HTTP/3

HTTP/1.1 ([HTTP/1.1]) uses whitespace-delimited text fields to convey HTTP messages. While these exchanges are human readable, using whitespace for message formatting leads to parsing complexity and excessive tolerance of variant behavior.

Because HTTP/1.1 does not include a multiplexing layer, multiple TCP connections are often used to service requests in parallel. However, that has a negative impact on congestion control and network efficiency, since TCP does not share congestion control across multiple connections.

HTTP/2 ([HTTP/2]) introduced a binary framing and multiplexing layer to improve latency without modifying the transport layer. However, because the parallel nature of HTTP/2's multiplexing is not visible to TCP's loss recovery mechanisms, a lost or reordered packet causes all active transactions to experience a stall regardless of whether that transaction was directly impacted by the lost packet.

The QUIC transport protocol incorporates stream multiplexing and per-stream flow control, similar to that provided by the HTTP/2 framing layer. By providing reliability at the stream level and congestion control across the entire connection, QUIC has the capability to improve the performance of HTTP compared to a TCP mapping. QUIC also incorporates TLS 1.3 ([TLS]) at the transport layer, offering comparable confidentiality and integrity to running TLS over TCP, with the improved connection setup latency of TCP Fast Open ([TFO]).

This document defines HTTP/3: a mapping of HTTP semantics over the QUIC transport protocol, drawing heavily on the design of HTTP/2. HTTP/3 relies on QUIC to provide confidentiality and integrity protection of data; peer authentication; and reliable, in-order, per-stream delivery. While delegating stream lifetime and flow-control issues to QUIC, a binary framing similar to the HTTP/2 framing is used on each stream. Some HTTP/2 features are subsumed by QUIC, while other features are implemented atop QUIC.

QUIC is described in [QUIC-TRANSPORT]. For a full description of HTTP/2, see [HTTP/2].

• "Connection Setup and Management" (Section 3) covers how an HTTP/3 endpoint is discovered and an HTTP/3 connection is established.
• "Expressing HTTP Semantics in HTTP/3" (Section 4) describes how HTTP semantics are expressed using frames.
• "Connection Closure" (Section 5) describes how HTTP/3 connections are terminated, either gracefully or abruptly.

• "Stream Mapping and Usage" (Section 6) describes the way QUIC streams are used.
• "HTTP Framing Layer" (Section 7) describes the frames used on most streams.
• "Error Handling" (Section 8) describes how error conditions are handled and expressed, either on a particular stream or for the connection as a whole.

• "Extensions to HTTP/3" (Section 9) describes how new capabilities can be added in future documents.
• A more detailed comparison between HTTP/2 and HTTP/3 can be found in Appendix A.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

This document uses the variable-length integer encoding from [QUIC-TRANSPORT].

abort:

An abrupt termination of a connection or stream, possibly due to an error condition.

client:

The endpoint that initiates an HTTP/3 connection. Clients send HTTP requests and receive HTTP responses.

connection:

A transport-layer connection between two endpoints using QUIC as the transport protocol.

connection error

:

An error that affects the entire HTTP/3 connection.

endpoint:

Either the client or server of the connection.

frame:

The smallest unit of communication on a stream in HTTP/3, consisting of a header and a variable-length sequence of bytes structured according to the frame type.

Protocol elements called "frames" exist in both this document and [QUIC-TRANSPORT]. Where frames from [QUIC-TRANSPORT] are referenced, the frame name will be prefaced with "QUIC". For example, "QUIC CONNECTION_CLOSE frames". References without this preface refer to frames defined in Section 7.2.

HTTP/3 connection:

A QUIC connection where the negotiated application protocol is HTTP/3.

peer:

An endpoint. When discussing a particular endpoint, "peer" refers to the endpoint that is remote to the primary subject of discussion.

receiver:

An endpoint that is receiving frames.

sender:

An endpoint that is transmitting frames.

server:

The endpoint that accepts an HTTP/3 connection. Servers receive HTTP requests and send HTTP responses.

stream:

A bidirectional or unidirectional bytestream provided by the QUIC transport. All streams within an HTTP/3 connection can be considered "HTTP/3 streams", but multiple stream types are defined within HTTP/3.

stream error

:

An application-level error on the individual stream.

The term "content" is defined in Section 6.4 of [HTTP].

Finally, the terms "resource", "message", "user agent", "origin server", "gateway", "intermediary", "proxy", and "tunnel" are defined in Section 3 of [HTTP].

Packet diagrams in this document use the format defined in Section 1.3 of [QUIC-TRANSPORT] to illustrate the order and size of fields.

HTTP relies on the notion of an authoritative response: a response that has been determined to be the most appropriate response for that request given the state of the target resource at the time of response message origination by (or at the direction of) the origin server identified within the target URI. Locating an authoritative server for an HTTP URI is discussed in Section 4.3 of [HTTP].

The "https" scheme associates authority with possession of a certificate that the client considers to be trustworthy for the host identified by the authority component of the URI. Upon receiving a server certificate in the TLS handshake, the client MUST verify that the certificate is an acceptable match for the URI's origin server using the process described in Section 4.3.4 of [HTTP]. If the certificate cannot be verified with respect to the URI's origin server, the client MUST NOT consider the server authoritative for that origin.

A client MAY attempt access to a resource with an "https" URI by resolving the host identifier to an IP address, establishing a QUIC connection to that address on the indicated port (including validation of the server certificate as described above), and sending an HTTP/3 request message targeting the URI to the server over that secured connection. Unless some other mechanism is used to select HTTP/3, the token "h3" is used in the Application-Layer Protocol Negotiation (ALPN; see [RFC7301]) extension during the TLS handshake.

Connectivity problems (e.g., blocking UDP) can result in a failure to establish a QUIC connection; clients SHOULD attempt to use TCP-based versions of HTTP in this case.

Servers MAY serve HTTP/3 on any UDP port; an alternative service advertisement always includes an explicit port, and URIs contain either an explicit port or a default port associated with the scheme.

HTTP/3 relies on QUIC version 1 as the underlying transport. The use of other QUIC transport versions with HTTP/3 MAY be defined by future specifications.

QUIC version 1 uses TLS version 1.3 or greater as its handshake protocol. HTTP/3 clients MUST support a mechanism to indicate the target host to the server during the TLS handshake. If the server is identified by a domain name ([DNS-TERMS]), clients MUST send the Server Name Indication (SNI; [RFC6066]) TLS extension unless an alternative mechanism to indicate the target host is used.

QUIC connections are established as described in [QUIC-TRANSPORT]. During connection establishment, HTTP/3 support is indicated by selecting the ALPN token "h3" in the TLS handshake. Support for other application-layer protocols MAY be offered in the same handshake.

While connection-level options pertaining to the core QUIC protocol are set in the initial crypto handshake, settings specific to HTTP/3 are conveyed in the SETTINGS frame. After the QUIC connection is established, a SETTINGS frame MUST be sent by each endpoint as the initial frame of their respective HTTP control stream.

HTTP/3 connections are persistent across multiple requests. For best performance, it is expected that clients will not close connections until it is determined that no further communication with a server is necessary (for example, when a user navigates away from a particular web page) or until the server closes the connection.

Once a connection to a server endpoint exists, this connection MAY be reused for requests with multiple different URI authority components. To use an existing connection for a new origin, clients MUST validate the certificate presented by the server for the new origin server using the process described in Section 4.3.4 of [HTTP]. This implies that clients will need to retain the server certificate and any additional information needed to verify that certificate; clients that do not do so will be unable to reuse the connection for additional origins.

If the certificate is not acceptable with regard to the new origin for any reason, the connection MUST NOT be reused and a new connection SHOULD be established for the new origin. If the reason the certificate cannot be verified might apply to other origins already associated with the connection, the client SHOULD revalidate the server certificate for those origins. For instance, if validation of a certificate fails because the certificate has expired or been revoked, this might be used to invalidate all other origins for which that certificate was used to establish authority.

Clients SHOULD NOT open more than one HTTP/3 connection to a given IP address and UDP port, where the IP address and port might be derived from a URI, a selected alternative service ([ALTSVC]), a configured proxy, or name resolution of any of these. A client MAY open multiple HTTP/3 connections to the same IP address and UDP port using different transport or TLS configurations but SHOULD avoid creating multiple connections with the same configuration.

Servers are encouraged to maintain open HTTP/3 connections for as long as possible but are permitted to terminate idle connections if necessary. When either endpoint chooses to close the HTTP/3 connection, the terminating endpoint SHOULD first send a GOAWAY frame (Section 5.2) so that both endpoints can reliably determine whether previously sent frames have been processed and gracefully complete or terminate any necessary remaining tasks.

A server that does not wish clients to reuse HTTP/3 connections for a particular origin can indicate that it is not authoritative for a request by sending a 421 (Misdirected Request) status code in response to the request; see Section 7.4 of [HTTP].

A client sends an HTTP request on a request stream, which is a client-initiated bidirectional QUIC stream; see Section 6.1. A client MUST send only a single request on a given stream. A server sends zero or more interim HTTP responses on the same stream as the request, followed by a single final HTTP response, as detailed below. See Section 15 of [HTTP] for a description of interim and final HTTP responses.

Pushed responses are sent on a server-initiated unidirectional QUIC stream; see Section 6.2.2. A server sends zero or more interim HTTP responses, followed by a single final HTTP response, in the same manner as a standard response. Push is described in more detail in Section 4.6.

On a given stream, receipt of multiple requests or receipt of an additional HTTP response following a final HTTP response MUST be treated as malformed.

1. the header section, including message control data, sent as a single HEADERS frame,
2. optionally, the content, if present, sent as a series of DATA frames, and
3. optionally, the trailer section, if present, sent as a single HEADERS frame.

Header and trailer sections are described in Sections 6.3 and 6.5 of [HTTP]; the content is described in Section 6.4 of [HTTP].

Receipt of an invalid sequence of frames MUST be treated as a connection error of type H3_FRAME_UNEXPECTED. In particular, a DATA frame before any HEADERS frame, or a HEADERS or DATA frame after the trailing HEADERS frame, is considered invalid. Other frame types, especially unknown frame types, might be permitted subject to their own rules; see Section 9.

A server MAY send one or more PUSH_PROMISE frames before, after, or interleaved with the frames of a response message. These PUSH_PROMISE frames are not part of the response; see Section 4.6 for more details. PUSH_PROMISE frames are not permitted on push streams; a pushed response that includes PUSH_PROMISE frames MUST be treated as a connection error of type H3_FRAME_UNEXPECTED.

Frames of unknown types (Section 9), including reserved frames (Section 7.2.8) MAY be sent on a request or push stream before, after, or interleaved with other frames described in this section.

The HEADERS and PUSH_PROMISE frames might reference updates to the QPACK dynamic table. While these updates are not directly part of the message exchange, they must be received and processed before the message can be consumed. See Section 4.2 for more details.

Transfer codings (see Section 7 of [HTTP/1.1]) are not defined for HTTP/3; the Transfer-Encoding header field MUST NOT be used.

A response MAY consist of multiple messages when and only when one or more interim responses (1xx; see Section 15.2 of [HTTP]) precede a final response to the same request. Interim responses do not contain content or trailer sections.

An HTTP request/response exchange fully consumes a client-initiated bidirectional QUIC stream. After sending a request, a client MUST close the stream for sending. Unless using the CONNECT method (see Section 4.4), clients MUST NOT make stream closure dependent on receiving a response to their request. After sending a final response, the server MUST close the stream for sending. At this point, the QUIC stream is fully closed.

When a stream is closed, this indicates the end of the final HTTP message. Because some messages are large or unbounded, endpoints SHOULD begin processing partial HTTP messages once enough of the message has been received to make progress. If a client-initiated stream terminates without enough of the HTTP message to provide a complete response, the server SHOULD abort its response stream with the error code H3_REQUEST_INCOMPLETE.

A server can send a complete response prior to the client sending an entire request if the response does not depend on any portion of the request that has not been sent and received. When the server does not need to receive the remainder of the request, it MAY abort reading the request stream, send a complete response, and cleanly close the sending part of the stream. The error code H3_NO_ERROR SHOULD be used when requesting that the client stop sending on the request stream. Clients MUST NOT discard complete responses as a result of having their request terminated abruptly, though clients can always discard responses at their discretion for other reasons. If the server sends a partial or complete response but does not abort reading the request, clients SHOULD continue sending the content of the request and close the stream normally.

HTTP messages carry metadata as a series of key-value pairs called "HTTP fields"; see Sections 6.3 and 6.5 of [HTTP]. For a listing of registered HTTP fields, see the "Hypertext Transfer Protocol (HTTP) Field Name Registry" maintained at <https://www.iana.org/assignments/http-fields/>. Like HTTP/2, HTTP/3 has additional considerations related to the use of characters in field names, the Connection header field, and pseudo-header fields.

Field names are strings containing a subset of ASCII characters. Properties of HTTP field names and values are discussed in more detail in Section 5.1 of [HTTP]. Characters in field names MUST be converted to lowercase prior to their encoding. A request or response containing uppercase characters in field names MUST be treated as malformed.

HTTP/3 does not use the Connection header field to indicate connection-specific fields; in this protocol, connection-specific metadata is conveyed by other means. An endpoint MUST NOT generate an HTTP/3 field section containing connection-specific fields; any message containing connection-specific fields MUST be treated as malformed.

The only exception to this is the TE header field, which MAY be present in an HTTP/3 request header; when it is, it MUST NOT contain any value other than "trailers".

An intermediary transforming an HTTP/1.x message to HTTP/3 MUST remove connection-specific header fields as discussed in Section 7.6.1 of [HTTP], or their messages will be treated by other HTTP/3 endpoints as malformed.

Like HTTP/2, HTTP/3 employs a series of pseudo-header fields, where the field name begins with the : character (ASCII 0x3a). These pseudo-header fields convey message control data; see Section 6.2 of [HTTP].

Pseudo-header fields are not HTTP fields. Endpoints MUST NOT generate pseudo-header fields other than those defined in this document. However, an extension could negotiate a modification of this restriction; see Section 9.

Pseudo-header fields are only valid in the context in which they are defined. Pseudo-header fields defined for requests MUST NOT appear in responses; pseudo-header fields defined for responses MUST NOT appear in requests. Pseudo-header fields MUST NOT appear in trailer sections. Endpoints MUST treat a request or response that contains undefined or invalid pseudo-header fields as malformed.

All pseudo-header fields MUST appear in the header section before regular header fields. Any request or response that contains a pseudo-header field that appears in a header section after a regular header field MUST be treated as malformed.

The CONNECT method requests that the recipient establish a tunnel to the destination origin server identified by the request-target; see Section 9.3.6 of [HTTP]. It is primarily used with HTTP proxies to establish a TLS session with an origin server for the purposes of interacting with "https" resources.

In HTTP/1.x, CONNECT is used to convert an entire HTTP connection into a tunnel to a remote host. In HTTP/2 and HTTP/3, the CONNECT method is used to establish a tunnel over a single stream.

• The :method pseudo-header field is set to "CONNECT"
• The :scheme and :path pseudo-header fields are omitted
• The :authority pseudo-header field contains the host and port to connect to (equivalent to the authority-form of the request-target of CONNECT requests; see Section 7.1 of [HTTP]).

The request stream remains open at the end of the request to carry the data to be transferred. A CONNECT request that does not conform to these restrictions is malformed.

A proxy that supports CONNECT establishes a TCP connection ([RFC0793]) to the server identified in the :authority pseudo-header field. Once this connection is successfully established, the proxy sends a HEADERS frame containing a 2xx series status code to the client, as defined in Section 15.3 of [HTTP].

All DATA frames on the stream correspond to data sent or received on the TCP connection. The payload of any DATA frame sent by the client is transmitted by the proxy to the TCP server; data received from the TCP server is packaged into DATA frames by the proxy. Note that the size and number of TCP segments is not guaranteed to map predictably to the size and number of HTTP DATA or QUIC STREAM frames.

Once the CONNECT method has completed, only DATA frames are permitted to be sent on the stream. Extension frames MAY be used if specifically permitted by the definition of the extension. Receipt of any other known frame type MUST be treated as a connection error of type H3_FRAME_UNEXPECTED.

The TCP connection can be closed by either peer. When the client ends the request stream (that is, the receive stream at the proxy enters the "Data Recvd" state), the proxy will set the FIN bit on its connection to the TCP server. When the proxy receives a packet with the FIN bit set, it will close the send stream that it sends to the client. TCP connections that remain half closed in a single direction are not invalid, but are often handled poorly by servers, so clients SHOULD NOT close a stream for sending while they still expect to receive data from the target of the CONNECT.

A TCP connection error is signaled by abruptly terminating the stream. A proxy treats any error in the TCP connection, which includes receiving a TCP segment with the RST bit set, as a stream error of type H3_CONNECT_ERROR.

Correspondingly, if a proxy detects an error with the stream or the QUIC connection, it MUST close the TCP connection. If the proxy detects that the client has reset the stream or aborted reading from the stream, it MUST close the TCP connection. If the stream is reset or reading is aborted by the client, a proxy SHOULD perform the same operation on the other direction in order to ensure that both directions of the stream are cancelled. In all these cases, if the underlying TCP implementation permits it, the proxy SHOULD send a TCP segment with the RST bit set.

Since CONNECT creates a tunnel to an arbitrary server, proxies that support CONNECT SHOULD restrict its use to a set of known ports or a list of safe request targets; see Section 9.3.6 of [HTTP] for more details.

HTTP/3 does not support the HTTP Upgrade mechanism (Section 7.8 of [HTTP]) or the 101 (Switching Protocols) informational status code (Section 15.2.2 of [HTTP]).

Server push is an interaction mode that permits a server to push a request-response exchange to a client in anticipation of the client making the indicated request. This trades off network usage against a potential latency gain. HTTP/3 server push is similar to what is described in Section 8.2 of [HTTP/2], but it uses different mechanisms.

Each server push is assigned a unique push ID by the server. The push ID is used to refer to the push in various contexts throughout the lifetime of the HTTP/3 connection.

The push ID space begins at zero and ends at a maximum value set by the MAX_PUSH_ID frame. In particular, a server is not able to push until after the client sends a MAX_PUSH_ID frame. A client sends MAX_PUSH_ID frames to control the number of pushes that a server can promise. A server SHOULD use push IDs sequentially, beginning from zero. A client MUST treat receipt of a push stream as a connection error of type H3_ID_ERROR when no MAX_PUSH_ID frame has been sent or when the stream references a push ID that is greater than the maximum push ID.

The push ID is used in one or more PUSH_PROMISE frames that carry the control data and header fields of the request message. These frames are sent on the request stream that generated the push. This allows the server push to be associated with a client request. When the same push ID is promised on multiple request streams, the decompressed request field sections MUST contain the same fields in the same order, and both the name and the value in each field MUST be identical.

The push ID is then included with the push stream that ultimately fulfills those promises. The push stream identifies the push ID of the promise that it fulfills, then contains a response to the promised request as described in Section 4.1.

Finally, the push ID can be used in CANCEL_PUSH frames; see Section 7.2.3. Clients use this frame to indicate they do not wish to receive a promised resource. Servers use this frame to indicate they will not be fulfilling a previous promise.

• cacheable; see Section 9.2.3 of [HTTP]
• safe; see Section 9.2.1 of [HTTP]
• does not include request content or a trailer section

The server MUST include a value in the :authority pseudo-header field for which the server is authoritative. If the client has not yet validated the connection for the origin indicated by the pushed request, it MUST perform the same verification process it would do before sending a request for that origin on the connection; see Section 3.3. If this verification fails, the client MUST NOT consider the server authoritative for that origin.

Clients SHOULD send a CANCEL_PUSH frame upon receipt of a PUSH_PROMISE frame carrying a request that is not cacheable, is not known to be safe, that indicates the presence of request content, or for which it does not consider the server authoritative. Any corresponding responses MUST NOT be used or cached.

Each pushed response is associated with one or more client requests. The push is associated with the request stream on which the PUSH_PROMISE frame was received. The same server push can be associated with additional client requests using a PUSH_PROMISE frame with the same push ID on multiple request streams. These associations do not affect the operation of the protocol, but they MAY be considered by user agents when deciding how to use pushed resources.

Ordering of a PUSH_PROMISE frame in relation to certain parts of the response is important. The server SHOULD send PUSH_PROMISE frames prior to sending HEADERS or DATA frames that reference the promised responses. This reduces the chance that a client requests a resource that will be pushed by the server.

Due to reordering, push stream data can arrive before the corresponding PUSH_PROMISE frame. When a client receives a new push stream with an as-yet-unknown push ID, both the associated client request and the pushed request header fields are unknown. The client can buffer the stream data in expectation of the matching PUSH_PROMISE. The client can use stream flow control (Section 4.1 of [QUIC-TRANSPORT]) to limit the amount of data a server may commit to the pushed stream. Clients SHOULD abort reading and discard data already read from push streams if no corresponding PUSH_PROMISE frame is processed in a reasonable amount of time.

Push stream data can also arrive after a client has cancelled a push. In this case, the client can abort reading the stream with an error code of H3_REQUEST_CANCELLED. This asks the server not to transfer additional data and indicates that it will be discarded upon receipt.

Pushed responses that are cacheable (see Section 3 of [HTTP-CACHING]) can be stored by the client, if it implements an HTTP cache. Pushed responses are considered successfully validated on the origin server (e.g., if the "no-cache" cache response directive is present; see Section 5.2.2.4 of [HTTP-CACHING]) at the time the pushed response is received.

Pushed responses that are not cacheable MUST NOT be stored by any HTTP cache. They MAY be made available to the application separately.

Each QUIC endpoint declares an idle timeout during the handshake. If the QUIC connection remains idle (no packets received) for longer than this duration, the peer will assume that the connection has been closed. HTTP/3 implementations will need to open a new HTTP/3 connection for new requests if the existing connection has been idle for longer than the idle timeout negotiated during the QUIC handshake, and they SHOULD do so if approaching the idle timeout; see Section 10.1 of [QUIC-TRANSPORT].

HTTP clients are expected to request that the transport keep connections open while there are responses outstanding for requests or server pushes, as described in Section 10.1.2 of [QUIC-TRANSPORT]. If the client is not expecting a response from the server, allowing an idle connection to time out is preferred over expending effort maintaining a connection that might not be needed. A gateway MAY maintain connections in anticipation of need rather than incur the latency cost of connection establishment to servers. Servers SHOULD NOT actively keep connections open.

Even when a connection is not idle, either endpoint can decide to stop using the connection and initiate a graceful connection close. Endpoints initiate the graceful shutdown of an HTTP/3 connection by sending a GOAWAY frame. The GOAWAY frame contains an identifier that indicates to the receiver the range of requests or pushes that were or might be processed in this connection. The server sends a client-initiated bidirectional stream ID; the client sends a push ID. Requests or pushes with the indicated identifier or greater are rejected (Section 4.1.1) by the sender of the GOAWAY. This identifier MAY be zero if no requests or pushes were processed.

The information in the GOAWAY frame enables a client and server to agree on which requests or pushes were accepted prior to the shutdown of the HTTP/3 connection. Upon sending a GOAWAY frame, the endpoint SHOULD explicitly cancel (see Sections 4.1.1 and 7.2.3) any requests or pushes that have identifiers greater than or equal to the one indicated, in order to clean up transport state for the affected streams. The endpoint SHOULD continue to do so as more requests or pushes arrive.

Endpoints MUST NOT initiate new requests or promise new pushes on the connection after receipt of a GOAWAY frame from the peer. Clients MAY establish a new connection to send additional requests.

• Upon receipt of a GOAWAY frame, if the client has already sent requests with a stream ID greater than or equal to the identifier contained in the GOAWAY frame, those requests will not be processed. Clients can safely retry unprocessed requests on a different HTTP connection. A client that is unable to retry requests loses all requests that are in flight when the server closes the connection.

  Requests on stream IDs less than the stream ID in a GOAWAY frame from the server might have been processed; their status cannot be known until a response is received, the stream is reset individually, another GOAWAY is received with a lower stream ID than that of the request in question, or the connection terminates.

  Servers MAY reject individual requests on streams below the indicated ID if these requests were not processed.

• If a server receives a GOAWAY frame after having promised pushes with a push ID greater than or equal to the identifier contained in the GOAWAY frame, those pushes will not be accepted.

Servers SHOULD send a GOAWAY frame when the closing of a connection is known in advance, even if the advance notice is small, so that the remote peer can know whether or not a request has been partially processed. For example, if an HTTP client sends a POST at the same time that a server closes a QUIC connection, the client cannot know if the server started to process that POST request if the server does not send a GOAWAY frame to indicate what streams it might have acted on.

An endpoint MAY send multiple GOAWAY frames indicating different identifiers, but the identifier in each frame MUST NOT be greater than the identifier in any previous frame, since clients might already have retried unprocessed requests on another HTTP connection. Receiving a GOAWAY containing a larger identifier than previously received MUST be treated as a connection error of type H3_ID_ERROR.

An endpoint that is attempting to gracefully shut down a connection can send a GOAWAY frame with a value set to the maximum possible value (2⁶²-4 for servers, 2⁶²-1 for clients). This ensures that the peer stops creating new requests or pushes. After allowing time for any in-flight requests or pushes to arrive, the endpoint can send another GOAWAY frame indicating which requests or pushes it might accept before the end of the connection. This ensures that a connection can be cleanly shut down without losing requests.

A client has more flexibility in the value it chooses for the Push ID field in a GOAWAY that it sends. A value of 2⁶²-1 indicates that the server can continue fulfilling pushes that have already been promised. A smaller value indicates the client will reject pushes with push IDs greater than or equal to this value. Like the server, the client MAY send subsequent GOAWAY frames so long as the specified push ID is no greater than any previously sent value.

Even when a GOAWAY indicates that a given request or push will not be processed or accepted upon receipt, the underlying transport resources still exist. The endpoint that initiated these requests can cancel them to clean up transport state.

Once all accepted requests and pushes have been processed, the endpoint can permit the connection to become idle, or it MAY initiate an immediate closure of the connection. An endpoint that completes a graceful shutdown SHOULD use the H3_NO_ERROR error code when closing the connection.

If a client has consumed all available bidirectional stream IDs with requests, the server need not send a GOAWAY frame, since the client is unable to make further requests.

An HTTP/3 implementation can immediately close the QUIC connection at any time. This results in sending a QUIC CONNECTION_CLOSE frame to the peer indicating that the application layer has terminated the connection. The application error code in this frame indicates to the peer why the connection is being closed. See Section 8 for error codes that can be used when closing a connection in HTTP/3.

Before closing the connection, a GOAWAY frame MAY be sent to allow the client to retry some requests. Including the GOAWAY frame in the same packet as the QUIC CONNECTION_CLOSE frame improves the chances of the frame being received by clients.

If there are open streams that have not been explicitly closed, they are implicitly closed when the connection is closed; see Section 10.2 of [QUIC-TRANSPORT].

For various reasons, the QUIC transport could indicate to the application layer that the connection has terminated. This might be due to an explicit closure by the peer, a transport-level error, or a change in network topology that interrupts connectivity.

If a connection terminates without a GOAWAY frame, clients MUST assume that any request that was sent, whether in whole or in part, might have been processed.

All client-initiated bidirectional streams are used for HTTP requests and responses. A bidirectional stream ensures that the response can be readily correlated with the request. These streams are referred to as request streams.

This means that the client's first request occurs on QUIC stream 0, with subsequent requests on streams 4, 8, and so on. In order to permit these streams to open, an HTTP/3 server SHOULD configure non-zero minimum values for the number of permitted streams and the initial stream flow-control window. So as to not unnecessarily limit parallelism, at least 100 request streams SHOULD be permitted at a time.

HTTP/3 does not use server-initiated bidirectional streams, though an extension could define a use for these streams. Clients MUST treat receipt of a server-initiated bidirectional stream as a connection error of type H3_STREAM_CREATION_ERROR unless such an extension has been negotiated.

Unidirectional streams, in either direction, are used for a range of purposes. The purpose is indicated by a stream type, which is sent as a variable-length integer at the start of the stream. The format and structure of data that follows this integer is determined by the stream type.

Two stream types are defined in this document: control streams (Section 6.2.1) and push streams (Section 6.2.2). [QPACK] defines two additional stream types. Other stream types can be defined by extensions to HTTP/3; see Section 9 for more details. Some stream types are reserved (Section 6.2.3).

The performance of HTTP/3 connections in the early phase of their lifetime is sensitive to the creation and exchange of data on unidirectional streams. Endpoints that excessively restrict the number of streams or the flow-control window of these streams will increase the chance that the remote peer reaches the limit early and becomes blocked. In particular, implementations should consider that remote peers may wish to exercise reserved stream behavior (Section 6.2.3) with some of the unidirectional streams they are permitted to use.

Each endpoint needs to create at least one unidirectional stream for the HTTP control stream. QPACK requires two additional unidirectional streams, and other extensions might require further streams. Therefore, the transport parameters sent by both clients and servers MUST allow the peer to create at least three unidirectional streams. These transport parameters SHOULD also provide at least 1,024 bytes of flow-control credit to each unidirectional stream.

Note that an endpoint is not required to grant additional credits to create more unidirectional streams if its peer consumes all the initial credits before creating the critical unidirectional streams. Endpoints SHOULD create the HTTP control stream as well as the unidirectional streams required by mandatory extensions (such as the QPACK encoder and decoder streams) first, and then create additional streams as allowed by their peer.

If the stream header indicates a stream type that is not supported by the recipient, the remainder of the stream cannot be consumed as the semantics are unknown. Recipients of unknown stream types MUST either abort reading of the stream or discard incoming data without further processing. If reading is aborted, the recipient SHOULD use the H3_STREAM_CREATION_ERROR error code or a reserved error code (Section 8.1). The recipient MUST NOT consider unknown stream types to be a connection error of any kind.

As certain stream types can affect connection state, a recipient SHOULD NOT discard data from incoming unidirectional streams prior to reading the stream type.

Implementations MAY send stream types before knowing whether the peer supports them. However, stream types that could modify the state or semantics of existing protocol components, including QPACK or other extensions, MUST NOT be sent until the peer is known to support them.

A sender can close or reset a unidirectional stream unless otherwise specified. A receiver MUST tolerate unidirectional streams being closed or reset prior to the reception of the unidirectional stream header.

Type:

A variable-length integer that identifies the frame type.

Length:

A variable-length integer that describes the length in bytes of the Frame Payload.

Frame Payload:

A payload, the semantics of which are determined by the Type field.

Each frame's payload MUST contain exactly the fields identified in its description. A frame payload that contains additional bytes after the identified fields or a frame payload that terminates before the end of the identified fields MUST be treated as a connection error of type H3_FRAME_ERROR. In particular, redundant length encodings MUST be verified to be self-consistent; see Section 10.8.

When a stream terminates cleanly, if the last frame on the stream was truncated, this MUST be treated as a connection error of type H3_FRAME_ERROR. Streams that terminate abruptly may be reset at any point in a frame.

The following error codes are defined for use when abruptly terminating streams, aborting reading of streams, or immediately closing HTTP/3 connections.

H3_NO_ERROR (0x0100)

:

No error. This is used when the connection or stream needs to be closed, but there is no error to signal.

H3_GENERAL_PROTOCOL_ERROR (0x0101)

:

Peer violated protocol requirements in a way that does not match a more specific error code or endpoint declines to use the more specific error code.

H3_INTERNAL_ERROR (0x0102)

:

An internal error has occurred in the HTTP stack.

H3_STREAM_CREATION_ERROR (0x0103)

:

The endpoint detected that its peer created a stream that it will not accept.

H3_CLOSED_CRITICAL_STREAM (0x0104)

:

A stream required by the HTTP/3 connection was closed or reset.

H3_FRAME_UNEXPECTED (0x0105)

:

A frame was received that was not permitted in the current state or on the current stream.

H3_FRAME_ERROR (0x0106)

:

A frame that fails to satisfy layout requirements or with an invalid size was received.

H3_EXCESSIVE_LOAD (0x0107)

:

The endpoint detected that its peer is exhibiting a behavior that might be generating excessive load.

H3_ID_ERROR (0x0108)

:

A stream ID or push ID was used incorrectly, such as exceeding a limit, reducing a limit, or being reused.

H3_SETTINGS_ERROR (0x0109)

:

An endpoint detected an error in the payload of a SETTINGS frame.

H3_MISSING_SETTINGS (0x010a)

:

No SETTINGS frame was received at the beginning of the control stream.

H3_REQUEST_REJECTED (0x010b)

:

A server rejected a request without performing any application processing.

H3_REQUEST_CANCELLED (0x010c)

:

The request or its response (including pushed response) is cancelled.

H3_REQUEST_INCOMPLETE (0x010d)

:

The client's stream terminated without containing a fully formed request.

H3_MESSAGE_ERROR (0x010e)

:

An HTTP message was malformed and cannot be processed.

H3_CONNECT_ERROR (0x010f)

:

The TCP connection established in response to a CONNECT request was reset or abnormally closed.

H3_VERSION_FALLBACK (0x0110)

:

The requested operation cannot be served over HTTP/3. The peer should retry over HTTP/1.1.

Error codes of the format 0x1f * N + 0x21 for non-negative integer values of N are reserved to exercise the requirement that unknown error codes be treated as equivalent to H3_NO_ERROR (Section 9). Implementations SHOULD select an error code from this space with some probability when they would have sent H3_NO_ERROR.

HTTP/3 relies on the HTTP definition of authority. The security considerations of establishing authority are discussed in Section 17.1 of [HTTP].

The use of ALPN in the TLS and QUIC handshakes establishes the target application protocol before application-layer bytes are processed. This ensures that endpoints have strong assurances that peers are using the same protocol.

This does not guarantee protection from all cross-protocol attacks. Section 21.5 of [QUIC-TRANSPORT] describes some ways in which the plaintext of QUIC packets can be used to perform request forgery against endpoints that don't use authenticated transports.

The HTTP/3 field encoding allows the expression of names that are not valid field names in the syntax used by HTTP (Section 5.1 of [HTTP]). Requests or responses containing invalid field names MUST be treated as malformed. Therefore, an intermediary cannot translate an HTTP/3 request or response containing an invalid field name into an HTTP/1.1 message.

Similarly, HTTP/3 can transport field values that are not valid. While most values that can be encoded will not alter field parsing, carriage return (ASCII 0x0d), line feed (ASCII 0x0a), and the null character (ASCII 0x00) might be exploited by an attacker if they are translated verbatim. Any request or response that contains a character not permitted in a field value MUST be treated as malformed. Valid characters are defined by the "field-content" ABNF rule in Section 5.5 of [HTTP].

Pushed responses do not have an explicit request from the client; the request is provided by the server in the PUSH_PROMISE frame.

Caching responses that are pushed is possible based on the guidance provided by the origin server in the Cache-Control header field. However, this can cause issues if a single server hosts more than one tenant. For example, a server might offer multiple users each a small portion of its URI space.

Where multiple tenants share space on the same server, that server MUST ensure that tenants are not able to push representations of resources that they do not have authority over. Failure to enforce this would allow a tenant to provide a representation that would be served out of cache, overriding the actual representation that the authoritative tenant provides.

Clients are required to reject pushed responses for which an origin server is not authoritative; see Section 4.6.

An HTTP/3 connection can demand a greater commitment of resources to operate than an HTTP/1.1 or HTTP/2 connection. The use of field compression and flow control depend on a commitment of resources for storing a greater amount of state. Settings for these features ensure that memory commitments for these features are strictly bounded.

The number of PUSH_PROMISE frames is constrained in a similar fashion. A client that accepts server push SHOULD limit the number of push IDs it issues at a time.

Processing capacity cannot be guarded as effectively as state capacity.

The ability to send undefined protocol elements that the peer is required to ignore can be abused to cause a peer to expend additional processing time. This might be done by setting multiple undefined SETTINGS parameters, unknown frame types, or unknown stream types. Note, however, that some uses are entirely legitimate, such as optional-to-understand extensions and padding to increase resistance to traffic analysis.

Compression of field sections also offers some opportunities to waste processing resources; see Section 7 of [QPACK] for more details on potential abuses.

All these features -- i.e., server push, unknown protocol elements, field compression -- have legitimate uses. These features become a burden only when they are used unnecessarily or to excess.

An endpoint that does not monitor such behavior exposes itself to a risk of denial-of-service attack. Implementations SHOULD track the use of these features and set limits on their use. An endpoint MAY treat activity that is suspicious as a connection error of type H3_EXCESSIVE_LOAD, but false positives will result in disrupting valid connections and requests.

Compression can allow an attacker to recover secret data when it is compressed in the same context as data under attacker control. HTTP/3 enables compression of fields (Section 4.2); the following concerns also apply to the use of HTTP compressed content-codings; see Section 8.4.1 of [HTTP].

There are demonstrable attacks on compression that exploit the characteristics of the web (e.g., [BREACH]). The attacker induces multiple requests containing varying plaintext, observing the length of the resulting ciphertext in each, which reveals a shorter length when a guess about the secret is correct.

Implementations communicating on a secure channel MUST NOT compress content that includes both confidential and attacker-controlled data unless separate compression contexts are used for each source of data. Compression MUST NOT be used if the source of data cannot be reliably determined.

Further considerations regarding the compression of field sections are described in [QPACK].

Padding can be used to obscure the exact size of frame content and is provided to mitigate specific attacks within HTTP, for example, attacks where compressed content includes both attacker-controlled plaintext and secret data (e.g., [BREACH]).

Where HTTP/2 employs PADDING frames and Padding fields in other frames to make a connection more resistant to traffic analysis, HTTP/3 can either rely on transport-layer padding or employ the reserved frame and stream types discussed in Sections 7.2.8 and 6.2.3. These methods of padding produce different results in terms of the granularity of padding, how padding is arranged in relation to the information that is being protected, whether padding is applied in the case of packet loss, and how an implementation might control padding.

Reserved stream types can be used to give the appearance of sending traffic even when the connection is idle. Because HTTP traffic often occurs in bursts, apparent traffic can be used to obscure the timing or duration of such bursts, even to the point of appearing to send a constant stream of data. However, as such traffic is still flow controlled by the receiver, a failure to promptly drain such streams and provide additional flow-control credit can limit the sender's ability to send real traffic.

To mitigate attacks that rely on compression, disabling or limiting compression might be preferable to padding as a countermeasure.

Use of padding can result in less protection than might seem immediately obvious. Redundant padding could even be counterproductive. At best, padding only makes it more difficult for an attacker to infer length information by increasing the number of frames an attacker has to observe. Incorrectly implemented padding schemes can be easily defeated. In particular, randomized padding with a predictable distribution provides very little protection; similarly, padding payloads to a fixed size exposes information as payload sizes cross the fixed-sized boundary, which could be possible if an attacker can control plaintext.

Several protocol elements contain nested length elements, typically in the form of frames with an explicit length containing variable-length integers. This could pose a security risk to an incautious implementer. An implementation MUST ensure that the length of a frame exactly matches the length of the fields it contains.

The use of 0-RTT with HTTP/3 creates an exposure to replay attack. The anti-replay mitigations in [HTTP-REPLAY] MUST be applied when using HTTP/3 with 0-RTT. When applying [HTTP-REPLAY] to HTTP/3, references to the TLS layer refer to the handshake performed within QUIC, while all references to application data refer to the contents of streams.

Certain HTTP implementations use the client address for logging or access-control purposes. Since a QUIC client's address might change during a connection (and future versions might support simultaneous use of multiple addresses), such implementations will need to either actively retrieve the client's current address or addresses when they are relevant or explicitly accept that the original address might change.

Several characteristics of HTTP/3 provide an observer an opportunity to correlate actions of a single client or server over time. These include the value of settings, the timing of reactions to stimulus, and the handling of any features that are controlled by settings.

As far as these create observable differences in behavior, they could be used as a basis for fingerprinting a specific client.

HTTP/3's preference for using a single QUIC connection allows correlation of a user's activity on a site. Reusing connections for different origins allows for correlation of activity across those origins.

Several features of QUIC solicit immediate responses and can be used by an endpoint to measure latency to their peer; this might have privacy implications in certain scenarios.

This document creates a new registration for the identification of HTTP/3 in the "TLS Application-Layer Protocol Negotiation (ALPN) Protocol IDs" registry established in [RFC7301].

Protocol:

HTTP/3

Identification Sequence:

0x68 0x33 ("h3")

Specification:

This document

New registries created in this document operate under the QUIC registration policy documented in Section 22.1 of [QUIC-TRANSPORT]. These registries all include the common set of fields listed in Section 22.1.1 of [QUIC-TRANSPORT]. These registries are collected under the "Hypertext Transfer Protocol version 3 (HTTP/3)" heading.

The initial allocations in these registries are all assigned permanent status and list a change controller of the IETF and a contact of the HTTP working group (ietf-http-wg@w3.org).

HTTP/3 permits use of a larger number of streams (2⁶²-1) than HTTP/2. The same considerations about exhaustion of stream identifier space apply, though the space is significantly larger such that it is likely that other limits in QUIC are reached first, such as the limit on the connection flow-control window.

In contrast to HTTP/2, stream concurrency in HTTP/3 is managed by QUIC. QUIC considers a stream closed when all data has been received and sent data has been acknowledged by the peer. HTTP/2 considers a stream closed when the frame containing the END_STREAM bit has been committed to the transport. As a result, the stream for an equivalent exchange could remain "active" for a longer period of time. HTTP/3 servers might choose to permit a larger number of concurrent client-initiated bidirectional streams to achieve equivalent concurrency to HTTP/2, depending on the expected usage patterns.

In HTTP/2, only request and response bodies (the frame payload of DATA frames) are subject to flow control. All HTTP/3 frames are sent on QUIC streams, so all frames on all streams are flow controlled in HTTP/3.

Due to the presence of other unidirectional stream types, HTTP/3 does not rely exclusively on the number of concurrent unidirectional streams to control the number of concurrent in-flight pushes. Instead, HTTP/3 clients use the MAX_PUSH_ID frame to control the number of pushes received from an HTTP/3 server.

Many framing concepts from HTTP/2 can be elided on QUIC, because the transport deals with them. Because frames are already on a stream, they can omit the stream number. Because frames do not block multiplexing (QUIC's multiplexing occurs below this layer), the support for variable-maximum-length packets can be removed. Because stream termination is handled by QUIC, an END_STREAM flag is not required. This permits the removal of the Flags field from the generic frame layout.

Frame payloads are largely drawn from [HTTP/2]. However, QUIC includes many features (e.g., flow control) that are also present in HTTP/2. In these cases, the HTTP mapping does not re-implement them. As a result, several HTTP/2 frame types are not required in HTTP/3. Where an HTTP/2-defined frame is no longer used, the frame ID has been reserved in order to maximize portability between HTTP/2 and HTTP/3 implementations. However, even frame types that appear in both mappings do not have identical semantics.

Many of the differences arise from the fact that HTTP/2 provides an absolute ordering between frames across all streams, while QUIC provides this guarantee on each stream only. As a result, if a frame type makes assumptions that frames from different streams will still be received in the order sent, HTTP/3 will break them.

Some examples of feature adaptations are described below, as well as general guidance to extension frame implementors converting an HTTP/2 extension to HTTP/3.

An important difference from HTTP/2 is that settings are sent once, as the first frame of the control stream, and thereafter cannot change. This eliminates many corner cases around synchronization of changes.

Some transport-level options that HTTP/2 specifies via the SETTINGS frame are superseded by QUIC transport parameters in HTTP/3. The HTTP-level setting that is retained in HTTP/3 has the same value as in HTTP/2. The superseded settings are reserved, and their receipt is an error. See Section 7.2.4.1 for discussion of both the retained and reserved values.

SETTINGS_HEADER_TABLE_SIZE (0x01):

See [QPACK].

SETTINGS_ENABLE_PUSH (0x02):

This is removed in favor of the MAX_PUSH_ID frame, which provides a more granular control over server push. Specifying a setting with the identifier 0x02 (corresponding to the SETTINGS_ENABLE_PUSH parameter) in the HTTP/3 SETTINGS frame is an error.

SETTINGS_MAX_CONCURRENT_STREAMS (0x03):

QUIC controls the largest open stream ID as part of its flow-control logic. Specifying a setting with the identifier 0x03 (corresponding to the SETTINGS_MAX_CONCURRENT_STREAMS parameter) in the HTTP/3 SETTINGS frame is an error.

SETTINGS_INITIAL_WINDOW_SIZE (0x04):

QUIC requires both stream and connection flow-control window sizes to be specified in the initial transport handshake. Specifying a setting with the identifier 0x04 (corresponding to the SETTINGS_INITIAL_WINDOW_SIZE parameter) in the HTTP/3 SETTINGS frame is an error.

SETTINGS_MAX_FRAME_SIZE (0x05):

This setting has no equivalent in HTTP/3. Specifying a setting with the identifier 0x05 (corresponding to the SETTINGS_MAX_FRAME_SIZE parameter) in the HTTP/3 SETTINGS frame is an error.

SETTINGS_MAX_HEADER_LIST_SIZE (0x06):

This setting identifier has been renamed SETTINGS_MAX_FIELD_SECTION_SIZE.

In HTTP/3, setting values are variable-length integers (6, 14, 30, or 62 bits long) rather than fixed-length 32-bit fields as in HTTP/2. This will often produce a shorter encoding, but can produce a longer encoding for settings that use the full 32-bit space. Settings ported from HTTP/2 might choose to redefine their value to limit it to 30 bits for more efficient encoding or to make use of the 62-bit space if more than 30 bits are required.

Settings need to be defined separately for HTTP/2 and HTTP/3. The IDs of settings defined in [HTTP/2] have been reserved for simplicity. Note that the settings identifier space in HTTP/3 is substantially larger (62 bits versus 16 bits), so many HTTP/3 settings have no equivalent HTTP/2 code point. See Section 11.2.2.

As QUIC streams might arrive out of order, endpoints are advised not to wait for the peers' settings to arrive before responding to other streams. See Section 7.2.4.2.

QUIC has the same concepts of "stream" and "connection" errors that HTTP/2 provides. However, the differences between HTTP/2 and HTTP/3 mean that error codes are not directly portable between versions.

NO_ERROR (0x00):

H3_NO_ERROR in Section 8.1.

PROTOCOL_ERROR (0x01):

This is mapped to H3_GENERAL_PROTOCOL_ERROR except in cases where more specific error codes have been defined. Such cases include H3_FRAME_UNEXPECTED, H3_MESSAGE_ERROR, and H3_CLOSED_CRITICAL_STREAM defined in Section 8.1.

INTERNAL_ERROR (0x02):

H3_INTERNAL_ERROR in Section 8.1.

FLOW_CONTROL_ERROR (0x03):

Not applicable, since QUIC handles flow control.

SETTINGS_TIMEOUT (0x04):

Not applicable, since no acknowledgment of SETTINGS is defined.

STREAM_CLOSED (0x05):

Not applicable, since QUIC handles stream management.

FRAME_SIZE_ERROR (0x06):

H3_FRAME_ERROR error code defined in Section 8.1.

REFUSED_STREAM (0x07):

H3_REQUEST_REJECTED (in Section 8.1) is used to indicate that a request was not processed. Otherwise, not applicable because QUIC handles stream management.

CANCEL (0x08):

H3_REQUEST_CANCELLED in Section 8.1.

COMPRESSION_ERROR (0x09):

Multiple error codes are defined in [QPACK].

CONNECT_ERROR (0x0a):

H3_CONNECT_ERROR in Section 8.1.

ENHANCE_YOUR_CALM (0x0b):

H3_EXCESSIVE_LOAD in Section 8.1.

INADEQUATE_SECURITY (0x0c):

Not applicable, since QUIC is assumed to provide sufficient security on all connections.

HTTP_1_1_REQUIRED (0x0d):

H3_VERSION_FALLBACK in Section 8.1.

Error codes need to be defined for HTTP/2 and HTTP/3 separately. See Section 11.2.3.