected to send another STOP_SENDING frame if a packet containing a previous STOP_SENDING is lost. However, once either all stream data or a RST_STREAM frame has been received for the stream - that is, the stream is in any state other than “Recv” or “Size Known” - sending a STOP_SENDING frame is unnecessary.¶

An endpoint limits the number of concurrently active incoming streams by adjusting the maximum stream ID. An initial value is set in the transport parameters (see Section 7.4.1) and is subsequently increased by MAX_STREAM_ID frames (see Section 8.8).¶

The maximum stream ID is specific to each endpoint and applies only to the peer that receives the setting. That is, clients specify the maximum stream ID the server can initiate, and servers specify the maximum stream ID the client can initiate. Each endpoint may respond on streams initiated by the other peer, regardless of whether it is permitted to initiated new streams.¶

Endpoints MUST NOT exceed the limit set by their peer. An endpoint that receives a STREAM frame with an ID greater than the limit it has sent MUST treat this as a stream error of type STREAM_ID_ERROR (Section 12), unless this is a result of a change in the initial offsets (see Section 7.4.2).¶

A receiver MUST NOT renege on an advertisement; that is, once a receiver advertises a stream ID via a MAX_STREAM_ID frame, it MUST NOT subsequently advertise a smaller maximum ID. A sender may receive MAX_STREAM_ID frames out of order; a sender MUST therefore ignore any MAX_STREAM_ID that does not increase the maximum.¶

Once a stream is created, endpoints may use the stream to send and receive data. Each endpoint may send a series of STREAM frames encapsulating data on a stream until the stream is terminated in that direction. Streams are an ordered byte-stream abstraction, and they have no other structure within them. STREAM frame boundaries are not expected to be preserved in retransmissions from the sender or during delivery to the application at the receiver.¶

When new data is to be sent on a stream, a sender MUST set the encapsulating STREAM frame’s offset field to the stream offset of the first byte of this new data. The first byte of data that is sent on a stream has the stream offset 0. The largest offset delivered on a stream MUST be less than 2^62. A receiver MUST ensure that received stream data is delivered to the application as an ordered byte-stream. Data received out of order MUST be buffered for later delivery, as long as it is not in violation of the receiver’s flow control limits.¶

An endpoint MUST NOT send data on any stream without ensuring that it is within the data limits set by its peer. The cryptographic handshake stream, Stream 0, is exempt from the connection-level data limits established by MAX_DATA. Data on stream 0 other than the initial cryptographic handshake message is still subject to stream-level data limits and MAX_STREAM_DATA. This message is exempt from flow control because it needs to be sent in a single packet regardless of the server’s flow control state. This rule applies even for 0-RTT handshakes where the remembered value of MAX_STREAM_DATA would not permit sending a full initial cryptographic handshake message.¶

Flow control is described in detail in Section 11, and congestion control is described in the companion document [QUIC-RECOVERY].¶

Stream multiplexing has a significant effect on application performance if resources allocated to streams are correctly prioritized. Experience with other multiplexed protocols, such as HTTP/2 [HTTP2], shows that effective prioritization strategies have a significant positive impact on performance.¶

QUIC does not provide frames for exchanging prioritization information. Instead it relies on receiving priority information from the application that uses QUIC. Protocols that use QUIC are able to define any prioritization scheme that suits their application semantics. A protocol might define explicit messages for signaling priority, such as those defined in HTTP/2; it could define rules that allow an endpoint to determine priority based on context; or it could leave the determination to the application.¶

A QUIC implementation SHOULD provide ways in which an application can indicate the relative priority of streams. When deciding which streams to dedicate resources to, QUIC SHOULD use the information provided by the application. Failure to account for priority of streams can result in suboptimal performance.¶

Stream priority is most relevant when deciding which stream data will be transmitted. Often, there will be limits on what can be transmitted as a result of connection flow control or the current congestion controller state.¶

Giving preference to the transmission of its own management frames ensures that the protocol functions efficiently. That is, prioritizing frames other than STREAM frames ensures that loss recovery, congestion control, and flow control operate effectively.¶

Stream 0 MUST be prioritized over other streams prior to the completion of the cryptographic handshake. This includes the retransmission of the second flight of client handshake messages, that is, the TLS Finished and any client authentication messages.¶

STREAM frames that are determined to be lost SHOULD be retransmitted before sending new data, unless application priorities indicate otherwise. Retransmitting lost stream data can fill in gaps, which allows the peer to consume already received data and free up flow control window.¶

There are some edge cases which must be considered when dealing with stream and connection level flow control. Given enough time, both endpoints must agree on flow control state. If one end believes it can send more than the other end is willing to receive, the connection will be torn down when too much data arrives.¶

Conversely if a sender believes it is blocked, while endpoint B expects more data can be received, then the connection can be in a deadlock, with the sender waiting for a MAX_DATA or MAX_STREAM_DATA frame which will never come.¶

On receipt of a RST_STREAM frame, an endpoint will tear down state for the matching stream and ignore further data arriving on that stream. This could result in the endpoints getting out of sync, since the RST_STREAM frame may have arrived out of order and there may be further bytes in flight. The data sender would have counted the data against its connection level flow control budget, but a receiver that has not received these bytes would not know to include them as well. The receiver must learn the number of bytes that were sent on the stream to make the same adjustment in its connection flow controller.¶

To avoid this de-synchronization, a RST_STREAM sender MUST include the final byte offset sent on the stream in the RST_STREAM frame. On receiving a RST_STREAM frame, a receiver definitively knows how many bytes were sent on that stream before the RST_STREAM frame, and the receiver MUST use the final offset to account for all bytes sent on the stream in its connection level flow controller.¶

As with flow control, this document leaves when and how many streams to make available to a peer via MAX_STREAM_ID to implementations, but offers a few considerations. MAX_STREAM_ID frames constitute minimal overhead, while withholding MAX_STREAM_ID frames can prevent the peer from using the available parallelism.¶

Implementations will likely want to increase the maximum stream ID as peer-initiated streams close. A receiver MAY also advance the maximum stream ID based on current activity, system conditions, and other environmental factors.¶

The final offset is the count of the number of octets that are transmitted on a stream. For a stream that is reset, the final offset is carried explicitly in a RST_STREAM frame. Otherwise, the final offset is the offset of the end of the data carried in a STREAM frame marked with a FIN flag, or 0 in the case of incoming unidirectional streams.¶

An endpoint will know the final offset for a stream when the receive stream enters the “Size Known” or “Reset Recvd” state.¶

An endpoint MUST NOT send data on a stream at or beyond the final offset.¶

Once a final offset for a stream is known, it cannot change. If a RST_STREAM or STREAM frame causes the final offset to change for a stream, an endpoint SHOULD respond with a FINAL_OFFSET_ERROR error (see Section 12). A receiver SHOULD treat receipt of data at or beyond the final offset as a FINAL_OFFSET_ERROR error, even after a stream is closed. Generating these errors is not mandatory, but only because requiring that an endpoint generate these errors also means that the endpoint needs to maintain the final offset state for closed streams, which could mean a significant state commitment.¶

Errors that result in the connection being unusable, such as an obvious violation of protocol semantics or corruption of state that affects an entire connection, MUST be signaled using a CONNECTION_CLOSE or APPLICATION_CLOSE frame (Section 8.4, Section 8.5). An endpoint MAY close the connection in this manner even if the error only affects a single stream.¶

Application protocols can signal application-specific protocol errors using the APPLICATION_CLOSE frame. Errors that are specific to the transport, including all those described in this document, are carried in a CONNECTION_CLOSE frame. Other than the type of error code they carry, these frames are identical in format and semantics.¶

A CONNECTION_CLOSE or APPLICATION_CLOSE frame could be sent in a packet that is lost. An endpoint SHOULD be prepared to retransmit a packet containing either frame type if it receives more packets on a terminated connection. Limiting the number of retransmissions and the time over which this final packet is sent limits the effort expended on terminated connections.¶

An endpoint that chooses not to retransmit packets containing CONNECTION_CLOSE or APPLICATION_CLOSE risks a peer missing the first such packet. The only mechanism available to an endpoint that continues to receive data for a terminated connection is to use the stateless reset process (Section 7.9.4).¶

An endpoint that receives an invalid CONNECTION_CLOSE or APPLICATION_CLOSE frame MUST NOT signal the existence of the error to its peer.¶

If the error affects a single stream, but otherwise leaves the connection in a recoverable state, the endpoint can send a RST_STREAM frame (Section 8.3) with an appropriate error code to terminate just the affected stream.¶

Stream 0 is critical to the functioning of the entire connection. If stream 0 is closed with either a RST_STREAM or STREAM frame bearing the FIN flag, an endpoint MUST generate a connection error of type PROTOCOL_VIOLATION.¶

RST_STREAM MUST be instigated by the application and MUST carry an application error code. Resetting a stream without knowledge of the application protocol could cause the protocol to enter an unrecoverable state. Application protocols might require certain streams to be reliably delivered in order to guarantee consistent state between endpoints.¶

QUIC error codes are 16-bit unsigned integers.¶

This section lists the defined QUIC transport error codes that may be used in a CONNECTION_CLOSE frame. These errors apply to the entire connection.¶

NO_ERROR (0x0):

An endpoint uses this with CONNECTION_CLOSE to signal that the connection is being closed abruptly in the absence of any error.

INTERNAL_ERROR (0x1):

The endpoint encountered an internal error and cannot continue with the connection.

FLOW_CONTROL_ERROR (0x3):

An endpoint received more data than it permitted in its advertised data limits (see Section 11).

STREAM_ID_ERROR (0x4):

An endpoint received a frame for a stream identifier that exceeded its advertised maximum stream ID.

STREAM_STATE_ERROR (0x5):

An endpoint received a frame for a stream that was not in a state that permitted that frame (see Section 10.2).

FINAL_OFFSET_ERROR (0x6):

An endpoint received a STREAM frame containing data that exceeded the previously established final offset. Or an endpoint received a RST_STREAM frame containing a final offset that was lower than the maximum offset of data that was already received. Or an endpoint received a RST_STREAM frame containing a different final offset to the one already established.

FRAME_FORMAT_ERROR (0x7):

An endpoint received a frame that was badly formatted. For instance, an empty STREAM frame that omitted the FIN flag, or an ACK frame that has more acknowledgment ranges than the remainder of the packet could carry. This is a generic error code; an endpoint SHOULD use the more specific frame format error codes (0x1XX) if possible.

TRANSPORT_PARAMETER_ERROR (0x8):

An endpoint received transport parameters that were badly formatted, included an invalid value, was absent even though it is mandatory, was present though it is forbidden, or is otherwise in error.

VERSION_NEGOTIATION_ERROR (0x9):

An endpoint received transport parameters that contained version negotiation parameters that disagreed with the version negotiation that it performed. This error code indicates a potential version downgrade attack.

PROTOCOL_VIOLATION (0xA):

An endpoint detected an error with protocol compliance that was not covered by more specific error codes.

UNSOLICITED_PONG (0xB):

An endpoint received a PONG frame that did not correspond to any PING frame that it previously sent.

FRAME_ERROR (0x1XX):

An endpoint detected an error in a specific frame type. The frame type is included as the last octet of the error code. For example, an error in a MAX_STREAM_ID frame would be indicated with the code (0x106).

See Section 14.2 for details of registering new error codes.¶

Application protocol error codes are 16-bit unsigned integers, but the management of application error codes are left to application protocols. Application protocol error codes are used for the RST_STREAM (Section 8.3) and APPLICATION_CLOSE (Section 8.5) frames.¶

There is no restriction on the use of the 16-bit error code space for application protocols. However, QUIC reserves the error code with a value of 0 to mean STOPPING. The application error code of STOPPING (0) is used by the transport to cancel a stream in response to receipt of a STOP_SENDING frame.¶

An attacker might be able to receive an address validation token (Section 7.6) from the server and then release the IP address it used to acquire that token. The attacker may, in the future, spoof this same address (which now presumably addresses a different endpoint), and initiate a 0-RTT connection with a server on the victim’s behalf. The attacker can then spoof ACK frames to the server which cause the server to send excessive amounts of data toward the new owner of the IP address.¶

There are two possible mitigations to this attack. The simplest one is that a server can unilaterally create a gap in packet-number space. In the non-attack scenario, the client will send an ACK frame with the larger value for largest acknowledged. In the attack scenario, the attacker could acknowledge a packet in the gap. If the server sees an acknowledgment for a packet that was never sent, the connection can be aborted.¶

The second mitigation is that the server can require that acknowledgments for sent packets match the encryption level of the sent packet. This mitigation is useful if the connection has an ephemeral forward-secure key that is generated and used for every new connection. If a packet sent is protected with a forward-secure key, then any acknowledgments that are received for them MUST also be forward-secure protected. Since the attacker will not have the forward secure key, the attacker will not be able to generate forward-secure protected packets with ACK frames.¶

The attacks commonly known as Slowloris [SLOWLORIS] try to keep many connections to the target endpoint open and hold them open as long as possible. These attacks can be executed against a QUIC endpoint by generating the minimum amount of activity necessary to avoid being closed for inactivity. This might involve sending small amounts of data, gradually opening flow control windows in order to control the sender rate, or manufacturing ACK frames that simulate a high loss rate.¶

QUIC deployments SHOULD provide mitigations for the Slowloris attacks, such as increasing the maximum number of clients the server will allow, limiting the number of connections a single IP address is allowed to make, imposing restrictions on the minimum transfer speed a connection is allowed to have, and restricting the length of time an endpoint is allowed to stay connected.¶

An adversarial endpoint might intentionally fragment the data on stream buffers in order to cause disproportionate memory commitment. An adversarial endpoint could open a stream and send some STREAM frames containing arbitrary fragments of the stream content.¶

The attack is mitigated if flow control windows correspond to available memory. However, some receivers will over-commit memory and advertise flow control offsets in the aggregate that exceed actual available memory. The over-commitment strategy can lead to better performance when endpoints are well behaved, but renders endpoints vulnerable to the stream fragmentation attack.¶

QUIC deployments SHOULD provide mitigations against the stream fragmentation attack. Mitigations could consist of avoiding over-committing memory, delaying reassembly of STREAM frames, implementing heuristics based on the age and duration of reassembly holes, or some combination.¶

An adversarial endpoint can open lots of streams, exhausting state on an endpoint. The adversarial endpoint could repeat the process on a large number of connections, in a manner similar to SYN flooding attacks in TCP.¶

Normally, clients will open streams sequentially, as explained in Section 10.1. However, when several streams are initiated at short intervals, transmission error may cause STREAM DATA frames opening streams to be received out of sequence. A receiver is obligated to open intervening streams if a higher-numbered stream ID is received. Thus, on a new connection, opening stream 2000001 opens 1 million streams, as required by the specification.¶

The number of active streams is limited by the concurrent stream limit transport parameter, as explained in Section 10.4. If chosen judisciously, this limit mitigates the effect of the stream commitment attack. However, setting the limit too low could affect performance when applications expect to open large number of streams.¶

IANA [SHALL add/has added] a registry for “QUIC Transport Parameters” under a “QUIC Protocol” heading.¶

The “QUIC Transport Parameters” registry governs a 16-bit space. This space is split into two spaces that are governed by different policies. Values with the first byte in the range 0x00 to 0xfe (in hexadecimal) are assigned via the Specification Required policy [RFC8126]. Values with the first byte 0xff are reserved for Private Use [RFC8126].¶

Value:

The numeric value of the assignment (registrations will be between 0x0000 and 0xfeff).

Parameter Name:

A short mnemonic for the parameter.

Specification:

A reference to a publicly available specification for the value.

The nominated expert(s) verify that a specification exists and is readily accessible. The expert(s) are encouraged to be biased towards approving registrations unless they are abusive, frivolous, or actively harmful (not merely aesthetically displeasing, or architecturally dubious).¶

The initial contents of this registry are shown in Table 7.¶

IANA [SHALL add/has added] a registry for “QUIC Transport Error Codes” under a “QUIC Protocol” heading.¶

The “QUIC Transport Error Codes” registry governs a 16-bit space. This space is split into two spaces that are governed by different policies. Values with the first byte in the range 0x00 to 0xfe (in hexadecimal) are assigned via the Specification Required policy [RFC8126]. Values with the first byte 0xff are reserved for Private Use [RFC8126].¶

Value:

The numeric value of the assignment (registrations will be between 0x0000 and 0xfeff).

Code:

A short mnemonic for the parameter.

Description:

A brief description of the error code semantics, which MAY be a summary if a specification reference is provided.

Specification:

A reference to a publicly available specification for the value.

The initial contents of this registry are shown in Table 8. Note that FRAME_ERROR takes the range from 0x100 to 0x1FF and private use occupies the range from 0xFE00 to 0xFFFF.¶

• Clarified requirements for BLOCKED usage (#65, #924)
• BLOCKED frame now includes reason for blocking (#452, #924, #927, #928)
• Cleartext integrity as version independent (#568)
• GAP limitation in ACK Frame (#613)
• Improved PMTUD description (#614, #1036)
• Clarified stream state machine (#634, #662, #894)
• Reserved versions don’t need to be generated deterministically (#831, #931)
• You don’t always need the draining period (#871)
• Stateless reset clarified as version-specific (#930, #986)
• initial_max_stream_id_x transport parameters are optional (#970, #971)
• Ack Delay assumes a default value during the handshake (#1007, #1009)
• Removed transport parameters from NewSessionTicket (#1015)

• The long header now has version before packet number (#926, #939)
• Rename and consolidate packet types (#846, #822, #847)
• Packet types are assigned new codepoints and the Connection ID Flag is inverted (#426, #956)
• Removed type for Version Negotiation and use Version 0 (#963, #968)
• Streams are split into unidirectional and bidirectional (#643, #656, #720, #872, #175, #885)
  • Stream limits now have separate uni- and bi-directinal transport parameters (#909, #958)
  • Stream limit transport parameters are now optional and default to 0 (#970, #971)
• The stream state machine has been split into read and write (#634, #894)
• Employ variable-length integer encodings throughout (#595)
• Improvements to connection close
  • Added distinct closing and draining states (#899, #871)
  • Draining period can terminate early (#869, #870)
  • Clarifications about stateless reset (#889, #890)
• Address validation for connection migration (#161, #732, #878)
• Clearly defined retransmission rules for BLOCKED (#452, #65, #924)
• negotiated_version is sent in server transport parameters (#710, #959)
• Increased the range over which packet numbers are randomized (#864, #850, #964)

• Replaced FNV-1a with AES-GCM for all “Cleartext” packets (#554)
• Split error code space between application and transport (#485)
• Stateless reset token moved to end (#820)
• 1-RTT-protected long header types removed (#848)
• No acknowledgments during draining period (#852)
• Remove “application close” as a separate close type (#854)
• Remove timestamps from the ACK frame (#841)
• Require transport parameters to only appear once (#792)

• Stateless token is server-only (#726)
• Refactor section on connection termination (#733, #748, #328, #177)
• Limit size of Version Negotiation packet (#585)
• Clarify when and what to ack (#736)
• Renamed STREAM_ID_NEEDED to STREAM_ID_BLOCKED
• Clarify Keep-alive requirements (#729)

• Introduce STOP_SENDING frame, RST_STREAM only resets in one direction (#165)
• Removed GOAWAY; application protocols are responsible for graceful shutdown (#696)
• Reduced the number of error codes (#96, #177, #184, #211)
• Version validation fields can’t move or change (#121)
• Removed versions from the transport parameters in a NewSessionTicket message (#547)
• Clarify the meaning of “bytes in flight” (#550)
• Public reset is now stateless reset and not visible to the path (#215)
• Reordered bits and fields in STREAM frame (#620)
• Clarifications to the stream state machine (#572, #571)
• Increased the maximum length of the Largest Acknowledged field in ACK frames to 64 bits (#629)
• truncate_connection_id is renamed to omit_connection_id (#659)
• CONNECTION_CLOSE terminates the connection like TCP RST (#330, #328)
• Update labels used in HKDF-Expand-Label to match TLS 1.3 (#642)

• Change STREAM and RST_STREAM layout
• Add MAX_STREAM_ID settings

• The size of the initial packet payload has a fixed minimum (#267, #472)
• Define when Version Negotiation packets are ignored (#284, #294, #241, #143, #474)
• The 64-bit FNV-1a algorithm is used for integrity protection of unprotected packets (#167, #480, #481, #517)
• Rework initial packet types to change how the connection ID is chosen (#482, #442, #493)
• No timestamps are forbidden in unprotected packets (#542, #429)
• Cryptographic handshake is now on stream 0 (#456)
• Remove congestion control exemption for cryptographic handshake (#248, #476)
• Version 1 of QUIC uses TLS; a new version is needed to use a different handshake protocol (#516)
• STREAM frames have a reduced number of offset lengths (#543, #430)
• Split some frames into separate connection- and stream- level frames (#443)
  • WINDOW_UPDATE split into MAX_DATA and MAX_STREAM_DATA (#450)
  • BLOCKED split to match WINDOW_UPDATE split (#454)
  • Define STREAM_ID_NEEDED frame (#455)
• A NEW_CONNECTION_ID frame supports connection migration without linkability (#232, #491, #496)
• Transport parameters for 0-RTT are retained from a previous connection (#405, #513, #512)
  • A client in 0-RTT no longer required to reset excess streams (#425, #479)
• Expanded security considerations (#440, #444, #445, #448)

• Defined short and long packet headers (#40, #148, #361)
• Defined a versioning scheme and stable fields (#51, #361)
• Define reserved version values for “greasing” negotiation (#112, #278)
• The initial packet number is randomized (#35, #283)
• Narrow the packet number encoding range requirement (#67, #286, #299, #323, #356)
• Defined client address validation (#52, #118, #120, #275)
• Define transport parameters as a TLS extension (#49, #122)
• SCUP and COPT parameters are no longer valid (#116, #117)
• Transport parameters for 0-RTT are either remembered from before, or assume default values (#126)
• The server chooses connection IDs in its final flight (#119, #349, #361)
• The server echoes the Connection ID and packet number fields when sending a Version Negotiation packet (#133, #295, #244)
• Defined a minimum packet size for the initial handshake packet from the client (#69, #136, #139, #164)
• Path MTU Discovery (#64, #106)
• The initial handshake packet from the client needs to fit in a single packet (#338)
• Forbid acknowledgment of packets containing only ACK and PADDING (#291)
• Require that frames are processed when packets are acknowledged (#381, #341)
• Removed the STOP_WAITING frame (#66)
• Don’t require retransmission of old timestamps for lost ACK frames (#308)
• Clarified that frames are not retransmitted, but the information in them can be (#157, #298)
• Error handling definitions (#335)
• Split error codes into four sections (#74)
• Forbid the use of Public Reset where CONNECTION_CLOSE is possible (#289)
• Define packet protection rules (#336)
• Require that stream be entirely delivered or reset, including acknowledgment of all STREAM frames or the RST_STREAM, before it closes (#381)
• Remove stream reservation from state machine (#174, #280)
• Only stream 1 does not contribute to connection-level flow control (#204)
• Stream 1 counts towards the maximum concurrent stream limit (#201, #282)
• Remove connection-level flow control exclusion for some streams (except 1) (#246)
• RST_STREAM affects connection-level flow control (#162, #163)
• Flow control accounting uses the maximum data offset on each stream, rather than bytes received (#378)
• Moved length-determining fields to the start of STREAM and ACK (#168, #277)
• Added the ability to pad between frames (#158, #276)
• Remove error code and reason phrase from GOAWAY (#352, #355)
• GOAWAY includes a final stream number for both directions (#347)
• Error codes for RST_STREAM and CONNECTION_CLOSE are now at a consistent offset (#249)
• Defined priority as the responsibility of the application protocol (#104, #303)

• Replaced DIVERSIFICATION_NONCE flag with KEY_PHASE flag
• Defined versioning
• Reworked description of packet and frame layout
• Error code space is divided into regions for each component
• Use big endian for all numeric values

• Adopted as base for draft-ietf-quic-tls
• Updated authors/editors list
• Added IANA Considerations section
• Moved Contributors and Acknowledgments to appendices