Skip to content
Switch branches/tags
Go to file
Cannot retrieve contributors at this time
// Copyright (c) 2017 Sandstorm Development Group, Inc. and contributors
// Licensed under the MIT License:
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
#pragma once
// The KJ HTTP client/server library.
// This is a simple library which can be used to implement an HTTP client or server. Properties
// of this library include:
// - Uses KJ async framework.
// - Agnostic to transport layer -- you can provide your own.
// - Header parsing is zero-copy -- it results in strings that point directly into the buffer
// received off the wire.
// - Application code which reads and writes headers refers to headers by symbolic names, not by
// string literals, with lookups being array-index-based, not map-based. To make this possible,
// the application announces what headers it cares about in advance, in order to assign numeric
// values to them.
// - Methods are identified by an enum.
#include <kj/string.h>
#include <kj/vector.h>
#include <kj/memory.h>
#include <kj/one-of.h>
#include <kj/async-io.h>
namespace kj {
/* standard methods */ \
/* */ \
/* (CONNECT is intentionally omitted since it should be handled specially in HttpServer) */ \
/* WebDAV */ \
/* Subversion */ \
/* UPnP */
enum class HttpMethod {
// Enum of known HTTP methods.
// We use an enum rather than a string to allow for faster parsing and switching and to reduce
// ambiguity.
#define DECLARE_METHOD(id) id,
kj::StringPtr KJ_STRINGIFY(HttpMethod method);
kj::Maybe<HttpMethod> tryParseHttpMethod(kj::StringPtr name);
class HttpHeaderTable;
class HttpHeaderId {
// Identifies an HTTP header by numeric ID that indexes into an HttpHeaderTable.
// The KJ HTTP API prefers that headers be identified by these IDs for a few reasons:
// - Integer lookups are much more efficient than string lookups.
// - Case-insensitivity is awkward to deal with when const strings are being passed to the lookup
// method.
// - Writing out strings less often means fewer typos.
// See HttpHeaderTable for usage hints.
HttpHeaderId() = default;
inline bool operator==(const HttpHeaderId& other) const { return id ==; }
inline bool operator!=(const HttpHeaderId& other) const { return id !=; }
inline bool operator< (const HttpHeaderId& other) const { return id <; }
inline bool operator> (const HttpHeaderId& other) const { return id >; }
inline bool operator<=(const HttpHeaderId& other) const { return id <=; }
inline bool operator>=(const HttpHeaderId& other) const { return id >=; }
inline size_t hashCode() const { return id; }
// Returned value is guaranteed to be small and never collide with other headers on the same
// table.
kj::StringPtr toString() const;
void requireFrom(const HttpHeaderTable& table) const;
// In debug mode, throws an exception if the HttpHeaderId is not from the given table.
// In opt mode, no-op.
/* Headers that are always read-only. */ \
MACRO(CONNECTION, "Connection") \
MACRO(KEEP_ALIVE, "Keep-Alive") \
MACRO(TRAILER, "Trailer") \
MACRO(UPGRADE, "Upgrade") \
/* Headers that are read-only except in the case of a response to a HEAD request. */ \
MACRO(CONTENT_LENGTH, "Content-Length") \
MACRO(TRANSFER_ENCODING, "Transfer-Encoding") \
/* Headers that are read-only for WebSocket handshakes. */ \
/* Headers that you can write. */ \
MACRO(HOST, "Host") \
MACRO(DATE, "Date") \
MACRO(LOCATION, "Location") \
// For convenience, these headers are valid for all HttpHeaderTables. You can refer to them like:
// HttpHeaderId::HOST
// TODO(someday): Fill this out with more common headers.
#define DECLARE_HEADER(id, name) \
static const HttpHeaderId id;
// Declare a constant for each builtin header, e.g.: HttpHeaderId::CONNECTION
const HttpHeaderTable* table;
uint id;
inline explicit constexpr HttpHeaderId(const HttpHeaderTable* table, uint id)
: table(table), id(id) {}
friend class HttpHeaderTable;
friend class HttpHeaders;
class HttpHeaderTable {
// Construct an HttpHeaderTable to declare which headers you'll be interested in later on, and
// to manufacture IDs for them.
// Example:
// // Build a header table with the headers we are interested in.
// kj::HttpHeaderTable::Builder builder;
// const HttpHeaderId accept = builder.add("Accept");
// const HttpHeaderId contentType = builder.add("Content-Type");
// kj::HttpHeaderTable table(kj::mv(builder));
// // Create an HTTP client.
// auto client = kj::newHttpClient(table, network);
// // Get
// HttpHeaders headers(table);
// headers.set(accept, "text/html");
// auto response = client->send(kj::HttpMethod::GET, "", headers)
// .wait(waitScope);
// auto msg = kj::str("Response content type: ", response.headers.get(contentType));
struct IdsByNameMap;
// Constructs a table that only contains the builtin headers.
class Builder {
HttpHeaderId add(kj::StringPtr name);
Own<HttpHeaderTable> build();
HttpHeaderTable& getFutureTable();
// Get the still-unbuilt header table. You cannot actually use it until build() has been
// called.
// This method exists to help when building a shared header table -- the Builder may be passed
// to several components, each of which will register the headers they need and get a reference
// to the future table.
kj::Own<HttpHeaderTable> table;
KJ_DISALLOW_COPY(HttpHeaderTable); // Can't copy because HttpHeaderId points to the table.
~HttpHeaderTable() noexcept(false);
uint idCount() const;
// Return the number of IDs in the table.
kj::Maybe<HttpHeaderId> stringToId(kj::StringPtr name) const;
// Try to find an ID for the given name. The matching is case-insensitive, per the HTTP spec.
// Note: if `name` contains characters that aren't allowed in HTTP header names, this may return
// a bogus value rather than null, due to optimizations used in case-insensitive matching.
kj::StringPtr idToString(HttpHeaderId id) const;
// Get the canonical string name for the given ID.
kj::Vector<kj::StringPtr> namesById;
kj::Own<IdsByNameMap> idsByName;
class HttpHeaders {
// Represents a set of HTTP headers.
// This class guards against basic HTTP header injection attacks: Trying to set a header name or
// value containing a newline, carriage return, or other invalid character will throw an
// exception.
explicit HttpHeaders(const HttpHeaderTable& table);
static bool isValidHeaderValue(kj::StringPtr value);
// This returns whether the value is a valid parameter to the set call. While the HTTP spec
// suggests that only printable ASCII characters are allowed in header values, in practice that
// turns out to not be the case. We follow the browser's lead in disallowing \r and \n.
// Use this if you want to validate the value before supplying it to set() if you want to avoid
// an exception being thrown (e.g. you have custom error reporting). NOTE that set will still
// validate the value. If performance is a problem this API needs to be adjusted to a
// `validateHeaderValue` function that returns a special type that set can be confident has
// already passed through the validation routine.
HttpHeaders(HttpHeaders&&) = default;
HttpHeaders& operator=(HttpHeaders&&) = default;
size_t size() const;
// Returns the number of headers that forEach() would iterate over.
void clear();
// Clears all contents, as if the object was freshly-allocated. However, calling this rather
// than actually re-allocating the object may avoid re-allocation of internal objects.
HttpHeaders clone() const;
// Creates a deep clone of the HttpHeaders. The returned object owns all strings it references.
HttpHeaders cloneShallow() const;
// Creates a shallow clone of the HttpHeaders. The returned object references the same strings
// as the original, owning none of them.
bool isWebSocket() const;
// Convenience method that checks for the presence of the header `Upgrade: websocket`.
// Note that this does not actually validate that the request is a complete WebSocket handshake
// with the correct version number -- such validation will occur if and when you call
// acceptWebSocket().
kj::Maybe<kj::StringPtr> get(HttpHeaderId id) const;
// Read a header.
template <typename Func>
void forEach(Func&& func) const;
// Calls `func(name, value)` for each header in the set -- including headers that aren't mapped
// to IDs in the header table. Both inputs are of type kj::StringPtr.
template <typename Func1, typename Func2>
void forEach(Func1&& func1, Func2&& func2) const;
// Calls `func1(id, value)` for each header in the set that has a registered HttpHeaderId, and
// `func2(name, value)` for each header that does not. All calls to func1() precede all calls to
// func2().
void set(HttpHeaderId id, kj::StringPtr value);
void set(HttpHeaderId id, kj::String&& value);
// Sets a header value, overwriting the existing value.
// The String&& version is equivalent to calling the other version followed by takeOwnership().
// WARNING: It is the caller's responsibility to ensure that `value` remains valid until the
// HttpHeaders object is destroyed. This allows string literals to be passed without making a
// copy, but complicates the use of dynamic values. Hint: Consider using `takeOwnership()`.
void add(kj::StringPtr name, kj::StringPtr value);
void add(kj::StringPtr name, kj::String&& value);
void add(kj::String&& name, kj::String&& value);
// Append a header. `name` will be looked up in the header table, but if it's not mapped, the
// header will be added to the list of unmapped headers.
// The String&& versions are equivalent to calling the other version followed by takeOwnership().
// WARNING: It is the caller's responsibility to ensure that `name` and `value` remain valid
// until the HttpHeaders object is destroyed. This allows string literals to be passed without
// making a copy, but complicates the use of dynamic values. Hint: Consider using
// `takeOwnership()`.
void unset(HttpHeaderId id);
// Removes a header.
// It's not possible to remove a header by string name because non-indexed headers would take
// O(n) time to remove. Instead, construct a new HttpHeaders object and copy contents.
void takeOwnership(kj::String&& string);
void takeOwnership(kj::Array<char>&& chars);
void takeOwnership(HttpHeaders&& otherHeaders);
// Takes ownership of a string so that it lives until the HttpHeaders object is destroyed. Useful
// when you've passed a dynamic value to set() or add() or parse*().
struct Request {
HttpMethod method;
kj::StringPtr url;
struct Response {
uint statusCode;
kj::StringPtr statusText;
struct ProtocolError {
// Represents a protocol error, such as a bad request method or invalid headers. Debugging such
// errors is difficult without a copy of the data which we tried to parse, but this data is
// sensitive, so we can't just lump it into the error description directly. ProtocolError
// provides this sensitive data separate from the error description.
// TODO(cleanup): Should maybe not live in HttpHeaders? HttpServerErrorHandler::ProtocolError?
// Or HttpProtocolError? Or maybe we need a more general way of attaching sensitive context to
// kj::Exceptions?
uint statusCode;
// Suggested HTTP status code that should be used when returning an error to the client.
// Most errors are 400. An unrecognized method will be 501.
kj::StringPtr statusMessage;
// HTTP status message to go with `statusCode`, e.g. "Bad Request".
kj::StringPtr description;
// An error description safe for all the world to see.
kj::ArrayPtr<char> rawContent;
// Unredacted data which led to the error condition. This may contain anything transported over
// HTTP, to include sensitive PII, so you must take care to sanitize this before using it in any
// error report that may leak to unprivileged eyes.
// This ArrayPtr is merely a copy of the `content` parameter passed to `tryParseRequest()` /
// `tryParseResponse()`, thus it remains valid for as long as a successfully-parsed HttpHeaders
// object would remain valid.
using RequestOrProtocolError = kj::OneOf<Request, ProtocolError>;
using ResponseOrProtocolError = kj::OneOf<Response, ProtocolError>;
RequestOrProtocolError tryParseRequest(kj::ArrayPtr<char> content);
ResponseOrProtocolError tryParseResponse(kj::ArrayPtr<char> content);
// Parse an HTTP header blob and add all the headers to this object.
// `content` should be all text from the start of the request to the first occurrance of two
// newlines in a row -- including the first of these two newlines, but excluding the second.
// The parse is performed with zero copies: The callee clobbers `content` with '\0' characters
// to split it into a bunch of shorter strings. The caller must keep `content` valid until the
// `HttpHeaders` is destroyed, or pass it to `takeOwnership()`.
bool tryParse(kj::ArrayPtr<char> content);
// Like tryParseRequest()/tryParseResponse(), but don't expect any request/response line.
kj::String serializeRequest(HttpMethod method, kj::StringPtr url,
kj::ArrayPtr<const kj::StringPtr> connectionHeaders = nullptr) const;
kj::String serializeResponse(uint statusCode, kj::StringPtr statusText,
kj::ArrayPtr<const kj::StringPtr> connectionHeaders = nullptr) const;
// **Most applications will not use these methods; they are called by the HTTP client and server
// implementations.**
// Serialize the headers as a complete request or response blob. The blob uses '\r\n' newlines
// and includes the double-newline to indicate the end of the headers.
// `connectionHeaders`, if provided, contains connection-level headers supplied by the HTTP
// implementation, in the order specified by the KJ_HTTP_FOR_EACH_BUILTIN_HEADER macro. These
// headers values override any corresponding header value in the HttpHeaders object. The
// CONNECTION_HEADERS_COUNT constants below can help you construct this `connectionHeaders` array.
enum class BuiltinIndicesEnum {
#define HEADER_ID(id, name) id,
#undef HEADER_ID
struct BuiltinIndices {
#define HEADER_ID(id, name) static constexpr uint id = static_cast<uint>(BuiltinIndicesEnum::id);
#undef HEADER_ID
static constexpr uint CONNECTION_HEADERS_COUNT = BuiltinIndices::SEC_WEBSOCKET_KEY;
static constexpr uint WEBSOCKET_CONNECTION_HEADERS_COUNT = BuiltinIndices::HOST;
// Constants for use with HttpHeaders::serialize*().
kj::String toString() const;
const HttpHeaderTable* table;
kj::Array<kj::StringPtr> indexedHeaders;
// Size is always table->idCount().
struct Header {
kj::StringPtr name;
kj::StringPtr value;
kj::Vector<Header> unindexedHeaders;
kj::Vector<kj::Array<char>> ownedStrings;
void addNoCheck(kj::StringPtr name, kj::StringPtr value);
kj::StringPtr cloneToOwn(kj::StringPtr str);
kj::String serialize(kj::ArrayPtr<const char> word1,
kj::ArrayPtr<const char> word2,
kj::ArrayPtr<const char> word3,
kj::ArrayPtr<const kj::StringPtr> connectionHeaders) const;
bool parseHeaders(char* ptr, char* end);
// TODO(perf): Arguably we should store a map, but header sets are never very long
// TODO(perf): We could optimize for common headers by storing them directly as fields. We could
// also add direct accessors for those headers.
class HttpInputStream {
// Low-level interface to receive HTTP-formatted messages (headers followed by body) from an
// input stream, without a paired output stream.
// Most applications will not use this. Regular HTTP clients and servers don't need this. This
// is mainly useful for apps implementing various protocols that look like HTTP but aren't
// really.
struct Request {
HttpMethod method;
kj::StringPtr url;
const HttpHeaders& headers;
kj::Own<kj::AsyncInputStream> body;
virtual kj::Promise<Request> readRequest() = 0;
// Reads one HTTP request from the input stream.
// The returned struct contains pointers directly into a buffer that is invalidated on the next
// message read.
struct Response {
uint statusCode;
kj::StringPtr statusText;
const HttpHeaders& headers;
kj::Own<kj::AsyncInputStream> body;
virtual kj::Promise<Response> readResponse(HttpMethod requestMethod) = 0;
// Reads one HTTP response from the input stream.
// You must provide the request method because responses to HEAD requests require special
// treatment.
// The returned struct contains pointers directly into a buffer that is invalidated on the next
// message read.
struct Message {
const HttpHeaders& headers;
kj::Own<kj::AsyncInputStream> body;
virtual kj::Promise<Message> readMessage() = 0;
// Reads an HTTP header set followed by a body, with no request or response line. This is not
// useful for HTTP but may be useful for other protocols that make the unfortunate choice to
// mimic HTTP message format, such as Visual Studio Code's JSON-RPC transport.
// The returned struct contains pointers directly into a buffer that is invalidated on the next
// message read.
virtual kj::Promise<bool> awaitNextMessage() = 0;
// Waits until more data is available, but doesn't consume it. Returns false on EOF.
class EntropySource {
// Interface for an object that generates entropy. Typically, cryptographically-random entropy
// is expected.
// TODO(cleanup): Put this somewhere more general.
virtual void generate(kj::ArrayPtr<byte> buffer) = 0;
class WebSocket {
// Interface representincg an open WebSocket session.
// Each side can send and receive data and "close" messages.
// Ping/Pong and message fragmentation are not exposed through this interface. These features of
// the underlying WebSocket protocol are not exposed by the browser-level JavaScript API either,
// and thus applications typically need to implement these features at the application protocol
// level instead. The implementation is, however, expected to reply to Ping messages it receives.
virtual kj::Promise<void> send(kj::ArrayPtr<const byte> message) = 0;
virtual kj::Promise<void> send(kj::ArrayPtr<const char> message) = 0;
// Send a message (binary or text). The underlying buffer must remain valid, and you must not
// call send() again, until the returned promise resolves.
virtual kj::Promise<void> close(uint16_t code, kj::StringPtr reason) = 0;
// Send a Close message.
// Note that the returned Promise resolves once the message has been sent -- it does NOT wait
// for the other end to send a Close reply. The application should await a reply before dropping
// the WebSocket object.
virtual kj::Promise<void> disconnect() = 0;
// Sends EOF on the underlying connection without sending a "close" message. This is NOT a clean
// shutdown, but is sometimes useful when you want the other end to trigger whatever behavior
// it normally triggers when a connection is dropped.
virtual void abort() = 0;
// Forcefully close this WebSocket, such that the remote end should get a DISCONNECTED error if
// it continues to write. This differs from disconnect(), which only closes the sending
// direction, but still allows receives.
virtual kj::Promise<void> whenAborted() = 0;
// Resolves when the remote side aborts the connection such that send() would throw DISCONNECTED,
// if this can be detected without actually writing a message. (If not, this promise never
// resolves, but send() or receive() will throw DISCONNECTED when appropriate. See also
// kj::AsyncOutputStream::whenWriteDisconnected().)
struct Close {
uint16_t code;
kj::String reason;
typedef kj::OneOf<kj::String, kj::Array<byte>, Close> Message;
static constexpr size_t SUGGESTED_MAX_MESSAGE_SIZE = 1u << 20; // 1MB
virtual kj::Promise<Message> receive(size_t maxSize = SUGGESTED_MAX_MESSAGE_SIZE) = 0;
// Read one message from the WebSocket and return it. Can only call once at a time. Do not call
// again after Close is received.
virtual kj::Promise<void> pumpTo(WebSocket& other);
// Continuously receives messages from this WebSocket and send them to `other`.
// On EOF, calls other.disconnect(), then resolves.
// On other read errors, calls other.close() with the error, then resolves.
// On write error, rejects with the error.
virtual kj::Maybe<kj::Promise<void>> tryPumpFrom(WebSocket& other);
// Either returns null, or performs the equivalent of other.pumpTo(*this). Only returns non-null
// if this WebSocket implementation is able to perform the pump in an optimized way, better than
// the default implementation of pumpTo(). The default implementation of pumpTo() always tries
// calling this first, and the default implementation of tryPumpFrom() always returns null.
virtual uint64_t sentByteCount() = 0;
virtual uint64_t receivedByteCount() = 0;
class HttpClient {
// Interface to the client end of an HTTP connection.
// There are two kinds of clients:
// * Host clients are used when talking to a specific host. The `url` specified in a request
// is actually just a path. (A `Host` header is still required in all requests.)
// * Proxy clients are used when the target could be any arbitrary host on the internet.
// The `url` specified in a request is a full URL including protocol and hostname.
struct Response {
uint statusCode;
kj::StringPtr statusText;
const HttpHeaders* headers;
kj::Own<kj::AsyncInputStream> body;
// `statusText` and `headers` remain valid until `body` is dropped or read from.
struct Request {
kj::Own<kj::AsyncOutputStream> body;
// Write the request entity body to this stream, then drop it when done.
// May be null for GET and HEAD requests (which have no body) and requests that have
// Content-Length: 0.
kj::Promise<Response> response;
// Promise for the eventual response.
virtual Request request(HttpMethod method, kj::StringPtr url, const HttpHeaders& headers,
kj::Maybe<uint64_t> expectedBodySize = nullptr) = 0;
// Perform an HTTP request.
// `url` may be a full URL (with protocol and host) or it may be only the path part of the URL,
// depending on whether the client is a proxy client or a host client.
// `url` and `headers` need only remain valid until `request()` returns (they can be
// stack-allocated).
// `expectedBodySize`, if provided, must be exactly the number of bytes that will be written to
// the body. This will trigger use of the `Content-Length` connection header. Otherwise,
// `Transfer-Encoding: chunked` will be used.
struct WebSocketResponse {
uint statusCode;
kj::StringPtr statusText;
const HttpHeaders* headers;
kj::OneOf<kj::Own<kj::AsyncInputStream>, kj::Own<WebSocket>> webSocketOrBody;
// `statusText` and `headers` remain valid until `webSocketOrBody` is dropped or read from.
virtual kj::Promise<WebSocketResponse> openWebSocket(
kj::StringPtr url, const HttpHeaders& headers);
// Tries to open a WebSocket. Default implementation calls send() and never returns a WebSocket.
// `url` and `headers` need only remain valid until `openWebSocket()` returns (they can be
// stack-allocated).
virtual kj::Promise<kj::Own<kj::AsyncIoStream>> connect(kj::StringPtr host);
// Handles CONNECT requests. Only relevant for proxy clients. Default implementation throws
class HttpService {
// Interface which HTTP services should implement.
// This interface is functionally equivalent to HttpClient, but is intended for applications to
// implement rather than call. The ergonomics and performance of the method signatures are
// optimized for the serving end.
// As with clients, there are two kinds of services:
// * Host services are used when talking to a specific host. The `url` specified in a request
// is actually just a path. (A `Host` header is still required in all requests, and the service
// may in fact serve multiple origins via this header.)
// * Proxy services are used when the target could be any arbitrary host on the internet, i.e. to
// implement an HTTP proxy. The `url` specified in a request is a full URL including protocol
// and hostname.
class Response {
virtual kj::Own<kj::AsyncOutputStream> send(
uint statusCode, kj::StringPtr statusText, const HttpHeaders& headers,
kj::Maybe<uint64_t> expectedBodySize = nullptr) = 0;
// Begin the response.
// `statusText` and `headers` need only remain valid until send() returns (they can be
// stack-allocated).
virtual kj::Own<WebSocket> acceptWebSocket(const HttpHeaders& headers) = 0;
// If headers.isWebSocket() is true then you can call acceptWebSocket() instead of send().
kj::Promise<void> sendError(uint statusCode, kj::StringPtr statusText,
const HttpHeaders& headers);
kj::Promise<void> sendError(uint statusCode, kj::StringPtr statusText,
const HttpHeaderTable& headerTable);
// Convenience wrapper around send() which sends a basic error. A generic error page specifying
// the error code is sent as the body.
// You must provide headers or a header table because downstream service wrappers may be
// expecting response headers built with a particular table so that they can insert additional
// headers.
virtual kj::Promise<void> request(
HttpMethod method, kj::StringPtr url, const HttpHeaders& headers,
kj::AsyncInputStream& requestBody, Response& response) = 0;
// Perform an HTTP request.
// `url` may be a full URL (with protocol and host) or it may be only the path part of the URL,
// depending on whether the service is a proxy service or a host service.
// `url` and `headers` are invalidated on the first read from `requestBody` or when the returned
// promise resolves, whichever comes first.
// Request processing can be canceled by dropping the returned promise. HttpServer may do so if
// the client disconnects prematurely.
virtual kj::Promise<kj::Own<kj::AsyncIoStream>> connect(kj::StringPtr host);
// Handles CONNECT requests. Only relevant for proxy services. Default implementation throws
class HttpClientErrorHandler {
virtual HttpClient::Response handleProtocolError(HttpHeaders::ProtocolError protocolError);
// Override this function to customize error handling when the client receives an HTTP message
// that fails to parse. The default implementations throws an exception.
// There are two main use cases for overriding this:
// 1. `protocolError` contains the actual header content that failed to parse, giving you the
// opportunity to log it for debugging purposes. The default implementation throws away this
// content.
// 2. You could potentially convert protocol errors into HTTP error codes, e.g. 502 Bad Gateway.
// Note that `protocolError` may contain pointers into buffers that are no longer valid once
// this method returns; you will have to make copies if you want to keep them.
virtual HttpClient::WebSocketResponse handleWebSocketProtocolError(
HttpHeaders::ProtocolError protocolError);
// Like handleProtocolError() but for WebSocket requests. The default implementation calls
// handleProtocolError() and converts the Response to WebSocketResponse. There is probably very
// little reason to override this.
struct HttpClientSettings {
kj::Duration idleTimeout = 5 * kj::SECONDS;
// For clients which automatically create new connections, any connection idle for at least this
// long will be closed. Set this to 0 to prevent connection reuse entirely.
kj::Maybe<EntropySource&> entropySource = nullptr;
// Must be provided in order to use `openWebSocket`. If you don't need WebSockets, this can be
// omitted. The WebSocket protocol uses random values to avoid triggering flaws (including
// security flaws) in certain HTTP proxy software. Specifically, entropy is used to generate the
// `Sec-WebSocket-Key` header and to generate frame masks. If you know that there are no broken
// or vulnerable proxies between you and the server, you can provide a dummy entropy source that
// doesn't generate real entropy (e.g. returning the same value every time). Otherwise, you must
// provide a cryptographically-random entropy source.
kj::Maybe<HttpClientErrorHandler&> errorHandler = nullptr;
// Customize how protocol errors are handled by the HttpClient. If null, HttpClientErrorHandler's
// default implementation will be used.
kj::Own<HttpClient> newHttpClient(kj::Timer& timer, const HttpHeaderTable& responseHeaderTable,
kj::Network& network, kj::Maybe<kj::Network&> tlsNetwork,
HttpClientSettings settings = HttpClientSettings());
// Creates a proxy HttpClient that connects to hosts over the given network. The URL must always
// be an absolute URL; the host is parsed from the URL. This implementation will automatically
// add an appropriate Host header (and convert the URL to just a path) once it has connected.
// Note that if you wish to route traffic through an HTTP proxy server rather than connect to
// remote hosts directly, you should use the form of newHttpClient() that takes a NetworkAddress,
// and supply the proxy's address.
// `responseHeaderTable` is used when parsing HTTP responses. Requests can use any header table.
// `tlsNetwork` is required to support HTTPS destination URLs. If null, only HTTP URLs can be
// fetched.
kj::Own<HttpClient> newHttpClient(kj::Timer& timer, const HttpHeaderTable& responseHeaderTable,
kj::NetworkAddress& addr,
HttpClientSettings settings = HttpClientSettings());
// Creates an HttpClient that always connects to the given address no matter what URL is requested.
// The client will open and close connections as needed. It will attempt to reuse connections for
// multiple requests but will not send a new request before the previous response on the same
// connection has completed, as doing so can result in head-of-line blocking issues. The client may
// be used as a proxy client or a host client depending on whether the peer is operating as
// a proxy. (Hint: This is the best kind of client to use when routing traffic through an HTTP
// proxy. `addr` should be the address of the proxy, and the proxy itself will resolve remote hosts
// based on the URLs passed to it.)
// `responseHeaderTable` is used when parsing HTTP responses. Requests can use any header table.
kj::Own<HttpClient> newHttpClient(const HttpHeaderTable& responseHeaderTable,
kj::AsyncIoStream& stream,
HttpClientSettings settings = HttpClientSettings());
// Creates an HttpClient that speaks over the given pre-established connection. The client may
// be used as a proxy client or a host client depending on whether the peer is operating as
// a proxy.
// Note that since this client has only one stream to work with, it will try to pipeline all
// requests on this stream. If one request or response has an I/O failure, all subsequent requests
// fail as well. If the destination server chooses to close the connection after a response,
// subsequent requests will fail. If a response takes a long time, it blocks subsequent responses.
// If a WebSocket is opened successfully, all subsequent requests fail.
kj::Own<HttpClient> newConcurrencyLimitingHttpClient(
HttpClient& inner, uint maxConcurrentRequests,
kj::Function<void(uint runningCount, uint pendingCount)> countChangedCallback);
// Creates an HttpClient that is limited to a maximum number of concurrent requests. Additional
// requests are queued, to be opened only after an open request completes. `countChangedCallback`
// is called when a new connection is opened or enqueued and when an open connection is closed,
// passing the number of open and pending connections.
kj::Own<HttpClient> newHttpClient(HttpService& service);
kj::Own<HttpService> newHttpService(HttpClient& client);
// Adapts an HttpClient to an HttpService and vice versa.
kj::Own<HttpInputStream> newHttpInputStream(
kj::AsyncInputStream& input, const HttpHeaderTable& headerTable);
// Create an HttpInputStream on top of the given stream. Normally applications would not call this
// directly, but it can be useful for implementing protocols that aren't quite HTTP but use similar
// message delimiting.
// The HttpInputStream implementation does read-ahead buffering on `input`. Therefore, when the
// HttpInputStream is destroyed, some data read from `input` may be lost, so it's not possible to
// continue reading from `input` in a reliable way.
kj::Own<WebSocket> newWebSocket(kj::Own<kj::AsyncIoStream> stream,
kj::Maybe<EntropySource&> maskEntropySource);
// Create a new WebSocket on top of the given stream. It is assumed that the HTTP -> WebSocket
// upgrade handshake has already occurred (or is not needed), and messages can immediately be
// sent and received on the stream. Normally applications would not call this directly.
// `maskEntropySource` is used to generate cryptographically-random frame masks. If null, outgoing
// frames will not be masked. Servers are required NOT to mask their outgoing frames, but clients
// ARE required to do so. So, on the client side, you MUST specify an entropy source. The mask
// must be crytographically random if the data being sent on the WebSocket may be malicious. The
// purpose of the mask is to prevent badly-written HTTP proxies from interpreting "things that look
// like HTTP requests" in a message as being actual HTTP requests, which could result in cache
// poisoning. See RFC6455 section 10.3.
struct WebSocketPipe {
kj::Own<WebSocket> ends[2];
WebSocketPipe newWebSocketPipe();
// Create a WebSocket pipe. Messages written to one end of the pipe will be readable from the other
// end. No buffering occurs -- a message send does not complete until a corresponding receive
// accepts the message.
class HttpServerErrorHandler;
class HttpServerCallbacks;
struct HttpServerSettings {
kj::Duration headerTimeout = 15 * kj::SECONDS;
// After initial connection open, or after receiving the first byte of a pipelined request,
// the client must send the complete request within this time.
kj::Duration pipelineTimeout = 5 * kj::SECONDS;
// After one request/response completes, we'll wait up to this long for a pipelined request to
// arrive.
kj::Duration canceledUploadGracePeriod = 1 * kj::SECONDS;
size_t canceledUploadGraceBytes = 65536;
// If the HttpService sends a response and returns without having read the entire request body,
// then we have to decide whether to close the connection or wait for the client to finish the
// request so that it can pipeline the next one. We'll give them a grace period defined by the
// above two values -- if they hit either one, we'll close the socket, but if the request
// completes, we'll let the connection stay open to handle more requests.
kj::Maybe<HttpServerErrorHandler&> errorHandler = nullptr;
// Customize how client protocol errors and service application exceptions are handled by the
// HttpServer. If null, HttpServerErrorHandler's default implementation will be used.
kj::Maybe<HttpServerCallbacks&> callbacks = nullptr;
// Additional optional callbacks used to control some server behavior.
class HttpServerErrorHandler {
virtual kj::Promise<void> handleClientProtocolError(
HttpHeaders::ProtocolError protocolError, kj::HttpService::Response& response);
virtual kj::Promise<void> handleApplicationError(
kj::Exception exception, kj::Maybe<kj::HttpService::Response&> response);
virtual kj::Promise<void> handleNoResponse(kj::HttpService::Response& response);
// Override these functions to customize error handling during the request/response cycle.
// Client protocol errors arise when the server receives an HTTP message that fails to parse. As
// such, HttpService::request() will not have been called yet, and the handler is always
// guaranteed an opportunity to send a response. The default implementation of
// handleClientProtocolError() replies with a 400 Bad Request response.
// Application errors arise when HttpService::request() throws an exception. The default
// implementation of handleApplicationError() maps the following exception types to HTTP statuses,
// and generates bodies from the stringified exceptions:
// - OVERLOADED: 503 Service Unavailable
// - UNIMPLEMENTED: 501 Not Implemented
// - DISCONNECTED: (no response)
// - FAILED: 500 Internal Server Error
// No-response errors occur when HttpService::request() allows its promise to settle before
// sending a response. The default implementation of handleNoResponse() replies with a 500
// Internal Server Error response.
// Unlike `HttpService::request()`, when calling `response.send()` in the context of one of these
// functions, a "Connection: close" header will be added, and the connection will be closed.
// Also unlike `HttpService::request()`, it is okay to return kj::READY_NOW without calling
// `response.send()`. In this case, no response will be sent, and the connection will be closed.
virtual void handleListenLoopException(kj::Exception&& exception);
// Override this function to customize error handling for individual connections in the
// `listenHttp()` overload which accepts a ConnectionReceiver reference.
// The default handler uses KJ_LOG() to log the exception as an error.
class HttpServerCallbacks {
virtual bool shouldClose() { return false; }
// Whenever the HttpServer begins response headers, it will check `shouldClose()` to decide
// whether to send a `Connection: close` header and close the connection.
// This can be useful e.g. if the server has too many connections open and wants to shed some
// of them. Note that to implement graceful shutdown of a server, you should use
// `HttpServer::drain()` instead.
class HttpServer final: private kj::TaskSet::ErrorHandler {
// Class which listens for requests on ports or connections and sends them to an HttpService.
typedef HttpServerSettings Settings;
typedef kj::Function<kj::Own<HttpService>(kj::AsyncIoStream&)> HttpServiceFactory;
class SuspendableRequest;
typedef kj::Function<kj::Maybe<kj::Own<HttpService>>(SuspendableRequest&)>
HttpServer(kj::Timer& timer, const HttpHeaderTable& requestHeaderTable, HttpService& service,
Settings settings = Settings());
// Set up an HttpServer that directs incoming connections to the given service. The service
// may be a host service or a proxy service depending on whether you are intending to implement
// an HTTP server or an HTTP proxy.
HttpServer(kj::Timer& timer, const HttpHeaderTable& requestHeaderTable,
HttpServiceFactory serviceFactory, Settings settings = Settings());
// Like the other constructor, but allows a new HttpService object to be used for each
// connection, based on the connection object. This is particularly useful for capturing the
// client's IP address and injecting it as a header.
kj::Promise<void> drain();
// Stop accepting new connections or new requests on existing connections. Finish any requests
// that are already executing, then close the connections. Returns once no more requests are
// in-flight.
kj::Promise<void> listenHttp(kj::ConnectionReceiver& port);
// Accepts HTTP connections on the given port and directs them to the handler.
// The returned promise never completes normally. It may throw if port.accept() throws. Dropping
// the returned promise will cause the server to stop listening on the port, but already-open
// connections will continue to be served. Destroy the whole HttpServer to cancel all I/O.
kj::Promise<void> listenHttp(kj::Own<kj::AsyncIoStream> connection);
// Reads HTTP requests from the given connection and directs them to the handler. A successful
// completion of the promise indicates that all requests received on the connection resulted in
// a complete response, and the client closed the connection gracefully or drain() was called.
// The promise throws if an unparseable request is received or if some I/O error occurs. Dropping
// the returned promise will cancel all I/O on the connection and cancel any in-flight requests.
kj::Promise<bool> listenHttpCleanDrain(kj::AsyncIoStream& connection);
// Like listenHttp(), but allows you to potentially drain the server without closing connections.
// The returned promise resolves to `true` if the connection has been left in a state where a
// new HttpServer could potentially accept further requests from it. If `false`, then the
// connection is either in an inconsistent state or already completed a closing handshake; the
// caller should close it without any further reads/writes. Note this only ever returns `true`
// if you called `drain()` -- otherwise this server would keep handling the connection.
class SuspendedRequest {
// SuspendedRequest is a representation of a request immediately after parsing the method line and
// headers. You can obtain one of these by suspending a request by calling
// SuspendableRequest::suspend(), then later resume the request with another call to
// listenHttpCleanDrain().
// Nothing, this is an opaque type.
SuspendedRequest(kj::Array<byte>, kj::ArrayPtr<byte>, HttpMethod, kj::StringPtr, HttpHeaders);
kj::Array<byte> buffer;
// A buffer containing at least the request's method, URL, and headers, and possibly content
// thereafter.
kj::ArrayPtr<byte> leftover;
// Pointer to the end of the request headers. If this has a non-zero length, then our buffer
// contains additional content, presumably the head of the request body.
HttpMethod method;
kj::StringPtr url;
HttpHeaders headers;
// Parsed request front matter. `url` and `headers` both store pointers into `buffer`.
friend class HttpServer;
kj::Promise<bool> listenHttpCleanDrain(kj::AsyncIoStream& connection,
SuspendableHttpServiceFactory factory,
kj::Maybe<SuspendedRequest> suspendedRequest = nullptr);
// Like listenHttpCleanDrain(), but allows you to suspend requests.
// When this overload is in use, the HttpServer's default HttpService or HttpServiceFactory is not
// used. Instead, the HttpServer reads the request method line and headers, then calls `factory`
// with a SuspendableRequest representing the request parsed so far. The factory may then return
// a kj::Own<HttpService> for that specific request, or it may call SuspendableRequest::suspend()
// and return nullptr. (It is an error for the factory to return nullptr without also calling
// suspend(); this will result in a rejected listenHttpCleanDrain() promise.)
// If the factory chooses to suspend, the listenHttpCleanDrain() promise is resolved with false
// at the earliest opportunity.
// SuspendableRequest::suspend() returns a SuspendedRequest. You can resume this request later by
// calling this same listenHttpCleanDrain() overload with the original connection stream, and the
// SuspendedRequest in question.
// This overload of listenHttpCleanDrain() implements draining, as documented above. Note that the
// returned promise will resolve to false (not clean) if a request is suspended.
class Connection;
kj::Timer& timer;
const HttpHeaderTable& requestHeaderTable;
kj::OneOf<HttpService*, HttpServiceFactory> service;
Settings settings;
bool draining = false;
kj::ForkedPromise<void> onDrain;
kj::Own<kj::PromiseFulfiller<void>> drainFulfiller;
uint connectionCount = 0;
kj::Maybe<kj::Own<kj::PromiseFulfiller<void>>> zeroConnectionsFulfiller;
kj::TaskSet tasks;
HttpServer(kj::Timer& timer, const HttpHeaderTable& requestHeaderTable,
kj::OneOf<HttpService*, HttpServiceFactory> service,
Settings settings, kj::PromiseFulfillerPair<void> paf);
kj::Promise<void> listenLoop(kj::ConnectionReceiver& port);
void taskFailed(kj::Exception&& exception) override;
class HttpServer::SuspendableRequest {
// Interface passed to the SuspendableHttpServiceFactory parameter of listenHttpCleanDrain().
HttpMethod method;
kj::StringPtr url;
const HttpHeaders& headers;
// Parsed request front matter, so the implementer can decide whether to suspend the request.
SuspendedRequest suspend();
// Signal to the HttpServer that the current request loop should be exited. Return a
// SuspendedRequest, containing HTTP method, URL, and headers access, along with the actual header
// buffer. The request can be later resumed with a call to listenHttpCleanDrain() using the same
// connection.
explicit SuspendableRequest(
Connection& connection, HttpMethod method, kj::StringPtr url, const HttpHeaders& headers)
: method(method), url(url), headers(headers), connection(connection) {}
Connection& connection;
friend class Connection;
// =======================================================================================
// inline implementation
inline void HttpHeaderId::requireFrom(const HttpHeaderTable& table) const {
KJ_IREQUIRE(this->table == nullptr || this->table == &table,
"the provided HttpHeaderId is from the wrong HttpHeaderTable");
inline kj::Own<HttpHeaderTable> HttpHeaderTable::Builder::build() { return kj::mv(table); }
inline HttpHeaderTable& HttpHeaderTable::Builder::getFutureTable() { return *table; }
inline uint HttpHeaderTable::idCount() const { return namesById.size(); }
inline kj::StringPtr HttpHeaderTable::idToString(HttpHeaderId id) const {
return namesById[];
inline kj::Maybe<kj::StringPtr> HttpHeaders::get(HttpHeaderId id) const {
auto result = indexedHeaders[];
return result == nullptr ? kj::Maybe<kj::StringPtr>(nullptr) : result;
inline void HttpHeaders::unset(HttpHeaderId id) {
indexedHeaders[] = nullptr;
template <typename Func>
inline void HttpHeaders::forEach(Func&& func) const {
for (auto i: kj::indices(indexedHeaders)) {
if (indexedHeaders[i] != nullptr) {
func(table->idToString(HttpHeaderId(table, i)), indexedHeaders[i]);
for (auto& header: unindexedHeaders) {
func(, header.value);
template <typename Func1, typename Func2>
inline void HttpHeaders::forEach(Func1&& func1, Func2&& func2) const {
for (auto i: kj::indices(indexedHeaders)) {
if (indexedHeaders[i] != nullptr) {
func1(HttpHeaderId(table, i), indexedHeaders[i]);
for (auto& header: unindexedHeaders) {
func2(, header.value);
} // namespace kj