KestrelServer详解[3]: 自定义一个迷你版的KestrelServer

和所有的服务器一样，KestrelServer最终需要解决的是网络传输的问题。在《网络连接的创建》，我们介绍了KestrelServer如何利用连接接听器的建立网络连接，并再次基础上演示了如何直接利用建立的连接接收请求和回复响应。本篇更进一步，我们根据其总体设计，定义了迷你版的KestrelServer让读者看看这个重要的服务器大体是如何实现的。

和所有的服务器一样，KestrelServer最终需要解决的是网络传输的问题。在《网络连接的创建》，我们介绍了KestrelServer如何利用连接接听器的建立网络连接，并再次基础上演示了如何直接利用建立的连接接收请求和回复响应。本篇更进一步，我们根据其总体设计，定义了迷你版的KestrelServer让读者看看这个重要的服务器大体是如何实现的。本文提供的示例演示已经同步到《ASP.NET Core 6框架揭秘-实例演示版》）

一、ConnectionDelegate
二、IConnectionBuilder
三、HTTP 1.x/HTTP 2.x V.S. HTTP 3
四、MiniKestrelServer

一、ConnectionDelegate

ASP.NET CORE在“应用”层将针对请求的处理抽象成由中间件构建的管道，实际上KestrelServer面向“传输”层的连接也采用了这样的设计。当代表连接的ConnectionContext上下文创建出来之后，后续的处理将交给由连接中间件构建的管道进行处理。我们可以根据需要注册任意的中间件来处理连接，比如可以将并发连结的控制实现在专门的连接中间件中。ASP.NET CORE管道利用RequestDelegate委托来表示请求处理器，连接管道同样定义了如下这个ConnectionDelegate委托。

public delegate Task ConnectionDelegate(ConnectionContext connection);

二、IConnectionBuilder

ASP.NET CORE管道中的中间件体现为一个Func<RequestDelegate, RequestDelegate>委托，连接管道的中间件同样可以利用Func<ConnectionDelegate, ConnectionDelegate>委托来表示。ASP.NET CORE管道中的中间件注册到IApplicationBuilder对象上并利用它将管道构建出来。连接管道依然具有如下这个IConnectionBuilder接口，ConnectionBuilder实现了该接口。

public interface IConnectionBuilder
{
    IServiceProvider ApplicationServices { get; }
    IConnectionBuilder Use(Func<ConnectionDelegate, ConnectionDelegate> middleware);
    ConnectionDelegate Build();
}

public class ConnectionBuilder : IConnectionBuilder
{
    public IServiceProvider ApplicationServices { get; }
    public ConnectionDelegate Build();
    public IConnectionBuilder Use(Func<ConnectionDelegate, ConnectionDelegate> middleware);
}

IConnectionBuilder接口还定义了如下三个扩展方法来注册连接中间件。第一个Use方法使用Func<ConnectionContext, Func<Task>, Task>委托来表示中间件。其余两个方法用来注册管道末端的中间件，这样的中间件本质上就是一个ConnectionDelegate委托，我们可以将其定义成一个派生于ConnectionHandler的类型。

public static class ConnectionBuilderExtensions
{
    public static IConnectionBuilder Use(this IConnectionBuilder connectionBuilder,Func<ConnectionContext, Func<Task>, Task> middleware);
    public static IConnectionBuilder Run(this IConnectionBuilder connectionBuilder,Func<ConnectionContext, Task> middleware);
    public static IConnectionBuilder UseConnectionHandler<TConnectionHandler>(this IConnectionBuilder connectionBuilder) where TConnectionHandler : ConnectionHandler;
}

public abstract class ConnectionHandler
{
    public abstract Task OnConnectedAsync(ConnectionContext connection);
}

三、HTTP 1.x/HTTP 2.x V.S. HTTP 3

KestrelServer针对HTTP 1.X/2和HTTP 3的设计和实现基本上独立的，这一点从监听器的定义就可以看出来。就连接管道来说，基于HTTP 3的多路复用连接通过MultiplexedConnectionContext表示，它也具有“配套”的MultiplexedConnectionDelegate委托和IMultiplexedConnectionBuilder接口。ListenOptions类型同时实现了IConnectionBuilder和IMultiplexedConnectionBuilder接口，意味着我们在注册终结点的时候还可以注册任意中间件。

public delegate Task MultiplexedConnectionDelegate(MultiplexedConnectionContext connection);

public interface IMultiplexedConnectionBuilder
{
    IServiceProvider ApplicationServices { get; }
    IMultiplexedConnectionBuilder Use(Func<MultiplexedConnectionDelegate, MultiplexedConnectionDelegate> middleware);
    MultiplexedConnectionDelegate Build();
}

public class MultiplexedConnectionBuilder : IMultiplexedConnectionBuilder
{
    public IServiceProvider ApplicationServices { get; }
    public IMultiplexedConnectionBuilder Use(Func<MultiplexedConnectionDelegate, MultiplexedConnectionDelegate> middleware);
    public MultiplexedConnectionDelegate Build();
}

public class ListenOptions : IConnectionBuilder, IMultiplexedConnectionBuilder

四、MiniKestrelServer

在了解了KestrelServer的连接管道后，我们来简单模拟一下这种服务器类型的实现，为此我们定义了一个名为MiniKestrelServer的服务器类型。简单起见，MiniKestrelServer只提供针对HTTP 1.1的支持。对于任何一个服务来说，它需要将请求交付给一个IHttpApplication<TContext>对象进行处理，MiniKestrelServer将这项工作实现在如下这个HostedApplication<TContext>类型中。

public class HostedApplication<TContext> : ConnectionHandler where TContext : notnull
{
    private readonly IHttpApplication<TContext> _application;
    public HostedApplication(IHttpApplication<TContext> application) => _application = application;

    public override async Task OnConnectedAsync(ConnectionContext connection)
    {
        var reader = connection!.Transport.Input;
        while (true)
        {
            var result = await reader.ReadAsync();
            using (var body = new MemoryStream())
            {
                var (features, request, response) = CreateFeatures(result, body);
                var closeConnection = request.Headers.TryGetValue("Connection", out var vallue) && vallue == "Close";
                reader.AdvanceTo(result.Buffer.End);

                var context = _application.CreateContext(features);
                Exception? exception = null;
                try
                {
                    await _application.ProcessRequestAsync(context);
                    await ApplyResponseAsync(connection, response, body);
                }
                catch (Exception ex)
                {
                    exception = ex;
                }
                finally
                {
                    _application.DisposeContext(context, exception);
                }
                if (closeConnection)
                {
                    await connection.DisposeAsync();
                    return;
                }
            }
            if (result.IsCompleted)
            {
                break;
            }
        }

        static (IFeatureCollection, IHttpRequestFeature, IHttpResponseFeature) CreateFeatures(ReadResult result, Stream body)
        {
            var handler = new HttpParserHandler();
            var parserHandler = new HttpParser(handler);
            var length = (int)result.Buffer.Length;
            var array = ArrayPool<byte>.Shared.Rent(length);
            try
            {
                result.Buffer.CopyTo(array);
                parserHandler.Execute(new ArraySegment<byte>(array, 0, length));
            }
            finally
            {
                ArrayPool<byte>.Shared.Return(array);
            }
            var bodyFeature = new StreamBodyFeature(body);

            var features = new FeatureCollection();
            var responseFeature = new HttpResponseFeature();
            features.Set<IHttpRequestFeature>(handler.Request);
            features.Set<IHttpResponseFeature>(responseFeature);
            features.Set<IHttpResponseBodyFeature>(bodyFeature);

            return (features, handler.Request, responseFeature);
        }

        static async Task ApplyResponseAsync(ConnectionContext connection, IHttpResponseFeature response, Stream body)
        {
            var builder = new StringBuilder();
            builder.AppendLine($"HTTP/1.1 {response.StatusCode} {response.ReasonPhrase}");
            foreach (var kv in response.Headers)
            {
                builder.AppendLine($"{kv.Key}: {kv.Value}");
            }
            builder.AppendLine($"Content-Length: {body.Length}");
            builder.AppendLine();
            var bytes = Encoding.UTF8.GetBytes(builder.ToString());

            var writer = connection.Transport.Output;
            await writer.WriteAsync(bytes);
            body.Position = 0;
            await body.CopyToAsync(writer);
        }
    }
}

HostedApplication<TContext>是对一个IHttpApplication<TContext>对象的封装。它派生于抽象类ConnectionHandler，重写的OnConnectedAsync方法将针对请求的读取和处理置于一个无限循环中。为了将读取的请求转交给IHostedApplication<TContext>对象进行处理，它需要根据特性集合将TContext上下文创建出来。这里提供的特性集合只包含三种核心的特性，一个是描述请求的HttpRequestFeature特性，它是利用HttpParser解析请求荷载内容得到的。另一个是描述响应的HttpResponseFeature特性，至于提供响应主体的特性由如下所示的StreamBodyFeature对象来表示。这三个特性的创建实现在CreateFeatures方法中。

public class StreamBodyFeature : IHttpResponseBodyFeature
{
    public Stream 	Stream { get; }
    public PipeWriter 	Writer { get; }

    public StreamBodyFeature(Stream stream)
    {
        Stream = stream;
        Writer = PipeWriter.Create(Stream);
    }

    public Task CompleteAsync() => Task.CompletedTask;
    public void DisableBuffering() { }
    public Task SendFileAsync(string path, long offset, long? count,
    CancellationToken cancellationToken = default)=> throw new NotImplementedException();
    public Task StartAsync(CancellationToken cancellationToken = default) => Task.CompletedTask;
}

包含三大特性的集合随后作为参数调用了IHostedApplication<TContext>对象的CreateContext方法将TContext上下文创建出来，此上下文作为参数传入了同一对象的ProcessRequestAsync方法，此时中间件管道接管请求。待中间件管道完成处理后， ApplyResponseAsync方法被调用以完成最终的响应工作。ApplyResponseAsync方法将响应状态从HttpResponseFeature特性中提取并生成首行响应内容（“HTTP/1.1 {StatusCode} {ReasonPhrase}”），然后再从这个特性中将响应报头提取出来并生成相应的文本。响应报文的首行内容和报头文本按照UTF-8编码生成二进制数组后利用ConnectionContext上下文的Transport属性返回的IDuplexPipe对象发送出去后，它再将StreamBodyFeature特性收集到的响应主体输出流“拷贝”到这个IDuplexPipe对象中，进而完成了针对响应主体内容的输出。

如下所示的是MiniKestrelServer类型的完整定义。该类型的构造函数中注入了用于提供配置选项的IOptions<KestrelServerOptions>特性和IConnectionListenerFactory工厂，并且创建了一个ServerAddressesFeature对象并注册到Features属性返回的特性集合中。

public class MiniKestrelServer : IServer
{
    private readonly KestrelServerOptions _options;
    private readonly IConnectionListenerFactory _factory;
    private readonly List<IConnectionListener> _listeners = new();

    public IFeatureCollection Features { get; } = new FeatureCollection();

    public MiniKestrelServer(IOptions<KestrelServerOptions> optionsAccessor, IConnectionListenerFactory factory)
    {
        _factory = factory;
        _options = optionsAccessor.Value;
        Features.Set<IServerAddressesFeature>(new ServerAddressesFeature());
    }

    public void Dispose() => StopAsync(CancellationToken.None).GetAwaiter().GetResult();
    public Task StartAsync<TContext>(IHttpApplication<TContext> application, CancellationToken cancellationToken) where TContext : notnull
    {
        var feature = Features.Get<IServerAddressesFeature>()!;
        IEnumerable<ListenOptions> listenOptions;
        if (feature.PreferHostingUrls)
        {
            listenOptions = BuildListenOptions(feature);
        }
        else
        {
            listenOptions = _options.GetListenOptions();
            if (!listenOptions.Any())
            {
                listenOptions = BuildListenOptions(feature);
            }
        }

        foreach (var options in listenOptions)
        {
            _ = StartAsync(options);
        }
        return Task.CompletedTask;

        async Task StartAsync(ListenOptions litenOptions)
        {
            var listener = await _factory.BindAsync(litenOptions.EndPoint,cancellationToken);
            _listeners.Add(listener!);

            var hostedApplication = new HostedApplication<TContext>(application);
            var pipeline = litenOptions.Use(next => context => hostedApplication.OnConnectedAsync(context)).Build();

            while (true)
            {
                var connection = await listener.AcceptAsync();
                if (connection != null)
                {
                    _ = pipeline(connection);
                }
            }
        }

        IEnumerable<ListenOptions> BuildListenOptions(IServerAddressesFeature feature)
        {
            var options = new KestrelServerOptions();
            foreach (var address in feature.Addresses)
            {
                var url = new Uri(address);
                if (string.Compare("localhost", url.Host, true) == 0)
                {
                    options.ListenLocalhost(url.Port);
                }
                else
                {
                    options.Listen(IPAddress.Parse(url.Host), url.Port);
                }

            }
            return options.GetListenOptions();
        }
    }

    public Task StopAsync(CancellationToken cancellationToken) => Task.WhenAll(_listeners.Select(it => it.DisposeAsync().AsTask()));
}

实现的StartAsync<TContext>方法先将IServerAddressesFeature特性提取出来，并利用其PreferHostingUrls属性决定应该使用直接注册到KestrelOptions配置选项上的终结点还是使用注册在该特定上的监听地址。如果使用后者，注册的监听地址会利用BuildListenOptions方法转换成对应的ListenOptions列表，否则直接从KestrelOptions对象的ListenOptions属性提取所有的ListenOptions列表，由于这是一个内部属性，不得不利用如下这个扩展方法以反射的方式获取这个列表。

public static class KestrelServerOptionsExtensions
{
    public static IEnumerable<ListenOptions> GetListenOptions(this KestrelServerOptions options)
    {
        var property = typeof(KestrelServerOptions).GetProperty("ListenOptions",BindingFlags.NonPublic | BindingFlags.Instance);
        return (IEnumerable<ListenOptions>)property!.GetValue(options)!;
    }
}

对于每一个表示注册终结点的ListenOptions配置选项，StartAsync<TContext>方法利用IConnectionListenerFactory工厂将对应的IConnectionListener监听器创建出来，并绑定到指定的终结点上监听连接请求。表示连接的ConnectionContext上下文一旦被创建出来后，该方法便会利用构建的连接管道对它进行处理。在调用ListenOptions配置选项的Build方法构建连接管道前，StartAsync<TContext>方法将HostedApplication<TContext>对象创建出来并作为中间件进行了注册。所以针对连接的处理将被这个HostedApplication<TContext>对象接管。

using App;
using Microsoft.AspNetCore.Hosting.Server;
using Microsoft.Extensions.DependencyInjection.Extensions;

var builder = WebApplication.CreateBuilder();
builder.WebHost.UseKestrel(kestrel => kestrel.ListenLocalhost(5000));
builder.Services.Replace(ServiceDescriptor.Singleton<IServer, MiniKestrelServer>());
var app = builder.Build();
app.Run(context => context.Response.WriteAsync("Hello World!"));
app.Run();

如上所示的演示程序将替换了针对IServer的服务注册，意味着默认的KestrelServer将被替换成自定义的MiniKestrelServer。启动该程序后，由浏览器发送的HTTP请求（不支持HTTPS）同样会被正常处理，并得到如图18-6所示的响应内容。需要强调一下，MiniKestrelServer仅仅用来模拟KestrelServer的实现原理，不要觉得真实的实现会如此简单。

图1 由MiniKestrelServer回复的响应内容