前言
本文编写时源码参考github仓库主分支。
aspnetcore提供了Use方法供开发者自定义中间件，该方法接收一个委托对象，该委托接收一个RequestDelegate对象，并返回一个RequestDelegate对象，方法定义如下：
IApplicationBuilder Use(Func<RequestDelegate, RequestDelegate> middleware);委托RequestDelegate的定义
/// <summary>
/// A function that can process an HTTP request.
/// </summary>
/// <param name="context">The <see cref="HttpContext"/> for the request.</param>
/// <returns>A task that represents the completion of request processing.</returns>
public delegate Task RequestDelegate(HttpContext context);如果我们直接使用IApplicationBuilder.Use来写中间件逻辑，可以使用lamda表达式来简化代码，如下：
app.Use((RequestDelegate next) =>
{return (HttpContext ctx) =>{// do your logicreturn next(ctx);};
});如果写一些简单的逻辑，这种方式最为方便，问题是如果需要写的中间件代码比较多，依然这样去写，会导致我们Program.cs文件代码非常多，如果有多个中间件，那么最后我们的的Program.cs文件包含多个中间件代码，看上去十分混乱。
将中间件逻辑独立出来
为了解决我们上面的代码不优雅，我们希望能将每个中间件业务独立成一个文件，多个中间件代码不混乱的搞到一起。我们需要这样做。
单独的中间件文件
// Middleware1.cs
public class Middleware1
{public static RequestDelegate Logic(RequestDelegate requestDelegate){return (HttpContext ctx) =>{// do your logicreturn requestDelegate(ctx);};}
}调用中间件
app.Use(Middleware1.Logic);
// 以下是其他中间件示例
app.Use(Middleware2.Logic);
app.Use(Middleware3.Logic);
app.Use(Middleware4.Logic);这种方式可以很好的将各个中间件逻辑独立出来，Program.cs此时变得十分简洁，然而我们还不满足这样，因为我们的Logic方法中直接返回一个lamada表达式（RequestDelegate对象），代码层级深了一层，每个中间件都多写这一层壳似乎不太优雅，能不能去掉这层lamada表达式呢？
UseMiddlewareExtensions
为了解决上面提到的痛点，UseMiddlewareExtensions扩展类应运而生，它在Aspnetcore底层大量使用,它主要提供一个泛型UseMiddleware<T>方法用来方便我们注册中间件，下面是该方法的定义
public static IApplicationBuilder UseMiddleware<TMiddleware>(this IApplicationBuilder app, params object?[] args)如果只看这个方法的声明，估计没人知道如何使用，因为该方法接收的泛型参数TMiddleware没有添加任何限制，而另一个args参数也是object类型，而且是可以不传的，也就是它只需要传任意一个类型都不会在编译时报错。 比如这样，完全不会报错：
当然，如果你这样就运行程序，一定会收到下面的异常
System.InvalidOperationException:“No public 'Invoke' or 'InvokeAsync' method found for middleware of type 'System.String'.”提示我们传的类型没有Invoke或InvokeAsync公共方法，这里大概能猜到，底层应该是通过反射进行动态调用Invoke或InvokeAsync公共方法的。
源码分析
想要知道其本质，唯有查看源码,以下源码来自UseMiddlewareExtensions
如下，该扩展类一共提供两个并且是重载的公共方法UseMiddleware，一般都只会使用第一个UseMiddleware,第一个UseMiddleware方法内部再去调用第二个UseMiddleware方法，源码中对类型前面添加的[DynamicallyAccessedMembers(MiddlewareAccessibility)]属性可以忽略，它的作用是为了告诉编译器我们通过反射访问的范围，以防止对程序集对我们可能调用的方法或属性等进行裁剪。
internal const string InvokeMethodName = "Invoke";
internal const string InvokeAsyncMethodName = "InvokeAsync";/// <summary>
/// Adds a middleware type to the application's request pipeline.
/// </summary>
/// <typeparam name="TMiddleware">The middleware type.</typeparam>
/// <param name="app">The <see cref="IApplicationBuilder"/> instance.</param>
/// <param name="args">The arguments to pass to the middleware type instance's constructor.</param>
/// <returns>The <see cref="IApplicationBuilder"/> instance.</returns>
public static IApplicationBuilder UseMiddleware<[DynamicallyAccessedMembers(MiddlewareAccessibility)] TMiddleware>(this IApplicationBuilder app, params object?[] args)
{return app.UseMiddleware(typeof(TMiddleware), args);
}/// <summary>
/// Adds a middleware type to the application's request pipeline.
/// </summary>
/// <param name="app">The <see cref="IApplicationBuilder"/> instance.</param>
/// <param name="middleware">The middleware type.</param>
/// <param name="args">The arguments to pass to the middleware type instance's constructor.</param>
/// <returns>The <see cref="IApplicationBuilder"/> instance.</returns>
public static IApplicationBuilder UseMiddleware(this IApplicationBuilder app,[DynamicallyAccessedMembers(MiddlewareAccessibility)] Type middleware,params object?[] args)
{if (typeof(IMiddleware).IsAssignableFrom(middleware)){// IMiddleware doesn't support passing args directly since it's// activated from the containerif (args.Length > 0){throw new NotSupportedException(Resources.FormatException_UseMiddlewareExplicitArgumentsNotSupported(typeof(IMiddleware)));}return UseMiddlewareInterface(app, middleware);}var applicationServices = app.ApplicationServices;var methods = middleware.GetMethods(BindingFlags.Instance | BindingFlags.Public);MethodInfo? invokeMethod = null;foreach (var method in methods){if (string.Equals(method.Name, InvokeMethodName, StringComparison.Ordinal) || string.Equals(method.Name, InvokeAsyncMethodName, StringComparison.Ordinal)){if (invokeMethod is not null){throw new InvalidOperationException(Resources.FormatException_UseMiddleMutlipleInvokes(InvokeMethodName, InvokeAsyncMethodName));}invokeMethod = method;}}if (invokeMethod is null){throw new InvalidOperationException(Resources.FormatException_UseMiddlewareNoInvokeMethod(InvokeMethodName, InvokeAsyncMethodName, middleware));}if (!typeof(Task).IsAssignableFrom(invokeMethod.ReturnType)){throw new InvalidOperationException(Resources.FormatException_UseMiddlewareNonTaskReturnType(InvokeMethodName, InvokeAsyncMethodName, nameof(Task)));}var parameters = invokeMethod.GetParameters();if (parameters.Length == 0 || parameters[0].ParameterType != typeof(HttpContext)){throw new InvalidOperationException(Resources.FormatException_UseMiddlewareNoParameters(InvokeMethodName, InvokeAsyncMethodName, nameof(HttpContext)));}var state = new InvokeMiddlewareState(middleware);return app.Use(next =>{var middleware = state.Middleware;var ctorArgs = new object[args.Length + 1];ctorArgs[0] = next;Array.Copy(args, 0, ctorArgs, 1, args.Length);var instance = ActivatorUtilities.CreateInstance(app.ApplicationServices, middleware, ctorArgs);if (parameters.Length == 1){return (RequestDelegate)invokeMethod.CreateDelegate(typeof(RequestDelegate), instance);}var factory = Compile<object>(invokeMethod, parameters);return context =>{var serviceProvider = context.RequestServices ?? applicationServices;if (serviceProvider == null){throw new InvalidOperationException(Resources.FormatException_UseMiddlewareIServiceProviderNotAvailable(nameof(IServiceProvider)));}return factory(instance, context, serviceProvider);};});
}第一个UseMiddleware可以直接跳过，看第二个UseMiddleware方法，该方法一上来就先判断我们传的泛型类型是不是IMiddleware接口的派生类，如果是，直接交给UseMiddlewareInterface方法。
if (typeof(IMiddleware).IsAssignableFrom(middleware)){// IMiddleware doesn't support passing args directly since it's// activated from the containerif (args.Length > 0){throw new NotSupportedException(Resources.FormatException_UseMiddlewareExplicitArgumentsNotSupported(typeof(IMiddleware)));}return UseMiddlewareInterface(app, middleware);}这里总算看到应该有的东西了，如果声明UseMiddleware<T>方法时，对泛型T添加IMiddleware限制,我们不看源码就知道如何编写我们的中间件逻辑了，只需要写一个类，继承IMiddleware并实现InvokeAsync方法即可， UseMiddlewareInterface方法的实现比较简单，因为我们继承了接口，逻辑相对会简单点。
private static IApplicationBuilder UseMiddlewareInterface(IApplicationBuilder app,Type middlewareType)
{return app.Use(next =>{return async context =>{var middlewareFactory = (IMiddlewareFactory?)context.RequestServices.GetService(typeof(IMiddlewareFactory));if (middlewareFactory == null){// No middleware factorythrow new InvalidOperationException(Resources.FormatException_UseMiddlewareNoMiddlewareFactory(typeof(IMiddlewareFactory)));}var middleware = middlewareFactory.Create(middlewareType);if (middleware == null){// The factory returned null, it's a broken implementationthrow new InvalidOperationException(Resources.FormatException_UseMiddlewareUnableToCreateMiddleware(middlewareFactory.GetType(), middlewareType));}try{await middleware.InvokeAsync(context, next);}finally{middlewareFactory.Release(middleware);}};});
}public interface IMiddleware
{/// <summary>/// Request handling method./// </summary>/// <param name="context">The <see cref="HttpContext"/> for the current request.</param>/// <param name="next">The delegate representing the remaining middleware in the request pipeline.</param>/// <returns>A <see cref="Task"/> that represents the execution of this middleware.</returns>Task InvokeAsync(HttpContext context, RequestDelegate next);
}如果我们的类不满足IMiddleware，继续往下看
通过反射查找泛型类中Invoke或InvokeAsync方法
var applicationServices = app.ApplicationServices;
var methods = middleware.GetMethods(BindingFlags.Instance | BindingFlags.Public);
MethodInfo? invokeMethod = null;
foreach (var method in methods)
{if (string.Equals(method.Name, InvokeMethodName, StringComparison.Ordinal) || string.Equals(method.Name, InvokeAsyncMethodName, StringComparison.Ordinal)){// 如果Invoke和InvokeAsync同时存在，则抛出异常，也就是，我们只能二选一if (invokeMethod is not null){throw new InvalidOperationException(Resources.FormatException_UseMiddleMutlipleInvokes(InvokeMethodName, InvokeAsyncMethodName));}invokeMethod = method;}
}// 如果找不到Invoke和InvokeAsync则抛出异常，上文提到的那个异常。
if (invokeMethod is null)
{throw new InvalidOperationException(Resources.FormatException_UseMiddlewareNoInvokeMethod(InvokeMethodName, InvokeAsyncMethodName, middleware));
}// 如果Invoke和InvokeAsync方法的返回值不是Task或Task的派生类，则抛出异常
if (!typeof(Task).IsAssignableFrom(invokeMethod.ReturnType))
{throw new InvalidOperationException(Resources.FormatException_UseMiddlewareNonTaskReturnType(InvokeMethodName, InvokeAsyncMethodName, nameof(Task)));
}
Snippet// 如果Invoke和InvokeAsync方法没有参数，或第一个参数不是HttpContext，抛异常
var parameters = invokeMethod.GetParameters();
if (parameters.Length == 0 || parameters[0].ParameterType != typeof(HttpContext))
{throw new InvalidOperationException(Resources.FormatException_UseMiddlewareNoParameters(InvokeMethodName, InvokeAsyncMethodName, nameof(HttpContext)));
}上面一堆逻辑主要就是检查我们的Invoke和InvokeAsync方法是否符合要求,即：必须是接收HttpContext参数，返回Task对象，这恰好就是委托RequestDelegate的定义。
构造RequestDelegate这部分源码的解读都注释到相应的位置了，如下
var state = new InvokeMiddlewareState(middleware);
// 调用Use函数，向管道中注册中间件
return app.Use(next =>
{var middleware = state.Middleware;var ctorArgs = new object[args.Length + 1];// next是RequestDelegate对象，作为构造函数的第一个参数传入ctorArgs[0] = next;Array.Copy(args, 0, ctorArgs, 1, args.Length);// 反射实例化我们传入的泛型类，并把next和args作为构造函数的参数传入var instance = ActivatorUtilities.CreateInstance(app.ApplicationServices, middleware, ctorArgs);// 如果我们的Invoke方法只有一个参数，则直接创建该方法的委托if (parameters.Length == 1){return (RequestDelegate)invokeMethod.CreateDelegate(typeof(RequestDelegate), instance);}// 当Invoke方法不止一个参数HttpContext，通过Compile函数创建动态表达式目录树，// 表达式目录树的构造此处略过，其目的是实现将除第一个参数的其他参数通过IOC注入var factory = Compile<object>(invokeMethod, parameters);return context =>{// 获取serviceProvider用于在上面构造的表达式目录树中实现依赖注入var serviceProvider = context.RequestServices ?? applicationServices;if (serviceProvider == null){throw new InvalidOperationException(Resources.FormatException_UseMiddlewareIServiceProviderNotAvailable(nameof(IServiceProvider)));}// 将所需的参数传入构造的表达式目录树工厂return factory(instance, context, serviceProvider);};
});至此，整个扩展类的源码就解读完了。
通过UseMiddleware注入自定义中间件
通过上面的源码解读，我们知道了其实我们传入的泛型类型是有严格的要求的，主要有两种
通过继承IMiddleware
继承IMiddleware并实现该接口的InvokeAsync函数
public class Middleware1 : IMiddleware
{public async Task InvokeAsync(HttpContext context, RequestDelegate next){// do your logicawait next(context);}
}通过反射
我们知道，在不继承IMiddleware的情况下，底层会通过反射实例化泛型类型，并通过构造函数传入RequestDelegate,而且要有一个公共函数Invoke或InvokeAsync，并且接收的第一个参数是HttpContext，返回Task,根据要求我们将Middleware1.cs改造如下
public class Middleware1
{RequestDelegate next;public Middleware1(RequestDelegate next){this.next = next;}public async Task Invoke(HttpContext httpContext){// do your logicawait this.next(httpContext);}
}总结
通过源码的学习，我们弄清楚底层注册中间件的来龙去脉，两种方式根据自己习惯进行使用，笔者认为通过接口的方式更加简洁直观简单，并且省去了反射带来的性能损失，推荐使用。既然通过继承接口那么爽，为啥还费那么大劲实现反射的方式呢？由源码可知，如果继承接口的话，就不能进行动态传参了。
if (typeof(IMiddleware).IsAssignableFrom(middleware)){// IMiddleware doesn't support passing args directly since it's// activated from the containerif (args.Length > 0){throw new NotSupportedException(Resources.FormatException_UseMiddlewareExplicitArgumentsNotSupported(typeof(IMiddleware)));}return UseMiddlewareInterface(app, middleware);}所以在需要传参的场景，则必须使用反射的方式，所以两种方式都有其存在的必要。
如果本文对您有帮助，还请点赞转发关注一波支持作者。
