Skip to content

Latest commit

 

History

History
538 lines (453 loc) · 26.9 KB

readme.md

File metadata and controls

538 lines (453 loc) · 26.9 KB

http2dotnet

Nuget package: http2dotnet

This library implements the HTTP/2 and HPACK protocol for .NET standard. The goal of the library is to cover all protocol handling parts of the HTTP/2 (RFC7540) and HPACK (RFC7541) specifications. The goal of this library is NOT to provide a ready-to-use HTTP/2 server or client framework with concrete Request/Response abstractions. Instead it should enable other .NET libraries and applications to easily integrate HTTP/2 support by encapsulating the protocol handling in easy to use and flexible .NET classes. Examples for building simple applications on top of the library are provided within the repository.

Current state

This library is currently in an experimental state. The majority of HTTP/2 features have been implemented and there is already a pretty good test coverage. It was tested in a h2c with prior knowledge configuration against nghttp2 and h2load. It was however not yet tested with an encrypted connection and a browser based client due a missing SSL connection with ALPN negotiation. Various features have not yet been implemented (see current limitations).

Design goals

  • Enable an easy integration of HTTP/2 into different frameworks and applications (ASP.NET core, other webservers, gRPC, etc.)
  • Provide an abstract interface for HTTP/2 Streams, on which custom Request and Response classes can be built. The Stream abstraction supports all features of the HTTP/2 specification, including sending HEADERS in both directions, flow-controlled DATA and trailing HEADERS. The Request and Response are handled fully independet, which enables features like pure server side data streaming.
  • IO layer independence:
    This library uses an abstract IO interface, which is defined in the ByteStreams.cs source file.
    The IO layer can be implemented by .NET System.IO.Streams, Pipelines, TCP sockets, TLS sockets and other IO sources.
    The abstract IO interface is inspired by the Go IO interfaces. It was chosen for simplicity, high performance (async reads/writes can be avoided) and because it can easily provide the necessary flow control behavior.
  • Correctness and reliability:
    All HTTP/2 frames are validated according to the HTTP/2 specification.
    Full backpressure behavior is provided for HTTP/2 DATA frames (through the specified flow control mechanism) as well as for other HTTP/2 frames (by suspending the reception of further frames until a suitable response for a frame could be sent).
  • High performance:
    To achieve a high performance and low overhead the library tries to minimize dynamic allocations whereever possible, e.g. through the use of ValueTask structures and reusing of buffers and data structures. If possible the library also avoids async operations by trying nonblocking operations first to avoid scheduling and continuation overhead.
  • Full support of all HTTP/2 features (however not yet everything is implemented)

Usage

The library exposes 2 important classes:

  • Connection represents a HTTP/2 connection between a client and the server.
    HTTP/2 connections are created on top of bidirectional streams. For encrypted HTTP/2 (the only variant support by browser - also called h2) the Connection must be created on top of SSL connection. For unencrypted connections (h2c) the Connection has to be created on top of a TCP stream.
    On top of the connection multiple bidirectional streams can be established.
  • IStream represents a single bidirectional HTTP/2 stream.
    Each Request/Response pair in HTTP/2 is represented by a single Stream.

In the most common lifecycle a HTTP/2 client will create a new Stream on top of the Connection, send all required headers for the Request through the Stream, and then send optional data through the stream. The server will receive a notification about a new create stream and invoke a request handler. The request handler can then read the received headers, read all incoming data from the stream and respond with headers and data.

Besides that there are some advanced features, e.g. both clients and servers can also Cancel/Reset the stream during processing and both are able to send trailing headers at the end of a Stream.

This library does not create the underlying TCP or SSL connections on server or client side. This is the responsibility of the library user. However this pattern allows to use the library on top of arbitrary IO stream types.

To create a Connection on top of IO streams the constructor of Connection can be used:

ConnectionConfiguration config =
    new ConnectionConfigurationBuilder(isServer: true)
    .UseStreamListener(AcceptIncomingStream)
    .Build();

Connection http2Connection = new Connection(
    config: config,
    inputStream: inputStream,
    outputStream: outputStream);

Configuration settings which are shared between multiple connections are configured through the ConnectionConfigurationBuilder which stores them in an ConnectionConfiguration instance. The remaining parameters which are unique for each Connection instance are configured directly thorugh the Connection constructor parameters.

The most important arguments for the Connection are:

  • inputStream/outputStream: The Connection will will use the IO streams to read data from the remote side and send data to the remote side. The connection will take ownership of these streams and close the OutputStream when done.
  • config: The configuration for the connection, which is possibly also shared with other Connection instances.

The most important arguments for the ConnectionConfiguration are:

  • isServer: Specifies whether the local side of the Connection represents a HTTP/2 server (true) or client (false).
  • UseStreamListener: This sets up the callback function that will be invoked every time a new Stream is initiated from the remote side. This means it is currently only required for HTTP/2 server applications. An IStream instance which represents the initiated stream is passed to the callback function as an argument. The callback function should either return true if it wants to process the new stream or false otherwise. If the application wants to process the new stream it must use it in another Task and must not block the callback.
  • UseSettings: The HTTP/2 settings that should be used. Settings.Default represents the default values from the HTTP/2 specification and should usually be a good choice. These are also utilized if not explicitely specified.

Besides that there are various optional arguments which allow further control about the desired behavior.

After the Connection was set up it will handle all HTTP/2 protocol related concerns up to the point where the underlying connection gets closed.

The IStream class allows to read and write headers and bytes for a single HTTP/2 stream. The methods ReadHeadersAsync() and WriteHeadersAsync() allow to read and write headers from a connection. The returned ValueTask from ReadHeadersAsync will only be fulfilled once headers from the remote side where received. According to the HTTP/2 request/reponse lifecycle applications have to send headers at the start of each Stream - which means it is not allowed to send data before any headers have been sent. The received headers will contain the pseudo-headers which are used in HTTP/2 for transmission of the HTTP method (:method), status code (:status), etc. These are not automatically extracted into seperated fields because IStream is used for requests and responses - depending on whether the library is used for a client or server. However higher level frameworks can easily extract the fields from the returned header lists. This library will always validate received and sent headers for correctness. This means a user of the library on the server side can rely on the fact that the :method, :scheme and :path pseudo headers are preset and that after the first real header field no pseudo header field follows.

After headers were sent for a Stream data can be written to the stream. As IStream implements the ByteStream abstractions in this library, the data can be written through the provided WriteAsync function. The returned Task will only be fulfilled after the data could be sent through the underlying connection. This will take the available flow control windows into account. The stream can be closed from the by calling the CloseAsync function. Alternativly the endOfStream parameter of WriteHeadersAsync can be used for closing the Stream without sending any data (e.g. for HTTP GET requests). The read and write directions of HTTP/2 are fully independet. This means closing the local side of the stream will still allow reading headers and data from the remote side. A stream is only fully closed after all headers and data were consumed from the remote side in addition to the local side having been closed. To read data from the stream the ReadAsync function can be used. This function will signal the end of stream if the stream was closed from the remote side. After a stream was closed from the remote side trailers can be read through the ReadTrailersAsync function. To send trailers data must be transferred and instead of closing the stream with CloseAsync or setting endOfStream to true the WriteTrailersAsync function must be used.

The Cancel function on the IStream can be used to reset the stream. This will move the Stream into the Closed state if it wasn't Closed before. In order to notify the remote side a RST_STREAM frame will be sent. All streams must be either fully processed (read side fully consumed - write side closed) or cancelled/reset by the application. Otherwise the Stream will be leaked. The Dispose method on stream will also cancel the stream.

The following example shows the handling of a stream on the server side:

static async void HandleIncomingStream(IStream stream)
{
    try
    {
        // Read the headers
        var headers = await stream.ReadHeadersAsync();
        var method = headers.First(h => h.Name == ":method").Value;
        var path = headers.First(h => h.Name == ":path").Value;
        // Print the request method and path
        Console.WriteLine("Method: {0}, Path: {1}", method, path);

        // Read the request body and write it to console
        var buf = new byte[2048];
        while (true)
        {
            var readResult = await stream.ReadAsync(new ArraySegment<byte>(buf));
            if (readResult.EndOfStream) break;
            // Print the received bytes
            Console.WriteLine(Encoding.ASCII.GetString(buf, 0, readResult.BytesRead));
        }

        // Send a response which consists of headers and a payload
        var responseHeaders = new HeaderField[] {
            new HeaderField { Name = ":status", Value = "200" },
            new HeaderField { Name = "content-type", Value = "text/html" },
        };
        await stream.WriteHeadersAsync(responseHeaders, false);
        await stream.WriteAsync(new ArraySegment<byte>(
            Encoding.ASCII.GetBytes("Hello World!")), true);

        // Request is fully handled here
    }
    catch (Exception)
    {
        stream.Cancel();
    }
}

All IStream APIs are completely thread-safe, which means they can be used from an arbitrary thread. However the user application must still follow the basic HTTP/2 contracts, which e.g. means that now trailers may be sent before data was sent, and that neither data nor trailers might be sent if the end of the stream was already indicated.

Client stream creation

The library can be used to build HTTP/2 clients. For this usecase outgoing IStreams need to be created, which represent the sent request. This can be achieved through the Connection.CreateStreamAsync() method, which creates a new outgoing stream. The user must pass all the required headers (including the pseudo-headers :method, :path, :scheme and optionally :authority) as an argument, as calling the method will directly lead to sending a HTTP/2 headers frame which the requested header fields. The method will return the newly created stream.

Example for creating and using client stream, in case a HTTP/2 connection has already been established and configured for client-side use:

HeaderField[] headers = new HeaderField[]
{
    new HeaderField { Name = ":method", Value = "GET" },
    new HeaderField { Name = ":scheme", Value = "http" },
    new HeaderField { Name = ":path", Value = "/" },
};

var stream = await http2Connection.CreateStreamAsync(
    headers, endOfStream: true);

// Wait for response headers
var reponseHeaders = await stream.ReadHeadersAsync();

// Read response data
var readDataResult = await stream.ReadAsync(buffer);

Checking for stream ID exhaustion

The HTTP/2 specification only allows 2^30 streams to be created on a single Connection. If a client requires to make more requests after some time it needs to create a new Connection. This library allows to check whether a new Connection needs to be created through the Connection.IsExhausted property, which will return true if no additional outgoing stream can be created.

Checking for the maximum concurrent streams limit of the remote side

The library currently does not take the maximum concurrent streams setting into account, which is indicated by a remote through a HTTP/2 settings frame. This means it will try to create any amount of streams that the user requests, with the risk of the remote side resetting those streams. A future update might provide an additional possibility to check the current maximum concurrent stream limit.

Server side connection upgrades

Upgrading over HTTPS

Over HTTPS, the server and client negotiate a protocol using Application Layer Protocol Negotiation. This is done before data is sent, so both the client and server know what protocol to use before handling the connection, so if we instruct the client to use HTTP/2, the connection is already HTTP/2 by the time we get it.

Example code for how an upgrade can be performed over HTTPS can be found in the HttpsExampleServer.

Upgrading from HTTP/1.1

The library supports HTTP connection upgrade requests on server side. Example code for how an upgrade can be performed can be found in the UpgradeExampleServer.

In order to upgrade from HTTP/1.1 to HTTP/2, an external HTTP/1.1 parser is needed which reads the complete upgrade request into memory (including an optional body).

If the headers of the received HTTP/1.1 request contain the related upgrade headers (connection: Upgrade, HTTP2-Settings), they payload can be checked for whether it is a valid upgrade request or not. In order to perform this step a ServerUpgradeRequestBuilder instance must be created, which gets fed all the information (headers, payload, request method, etc.) from the HTTP/1.1 request. With the ServerUpgradeRequestBuilder.build() method a ServerUpgradeRequest can be created, which exposes an IsValid property This property reflects whether the incoming request is a fully valid upgrade request.

Only if the upgrade request is valid a HTTP/2 Connection object may be created, which get's the inital upgrade request injected into the constructor as part of the Connection.Options struct. In this case the Connection will create a HTTP/2 Stream from the content of the initial upgrade request, which can be handled like any other HTTP/2 Stream/Request.

In case the upgrade request is valid, it is the applications responsibility to send HTTP/1.1 101 Switching Protocols response before creating the HTTP/2 connection object - which won't perform this step, and instead directly speak HTTP/2.

In case the upgrade request is not valid, the application may choose to either treat the request as a normal HTTP/1.1 request (by filtering all ugprade related) headers, or by responding with an HTTP error.

Even if the upgrade headers are available and valid in the initial HTTP/1.1 request, applications may choose not to perform the upgrade but instead handle the request as a normal HTTP/1.1 request. An example scenario for this is when the upgrade request indicates a large body (through content-length header), which would needed to be completely read into memory before the upgrade can be performed: If the application doesn't perform the upgrade and just handles the request in traditional fashion, the request body can be streamed in normal HTTP/1.1 fashion. It is generelly not recommended to accept upgrade requests which contain a body payload. However most client side libraries that perform upgrades (e.g. nghttp), will perform the upgrade only via GET and HEAD requests anyway.

Handling HTTP/2 (with and without upgrade) and HTTP/1 in parallel

In some scenarios it might be required to implement a HTTP server that supports all of the following scenarios:

  • Support for standard HTTP/1.1 requests
  • Support for HTTP/2 requests with prior knowledge (The client knows that the server speaks HTTP/2 and no upgrade is required).
  • Support for upgrading from HTTP/1.1 to HTTP/2

This library supports this scenario, and a suitable example implementation can be found in the UpgradeExampleServer example.

The idea is that the application (or web-server which is built around the library) starts off by reading and parsing an initial HTTP/1.1 request header through it's preferred HTTP parser.

  • If the request header exactly equals PRI * HTTP/2.0\r\n\r\n it's the start of a HTTP/2 connection with prior knowledge. In this case the application can directly create a HTTP/2 Connection object on top of the incoming stream and delegate all further processing to it. In this special case, the application must make sure that the stream that is passed to the Connection object still emits the PRI * HTTP/2.0\r\n\r\n header when the Connection starts reading from it. Otherwise the Connection establishment will fail. This means the application must inspect, but not consume this header.
  • If the request is another HTTP/1.1 request which contains ugprade headers, the application can handle it as described in the Upgrading from HTTP/1.1 on server side section. In this case it's the applications responsibility to consume the ugprade request from the underlying stream and send the necessary upgrade status response.
  • If the request header is an HTTP/1.1 request without upgrade headers, the application can directly handle it in it's preferred way.
  • If the incoming data resembles no HTTP request at all, the application can also handle it in it's preferred way, e.g. by optionally sending an HTTP error code and closing the connection.

In the first two cases the actual HTTP requests from the connection will further flow to the application through by the means of this libraries IStream abstraction. In the third case it depends on the approach for HTTP/1.1 parrsing and processing. It's the applications responsibility to transform all request kinds into a common Request/Response representation if needed.

Client side connection upgrades

Upgrading from HTTP/1.1

The library supports HTTP connection upgrade requests on client side in a similar fashion as they are supported on the server side. Example code for how an upgrade can be performed can be found in the CliExample.

In order to upgrade from HTTP/1.1 to HTTP/2, the user has to send a normal HTTP/1.1 request which contains the required Upgrade headers to the server. Sending this request over the underyling TCP connection is outside of the scope of this library. When the HTTP/1.1 response status line and headers are received the user code must check whether the upgrade was successful or not. In the success case a HTTP/2 Connection object can be constructed on top of the Connection. This Connection object obtain the ownership to the underlying streams (from which the HTTP/1.1 data already has been consumed) and an information about that the fact that the Connection was created as part of an Upgrade, because in this case one Stream will already be existing at startup.

For performing client side upgrades, the first step is to create a ClientUpgradeRequestBuilder, which allows to configure the HTTP/2 settings which will be later on used. Through the ClientUpgradeRequestBuilder.build() method a ClientUpgradeRequest can be created. This ClientUpgradeRequest also exposes an IsValid property which reflects whether the upgrade request is valid. Only if the upgrade request is valid a HTTP/2 Connection object may be created.

The ClientUpgradeRequest contains a Base64EncodedSettings property, which returns the base64 encoded settings which the user must send to the server inside the Http2-Settings header during the upgrade attempt.

In case the user received a HTTP/1.1 101 Switching Protocols status a Connection object can be created on top of the underlying streams. The Connection must be informed about the pending upgrade attempt through the Connection.Options.ClientUpgradeRequest constructor option.

The response for the HTTP/1.1 request which triggered the upgrade will in this case be delivered through a HTTP/2 stream, which will always utilize the Stream ID 1. As the stream was created implicitly through the upgrade and not due to calling CreateStreamAsync() the created IStream must be returned to the user in another fashion: The user can retrieve a reference to this first stream through the ClientUpgradeRequest.UpgradeRequestStream property, which returns a Task<IStream> which will be fulfilled once a Connection has been created which used the UpgradeRequest. Due to this fact a ClientUpgradeRequest instance may be not be reused for performing multiple connection upgrade requests. It will be bound to the first Connection to which it gets handed over.

Handling HTTP/2 (with and without upgrade) and HTTP/1 in parallel

In order to handle HTTP/2 on client side with a possible fallback to HTTP/1.1 for servers which do not support HTTP/1.1 the upgrade mechanism can be used: The first request form the client can be started as a HTTP/1.1 request, which contains the connection upgrade request in it's headers. If the server responds with 101 Switching Protocols then a HTTP/2 connection can be established on top of this connection and the response of this request as well as further requests can be performed in HTTP/2 fashion. If the server ignores the upgrade the response must be normally processed with a HTTP/1.1 reader.

Ping handling

The library will automatically respond to received PING frames by sending the associated PING ACKs.

The user can also issue PINGs from the local side of the connection through the PingAsync method of the Connection class. Calling this method will issue sending a PING frame to the remote side of the connection. It returns a task that will be completed once either the remote side acknowledges the PING or the connection was closed. If the connection closes before an ACK was received the returned Task will fail with a ConnectionClosedException.

Example for measuring the connection latency through the ping mechanism:

try
{
    var stopWatch = new Stopwatch();
    stopWatch.Start();
    await http2Connection.PingAsync();
    // Ping request was sent and acknowledge has been received
    stopWatch.Stop();
    var latency = stopWatch.Elapsed;
}
catch (ConnectionClosedException)
{
    // The connection was closed before an acknowledge to the PING
    // was received
}

Informational headers

The library supports sending informational headers (HTTP status code in the 1xy range) on server side and receiving them on client side.

To send informational headers on server side the user can simply call stream.WriteHeadersAsync multiple times. It is possible to send multiple sets of informational headers before the set of headers with the final status code is written. Only the final header set may set endOfStream to true. Users should not write data before a header with a final :status code (outside of the 100 range) was sent - because this would be detected as a protocol error on the client side.

To receive informational headers on the client side the stream.ReadHeadersAsync() may be called multiple times on the client side in the special case where one call returns headers with a status: 1xy field. In these cases the next call to ReadHeadersAsync() will block until the next set of headers (either informational or of the final response) will be received.

Example:

var headers = await stream.ReadHeadersAsync();
if (headers.First(h => h.Name == ":status").Value == "100")
{
    // Received only informational headers
    headers = await stream.ReadHeadersAsync();
    // headers now contains the next set of received headers.
}

Remark:

If the server sends multiple sets of informational headers and possibly the final headers before the user calls stream.ReadHeadersAsync(), this call will only deliver the last delivered set of headers to the user. The headers are not queued internally but overwritten. However for known use-cases of informational headers (the 100 Continue header) this causes no problem, since the server will only send more headers after user interaction (sending data).

Performance considerations

For achieving the best possible performance the following things should be considered:

  • The IO stream abstractions which are handed over to the Connection should be optimized for handling small writes and reads in a fast way. E.g. the library might often try to read or write only a few bytes, like the header of a HTTP/2 frame. In this case the IO operation should operate synchronously if possible. The TCP socket wrappers which are provided inside this library (see SocketExtensions.CreateStreams()) e.g. provide this behavior by trying to perform synchronous nonblocking read and write operations before falling back into async behavior.
  • The library will try to allocate all send and receive buffers from a custom ArrayPool<byte> allocator which can be provided by the user through the ConnectionConfigurationBuilder.UseBufferPool() setting. If no allocator is specified the ArrayPool<byte>.Shared is used, which might not be optimal for HTTP/2 operations. The library will use the allocator to allocate buffers of up to Settings.MaxFrameSize bytes in normal operation. Therefore the configured allocator can be optimized for these kind of buffer sizes.

Current limitations

The library currently faces the following limitations:

  • Missing support for push promises.
  • Missing support for reading the remote SETTINGS from application side.
  • The scheduling of outgoing DATA frames is very basic and relies mostly on flow control windows and the maximum supported frame size. It is currently not guaranteed that concurrent streams with equal flow control windows will get the same amount of bandwith. HTTP/2 priorization features are also not supported - however these are optional according to the HTTP/2 specification and may not be required for lots of applications.