| Internet-Draft | HTTP/3 | June 2020 |
| Bishop | Expires 11 December 2020 | [Page] |
The QUIC transport protocol has several features that are desirable in atransport for HTTP, such as stream multiplexing, per-stream flow control, andlow-latency connection establishment. This document describes a mapping of HTTPsemantics over QUIC. This document also identifies HTTP/2 features that aresubsumed by QUIC, and describes how HTTP/2 extensions can be ported to HTTP/3.¶
Discussion of this draft takes place on the QUIC working group mailing list(quic@ietf.org), which is archived athttps://mailarchive.ietf.org/arch/search/?email_list=quic.¶
Working Group information can be found athttps://github.com/quicwg; sourcecode and issues list for this draft can be found athttps://github.com/quicwg/base-drafts/labels/-http.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is athttps://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 11 December 2020.¶
Copyright (c) 2020 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.¶
HTTP semantics[SEMANTICS] are used for a broadrange of services on the Internet. These semantics have most commonly been usedwith two different TCP mappings, HTTP/1.1 and HTTP/2. HTTP/3 supports the samesemantics over a new transport protocol, QUIC.¶
HTTP/1.1[HTTP11] is a TCP mapping which useswhitespace-delimited text fields to convey HTTP messages. While these exchangesare human-readable, using whitespace for message formatting leads to parsingcomplexity and motivates tolerance of variant behavior. Because eachconnection can transfer only a single HTTP request or response at a time in eachdirection, multiple parallel TCP connections are often used, reducing theability of the congestion controller to effectively manage traffic betweenendpoints.¶
HTTP/2[HTTP2] introduced a binary framing and multiplexing layer toimprove latency without modifying the transport layer. However, because theparallel nature of HTTP/2's multiplexing is not visible to TCP's loss recoverymechanisms, a lost or reordered packet causes all active transactions toexperience a stall regardless of whether that transaction was directly impactedby the lost packet.¶
The QUIC transport protocol incorporates stream multiplexing and per-stream flowcontrol, similar to that provided by the HTTP/2 framing layer. By providingreliability at the stream level and congestion control across the entireconnection, it has the capability to improve the performance of HTTP compared toa TCP mapping. QUIC also incorporates TLS 1.3[TLS13] at thetransport layer, offering comparable security to running TLS over TCP, withthe improved connection setup latency of TCP Fast Open[TFO].¶
This document defines a mapping of HTTP semantics over the QUIC transportprotocol, drawing heavily on the design of HTTP/2. While delegating streamlifetime and flow control issues to QUIC, a similar binary framing is used oneach stream. Some HTTP/2 features are subsumed by QUIC, while other features areimplemented atop QUIC.¶
QUIC is described in[QUIC-TRANSPORT]. For a full description of HTTP/2, see[HTTP2].¶
HTTP/3 provides a transport for HTTP semantics using the QUIC transport protocoland an internal framing layer similar to HTTP/2.¶
Once a client knows that an HTTP/3 server exists at a certain endpoint, it opensa QUIC connection. QUIC provides protocol negotiation, stream-basedmultiplexing, and flow control. Discovery of an HTTP/3 endpoint is described inSection 3.2.¶
Within each stream, the basic unit of HTTP/3 communication is a frame(Section 7.2). Each frame type serves a different purpose. For example, HEADERSand DATA frames form the basis of HTTP requests and responses(Section 4.1).¶
Multiplexing of requests is performed using the QUIC stream abstraction,described in Section 2 of[QUIC-TRANSPORT]. Each request-response pairconsumes a single QUIC stream. Streams are independent of each other, so onestream that is blocked or suffers packet loss does not prevent progress on otherstreams.¶
Server push is an interaction mode introduced in HTTP/2[HTTP2] which permitsa server to push a request-response exchange to a client in anticipation of theclient making the indicated request. This trades off network usage against apotential latency gain. Several HTTP/3 frames are used to manage server push,such as PUSH_PROMISE, MAX_PUSH_ID, and CANCEL_PUSH.¶
As in HTTP/2, request and response fields are compressed for transmission.Because HPACK[HPACK] relies on in-order transmission of compressedfield sections (a guarantee not provided by QUIC), HTTP/3 replaces HPACK withQPACK[QPACK]. QPACK uses separate unidirectional streams to modify and trackfield table state, while encoded field sections refer to the state of the tablewithout modifying it.¶
The following sections provide a detailed overview of the connection lifecycleand key concepts:¶
The details of the wire protocol and interactions with the transport aredescribed in subsequent sections:¶
Additional resources are provided in the final sections:¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALLNOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED","MAY", and "OPTIONAL" in this document are to be interpreted asdescribed in BCP 14[RFC2119][RFC8174] when, and only when, theyappear in all capitals, as shown here.¶
Field definitions are given in Augmented Backus-Naur Form (ABNF), as defined in[RFC5234].¶
This document uses the variable-length integer encoding from[QUIC-TRANSPORT].¶
The following terms are used:¶
The term "payload body" is defined in Section 6.3.3 of[SEMANTICS].¶
Finally, the terms "gateway", "intermediary", "proxy", and "tunnel" are definedin Section 2.2 of[SEMANTICS]. Intermediaries act as both client and serverat different times.¶
HTTP/3 uses the token "h3" to identify itself in ALPN and Alt-Svc. Onlyimplementations of the final, published RFC can identify themselves as "h3".Until such an RFC exists, implementations MUST NOT identify themselves usingthis string.¶
Implementations of draft versions of the protocol MUST add the string "-" andthe corresponding draft number to the identifier. For example,draft-ietf-quic-http-01 is identified using the string "h3-01".¶
Draft versions MUST use the corresponding draft transport version as theirtransport. For example, the application protocol defined indraft-ietf-quic-http-25 uses the transport defined indraft-ietf-quic-transport-25.¶
Non-compatible experiments that are based on these draft versions MUST appendthe string "-" and an experiment name to the identifier. For example, anexperimental implementation based on draft-ietf-quic-http-09 which reserves anextra stream for unsolicited transmission of 1980s pop music might identifyitself as "h3-09-rickroll". Note that any label MUST conform to the "token"syntax defined in Section 4.4.1.1 of[SEMANTICS]. Experimenters areencouraged to coordinate their experiments on thequic@ietf.org mailing list.¶
HTTP relies on the notion of an authoritative response: a response that has beendetermined to be the most appropriate response for that request given the stateof the target resource at the time of response message origination by (or at thedirection of) the origin server identified within the target URI. Locating anauthoritative server for an HTTP URL is discussed in Section 5.4 of[SEMANTICS].¶
The "https" scheme associates authority with possession of a certificate thatthe client considers to be trustworthy for the host identified by the authoritycomponent of the URL. If a server presents a certificate and proof that itcontrols the corresponding private key, then a client will accept a securedconnection to that server as being authoritative for all origins with the"https" scheme and a host identified in the certificate.¶
A client MAY attempt access to a resource with an "https" URI by resolving thehost identifier to an IP address, establishing a QUIC connection to that addresson the indicated port, and sending an HTTP/3 request message targeting the URIto the server over that secured connection.¶
Connectivity problems (e.g., blocking UDP) can result in QUIC connectionestablishment failure; clients SHOULD attempt to use TCP-based versions of HTTPin this case.¶
Servers MAY serve HTTP/3 on any UDP port; an alternative service advertisementalways includes an explicit port, and URLs contain either an explicit port or adefault port associated with the scheme.¶
An HTTP origin advertises the availability of an equivalent HTTP/3 endpoint viathe Alt-Svc HTTP response header field or the HTTP/2 ALTSVC frame ([ALTSVC]),using the ALPN token defined inSection 3.3.¶
For example, an origin could indicate in an HTTP response that HTTP/3 wasavailable on UDP port 50781 at the same hostname by including the followingheader field:¶
Alt-Svc: h3=":50781"¶
On receipt of an Alt-Svc record indicating HTTP/3 support, a client MAY attemptto establish a QUIC connection to the indicated host and port and, ifsuccessful, send HTTP requests using the mapping described in this document.¶
Although HTTP is independent of the transport protocol, the "http" schemeassociates authority with the ability to receive TCP connections on theindicated port of whatever host is identified within the authority component.Because HTTP/3 does not use TCP, HTTP/3 cannot be used for direct access to theauthoritative server for a resource identified by an "http" URI. However,protocol extensions such as[ALTSVC] permit the authoritative serverto identify other services which are also authoritative and which might bereachable over HTTP/3.¶
Prior to making requests for an origin whose scheme is not "https", the clientMUST ensure the server is willing to serve that scheme. If the client intendsto make requests for an origin whose scheme is "http", this means that it MUSTobtain a validhttp-opportunistic response for the origin as described in[RFC8164] prior to making any such requests. Other schemes might defineother mechanisms.¶
HTTP/3 relies on QUIC version 1 as the underlying transport. The use of otherQUIC 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 theserver during the TLS handshake. If the server is identified by a DNS name,clients MUST send the Server Name Indication (SNI)[RFC6066] TLS extensionunless an alternative mechanism to indicate the target host is used.¶
QUIC connections are established as described in[QUIC-TRANSPORT]. Duringconnection establishment, HTTP/3 support is indicated by selecting the ALPNtoken "h3" in the TLS handshake. Support for other application-layer protocolsMAY be offered in the same handshake.¶
While connection-level options pertaining to the core QUIC protocol are set inthe initial crypto handshake, HTTP/3-specific settings are conveyed in theSETTINGS frame. After the QUIC connection is established, a SETTINGS frame(Section 7.2.4) MUST be sent by each endpoint as the initial frame of theirrespective HTTP control stream; seeSection 6.2.1.¶
HTTP/3 connections are persistent across multiple requests. For bestperformance, it is expected that clients will not close connections until it isdetermined that no further communication with a server is necessary (forexample, when a user navigates away from a particular web page) or until theserver closes the connection.¶
Once a connection exists to a server endpoint, this connection MAY be reused forrequests with multiple different URI authority components. In general, a serveris considered authoritative for all URIs with the "https" scheme for which thehostname in the URI is present in the authenticated certificate provided by theserver, either as the CN field of the certificate subject or as a dNSName in thesubjectAltName field of the certificate; see[RFC6125]. For a host that isan IP address, the client MUST verify that the address appears as an iPAddressin the subjectAltName field of the certificate. If the hostname or address isnot present in the certificate, the client MUST NOT consider the serverauthoritative for origins containing that hostname or address. See Section 5.4of[SEMANTICS] for more detail on authoritative access.¶
Clients SHOULD NOT open more than one HTTP/3 connection to a given host and portpair, where the host is derived from a URI, a selected alternative service[ALTSVC], or a configured proxy. A client MAY open multiple connections tothe same IP address and UDP port using different transport or TLS configurationsbut SHOULD avoid creating multiple connections with the same configuration.¶
Servers are encouraged to maintain open connections for as long as possible butare permitted to terminate idle connections if necessary. When either endpointchooses to close the HTTP/3 session, the terminating endpoint SHOULD first senda GOAWAY frame (Section 5.2) so that both endpoints can reliablydetermine whether previously sent frames have been processed and gracefullycomplete or terminate any necessary remaining tasks.¶
A server that does not wish clients to reuse connections for a particular origincan indicate that it is not authoritative for a request by sending a 421(Misdirected Request) status code in response to the request; see Section 9.1.2of[HTTP2].¶
A client sends an HTTP request on a client-initiated bidirectional QUIC stream.A client MUST send only a single request on a given stream. A server sends zeroor more interim HTTP responses on the same stream as the request, followed by asingle final HTTP response, as detailed below.¶
Pushed responses are sent on a server-initiated unidirectional QUIC stream; seeSection 6.2.2. A server sends zero or more interim HTTP responses, followedby a single final HTTP response, in the same manner as a standard response.Push is described in more detail inSection 4.4.¶
On a given stream, receipt of multiple requests or receipt of an additional HTTPresponse following a final HTTP response MUST be treated as malformed(Section 4.1.3).¶
An HTTP message (request or response) consists of:¶
Receipt of an invalid sequence of frames MUST be treated as a connection errorof type H3_FRAME_UNEXPECTED (Section 8). In particular, a DATA frame beforeany HEADERS frame, or a HEADERS or DATA frame after the trailing HEADERS frameis considered invalid.¶
A server MAY send one or more PUSH_PROMISE frames (seeSection 7.2.5)before, after, or interleaved with the frames of a response message. ThesePUSH_PROMISE frames are not part of the response; seeSection 4.4 for moredetails. These frames are not permitted in pushed responses; a pushed responsewhich includes PUSH_PROMISE frames MUST be treated as a connection error of typeH3_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, orinterleaved with other frames described in this section.¶
The HEADERS and PUSH_PROMISE frames might reference updates to the QPACK dynamictable. While these updates are not directly part of the message exchange, theymust be received and processed before the message can be consumed. SeeSection 4.1.1 for more details.¶
The "chunked" transfer encoding defined in Section 7.1 of[HTTP11] MUST NOTbe used.¶
A response MAY consist of multiple messages when and only when one or moreinformational responses (1xx; see Section 9.2 of[SEMANTICS]) precede a finalresponse to the same request. Interim responses do not contain a payload bodyor trailers.¶
An HTTP request/response exchange fully consumes a client-initiatedbidirectional QUIC stream. After sending a request, a client MUST close thestream for sending. Unless using the CONNECT method (seeSection 4.2), clientsMUST 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. Atthis point, the QUIC stream is fully closed.¶
When a stream is closed, this indicates the end of an HTTP message. Because somemessages are large or unbounded, endpoints SHOULD begin processing partial HTTPmessages once enough of the message has been received to make progress. If aclient stream terminates without enough of the HTTP message to provide acomplete response, the server SHOULD abort its response with the error codeH3_REQUEST_INCOMPLETE.¶
A server can send a complete response prior to the client sending an entirerequest if the response does not depend on any portion of the request that hasnot been sent and received. When the server does not need to receive theremainder of the request, it MAY abort reading the request stream, send acomplete response, and cleanly close the sending part of the stream. The errorcode H3_NO_ERROR SHOULD be used when requesting that the client stop sending onthe request stream. Clients MUST NOT discard complete responses as a result ofhaving their request terminated abruptly, though clients can always discardresponses at their discretion for other reasons. If the server sends a partialor complete response but does not abort reading, clients SHOULD continue sendingthe body of the request and close the stream normally.¶
HTTP messages carry metadata as a series of key-value pairs, called HTTP fields.For a listing of registered HTTP fields, see the "Hypertext Transfer Protocol(HTTP) Field Name Registry" maintained athttps://www.iana.org/assignments/http-fields/.¶
As in previous versions of HTTP, field names are strings containing a subset ofASCII characters that are compared in a case-insensitive fashion. Properties ofHTTP field names and values are discussed in more detail in Section 4.3 of[SEMANTICS]. As in HTTP/2, characters in field names MUST be converted tolowercase prior to their encoding. A request or response containing uppercasecharacters in field names MUST be treated as malformed (Section 4.1.3).¶
Like HTTP/2, HTTP/3 does not use the Connection header field to indicateconnection-specific fields; in this protocol, connection-specific metadata isconveyed by other means. An endpoint MUST NOT generate an HTTP/3 field sectioncontaining connection-specific fields; any message containingconnection-specific fields MUST be treated as malformed (Section 4.1.3).¶
The only exception to this is the TE header field, which MAY be present in anHTTP/3 request header; when it is, it MUST NOT contain any value other than"trailers".¶
This means that an intermediary transforming an HTTP/1.x message to HTTP/3 willneed to remove any fields nominated by the Connection field, along with theConnection field itself. Such intermediaries SHOULD also remove otherconnection-specific fields, such as Keep-Alive, Proxy-Connection,Transfer-Encoding, and Upgrade, even if they are not nominated by the Connectionfield.¶
Like HTTP/2, HTTP/3 employs a series of pseudo-header fields where the fieldname begins with the ':' character (ASCII 0x3a). These pseudo-header fieldsconvey the target URI, the method of the request, and the status code for theresponse.¶
Pseudo-header fields are not HTTP fields. Endpoints MUST NOT generatepseudo-header fields other than those defined in this document, except asnegotiated via an extension; seeSection 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 trailers. Endpoints MUST treat arequest or response that contains undefined or invalid pseudo-header fields asmalformed (Section 4.1.3).¶
All pseudo-header fields MUST appear in the header field section before regularheader fields. Any request or response that contains a pseudo-header field thatappears in a header field section after a regular header field MUST be treatedas malformed (Section 4.1.3).¶
The following pseudo-header fields are defined for requests:¶
All HTTP/3 requests MUST include exactly one value for the ":method", ":scheme",and ":path" pseudo-header fields, unless it is a CONNECT request; seeSection 4.2.¶
If the ":scheme" pseudo-header field identifies a scheme which has a mandatoryauthority component (including "http" and "https"), the request MUST containeither an ":authority" pseudo-header field or a "Host" header field. If thesefields are present, they MUST NOT be empty. If both fields are present, theyMUST contain the same value. If the scheme does not have a mandatory authoritycomponent and none is provided in the request target, the request MUST NOTcontain the ":authority" pseudo-header and "Host" header fields.¶
An HTTP request that omits mandatory pseudo-header fields or contains invalidvalues for those pseudo-header fields is malformed (Section 4.1.3).¶
HTTP/3 does not define a way to carry the version identifier that is included inthe HTTP/1.1 request line.¶
For responses, a single ":status" pseudo-header field is defined that carriesthe HTTP status code; see Section 9 of[SEMANTICS]. This pseudo-headerfield MUST be included in all responses; otherwise, the response is malformed(Section 4.1.3).¶
HTTP/3 does not define a way to carry the version or reason phrase that isincluded in an HTTP/1.1 status line.¶
HTTP/3 uses QPACK field compression as described in[QPACK], a variation ofHPACK which allows the flexibility to avoid compression-induced head-of-lineblocking. See that document for additional details.¶
To allow for better compression efficiency, the "Cookie" field[RFC6265] MAYbe split into separate field lines, each with one or more cookie-pairs, beforecompression. If a decompressed field section contains multiple cookie fieldlines, these MUST be concatenated into a single octet string using the two-octetdelimiter of 0x3B, 0x20 (the ASCII string "; ") before being passed into acontext other than HTTP/2 or HTTP/3, such as an HTTP/1.1 connection, or ageneric HTTP server application.¶
An HTTP/3 implementation MAY impose a limit on the maximum size of the messageheader it will accept on an individual HTTP message. A server that receives alarger header section than it is willing to handle can send an HTTP 431 (RequestHeader Fields Too Large) status code ([RFC6585]). A client can discardresponses that it cannot process. The size of a field list is calculated basedon the uncompressed size of fields, including the length of the name and valuein bytes plus an overhead of 32 bytes for each field.¶
If an implementation wishes to advise its peer of this limit, it can be conveyedas a number of bytes in the SETTINGS_MAX_FIELD_SECTION_SIZE parameter. Animplementation which has received this parameter SHOULD NOT send an HTTP messageheader which exceeds the indicated size, as the peer will likely refuse toprocess it. However, because this limit is applied at each hop, messages belowthis limit are not guaranteed to be accepted.¶
Clients can cancel requests by resetting and aborting the request stream with anerror code of H3_REQUEST_CANCELLED (Section 8.1). When the clientaborts reading a response, it indicates that this response is no longer ofinterest. Implementations SHOULD cancel requests by abruptly terminating anydirections of a stream that are still open.¶
When the server rejects a request without performing any application processing,it SHOULD abort its response stream with the error code H3_REQUEST_REJECTED.In this context, "processed" means that some data from the stream was passed tosome higher layer of software that might have taken some action as a result. Theclient can treat requests rejected by the server as though they had never beensent at all, thereby allowing them to be retried later on a new connection.Servers MUST NOT use the H3_REQUEST_REJECTED error code for requests whichwere partially or fully processed. When a server abandons a response afterpartial processing, it SHOULD abort its response stream with the error codeH3_REQUEST_CANCELLED.¶
When a client resets a request with the error code H3_REQUEST_CANCELLED, aserver MAY abruptly terminate the response using the error codeH3_REQUEST_REJECTED if no processing was performed. Clients MUST NOT use theH3_REQUEST_REJECTED error code, except when a server has requested closure ofthe request stream with this error code.¶
If a stream is cancelled after receiving a complete response, the client MAYignore the cancellation and use the response. However, if a stream is cancelledafter receiving a partial response, the response SHOULD NOT be used.Automatically retrying such requests is not possible, unless this is otherwisepermitted (e.g., idempotent actions like GET, PUT, or DELETE).¶
A malformed request or response is one that is an otherwise valid sequence offrames but is invalid due to:¶
A request or response that includes a payload body can include aContent-Length header field. A request or response is also malformed if thevalue of a content-length header field does not equal the sum of the DATA framepayload lengths that form the body. A response that is defined to have nopayload, as described in Section 6.3.3 of[SEMANTICS] can have a non-zerocontent-length field, even though no content is included in DATA frames.¶
Intermediaries that process HTTP requests or responses (i.e., any intermediarynot acting as a tunnel) MUST NOT forward a malformed request or response.Malformed requests or responses that are detected MUST be treated as a streamerror (Section 8) of type H3_GENERAL_PROTOCOL_ERROR.¶
For malformed requests, a server MAY send an HTTP response prior to closing orresetting the stream. Clients MUST NOT accept a malformed response. Note thatthese requirements are intended to protect against several types of commonattacks against HTTP; they are deliberately strict because being permissive canexpose implementations to these vulnerabilities.¶
The CONNECT method requests that the recipient establish a tunnel to thedestination origin server identified by the request-target (Section 3.2 of[HTTP11]). It is primarily used with HTTP proxies to establish a TLSsession 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 tunnelto a remote host. In HTTP/2 and HTTP/3, the CONNECT method is used to establisha tunnel over a single stream.¶
A CONNECT request MUST be constructed as follows:¶
The request stream remains open at the end of the request to carry the data tobe transferred. A CONNECT request that does not conform to these restrictionsis malformed; seeSection 4.1.3.¶
A proxy that supports CONNECT establishes a TCP connection ([RFC0793]) to theserver identified in the ":authority" pseudo-header field. Once this connectionis successfully established, the proxy sends a HEADERS frame containing a 2xxseries status code to the client, as defined in Section 9.3 of[SEMANTICS].¶
All DATA frames on the stream correspond to data sent or received on the TCPconnection. Any DATA frame sent by the client is transmitted by the proxy to theTCP server; data received from the TCP server is packaged into DATA frames bythe proxy. Note that the size and number of TCP segments is not guaranteed tomap predictably to the size and number of HTTP DATA or QUIC STREAM frames.¶
Once the CONNECT method has completed, only DATA frames are permittedto be sent on the stream. Extension frames MAY be used if specificallypermitted by the definition of the extension. Receipt of any other frame typeMUST be treated as a connection error of type H3_FRAME_UNEXPECTED.¶
The TCP connection can be closed by either peer. When the client ends therequest 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. Whenthe proxy receives a packet with the FIN bit set, it will terminate the sendstream that it sends to the client. TCP connections which remain half-closed ina single direction are not invalid, but are often handled poorly by servers, soclients SHOULD NOT close a stream for sending while they still expect to receivedata from the target of the CONNECT.¶
A TCP connection error is signaled by abruptly terminating the stream. A proxytreats any error in the TCP connection, which includes receiving a TCP segmentwith the RST bit set, as a stream error of type H3_CONNECT_ERROR(Section 8.1). Correspondingly, if a proxy detects an error with thestream or the QUIC connection, it MUST close the TCP connection. If theunderlying TCP implementation permits it, the proxy SHOULD send a TCP segmentwith the RST bit set.¶
HTTP/3 does not support the HTTP Upgrade mechanism (Section 9.9 of[HTTP11])or 101 (Switching Protocols) informational status code (Section 9.2.2 of[SEMANTICS]).¶
Server push is an interaction mode which permits a server to push arequest-response exchange to a client in anticipation of the client making theindicated request. This trades off network usage against a potential latencygain. HTTP/3 server push is similar to what is described in HTTP/2[HTTP2],but uses different mechanisms.¶
Each server push is identified by a unique Push ID. This Push ID is used in oneor more PUSH_PROMISE frames (seeSection 7.2.5) that carry the requestfields, then included with the push stream which ultimately fulfills thosepromises. When the same Push ID is promised on multiple request streams, thedecompressed request field sections MUST contain the same fields in thesame order, and both the name and the value in each field MUST be exactmatches.¶
Server push is only enabled on a connection when a client sends a MAX_PUSH_IDframe; seeSection 7.2.7. A server cannot use server push until itreceives a MAX_PUSH_ID frame. A client sends additional MAX_PUSH_ID frames tocontrol the number of pushes that a server can promise. A server SHOULD use PushIDs sequentially, starting at 0. A client MUST treat receipt of a push streamwith a Push ID that is greater than the maximum Push ID as a connection error oftype H3_ID_ERROR.¶
The header section of the request message is carried by a PUSH_PROMISE frame(seeSection 7.2.5) on the request stream which generated the push.This allows the server push to be associated with a client request.¶
Not all requests can be pushed. A server MAY push requests which have thefollowing properties:¶
The server MUST include a value in the ":authority" pseudo-header field forwhich the server is authoritative; seeSection 3.4.¶
Clients SHOULD send a CANCEL_PUSH frame upon receipt of a PUSH_PROMISE framecarrying a request which is not cacheable, is not known to be safe, thatindicates the presence of a request body, or for which it does not consider theserver authoritative.¶
Each pushed response is associated with one or more client requests. The pushis associated with the request stream on which the PUSH_PROMISE frame wasreceived. The same server push can be associated with additional clientrequests using a PUSH_PROMISE frame with the same Push ID on multiple requeststreams. These associations do not affect the operation of the protocol, butMAY be considered by user agents when deciding how to use pushed resources.¶
Ordering of a PUSH_PROMISE in relation to certain parts of the response isimportant. The server SHOULD send PUSH_PROMISE frames prior to sending HEADERSor DATA frames that reference the promised responses. This reduces the chancethat a client requests a resource that will be pushed by the server.¶
When a server later fulfills a promise, the server push response is conveyed ona push stream; seeSection 6.2.2. The push stream identifies the Push ID ofthe promise that it fulfills, then contains a response to the promised requestusing the same format described for responses inSection 4.1.¶
Due to reordering, push stream data can arrive before the correspondingPUSH_PROMISE frame. When a client receives a new push stream with anas-yet-unknown Push ID, both the associated client request and the pushedrequest header fields are unknown. The client can buffer the stream data inexpectation of the matching PUSH_PROMISE. The client can use stream flow control(see section 4.1 of[QUIC-TRANSPORT]) to limit the amount of data a server maycommit to the pushed stream.¶
If a promised server push is not needed by the client, the client SHOULD send aCANCEL_PUSH frame. If the push stream is already open or opens after sending theCANCEL_PUSH frame, the client can abort reading the stream with an error code ofH3_REQUEST_CANCELLED. This asks the server not to transfer additional data andindicates that it will be discarded upon receipt.¶
Pushed responses that are cacheable (see Section 3 of[CACHING]) can be stored by the client, if itimplements an HTTP cache. Pushed responses are considered successfullyvalidated on the origin server (e.g., if the "no-cache" cache response directiveis present (Section 5.2.2.3 of[CACHING])) at the time the pushed response isreceived.¶
Pushed responses that are not cacheable MUST NOT be stored by any HTTP cache.They MAY be made available to the application separately.¶
Once established, an HTTP/3 connection can be used for many requests andresponses over time until the connection is closed. Connection closure canhappen in any of several different ways.¶
Each QUIC endpoint declares an idle timeout during the handshake. If theconnection remains idle (no packets received) for longer than this duration, thepeer will assume that the connection has been closed. HTTP/3 implementationswill need to open a new connection for new requests if the existing connectionhas been idle for longer than the server's advertised idle timeout, and SHOULDdo so if approaching the idle timeout.¶
HTTP clients are expected to request that the transport keep connections openwhile there are responses outstanding for requests or server pushes, asdescribed in Section 10.2.2 of[QUIC-TRANSPORT]. If the client is notexpecting a response from the server, allowing an idle connection to time out ispreferred over expending effort maintaining a connection that might not beneeded. A gateway MAY maintain connections in anticipation of need rather thanincur the latency cost of connection establishment to servers. Servers SHOULDNOT actively keep connections open.¶
Even when a connection is not idle, either endpoint can decide to stop using theconnection and initiate a graceful connection close. Endpoints initiate thegraceful shutdown of a connection by sending a GOAWAY frame (Section 7.2.6).The GOAWAY frame contains an identifier that indicates to the receiver the rangeof requests or pushes that were or might be processed in this connection. Theserver sends a client-initiated bidirectional Stream ID; the client sends a PushID. Requests or pushes with the indicated identifier or greater are rejected bythe sender of the GOAWAY. This identifier MAY be zero if no requests or pusheswere processed.¶
The information in the GOAWAY frame enables a client and server to agree onwhich requests or pushes were accepted prior to the connection shutdown. Uponsending a GOAWAY frame, the endpoint SHOULD explicitly cancel (seeSection 4.1.2 andSection 7.2.3) any requests or pushes thathave identifiers greater than or equal to that indicated, in order to clean uptransport state for the affected streams. The endpoint SHOULD continue to do soas more requests or pushes arrive.¶
Endpoints MUST NOT initiate new requests or promise new pushes on the connectionafter receipt of a GOAWAY frame from the peer. Clients MAY establish a newconnection to send additional requests.¶
Some requests or pushes might already be in transit:¶
Upon receipt of a GOAWAY frame, if the client has already sent requests witha Stream ID greater than or equal to the identifier received in a GOAWAYframe, those requests will not be processed. Clients can safely retryunprocessed requests on a different connection.¶
Requests on Stream IDs less than the Stream ID in a GOAWAY frame from theserver might have been processed; their status cannot be known until aresponse is received, the stream is reset individually, another GOAWAY isreceived, or the connection terminates.¶
Servers MAY reject individual requests on streams below the indicated ID ifthese requests were not processed.¶
Servers SHOULD send a GOAWAY frame when the closing of a connection is knownin advance, even if the advance notice is small, so that the remote peer canknow whether a request has been partially processed or not. For example, if anHTTP client sends a POST at the same time that a server closes a QUICconnection, the client cannot know if the server started to process that POSTrequest if the server does not send a GOAWAY frame to indicate what streams itmight have acted on.¶
A client that is unable to retry requests loses all requests that are in flightwhen the server closes the connection. An endpoint MAY send multiple GOAWAYframes indicating different identifiers, but the identifier in each frame MUSTNOT be greater than the identifier in any previous frame, since clients mightalready have retried unprocessed requests on another connection. Receiving aGOAWAY containing a larger identifier than previously received MUST be treatedas a connection error of type H3_ID_ERROR.¶
An endpoint that is attempting to gracefully shut down a connection can send aGOAWAY frame with a value set to the maximum possible value (2^62-4 for servers,2^62-1 for clients). This ensures that the peer stops creating new requests orpushes. After allowing time for any in-flight requests or pushes to arrive, theendpoint can send another GOAWAY frame indicating which requests or pushes itmight accept before the end of the connection. This ensures that a connectioncan be cleanly shut down without losing requests.¶
A client has more flexibility in the value it chooses for the Push ID in aGOAWAY that it sends. A value of 2^62 - 1 indicates that the server cancontinue fulfilling pushes which have already been promised, and the client cancontinue granting push credit as needed; seeSection 7.2.7. A smallervalue indicates the client will reject pushes with Push IDs greater than orequal to this value. Like the server, the client MAY send subsequent GOAWAYframes so long as the specified Push ID is strictly smaller than all previouslysent values.¶
Even when a GOAWAY indicates that a given request or push will not be processedor accepted upon receipt, the underlying transport resources still exist. Theendpoint that initiated these requests can cancel them to clean up transportstate.¶
Once all accepted requests and pushes have been processed, the endpoint canpermit the connection to become idle, or MAY initiate an immediate closure ofthe connection. An endpoint that completes a graceful shutdown SHOULD use theH3_NO_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 makefurther 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 indicatingthat the application layer has terminated the connection. The application errorcode in this frame indicates to the peer why the connection is being closed.SeeSection 8 for error codes which can be used when closing a connection inHTTP/3.¶
Before closing the connection, a GOAWAY frame MAY be sent to allow the client toretry some requests. Including the GOAWAY frame in the same packet as the QUICCONNECTION_CLOSE frame improves the chances of the frame being received byclients.¶
For various reasons, the QUIC transport could indicate to the application layerthat the connection has terminated. This might be due to an explicit closureby the peer, a transport-level error, or a change in network topology whichinterrupts connectivity.¶
If a connection terminates without a GOAWAY frame, clients MUST assume that anyrequest which was sent, whether in whole or in part, might have been processed.¶
A QUIC stream provides reliable in-order delivery of bytes, but makes noguarantees about order of delivery with regard to bytes on other streams. On thewire, data is framed into QUIC STREAM frames, but this framing is invisible tothe HTTP framing layer. The transport layer buffers and orders received QUICSTREAM frames, exposing the data contained within as a reliable byte stream tothe application. Although QUIC permits out-of-order delivery within a stream,HTTP/3 does not make use of this feature.¶
QUIC streams can be either unidirectional, carrying data only from initiator toreceiver, or bidirectional. Streams can be initiated by either the client orthe server. For more detail on QUIC streams, see Section 2 of[QUIC-TRANSPORT].¶
When HTTP fields and data are sent over QUIC, the QUIC layer handles most ofthe stream management. HTTP does not need to do any separate multiplexing whenusing QUIC - data sent over a QUIC stream always maps to a particular HTTPtransaction or connection context.¶
All client-initiated bidirectional streams are used for HTTP requests andresponses. A bidirectional stream ensures that the response can be readilycorrelated with the request. This means that the client's first request occurson QUIC stream 0, with subsequent requests on stream 4, 8, and so on. In orderto permit these streams to open, an HTTP/3 server SHOULD configure non-zerominimum values for the number of permitted streams and the initial stream flowcontrol window. So as to not unnecessarily limit parallelism, at least 100requests SHOULD be permitted at a time.¶
HTTP/3 does not use server-initiated bidirectional streams, though an extensioncould define a use for these streams. Clients MUST treat receipt of aserver-initiated bidirectional stream as a connection error of typeH3_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-lengthinteger at the start of the stream. The format and structure of data thatfollows this integer is determined by the stream type.¶
Unidirectional Stream Header { Stream Type (i),}Some stream types are reserved (Section 6.2.3). Two stream types aredefined in this document: control streams (Section 6.2.1) and push streams(Section 6.2.2).[QPACK] defines two additional stream types. Other streamtypes can be defined by extensions to HTTP/3; seeSection 9 for moredetails.¶
The performance of HTTP/3 connections in the early phase of their lifetime issensitive to the creation and exchange of data on unidirectional streams.Endpoints that excessively restrict the number of streams or the flow controlwindow of these streams will increase the chance that the remote peer reachesthe limit early and becomes blocked. In particular, implementations shouldconsider that remote peers may wish to exercise reserved stream behavior(Section 6.2.3) with some of the unidirectional streams they are permittedto use. To avoid blocking, the transport parameters sent by both clients andservers MUST allow the peer to create at least one unidirectional stream for theHTTP control stream plus the number of unidirectional streams required bymandatory extensions (three being the minimum number required for the baseHTTP/3 protocol and QPACK), and SHOULD provide at least 1,024 bytes of flowcontrol credit to each stream.¶
Note that an endpoint is not required to grant additional credits to create moreunidirectional streams if its peer consumes all the initial credits beforecreating the critical unidirectional streams. Endpoints SHOULD create the HTTPcontrol stream as well as the unidirectional streams required by mandatoryextensions (such as the QPACK encoder and decoder streams) first, and thencreate additional streams as allowed by their peer.¶
If the stream header indicates a stream type which is not supported by therecipient, the remainder of the stream cannot be consumed as the semantics areunknown. Recipients of unknown stream types MAY abort reading of the stream withan error code of H3_STREAM_CREATION_ERROR, but MUST NOT consider such streamsto be a connection error of any kind.¶
Implementations MAY send stream types before knowing whether the peer supportsthem. However, stream types which could modify the state or semantics ofexisting protocol components, including QPACK or other extensions, MUST NOT besent 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 tothe reception of the unidirectional stream header.¶
A control stream is indicated by a stream type of 0x00. Data on this streamconsists of HTTP/3 frames, as defined inSection 7.2.¶
Each side MUST initiate a single control stream at the beginning of theconnection and send its SETTINGS frame as the first frame on this stream. Ifthe first frame of the control stream is any other frame type, this MUST betreated as a connection error of type H3_MISSING_SETTINGS. Only one controlstream per peer is permitted; receipt of a second stream which claims to be acontrol stream MUST be treated as a connection error of typeH3_STREAM_CREATION_ERROR. The sender MUST NOT close the control stream, andthe receiver MUST NOT request that the sender close the control stream. Ifeither control stream is closed at any point, this MUST be treated as aconnection error of type H3_CLOSED_CRITICAL_STREAM.¶
A pair of unidirectional streams is used rather than a single bidirectionalstream. This allows either peer to send data as soon as it is able. Dependingon whether 0-RTT is enabled on the connection, either client or server might beable to send stream data first after the cryptographic handshake completes.¶
Server push is an optional feature introduced in HTTP/2 that allows a server toinitiate a response before a request has been made. SeeSection 4.4 formore details.¶
A push stream is indicated by a stream type of 0x01, followed by the Push IDof the promise that it fulfills, encoded as a variable-length integer. Theremaining data on this stream consists of HTTP/3 frames, as defined inSection 7.2, and fulfills a promised server push by zero or more interim HTTPresponses followed by a single final HTTP response, as defined inSection 4.1. Server push and Push IDs are described inSection 4.4.¶
Only servers can push; if a server receives a client-initiated push stream, thisMUST be treated as a connection error of type H3_STREAM_CREATION_ERROR.¶
Push Stream Header { Stream Type (i) = 0x01, Push ID (i),}Each Push ID MUST only be used once in a push stream header. If a push streamheader includes a Push ID that was used in another push stream header, theclient MUST treat this as a connection error of type H3_ID_ERROR.¶
Stream types of the format0x1f * N + 0x21 for non-negative integer values ofN are reserved to exercise the requirement that unknown types be ignored. Thesestreams have no semantics, and can be sent when application-layer padding isdesired. They MAY also be sent on connections where no data is currently beingtransferred. Endpoints MUST NOT consider these streams to have any meaning uponreceipt.¶
The payload and length of the stream are selected in any manner theimplementation chooses. Implementations MAY terminate these streams cleanly, orMAY abruptly terminate them. When terminating abruptly, the error codeH3_NO_ERROR or a reserved error code (Section 8.1) SHOULD be used.¶
HTTP frames are carried on QUIC streams, as described inSection 6.HTTP/3 defines three stream types: control stream, request stream, and pushstream. This section describes HTTP/3 frame formats and the streams types onwhich they are permitted; seeTable 1 for an overview. Acomparison between HTTP/2 and HTTP/3 frames is provided inAppendix A.2.¶
| Frame | Control Stream | Request Stream | Push Stream | Section |
|---|---|---|---|---|
| DATA | No | Yes | Yes | Section 7.2.1 |
| HEADERS | No | Yes | Yes | Section 7.2.2 |
| CANCEL_PUSH | Yes | No | No | Section 7.2.3 |
| SETTINGS | Yes (1) | No | No | Section 7.2.4 |
| PUSH_PROMISE | No | Yes | No | Section 7.2.5 |
| GOAWAY | Yes | No | No | Section 7.2.6 |
| MAX_PUSH_ID | Yes | No | No | Section 7.2.7 |
| Reserved | Yes | Yes | Yes | Section 7.2.8 |
Certain frames can only occur as the first frame of a particular stream type;these are indicated inTable 1 with a (1). Specific guidanceis provided in the relevant section.¶
Note that, unlike QUIC frames, HTTP/3 frames can span multiple packets.¶
All frames have the following format:¶
HTTP/3 Frame Format { Type (i), Length (i), Frame Payload (..),}A frame includes the following fields:¶
Each frame's payload MUST contain exactly the fields identified in itsdescription. A frame payload that contains additional bytes after theidentified fields or a frame payload that terminates before the end of theidentified fields MUST be treated as a connection error of typeH3_FRAME_ERROR.¶
When a stream terminates cleanly, if the last frame on the stream was truncated,this MUST be treated as a connection error (Section 8) of typeH3_FRAME_ERROR. Streams which terminate abruptly may be reset at any point ina frame.¶
DATA frames (type=0x0) convey arbitrary, variable-length sequences of bytesassociated with an HTTP request or response payload.¶
DATA frames MUST be associated with an HTTP request or response. If a DATAframe is received on a control stream, the recipient MUST respond with aconnection error (Section 8) of type H3_FRAME_UNEXPECTED.¶
DATA Frame { Type (i) = 0x0, Length (i), Data (..),}The HEADERS frame (type=0x1) is used to carry an HTTP field section, encodedusing QPACK. See[QPACK] for more details.¶
HEADERS Frame { Type (i) = 0x1, Length (i), Encoded Field Section (..),}HEADERS frames can only be sent on request / push streams. If a HEADERS frameis received on a control stream, the recipient MUST respond with a connectionerror (Section 8) of type H3_FRAME_UNEXPECTED.¶
The CANCEL_PUSH frame (type=0x3) is used to request cancellation of a serverpush prior to the push stream being received. The CANCEL_PUSH frame identifiesa server push by Push ID (seeSection 7.2.5), encoded as avariable-length integer.¶
When a client sends CANCEL_PUSH, it is indicating that it does not wish toreceive the promised resource. The server SHOULD abort sending the resource,but the mechanism to do so depends on the state of the corresponding pushstream. If the server has not yet created a push stream, it does not createone. If the push stream is open, the server SHOULD abruptly terminate thatstream. If the push stream has already ended, the server MAY still abruptlyterminate the stream or MAY take no action.¶
When a server sends CANCEL_PUSH, it is indicating that it will not be fulfillinga promise. The client cannot expect the corresponding promise to be fulfilled,unless it has already received and processed the promised response. A serverSHOULD send a CANCEL_PUSH even if it has opened the corresponding stream.¶
Sending CANCEL_PUSH has no direct effect on the state of existing push streams.A client SHOULD NOT send a CANCEL_PUSH when it has already received acorresponding push stream. A push stream could arrive after a client has sentCANCEL_PUSH, because a server might not have processed the CANCEL_PUSH. Theclient SHOULD abort reading the stream with an error code ofH3_REQUEST_CANCELLED.¶
A CANCEL_PUSH frame is sent on the control stream. Receiving a CANCEL_PUSHframe on a stream other than the control stream MUST be treated as a connectionerror of type H3_FRAME_UNEXPECTED.¶
CANCEL_PUSH Frame { Type (i) = 0x3, Length (i), Push ID (..),}The CANCEL_PUSH frame carries a Push ID encoded as a variable-length integer.The Push ID identifies the server push that is being cancelled; seeSection 7.2.5. If a CANCEL_PUSH frame is received which references aPush ID greater than currently allowed on the connection, this MUST be treatedas a connection error of type H3_ID_ERROR.¶
If the client receives a CANCEL_PUSH frame, that frame might identify a Push IDthat has not yet been mentioned by a PUSH_PROMISE frame due to reordering. If aserver receives a CANCEL_PUSH frame for a Push ID that has not yet beenmentioned by a PUSH_PROMISE frame, this MUST be treated as a connection error oftype H3_ID_ERROR.¶
The SETTINGS frame (type=0x4) conveys configuration parameters that affect howendpoints communicate, such as preferences and constraints on peer behavior.Individually, a SETTINGS parameter can also be referred to as a "setting"; theidentifier and value of each setting parameter can be referred to as a "settingidentifier" and a "setting value".¶
SETTINGS frames always apply to a connection, never a single stream. A SETTINGSframe MUST be sent as the first frame of each control stream (seeSection 6.2.1) by each peer, and MUST NOT be sent subsequently. Ifan endpoint receives a second SETTINGS frame on the control stream, the endpointMUST respond with a connection error of type H3_FRAME_UNEXPECTED.¶
SETTINGS frames MUST NOT be sent on any stream other than the control stream.If an endpoint receives a SETTINGS frame on a different stream, the endpointMUST respond with a connection error of type H3_FRAME_UNEXPECTED.¶
SETTINGS parameters are not negotiated; they describe characteristics of thesending peer, which can be used by the receiving peer. However, a negotiationcan be implied by the use of SETTINGS - each peer uses SETTINGS to advertise aset of supported values. The definition of the setting would describe how eachpeer combines the two sets to conclude which choice will be used. SETTINGS doesnot provide a mechanism to identify when the choice takes effect.¶
Different values for the same parameter can be advertised by each peer. Forexample, a client might be willing to consume a very large response fieldsection, while servers are more cautious about request size.¶
The same setting identifier MUST NOT occur more than once in the SETTINGS frame.A receiver MAY treat the presence of duplicate setting identifiers as aconnection error of type H3_SETTINGS_ERROR.¶
The payload of a SETTINGS frame consists of zero or more parameters. Eachparameter consists of a setting identifier and a value, both encoded as QUICvariable-length integers.¶
Setting { Identifier (i), Value (i),}SETTINGS Frame { Type (i) = 0x4, Length (i), Setting (..) ...,}An implementation MUST ignore the contents for any SETTINGS identifier it doesnot understand.¶
The following settings are defined in HTTP/3:¶
Setting identifiers of the format0x1f * N + 0x21 for non-negative integervalues of N are reserved to exercise the requirement that unknown identifiers beignored. Such settings have no defined meaning. Endpoints SHOULD include atleast one such setting in their SETTINGS frame. Endpoints MUST NOT consider suchsettings to have any meaning upon receipt.¶
Because the setting has no defined meaning, the value of the setting can be anyvalue the implementation selects.¶
Additional settings can be defined by extensions to HTTP/3; seeSection 9for more details.¶
An HTTP implementation MUST NOT send frames or requests which would be invalidbased on its current understanding of the peer's settings.¶
All settings begin at an initial value. Each endpoint SHOULD use these initialvalues to send messages before the peer's SETTINGS frame has arrived, as packetscarrying the settings can be lost or delayed. When the SETTINGS frame arrives,any settings are changed to their new values.¶
This removes the need to wait for the SETTINGS frame before sending messages.Endpoints MUST NOT require any data to be received from the peer prior tosending the SETTINGS frame; settings MUST be sent as soon as the transport isready to send data.¶
For servers, the initial value of each client setting is the default value.¶
For clients using a 1-RTT QUIC connection, the initial value of each serversetting is the default value. 1-RTT keys will always become available prior toSETTINGS arriving, even if the server sends SETTINGS immediately. Clients SHOULDNOT wait indefinitely for SETTINGS to arrive before sending requests, but SHOULDprocess received datagrams in order to increase the likelihood of processingSETTINGS before sending the first request.¶
When a 0-RTT QUIC connection is being used, the initial value of each serversetting is the value used in the previous session. Clients SHOULD store thesettings the server provided in the connection where resumption information wasprovided, but MAY opt not to store settings in certain cases (e.g., if thesession ticket is received before the SETTINGS frame). A client MUST comply withstored settings - or default values, if no values are stored - when attempting0-RTT. Once a server has provided new settings, clients MUST comply with thosevalues.¶
A server can remember the settings that it advertised, or store anintegrity-protected copy of the values in the ticket and recover the informationwhen accepting 0-RTT data. A server uses the HTTP/3 settings values indetermining whether to accept 0-RTT data. If the server cannot determine thatthe settings remembered by a client are compatible with its current settings, itMUST NOT accept 0-RTT data. Remembered settings are compatible if a clientcomplying with those settings would not violate the server's current settings.¶
A server MAY accept 0-RTT and subsequently provide different settings in itsSETTINGS frame. If 0-RTT data is accepted by the server, its SETTINGS frame MUSTNOT reduce any limits or alter any values that might be violated by the clientwith its 0-RTT data. The server MUST include all settings which differ fromtheir default values. If a server accepts 0-RTT but then sends settings thatare not compatible with the previously specified settings, this MUST be treatedas a connection error of type H3_SETTINGS_ERROR. If a server accepts 0-RTT butthen sends a SETTINGS frame that omits a setting value that the clientunderstands (apart from reserved setting identifiers) that was previouslyspecified to have a non-default value, this MUST be treated as a connectionerror of type H3_SETTINGS_ERROR.¶
The PUSH_PROMISE frame (type=0x5) is used to carry a promised request headerfield section from server to client on a request stream, as in HTTP/2.¶
PUSH_PROMISE Frame { Type (i) = 0x5, Length (i), Push ID (i), Encoded Field Section (..),}The payload consists of:¶
A server MUST NOT use a Push ID that is larger than the client has provided in aMAX_PUSH_ID frame (Section 7.2.7). A client MUST treat receipt of aPUSH_PROMISE frame that contains a larger Push ID than the client has advertisedas a connection error of H3_ID_ERROR.¶
A server MAY use the same Push ID in multiple PUSH_PROMISE frames. If so, thedecompressed request header sets MUST contain the same fields in the same order,and both the name and the value in each field MUST be exact matches. ClientsSHOULD compare the request header sections for resources promised multipletimes. If a client receives a Push ID that has already been promised and detectsa mismatch, it MUST respond with a connection error of typeH3_GENERAL_PROTOCOL_ERROR. If the decompressed field sections match exactly, theclient SHOULD associate the pushed content with each stream on which aPUSH_PROMISE was received.¶
Allowing duplicate references to the same Push ID is primarily to reduceduplication caused by concurrent requests. A server SHOULD avoid reusing a PushID over a long period. Clients are likely to consume server push responses andnot retain them for reuse over time. Clients that see a PUSH_PROMISE that usesa Push ID that they have already consumed and discarded are forced to ignore thePUSH_PROMISE.¶
If a PUSH_PROMISE frame is received on the control stream, the client MUSTrespond with a connection error (Section 8) of type H3_FRAME_UNEXPECTED.¶
A client MUST NOT send a PUSH_PROMISE frame. A server MUST treat the receiptof a PUSH_PROMISE frame as a connection error of type H3_FRAME_UNEXPECTED.¶
SeeSection 4.4 for a description of the overall server push mechanism.¶
The GOAWAY frame (type=0x7) is used to initiate graceful shutdown of aconnection by either endpoint. GOAWAY allows an endpoint to stop accepting newrequests or pushes while still finishing processing of previously receivedrequests and pushes. This enables administrative actions, like servermaintenance. GOAWAY by itself does not close a connection.¶
GOAWAY Frame { Type (i) = 0x7, Length (i), Stream ID/Push ID (..),}The GOAWAY frame is always sent on the control stream. In the server to clientdirection, it carries a QUIC Stream ID for a client-initiated bidirectionalstream encoded as a variable-length integer. A client MUST treat receipt of aGOAWAY frame containing a Stream ID of any other type as a connection error oftype H3_ID_ERROR.¶
In the client to server direction, the GOAWAY frame carries a Push ID encoded asa variable-length integer.¶
The GOAWAY frame applies to the connection, not a specific stream. A clientMUST treat a GOAWAY frame on a stream other than the control stream as aconnection error (Section 8) of type H3_FRAME_UNEXPECTED.¶
SeeSection 5.2 for more information on the use of the GOAWAY frame.¶
The MAX_PUSH_ID frame (type=0xD) is used by clients to control the number ofserver pushes that the server can initiate. This sets the maximum value for aPush ID that the server can use in PUSH_PROMISE and CANCEL_PUSH frames.Consequently, this also limits the number of push streams that the server caninitiate in addition to the limit maintained by the QUIC transport.¶
The MAX_PUSH_ID frame is always sent on the control stream. Receipt of aMAX_PUSH_ID frame on any other stream MUST be treated as a connection error oftype H3_FRAME_UNEXPECTED.¶
A server MUST NOT send a MAX_PUSH_ID frame. A client MUST treat the receipt ofa MAX_PUSH_ID frame as a connection error of type H3_FRAME_UNEXPECTED.¶
The maximum Push ID is unset when a connection is created, meaning that a servercannot push until it receives a MAX_PUSH_ID frame. A client that wishes tomanage the number of promised server pushes can increase the maximum Push ID bysending MAX_PUSH_ID frames as the server fulfills or cancels server pushes.¶
MAX_PUSH_ID Frame { Type (i) = 0x1, Length (i), Push ID (..),}The MAX_PUSH_ID frame carries a single variable-length integer that identifiesthe maximum value for a Push ID that the server can use; seeSection 7.2.5. A MAX_PUSH_ID frame cannot reduce the maximum Push ID;receipt of a MAX_PUSH_ID that contains a smaller value than previously receivedMUST be treated as a connection error of type H3_ID_ERROR.¶
Frame types of the format0x1f * N + 0x21 for non-negative integer values of Nare reserved to exercise the requirement that unknown types be ignored(Section 9). These frames have no semantics, and can be sent on any openstream when application-layer padding is desired. They MAY also be sent onconnections where no data is currently being transferred. Endpoints MUST NOTconsider these frames to have any meaning upon receipt.¶
The payload and length of the frames are selected in any manner theimplementation chooses.¶
Frame types which were used in HTTP/2 where there is no corresponding HTTP/3frame have also been reserved (Section 11.2.1). These frame types MUST NOT besent, and receipt MAY be treated as an error of type H3_FRAME_UNEXPECTED.¶
QUIC allows the application to abruptly terminate (reset) individual streams orthe entire connection when an error is encountered. These are referred to as"stream errors" or "connection errors" and are described in more detail in[QUIC-TRANSPORT].¶
An endpoint MAY choose to treat a stream error as a connection error undercertain circumstances. Implementations need to consider the impact onoutstanding requests before making this choice.¶
Because new error codes can be defined without negotiation (seeSection 9),use of an error code in an unexpected context or receipt of an unknown errorcode MUST be treated as equivalent to H3_NO_ERROR. However, closing a streamcan have other effects regardless of the error code; seeSection 4.1.¶
This section describes HTTP/3-specific error codes which can be used to expressthe cause of a connection or stream error.¶
The following error codes are defined for use when abruptly terminating streams,aborting reading of streams, or immediately closing connections.¶
Error codes of the format0x1f * N + 0x21 for non-negative integer values of Nare reserved to exercise the requirement that unknown error codes be treated asequivalent to H3_NO_ERROR (Section 9). Implementations SHOULD select anerror code from this space with some probability when they would have sentH3_NO_ERROR.¶
HTTP/3 permits extension of the protocol. Within the limitations described inthis section, protocol extensions can be used to provide additional services oralter any aspect of the protocol. Extensions are effective only within thescope of a single HTTP/3 connection.¶
This applies to the protocol elements defined in this document. This does notaffect the existing options for extending HTTP, such as defining new methods,status codes, or fields.¶
Extensions are permitted to use new frame types (Section 7.2), new settings(Section 7.2.4.1), new error codes (Section 8), or new unidirectionalstream types (Section 6.2). Registries are established formanaging these extension points: frame types (Section 11.2.1), settings(Section 11.2.2), error codes (Section 11.2.3), and stream types(Section 11.2.4).¶
Implementations MUST ignore unknown or unsupported values in all extensibleprotocol elements. Implementations MUST discard frames and unidirectionalstreams that have unknown or unsupported types. This means that any of theseextension points can be safely used by extensions without prior arrangement ornegotiation. However, where a known frame type is required to be in a specificlocation, such as the SETTINGS frame as the first frame of the control stream(seeSection 6.2.1), an unknown frame type does not satisfy thatrequirement and SHOULD be treated as an error.¶
Extensions that could change the semantics of existing protocol components MUSTbe negotiated before being used. For example, an extension that changes thelayout of the HEADERS frame cannot be used until the peer has given a positivesignal that this is acceptable. Coordinating when such a revised layout comesinto effect could prove complex. As such, allocating new identifiers fornew definitions of existing protocol elements is likely to be more effective.¶
This document doesn't mandate a specific method for negotiating the use of anextension but notes that a setting (Section 7.2.4.1) could be used forthat purpose. If both peers set a value that indicates willingness to use theextension, then the extension can be used. If a setting is used for extensionnegotiation, the default value MUST be defined in such a fashion that theextension is disabled if the setting is omitted.¶
The security considerations of HTTP/3 should be comparable to those of HTTP/2with TLS. However, many of the considerations from Section 10 of[HTTP2]apply to[QUIC-TRANSPORT] and are discussed in that document.¶
HTTP/3 relies on the HTTP definition of authority. The security considerationsof establishing authority are discussed in Section 11.1 of[SEMANTICS].¶
The use of ALPN in the TLS and QUIC handshakes establishes the targetapplication protocol before application-layer bytes are processed. Because allQUIC packets are encrypted, it is difficult for an attacker to control theplaintext bytes of an HTTP/3 connection which could be used in a cross-protocolattack on a plaintext protocol.¶
The HTTP/3 field encoding allows the expression of names that are not validfield names in the syntax used by HTTP (Section 4.3 of[SEMANTICS]). Requestsor responses containing invalid field names MUST be treated as malformed(Section 4.1.3). An intermediary therefore cannot translate an HTTP/3 requestor response containing an invalid field name into an HTTP/1.1 message.¶
Similarly, HTTP/3 allows field values that are not valid. While most of thevalues that can be encoded will not alter field parsing, carriage return (CR,ASCII 0xd), line feed (LF, ASCII 0xa), and the zero character (NUL, ASCII 0x0)might be exploited by an attacker if they are translated verbatim. Any requestor response that contains a character not permitted in a field value MUST betreated as malformed (Section 4.1.3). Valid characters are defined by the"field-content" ABNF rule in Section 4.4 of[SEMANTICS].¶
Pushed responses do not have an explicit request from the client; the request isprovided by the server in the PUSH_PROMISE frame.¶
Caching responses that are pushed is possible based on the guidance provided bythe origin server in the Cache-Control header field. However, this can causeissues if a single server hosts more than one tenant. For example, a servermight offer multiple users each a small portion of its URI space.¶
Where multiple tenants share space on the same server, that server MUST ensurethat tenants are not able to push representations of resources that they do nothave authority over. Failure to enforce this would allow a tenant to provide arepresentation that would be served out of cache, overriding the actualrepresentation that the authoritative tenant provides.¶
Pushed responses for which an origin server is not authoritative (seeSection 3.4) MUST NOT be used or cached.¶
An HTTP/3 connection can demand a greater commitment of resources to operatethan an HTTP/1.1 or HTTP/2 connection. The use of field compression and flowcontrol depend on a commitment of resources for storing a greater amount ofstate. Settings for these features ensure that memory commitments for thesefeatures are strictly bounded.¶
The number of PUSH_PROMISE frames is constrained in a similar fashion. A clientthat accepts server push SHOULD limit the number of Push IDs it issues at atime.¶
Processing capacity cannot be guarded as effectively as state capacity.¶
The ability to send undefined protocol elements which the peer is required toignore can be abused to cause a peer to expend additional processing time. Thismight be done by setting multiple undefined SETTINGS parameters, unknown frametypes, or unknown stream types. Note, however, that some uses are entirelylegitimate, such as optional-to-understand extensions and padding to increaseresistance to traffic analysis.¶
Compression of field sections also offers some opportunities to waste processingresources; see Section 7 of[QPACK] for more details on potential abuses.¶
All these features - i.e., server push, unknown protocol elements, fieldcompression - have legitimate uses. These features become a burden only whenthey are used unnecessarily or to excess.¶
An endpoint that doesn't monitor this behavior exposes itself to a risk ofdenial-of-service attack. Implementations SHOULD track the use of thesefeatures and set limits on their use. An endpoint MAY treat activity that issuspicious as a connection error (Section 8) of type H3_EXCESSIVE_LOAD, butfalse positives will result in disrupting valid connections and requests.¶
A large field section (Section 4.1) can cause an implementation tocommit a large amount of state. Header fields that are critical for routing canappear toward the end of a header field section, which prevents streaming of theheader field section to its ultimate destination. This ordering and otherreasons, such as ensuring cache correctness, mean that an endpoint likely needsto buffer the entire header field section. Since there is no hard limit to thesize of a field section, some endpoints could be forced to commit a large amountof available memory for header fields.¶
An endpoint can use the SETTINGS_MAX_HEADER_LIST_SIZE (Section 7.2.4.1)setting to advise peers of limits that might apply on the size of fieldsections. This setting is only advisory, so endpoints MAY choose to send fieldsections that exceed this limit and risk having the request or response beingtreated as malformed. This setting is specific to a connection, so any requestor response could encounter a hop with a lower, unknown limit. An intermediarycan attempt to avoid this problem by passing on values presented by differentpeers, but they are not obligated to do so.¶
A server that receives a larger field section than it is willing to handle cansend an HTTP 431 (Request Header Fields Too Large) status code[RFC6585]. Aclient can discard responses that it cannot process.¶
The CONNECT method can be used to create disproportionate load on an proxy,since stream creation is relatively inexpensive when compared to the creationand maintenance of a TCP connection. A proxy might also maintain some resourcesfor a TCP connection beyond the closing of the stream that carries the CONNECTrequest, since the outgoing TCP connection remains in the TIME_WAIT state.Therefore, a proxy cannot rely on QUIC stream limits alone to control theresources consumed by CONNECT requests.¶
Compression can allow an attacker to recover secret data when it is compressedin the same context as data under attacker control. HTTP/3 enables compressionof fields (Section 4.1.1); the following concerns also apply to the useof HTTP compressed content-codings; see Section 6.1.2 of[SEMANTICS].¶
There are demonstrable attacks on compression that exploit the characteristicsof the web (e.g.,[BREACH]). The attacker induces multiple requestscontaining varying plaintext, observing the length of the resulting ciphertextin each, which reveals a shorter length when a guess about the secret iscorrect.¶
Implementations communicating on a secure channel MUST NOT compress content thatincludes both confidential and attacker-controlled data unless separatecompression dictionaries are used for each source of data. Compression MUST NOTbe used if the source of data cannot be reliably determined.¶
Further considerations regarding the compression of fields sections aredescribed in[QPACK].¶
Padding can be used to obscure the exact size of frame content and is providedto mitigate specific attacks within HTTP, for example, attacks where compressedcontent 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 aconnection more resistant to traffic analysis, HTTP/3 can either rely ontransport-layer padding or employ the reserved frame and stream types discussedinSection 7.2.8 andSection 6.2.3. These methods of padding producedifferent results in terms of the granularity of padding, how padding isarranged in relation to the information that is being protected, whether paddingis applied in the case of packet loss, and how an implementation might controlpadding. Redundant padding could even be counterproductive.¶
To mitigate attacks that rely on compression, disabling or limiting compressionmight be preferable to padding as a countermeasure.¶
Use of padding can result in less protection than might seem immediatelyobvious. At best, padding only makes it more difficult for an attacker to inferlength information by increasing the number of frames an attacker has toobserve. Incorrectly implemented padding schemes can be easily defeated. Inparticular, randomized padding with a predictable distribution provides verylittle protection; similarly, padding payloads to a fixed size exposesinformation as payload sizes cross the fixed-sized boundary, which could bepossible if an attacker can control plaintext.¶
Several protocol elements contain nested length elements, typically in the formof frames with an explicit length containing variable-length integers. Thiscould pose a security risk to an incautious implementer. An implementation MUSTensure that the length of a frame exactly matches the length of the fields itcontains.¶
The use of 0-RTT with HTTP/3 creates an exposure to replay attack. Theanti-replay mitigations in[HTTP-REPLAY] MUST be applied when usingHTTP/3 with 0-RTT.¶
Certain HTTP implementations use the client address for logging oraccess-control purposes. Since a QUIC client's address might change during aconnection (and future versions might support simultaneous use of multipleaddresses), such implementations will need to either actively retrieve theclient's current address or addresses when they are relevant or explicitlyaccept that the original address might change.¶
Several characteristics of HTTP/3 provide an observer an opportunity tocorrelate actions of a single client or server over time. These include thevalue of settings, the timing of reactions to stimulus, and the handling of anyfeatures that are controlled by settings.¶
As far as these create observable differences in behavior, they could be used asa basis for fingerprinting a specific client.¶
HTTP/3's preference for using a single QUIC connection allows correlation of auser's activity on a site. Reusing connections for different origins allowsfor correlation of activity across those origins.¶
Several features of QUIC solicit immediate responses and can be used by anendpoint to measure latency to their peer; this might have privacy implicationsin certain scenarios.¶
This document registers a new ALPN protocol ID (Section 11.1) and creates newregistries that manage the assignment of codepoints in HTTP/3.¶
This document creates a new registration for the identification ofHTTP/3 in the "Application Layer Protocol Negotiation (ALPN)Protocol IDs" registry established in[RFC7301].¶
The "h3" string identifies HTTP/3:¶
New registries created in this document operate under the QUIC registrationpolicy documented in Section 22.1 of[QUIC-TRANSPORT]. These registries allinclude the common set of fields listed in Section 22.1.1 of[QUIC-TRANSPORT].¶
The initial allocations in these registries created in this document are allassigned permanent status and list as contact both the IESG (iesg@ietf.org) andthe HTTP working group (ietf-http-wg@w3.org).¶
This document establishes a registry for HTTP/3 frame type codes. The "HTTP/3Frame Type" registry governs a 62-bit space. This registry follows the QUICregistry policy; seeSection 11.2. Permanent registrations in this registryare assigned using the Specification Required policy[RFC8126], except forvalues between 0x00 and 0x3f (in hexadecimal; inclusive), which are assignedusing Standards Action or IESG Approval as defined in Section 4.9 and 4.10 of[RFC8126].¶
While this registry is separate from the "HTTP/2 Frame Type" registry defined in[HTTP2], it is preferable that the assignments parallel each other where thecode spaces overlap. If an entry is present in only one registry, every effortSHOULD be made to avoid assigning the corresponding value to an unrelatedoperation.¶
In addition to common fields as described inSection 11.2, permanentregistrations in this registry MUST include the following field:¶
Specifications of frame types MUST include a description of the frame layout andits semantics, including any parts of the frame that are conditionally present.¶
The entries inTable 2 are registered by this document.¶
| Frame Type | Value | Specification |
|---|---|---|
| DATA | 0x0 | Section 7.2.1 |
| HEADERS | 0x1 | Section 7.2.2 |
| Reserved | 0x2 | N/A |
| CANCEL_PUSH | 0x3 | Section 7.2.3 |
| SETTINGS | 0x4 | Section 7.2.4 |
| PUSH_PROMISE | 0x5 | Section 7.2.5 |
| Reserved | 0x6 | N/A |
| GOAWAY | 0x7 | Section 7.2.6 |
| Reserved | 0x8 | N/A |
| Reserved | 0x9 | N/A |
| MAX_PUSH_ID | 0xD | Section 7.2.7 |
Additionally, each code of the format0x1f * N + 0x21 for non-negative integervalues of N (that is, 0x21, 0x40, ..., through 0x3FFFFFFFFFFFFFFE) MUSTNOT be assigned by IANA.¶
This document establishes a registry for HTTP/3 settings. The "HTTP/3 Settings"registry governs a 62-bit space. This registry follows the QUIC registrypolicy; seeSection 11.2. Permanent registrations in this registry areassigned using the Specification Required policy[RFC8126], except for valuesbetween 0x00 and 0x3f (in hexadecimal; inclusive), which are assigned usingStandards Action or IESG Approval as defined in Section 4.9 and 4.10 of[RFC8126].¶
While this registry is separate from the "HTTP/2 Settings" registry defined in[HTTP2], it is preferable that the assignments parallel each other. If anentry is present in only one registry, every effort SHOULD be made to avoidassigning the corresponding value to an unrelated operation.¶
In addition to common fields as described inSection 11.2, permanentregistrations in this registry MUST include the following fields:¶
The entries inTable 3 are registered by this document.¶
| Setting Name | Value | Specification | Default |
|---|---|---|---|
| Reserved | 0x2 | N/A | N/A |
| Reserved | 0x3 | N/A | N/A |
| Reserved | 0x4 | N/A | N/A |
| Reserved | 0x5 | N/A | N/A |
| MAX_FIELD_SECTION_SIZE | 0x6 | Section 7.2.4.1 | Unlimited |
Additionally, each code of the format0x1f * N + 0x21 for non-negative integervalues of N (that is, 0x21, 0x40, ..., through 0x3FFFFFFFFFFFFFFE) MUSTNOT be assigned by IANA.¶
This document establishes a registry for HTTP/3 error codes. The "HTTP/3 ErrorCode" registry manages a 62-bit space. This registry follows the QUIC registrypolicy; seeSection 11.2. Permanent registrations in this registry areassigned using the Specification Required policy[RFC8126], except for valuesbetween 0x00 and 0x3f (in hexadecimal; inclusive), which are assigned usingStandards Action or IESG Approval as defined in Section 4.9 and 4.10 of[RFC8126].¶
Registrations for error codes are required to include a descriptionof the error code. An expert reviewer is advised to examine newregistrations for possible duplication with existing error codes.Use of existing registrations is to be encouraged, but not mandated.¶
In addition to common fields as described inSection 11.2, permanentregistrations in this registry MUST include the following fields:¶
The entries in theTable 4 are registered by this document.¶
| Name | Value | Description | Specification |
|---|---|---|---|
| H3_NO_ERROR | 0x0100 | No error | Section 8.1 |
| H3_GENERAL_PROTOCOL_ERROR | 0x0101 | General protocol error | Section 8.1 |
| H3_INTERNAL_ERROR | 0x0102 | Internal error | Section 8.1 |
| H3_STREAM_CREATION_ERROR | 0x0103 | Stream creation error | Section 8.1 |
| H3_CLOSED_CRITICAL_STREAM | 0x0104 | Critical stream was closed | Section 8.1 |
| H3_FRAME_UNEXPECTED | 0x0105 | Frame not permitted in the current state | Section 8.1 |
| H3_FRAME_ERROR | 0x0106 | Frame violated layout or size rules | Section 8.1 |
| H3_EXCESSIVE_LOAD | 0x0107 | Peer generating excessive load | Section 8.1 |
| H3_ID_ERROR | 0x0108 | An identifier was used incorrectly | Section 8.1 |
| H3_SETTINGS_ERROR | 0x0109 | SETTINGS frame contained invalid values | Section 8.1 |
| H3_MISSING_SETTINGS | 0x010A | No SETTINGS frame received | Section 8.1 |
| H3_REQUEST_REJECTED | 0x010B | Request not processed | Section 8.1 |
| H3_REQUEST_CANCELLED | 0x010C | Data no longer needed | Section 8.1 |
| H3_REQUEST_INCOMPLETE | 0x010D | Stream terminated early | Section 8.1 |
| H3_CONNECT_ERROR | 0x010F | TCP reset or error on CONNECT request | Section 8.1 |
| H3_VERSION_FALLBACK | 0x0110 | Retry over HTTP/1.1 | Section 8.1 |
Additionally, each code of the format0x1f * N + 0x21 for non-negative integervalues of N (that is, 0x21, 0x40, ..., through 0x3FFFFFFFFFFFFFFE) MUSTNOT be assigned by IANA.¶
This document establishes a registry for HTTP/3 unidirectional stream types. The"HTTP/3 Stream Type" registry governs a 62-bit space. This registry follows theQUIC registry policy; seeSection 11.2. Permanent registrations in thisregistry are assigned using the Specification Required policy[RFC8126],except for values between 0x00 and 0x3f (in hexadecimal; inclusive), which areassigned using Standards Action or IESG Approval as defined in Section 4.9 and4.10 of[RFC8126].¶
In addition to common fields as described inSection 11.2, permanentregistrations in this registry MUST include the following fields:¶
Specifications for permanent registrations MUST include a description of thestream type, including the layout semantics of the stream contents.¶
The entries in the following table are registered by this document.¶
| Stream Type | Value | Specification | Sender |
|---|---|---|---|
| Control Stream | 0x00 | Section 6.2.1 | Both |
| Push Stream | 0x01 | Section 4.4 | Server |
Additionally, each code of the format0x1f * N + 0x21 for non-negative integervalues of N (that is, 0x21, 0x40, ..., through 0x3FFFFFFFFFFFFFFE) MUSTNOT be assigned by IANA.¶
HTTP/3 is strongly informed by HTTP/2, and bears many similarities. Thissection describes the approach taken to design HTTP/3, points out importantdifferences from HTTP/2, and describes how to map HTTP/2 extensions into HTTP/3.¶
HTTP/3 begins from the premise that similarity to HTTP/2 is preferable, but nota hard requirement. HTTP/3 departs from HTTP/2 where QUIC differs from TCP,either to take advantage of QUIC features (like streams) or to accommodateimportant shortcomings (such as a lack of total ordering). These differencesmake HTTP/3 similar to HTTP/2 in key aspects, such as the relationship ofrequests and responses to streams. However, the details of the HTTP/3 design aresubstantially different than HTTP/2.¶
These departures are noted in this section.¶
HTTP/3 permits use of a larger number of streams (2^62-1) than HTTP/2. Theconsiderations about exhaustion of stream identifier space apply, though thespace is significantly larger such that it is likely that other limits in QUICare 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. QUICconsiders a stream closed when all data has been received and sent data has beenacknowledged by the peer. HTTP/2 considers a stream closed when the framecontaining 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 periodof time. HTTP/3 servers might choose to permit a larger number of concurrentclient-initiated bidirectional streams to achieve equivalent concurrency toHTTP/2, depending on the expected usage patterns.¶
Due to the presence of other unidirectional stream types, HTTP/3 does not relyexclusively on the number of concurrent unidirectional streams to control thenumber of concurrent in-flight pushes. Instead, HTTP/3 clients use theMAX_PUSH_ID frame to control the number of pushes received from an HTTP/3server.¶
Many framing concepts from HTTP/2 can be elided on QUIC, because the transportdeals with them. Because frames are already on a stream, they can omit thestream number. Because frames do not block multiplexing (QUIC's multiplexingoccurs below this layer), the support for variable-maximum-length packets can beremoved. Because stream termination is handled by QUIC, an END_STREAM flag isnot required. This permits the removal of the Flags field from the genericframe layout.¶
Frame payloads are largely drawn from[HTTP2]. However, QUIC includes manyfeatures (e.g., flow control) which are also present in HTTP/2. In these cases,the HTTP mapping does not re-implement them. As a result, several HTTP/2 frametypes are not required in HTTP/3. Where an HTTP/2-defined frame is no longerused, the frame ID has been reserved in order to maximize portability betweenHTTP/2 and HTTP/3 implementations. However, even equivalent frames between thetwo mappings are not identical.¶
Many of the differences arise from the fact that HTTP/2 provides an absoluteordering between frames across all streams, while QUIC provides this guaranteeon each stream only. As a result, if a frame type makes assumptions that framesfrom different streams will still be received in the order sent, HTTP/3 willbreak them.¶
Some examples of feature adaptations are described below, as well as generalguidance to extension frame implementors converting an HTTP/2 extension toHTTP/3.¶
HTTP/2 specifies priority assignments in PRIORITY frames and (optionally) inHEADERS frames. HTTP/3 does not provide a means of signaling priority.¶
Note that while there is no explicit signaling for priority, this does not meanthat prioritization is not important for achieving good performance.¶
HPACK was designed with the assumption of in-order delivery. A sequence ofencoded field sections must arrive (and be decoded) at an endpoint in the sameorder in which they were encoded. This ensures that the dynamic state at the twoendpoints remains in sync.¶
Because this total ordering is not provided by QUIC, HTTP/3 uses a modifiedversion of HPACK, called QPACK. QPACK uses a single unidirectional stream tomake all modifications to the dynamic table, ensuring a total order of updates.All frames which contain encoded fields merely reference the table state at agiven time without modifying it.¶
Frame type definitions in HTTP/3 often use the QUIC variable-length integerencoding. In particular, Stream IDs use this encoding, which allows for alarger range of possible values than the encoding used in HTTP/2. Some framesin HTTP/3 use an identifier rather than a Stream ID (e.g., PushIDs). Redefinition of the encoding of extension frame types might be necessaryif the encoding includes a Stream ID.¶
Because the Flags field is not present in generic HTTP/3 frames, those frameswhich depend on the presence of flags need to allocate space for flags as partof their frame payload.¶
Other than this issue, frame type HTTP/2 extensions are typically portable toQUIC simply by replacing Stream 0 in HTTP/2 with a control stream in HTTP/3.HTTP/3 extensions will not assume ordering, but would not be harmed by ordering,and would be portable to HTTP/2 in the same manner.¶
Frame types defined by extensions to HTTP/2 need to be separately registered forHTTP/3 if still applicable. The IDs of frames defined in[HTTP2] have beenreserved for simplicity. Note that the frame type space in HTTP/3 issubstantially larger (62 bits versus 8 bits), so many HTTP/3 frame types have noequivalent HTTP/2 code points. SeeSection 11.2.1.¶
An important difference from HTTP/2 is that settings are sent once, as the firstframe of the control stream, and thereafter cannot change. This eliminates manycorner cases around synchronization of changes.¶
Some transport-level options that HTTP/2 specifies via the SETTINGS frame aresuperseded by QUIC transport parameters in HTTP/3. The HTTP-level options thatare retained in HTTP/3 have the same value as in HTTP/2.¶
Below is a listing of how each HTTP/2 SETTINGS parameter is mapped:¶
In HTTP/3, setting values are variable-length integers (6, 14, 30, or 62 bitslong) rather than fixed-length 32-bit fields as in HTTP/2. This will oftenproduce a shorter encoding, but can produce a longer encoding for settings whichuse the full 32-bit space. Settings ported from HTTP/2 might choose to redefinetheir value to limit it to 30 bits for more efficient encoding, or to make useof 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 ofsettings defined in[HTTP2] have been reserved for simplicity. Note thatthe settings identifier space in HTTP/3 is substantially larger (62 bits versus16 bits), so many HTTP/3 settings have no equivalent HTTP/2 code point. SeeSection 11.2.2.¶
As QUIC streams might arrive out-of-order, endpoints are advised to not wait forthe peers' settings to arrive before responding to other streams. SeeSection 7.2.4.2.¶
QUIC has the same concepts of "stream" and "connection" errors that HTTP/2provides. However, the differences between HTTP/2 and HTTP/3 mean that errorcodes are not directly portable between versions.¶
The HTTP/2 error codes defined in Section 7 of[HTTP2] logically map tothe HTTP/3 error codes as follows:¶
Error codes need to be defined for HTTP/2 and HTTP/3 separately. SeeSection 11.2.3.¶
An intermediary that converts between HTTP/2 and HTTP/3 may encounter errorconditions from either upstream. It is useful to communicate the occurrence oferror to the downstream but error codes largely reflect connection-localproblems that generally do not make sense to propagate.¶
An intermediary that encounters an error from an upstream origin can indicatethis by sending an HTTP status code such as 502, which is suitable for a broadclass of errors.¶
There are some rare cases where it is beneficial to propagate the error bymapping it to the closest matching error type to the receiver. For example, anintermediary that receives an HTTP/2 stream error of type REFUSED_STREAM fromthe origin has a clear signal that the request was not processed and that therequest is safe to retry. Propagating this error condition to the client as anHTTP/3 stream error of type H3_REQUEST_REJECTED allows the client to take theaction it deems most appropriate. In the reverse direction the intermediarymight deem it beneficial to pass on client request cancellations that areindicated by terminating a stream with H3_REQUEST_CANCELLED.¶
Conversion between errors is described in the logical mapping. The error codesare defined in non-overlapping spaces in order to protect against accidentalconversion that could result in the use of inappropriate or unknown error codesfor the target version. An intermediary is permitted to promote stream errors toconnection errors but they should be aware of the cost to the connection forwhat might be a temporary or intermittent error.¶
Further changes to error codes (#2662,#2551):¶
http-opportunistic resource (RFC 8164) when scheme ishttp (#2439,#2973)¶Changes to SETTINGS frames in 0-RTT (#2972,#2790,#2945):¶
No changes¶
Extensive changes to error codes and conditions of their sending¶
Use variable-length integers throughout (#2437,#2233,#2253,#2275)¶
Changes to PRIORITY frame (#1865, #2075)¶
Substantial editorial reorganization; no technical changes.¶
SETTINGS changes (#181):¶
The original authors of this specification were Robbie Shade and Mike Warres.¶
The IETF QUIC Working Group received an enormous amount of support from manypeople. Among others, the following people provided substantial contributions tothis document:¶
奥 一穂 (Kazuho Oku)¶
A portion of Mike's contribution was supported by Microsoft during hisemployment there.¶
draft-ietf-quic-http-29
| Document | Document type | This is an older version of an Internet-Draft that was ultimately published asRFC 9114. | |
|---|---|---|---|
| Select version | |||
| Compare versions | |||
| Author | Mike Bishop | ||
| Replaces | draft-shade-quic-http2-mapping | ||
| RFC stream | |||
| Other formats | |||
| Additional resources | Mailing list discussion |