kestrel Server的源码分析
今天这一篇博客讲的是.net core 自带的kestrel server,当你开发微服务k8s部署在linux环境下,一般默认开启这个高性能服务,如果大家之前看过我的owin katana的博客,会发现.net core 的好多实现在之前.net standard 的版本已经实现过了,当时开发的asp.net 程序与IIS紧紧耦合在一起,后来的微软团队意识到这个问题并尝试将asp.net 解耦server,制定了owin标准并启动了一个开源项目katana,这个项目的结果并没有带动社区效应,但是此时微软已经制订了.net core的开发,并在katana文档暗示了.net vnext 版本,这个就是。net core 与owin katana 的故事。强烈建议大家有时间看看owin katana,里面有一些 dependency inject, hash map, raw http 协议等等实现。非常收益。说到这些我们开始步入正题吧。原代码在github上的asp.net core 源码。
上图大致地描述了一个asp.net core 的请求过程,但是我们会发现appication 依赖了server,所以我们需要一个Host 的去解耦server 和aplication 的实现,只要server符合host标准可以任意更换,解耦之后的代码与下图所示。
所以我们的代码都是创建一个web host然后使用usekestrel,如下所示。
var host = new WebHostBuilder() .UseKestrel(options => { options.Listen(IPAddress.Loopback, 5001); }) .UseStartup<Startup>();
首先我们知道一个server 实现需要网络编程,所以我们需要socket库来快速编程,它已经帮你实现了tcp与udp协议,不需要自己重复的造轮子。首先我们需要看UseKestrel的方法做了什么。
1 public static IWebHostBuilder UseKestrel(this IWebHostBuilder hostBuilder) 2 { 3 return hostBuilder.ConfigureServices(services => 4 { 5 // Don't override an already-configured transport 6 services.TryAddSingleton<IConnectionListenerFactory, SocketTransportFactory>(); 7 8 services.AddTransient<IConfigureOptions<KestrelServerOptions>, KestrelServerOptionsSetup>(); 9 services.AddSingleton<IServer, KestrelServer>(); 10 }); 11 }
依赖注入注册了三个对象,一个连接池,一个配置类还有一个是server,会和web host注册了IServer 的实现类,然后我们继续看一下,当你调用run的时候会将控制权从web host 转移给server,如下代码第18行所示。
1 public virtual async Task StartAsync(CancellationToken cancellationToken = default) 2 { 3 HostingEventSource.Log.HostStart(); 6 7 var application = BuildApplication(); 8 12 // Fire IHostedService.Start 13 await _hostedServiceExecutor.StartAsync(cancellationToken).ConfigureAwait(false);//启动后台服务 14 15 var diagnosticSource = _applicationServices.GetRequiredService<DiagnosticListener>(); 16 var httpContextFactory = _applicationServices.GetRequiredService<IHttpContextFactory>(); 17 var hostingApp = new HostingApplication(application, _logger, diagnosticSource, httpContextFactory); 18 await Server.StartAsync(hostingApp, cancellationToken).ConfigureAwait(false);//socket 启动 19 _startedServer = true; 20 21 // Fire IApplicationLifetime.Started 22 _applicationLifetime?.NotifyStarted(); 23 24 25 _logger.Started(); 26 27 // Log the fact that we did load hosting startup assemblies. 28 if (_logger.IsEnabled(LogLevel.Debug)) 29 { 30 foreach (var assembly in _options.GetFinalHostingStartupAssemblies()) 31 { 32 _logger.LogDebug("Loaded hosting startup assembly {assemblyName}", assembly); 33 } 34 } 35 36 if (_hostingStartupErrors != null) 37 { 38 foreach (var exception in _hostingStartupErrors.InnerExceptions) 39 { 40 _logger.HostingStartupAssemblyError(exception); 41 } 42 } 43 }
当我们转进到StartAsync方法时会看到如下代码
1 public async Task StartAsync<TContext>(IHttpApplication<TContext> application, CancellationToken cancellationToken) 2 { 3 try 4 { 19 // ServiceContext.Heartbeat?.Start();//一个是连接池一个是日期时间 20 21 async Task OnBind(ListenOptions options) 22 { 23 // Add the HTTP middleware as the terminal connection middleware 24 options.UseHttpServer(ServiceContext, application, options.Protocols);//注册中间件 25 26 var connectionDelegate = options.Build(); 27 28 // Add the connection limit middleware 29 if (Options.Limits.MaxConcurrentConnections.HasValue) 30 { 31 connectionDelegate = new ConnectionLimitMiddleware(connectionDelegate, Options.Limits.MaxConcurrentConnections.Value, Trace).OnConnectionAsync; 32 } 33 34 var connectionDispatcher = new ConnectionDispatcher(ServiceContext, connectionDelegate); 35 var transport = await _transportFactory.BindAsync(options.EndPoint).ConfigureAwait(false); 36 37 // Update the endpoint 38 options.EndPoint = transport.EndPoint; 39 var acceptLoopTask = connectionDispatcher.StartAcceptingConnections(transport); 40 41 _transports.Add((transport, acceptLoopTask)); 42 } 43 44 await AddressBinder.BindAsync(_serverAddresses, Options, Trace, OnBind).ConfigureAwait(false); 45 } 46 catch (Exception ex) 47 {51 } 52 }
AddressBinder就是server绑定的ip地址,这个可以在StartUp方法或者环境变量里面配置,里面传了一个回调方法OnBind, 在第24行的UseHttpServer会注册server 内部的中间件去处理这个请求,在第35行socet会绑定地址,用tcp协议,默认使用512个最大pending队列,在接受socket会有多处异步编程和开启线程,建议大家在调试的时候可以修改代码用尽可能少的线程来进行调试。accept 的代码如下图所示
1 private void StartAcceptingConnectionsCore(IConnectionListener listener) 2 { 3 // REVIEW: Multiple accept loops in parallel? 4 _ = AcceptConnectionsAsync(); 5 6 async Task AcceptConnectionsAsync() 7 { 8 try 9 { 10 while (true) 11 { 12 var connection = await listener.AcceptAsync(); 13 19 20 // Add the connection to the connection manager before we queue it for execution 21 var id = Interlocked.Increment(ref _lastConnectionId); 22 var kestrelConnection = new KestrelConnection(id, _serviceContext, _connectionDelegate, connection, Log); 23 24 _serviceContext.ConnectionManager.AddConnection(id, kestrelConnection);27 28 ThreadPool.UnsafeQueueUserWorkItem(kestrelConnection, preferLocal: false); 29 } 30 } 31 catch (Exception ex) 32 { 33 // REVIEW: If the accept loop ends should this trigger a server shutdown? It will manifest as a hang 34 Log.LogCritical(0, ex, "The connection listener failed to accept any new connections."); 35 } 36 finally 37 { 38 _acceptLoopTcs.TrySetResult(null); 39 } 40 } 41 }
接收到accept socket的时候,会创建一个kestrelconnection 对象,这个对象实现线程方法,然后它会重新去等待一个请求的到来,而用户代码的执行则交给线程池执行。在第14行就是之前kerstrel server 内部的中间件build生成的方法,他的主要功能就是解析socket的携带http信息。
1 internal async Task ExecuteAsync() 2 { 3 var connectionContext = TransportConnection; 4 5 try 6 { 10 using (BeginConnectionScope(connectionContext)) 11 { 12 try 13 { 14 await _connectionDelegate(connectionContext); 15 } 16 catch (Exception ex) 17 { 18 Logger.LogError(0, ex, "Unhandled exception while processing {ConnectionId}.", connectionContext.ConnectionId); 19 } 20 } 21 } 22 finally 23 { 34 _serviceContext.ConnectionManager.RemoveConnection(_id); 35 } 36 }
由于http协议版本的不一致导致解析方式的不同,如果有兴趣的小伙伴可以具体查看这一块的逻辑。
1 switch (SelectProtocol()) 2 { 3 case HttpProtocols.Http1: 4 // _http1Connection must be initialized before adding the connection to the connection manager 5 requestProcessor = _http1Connection = new Http1Connection<TContext>(_context); 6 _protocolSelectionState = ProtocolSelectionState.Selected; 7 break; 8 case HttpProtocols.Http2: 9 // _http2Connection must be initialized before yielding control to the transport thread, 10 // to prevent a race condition where _http2Connection.Abort() is called just as 11 // _http2Connection is about to be initialized. 12 requestProcessor = new Http2Connection(_context); 13 _protocolSelectionState = ProtocolSelectionState.Selected; 14 break; 15 case HttpProtocols.None: 16 // An error was already logged in SelectProtocol(), but we should close the connection. 17 break; 18 default: 19 // SelectProtocol() only returns Http1, Http2 or None. 20 throw new NotSupportedException($"{nameof(SelectProtocol)} returned something other than Http1, Http2 or None."); 21 }
然后server解析完请求之后所做的重要的一步就是创建httpContext,然后server在第40行将控制权转给web host,web host 会自动调用application code 也就是用户代码。
1 private async Task ProcessRequests<TContext>(IHttpApplication<TContext> application) 2 { 3 while (_keepAlive) 4 { 33 var context = application.CreateContext(this); 34 35 try 36 { 37 KestrelEventSource.Log.RequestStart(this); 38 39 // Run the application code for this request 40 await application.ProcessRequestAsync(context); 41 55 } 56 catch (BadHttpRequestException ex) 57 { 58 // Capture BadHttpRequestException for further processing 59 // This has to be caught here so StatusCode is set properly before disposing the HttpContext 60 // (DisposeContext logs StatusCode). 61 SetBadRequestState(ex); 62 ReportApplicationError(ex); 63 } 64 catch (Exception ex) 65 { 66 ReportApplicationError(ex); 67 } 68 69 KestrelEventSource.Log.RequestStop(this);129 } 130 }
到这里server 的工作大部分都结束了,在之前的描述中我们看到web host 怎么将控制权给到server 的, server 创建好httpContext规则后又是如何将控制权给到web host , web host 又如何去调用application code的, web host 实际上build 的时候将用户的中间件定义为链表结构暴露一个入口供web host调用,其他的有时间我会再写博客描述这一块。谢谢大家今天的阅读了。欢迎大家能够留言一起讨论。最后谢谢大家的阅读,如果有任何不懂的地方可以私信我。