ASP.NET Core 5-Kestrel源码解读（asp.net core cors）

上节讲到了kestrel服务器的配置及使用，相信很多同学已经对kestrel服务器有了初步的了解，那么有的同学可能会想更加深入的了解一下Kestrel服务器的是怎么实现监听和接收http请求的，今天我们来看下Kestrel服务器的源码，相信看完这些，你一定会对Kestrel服务器的运行机制有更深入的了解。

首先，让我们从程序启动类Program.cs开始分析。

public class Program
{
    public static void Main(string[] args)
    {
        CreateHostBuilder(args).Build().Run();
    }
 
    public static IHostBuilder CreateHostBuilder(string[] args) =>
        Host.CreateDefaultBuilder(args)
            .ConfigureWebHostDefaults(webBuilder =>
             {
                 webBuilder.UseStartup<Startup>();
             });
}

其中，Host类链式调用了两个方法：

• CreateDefaultBuilder
• ConfigureWebHostDefaults

首先我们来看下CreateDefaultBuidler方法：

public static IHostBuilder CreateDefaultBuilder(string[] args)
{
      HostBuilder hostBuilder = new HostBuilder();
      hostBuilder.UseContentRoot(Directory.GetCurrentDirectory());
      hostBuilder.ConfigureHostConfiguration((Action<IConfigurationBuilder>) (config =>
      {
        ...
      }));
      hostBuilder.ConfigureAppConfiguration((Action<HostBuilderContext, IConfigurationBuilder>) ((hostingContext, config) =>
      {
        ...
      })).ConfigureLogging((Action<HostBuilderContext, ILoggingBuilder>) ((hostingContext, logging) =>
      {
        ...
      })).UseDefaultServiceProvider((Action<HostBuilderContext, ServiceProviderOptions>) ((context, options) =>
      {
        ...
      }));
      return (IHostBuilder) hostBuilder;
    }
 }

从上述代码可以看出，CreateDefaultBuilder并未涉及Kestrel服务器相关代码，仅仅是进行一些应用的初始化配置，例如，设置应用程序目录，设置配置文件等操作。

我们再来看下ConfigureWebHostDefaults方法：

public static IHostBuilder ConfigureWebHostDefaults(
      this IHostBuilder builder,
      Action<IWebHostBuilder> configure)
{
      if (configure == null)
        throw new ArgumentNullException(nameof (configure));
      return builder.ConfigureWebHost((Action<IWebHostBuilder>) (webHostBuilder =>
      {
        Microsoft.AspNetCore.WebHost.ConfigureWebDefaults(webHostBuilder);
        configure(webHostBuilder);
      }));
}

通过阅读源码可以发现： ConfigureWebHostDefaults方法中的Microsoft.AspNetCore.WebHost.ConfigureWebDefaults(IWebHostBuilder)为实际执行初始化Kestrel服务器的代码。

internal static void ConfigureWebDefaults(IWebHostBuilder builder)
{
     ...
     builder.UseKestrel((Action<WebHostBuilderContext, KestrelServerOptions>) ((builderContext, options) => options.Configure((IConfiguration) builderContext.Configuration.GetSection("Kestrel"), true))).ConfigureServices((Action<WebHostBuilderContext, IServiceCollection>) ((hostingContext, services) =>
     {
       services.PostConfigure<HostFilteringOptions>((Action<HostFilteringOptions>) (options =>
       {
        ...
       }
     })).UseIIS().UseIISIntegration();
}

看到这里，可能有的同学已经的迫不及待的想要看下Kestrel初始化流程相关的代码了。别着急，我们一步一步来。

首先我们查看一下上面的UseKestrel扩展方法：

public static IWebHostBuilder UseKestrel(
      this IWebHostBuilder hostBuilder,
      Action<WebHostBuilderContext, KestrelServerOptions> configureOptions)
    {
      return hostBuilder.UseKestrel().ConfigureKestrel(configureOptions);
    }

发现该方法只是对传入的配置项KestrelServerOptions做了封装，最终是调用了IWebHostBuilder的扩展方法UseKestrel和ConfigureKestrel(Action<WebHostBuilderContext, KestrelServerOptions> configureOptions)扩展方法来初始化Kestrel服务器配置，同样是链式调用。

现在我们来看下UseKestrel()这个扩展方法：

public static IWebHostBuilder UseKestrel(this IWebHostBuilder hostBuilder)
{
    return hostBuilder.ConfigureServices((Action<IServiceCollection>) (services =>
    {
      services.TryAddSingleton<IConnectionListenerFactory, SocketTransportFactory>();
      services.AddTransient<IConfigureOptions<KestrelServerOptions>, KestrelServerOptionsSetup>();
      services.AddSingleton<IServer, KestrelServerImpl>();
    }));
}

细心的同学可能会发现，配置一个Kestrel服务器居然只需要仅仅三行代码？是不是感觉有些不可思议？Kestrel服务器这么简单？是的，Kestrel服务器就是这么简单。那么，Kestrel服务器是如何实现监听和接收请求的呢？

首先看下IConnectionListenerFactory接口类：

public interface IConnectionListenerFactory
{
    ValueTask<IConnectionListener> BindAsync(
      EndPoint endpoint,
      CancellationToken cancellationToken = default (CancellationToken));
}

这个接口职责只有一个，就是执行Sokcert的绑定EndPoint操作，然后返回一个IConnectionListener对象。EndPoint可以有三种实现：

• FileHandleEndPoint
• UnixDomainSocketEndPoint
• IPEndPoint

我们再来看下实现类SocketTransportFactory：

public sealed class SocketTransportFactory : IConnectionListenerFactory
{
    private readonly SocketTransportOptions _options;
    private readonly SocketsTrace _trace;
    public SocketTransportFactory(
      IOptions<SocketTransportOptions> options,
      ILoggerFactory loggerFactory)
    {
      if (options == null)
        throw new ArgumentNullException(nameof (options));
      if (loggerFactory == null)
        throw new ArgumentNullException(nameof (loggerFactory));
      this._options = options.Value;
      this._trace = new SocketsTrace(loggerFactory.CreateLogger("Microsoft.AspNetCore.Server.Kestrel.Transport.Sockets"));
    }
    public ValueTask<IConnectionListener> BindAsync(
      EndPoint endpoint,
      CancellationToken cancellationToken = default (CancellationToken))
    {
      SocketConnectionListener connectionListener = new SocketConnectionListener(endpoint, this._options, (ISocketsTrace) this._trace);
      connectionListener.Bind();
      return new ValueTask<IConnectionListener>((IConnectionListener) connectionListener);
    }
}

代码非常简单，先实例化SocketConnectionListener对象，然后调用SocketConnectionListener的Bind方法并根据传入的EndPoint类型来创建Socket对象，来实现对EndPoint的监听和绑定操作。

internal void Bind()
{
    if (this._listenSocket != null)
      throw new InvalidOperationException(SocketsStrings.TransportAlreadyBound);
    Socket listenSocket;
    switch (this.EndPoint)
    {
      case FileHandleEndPoint fileHandleEndPoint:
        this._socketHandle = new SafeSocketHandle((IntPtr) (long) fileHandleEndPoint.FileHandle, true);
        listenSocket = new Socket(this._socketHandle);
        break;
      case UnixDomainSocketEndPoint domainSocketEndPoint:
        listenSocket = new Socket(domainSocketEndPoint.AddressFamily, SocketType.Stream, ProtocolType.IP);
        BindSocket();
        break;
      case IPEndPoint ipEndPoint:
        listenSocket = new Socket(ipEndPoint.AddressFamily, SocketType.Stream, ProtocolType.Tcp);
        if (ipEndPoint.Address == IPAddress.IPv6Any)
          listenSocket.DualMode = true;
        BindSocket();
        break;
      default:
        listenSocket = new Socket(this.EndPoint.AddressFamily, SocketType.Stream, ProtocolType.Tcp);
        BindSocket();
        break;
    }
    this.EndPoint = listenSocket.LocalEndPoint;
    listenSocket.Listen(this._options.Backlog);
    this._listenSocket = listenSocket;
    void BindSocket()
    {
      try
      {
        listenSocket.Bind(this.EndPoint);
      }
      catch (SocketException ex) when (ex.SocketErrorCode == SocketError.AddressAlreadyInUse)
      {
        throw new AddressInUseException(ex.Message, (Exception) ex);
      }
    }
}

现在我们已经知道了Kestrel服务器内部是如何进行绑定和监听操作。那么Kestrel服务器是如何对http请求进行接收处理的呢？

接下来我们来看IServer接口：

public interface IServer : IDisposable
{
    IFeatureCollection Features { get; }
    Task StartAsync<TContext>(IHttpApplication<TContext> application, CancellationToken cancellationToken) where TContext : notnull;
    Task StopAsync(CancellationToken cancellationToken);
}

IServer接口也非常简单，定义了一个Server最基本的有两个功能：启动和停止。那么Kestrel服务器是怎么实现的这个接口呢？

下面我们来看下微软官方为IServer注入的实现类KestrelServerImpl:

internal class KestrelServerImpl : IServer
{
    ...
    public IFeatureCollection Features { get; }
    public KestrelServerOptions Options => ServiceContext.ServerOptions;
    private ServiceContext ServiceContext { get; }
    private IKestrelTrace Trace => ServiceContext.Log;
    private AddressBindContext AddressBindContext { get; set; }
    public async Task StartAsync<TContext>(IHttpApplication<TContext> application, CancellationToken cancellationToken)
    {
        ...
        async Task OnBind(ListenOptions options)
        {
            if (!BitConverter.IsLittleEndian)
            {
                throw new PlatformNotSupportedException(CoreStrings.BigEndianNotSupported);
            }
            ValidateOptions();
            if (_hasStarted)
            {
                    // The server has already started and/or has not been cleaned up yet
                throw new InvalidOperationException(CoreStrings.ServerAlreadyStarted);
            }
            _hasStarted = true;

            ServiceContext.Heartbeat?.Start();
            if ((options.Protocols & HttpProtocols.Http3) == HttpProtocols.Http3)
            {
                if (_multiplexedTransportFactory is null)
                {
                    throw new InvalidOperationException(#34;Cannot start HTTP/3 server if no {nameof(IMultiplexedConnectionListenerFactory)} is registered.");
                }
 
                options.UseHttp3Server(ServiceContext, application, options.Protocols);
                var multiplexedConnectionDelegate = ((IMultiplexedConnectionBuilder)options).Build();
 
                multiplexedConnectionDelegate = EnforceConnectionLimit(multiplexedConnectionDelegate, Options.Limits.MaxConcurrentConnections, Trace);
                options.EndPoint = await _transportManager.BindAsync(options.EndPoint, multiplexedConnectionDelegate, options.EndpointConfig).ConfigureAwait(false);
            }
 
            if ((options.Protocols & HttpProtocols.Http1) == HttpProtocols.Http1
                || (options.Protocols & HttpProtocols.Http2) == HttpProtocols.Http2
                || options.Protocols == HttpProtocols.None) // TODO a test fails because it doesn't throw an exception in the right place
                                                                // when there is no HttpProtocols in KestrelServer, can we remove/change the test?
            {
               if (_transportFactory is null)
                {
                    throw new InvalidOperationException(#34;Cannot start HTTP/1.x or HTTP/2 server if no {nameof(IConnectionListenerFactory)} is registered.");
                }
                options.UseHttpServer(ServiceContext, application, options.Protocols);
                var connectionDelegate = options.Build();
                connectionDelegate = EnforceConnectionLimit(connectionDelegate, Options.Limits.MaxConcurrentConnections, Trace);
                options.EndPoint = await _transportManager.BindAsync(options.EndPoint, connectionDelegate, options.EndpointConfig).ConfigureAwait(false);
            }
         }

         AddressBindContext = new AddressBindContext
         {
             ServerAddressesFeature = _serverAddresses,
             ServerOptions = Options,
             Logger = Trace,
             CreateBinding = OnBind,
         };
         await BindAsync(cancellationToken).ConfigureAwait(false);
         ...
    }
 
    public async Task StopAsync(CancellationToken cancellationToken)
    {
            ...
    }
    ...
    private async Task BindAsync(CancellationToken cancellationToken)
    {
             ...
         await AddressBinder.BindAsync(Options.ListenOptions, AddressBindContext).ConfigureAwait(false);
             ...
    }
    ...
}

我们来整理一下StartAsync方法的流程：

1. 字节序校验：不支持BigEndian
2. 请求参数长度校验，最大8kb
3. 判断服务器是否已经启动过
4. 启动心跳检测
5. 实例化AddressBindContext用于BindAsync方法使用
6. 执行BindAsync方法来绑定地址操作

BindAsync调用了AddressBindContext的OnBind方法。OnBind方法会根据使用的http协议类型创建不同的HttpConnectionMiddleware中间件并加入到connection管道中，用于处理Http请求。

具体规则如下：

• 当协议是HttpProtocols.Http1/2时，创建HttpConnectionMiddleware中间件
• 当协议是HttpProtocols.Http3时，创建Http3ConnectionMiddleware中间件

目前常用的是HttpConnectionMiddleware：

IConnectionBuilder UseHttpServer<TContext>(
      this IConnectionBuilder builder,
      ServiceContext serviceContext,
      IHttpApplication<TContext> application,
      HttpProtocols protocols)
    {
      HttpConnectionMiddleware<TContext> middleware = new HttpConnectionMiddleware<TContext>(serviceContext, application, protocols);
      return builder.Use((Func<ConnectionDelegate, ConnectionDelegate>) (next => new ConnectionDelegate(middleware.OnConnectionAsync)));
    }

UseHttpServer方法为connection管道(注意不是IApplicationBuilder中的请求管道)添加了一个HttpConnectionmiddleware中间件，当请求到达时，会执行OnConnectionAsync方法来创建HttpConnection对象，然后通过该对象处理http请求：

public Task OnConnectionAsync(ConnectionContext connectionContext)
{
     IMemoryPoolFeature memoryPoolFeature = connectionContext.Features.Get<IMemoryPoolFeature>();
     HttpConnectionContext context = new HttpConnectionContext();
     context.ConnectionId = connectionContext.ConnectionId;
     context.ConnectionContext = connectionContext;
     HttpProtocolsFeature protocolsFeature = connectionContext.Features.Get<HttpProtocolsFeature>();
     context.Protocols = protocolsFeature != null ? protocolsFeature.HttpProtocols : this._endpointDefaultProtocols;
     context.ServiceContext = this._serviceContext;
     context.ConnectionFeatures = connectionContext.Features;
     context.MemoryPool = memoryPoolFeature?.MemoryPool ?? MemoryPool<byte>.Shared;
     context.Transport = connectionContext.Transport;
     context.LocalEndPoint = connectionContext.LocalEndPoint as IPEndPoint;
     context.RemoteEndPoint = connectionContext.RemoteEndPoint as IPEndPoint;
     return new HttpConnection(context).ProcessRequestsAsync<TContext>(this._application);
}

ProcessRequestsAsync为具体的处理请求的方法，此方法会根据使用的http协议版本来创建Http1Connection还是Http2Connection，然后使用此httpConnection来创建context对象(注意不是HttpContext对象)。

Kestrel服务器对请求的接收是通过OnBind里面的TransportManager.BindAsync来实现的。

public async Task<EndPoint> BindAsync(
      EndPoint endPoint,
      ConnectionDelegate connectionDelegate,
      EndpointConfig? endpointConfig)
{
     if (this._transportFactory == null)
       throw new InvalidOperationException("Cannot bind with ConnectionDelegate no IConnectionListenerFactory is registered.");
     IConnectionListener connectionListener = await this._transportFactory.BindAsync(endPoint).ConfigureAwait(false);
     this.StartAcceptLoop<ConnectionContext>((IConnectionListener<ConnectionContext>) new TransportManager.GenericConnectionListener(connectionListener), (Func<ConnectionContext, Task>) (c => connectionDelegate(c)), endpointConfig);
     return connectionListener.EndPoint;

其中StartAcceptLoop方法为实际接收数据的方法，通过方法名“开始循环接收”，我们猜测，是不是Kestrel服务器是通过对Socket的Accept方法进行循环监听来接收数据的？那么到底是不是呢？让我们来继续跟踪一下connectionDispatcher.StartAcceptingConnections方法：

public Task StartAcceptingConnections(IConnectionListener<T> listener)
{
     ThreadPool.UnsafeQueueUserWorkItem<IConnectionListener<T>>(new Action<IConnectionListener<T>>(this.StartAcceptingConnectionsCore), listener, false);
     return this._acceptLoopTcs.Task;
}
private void StartAcceptingConnectionsCore(IConnectionListener<T> listener)
{
     AcceptConnectionsAsync();
 
     async Task AcceptConnectionsAsync()
     {
       try
       {
         while (true)
         {
           T connectionContext = await listener.AcceptAsync(new CancellationToken());
           if ((object) connectionContext != null)
           {
             long id = Interlocked.Increment(ref ConnectionDispatcher<T>._lastConnectionId);
             KestrelConnection<T> kestrelConnection = new KestrelConnection<T>(id, this._serviceContext, this._transportConnectionManager, this._connectionDelegate, connectionContext, this.Log);
             this._transportConnectionManager.AddConnection(id, (KestrelConnection) kestrelConnection);
             this.Log.ConnectionAccepted(connectionContext.ConnectionId);
             KestrelEventSource.Log.ConnectionQueuedStart((BaseConnectionContext) connectionContext);
             ThreadPool.UnsafeQueueUserWorkItem((IThreadPoolWorkItem) kestrelConnection, false);
           }
           else
             break;
         }
       }
       catch (Exception ex)
       {
         this.Log.LogCritical((EventId) 0, ex, "The connection listener failed to accept any new connections.");
       }
       finally
       {
         this._acceptLoopTcs.TrySetResult();
       }
  }
}

相信现在大家已经了解是怎么回事了吧？原来Kestrel服务器是通过while(true)循环接收的方式接收用户请求数据，然后通过线程池的ThreadPool.UnsafeQueueUserWorkItem方法将请求分发到CLR线程池来处理的。换句话说，在请求到来时，TransportManager将OnBind方法加入线程池并待CLR线程池调度。

那么回到开始的时候，Kestrel服务器是如何启动的呢？

让我们再回顾一下Program.cs中的方法

public static void Main(string[] args)
{
   CreateHostBuilder(args).Build().Run();
}

相信聪明的同学已经猜到了，是通过Run()方法来执行的，Run()方法做了些什么呢？

Run方法实际上是执行了Host类中的StartAsync方法，此方法通过获取预先注入的GenericeWebHostService类中注入的IServer类来最终调用到IServer实现类的StartAsnyc方法的。

internal class GenericWebHostService : IHostedService
{
  ...
  public IServer Server { get; }
  ...
	public async Task StartAsync(CancellationToken cancellationToken)
	{
     ...
     var httpApplication = new HostingApplication(application, Logger, DiagnosticListener, HttpContextFactory);
     await Server.StartAsync(httpApplication, cancellationToken);
     ...
	}
}

至此，Kestrel成功启动并开始监听用户请求。

一句话总结：其实ASP.NET Core 5中的Kestrel服务器只是对Socket的简单封装，简单到直接用socket通过while(true)的方式来循环接收socket请求，并直接放入clr线程池中来等待线程池调度处理。

原来，Kestrel服务器这么简单~

相信通过本文的介绍，大家已经对ASP.NET Core 5中的Kestrel服务器有了解了吧？

还在等什么，赶快收藏加关注吧！更多精彩内容还在等着你！