.NET6中的await原理浅析

前言

看过不少关于 await 的原理的文章,也知道背后是编译器给转成了状态机实现的,但是具体是怎么完成的,回调又是如何衔接的,一直都没有搞清楚,这次下定决心把源码自己跑了下,终于豁然开朗了
本文的演示代码基于 VS2022 + .NET 6

示例

public class Program
{
    static int Work()
    {
        Console.WriteLine("In Task.Run");
        return 1;
    }

    static async Task TestAsync()
    {
        Console.WriteLine("Before Task.Run");
        await Task.Run(Work);
        Console.WriteLine("After Task.Run");
    }

    static void Main()
    {
        _ = TestAsync();
        Console.WriteLine("End");
        Console.ReadKey();
    }
}

• 很简单的异步代码,我们来看下,编译器把它变成了啥

class Program
{
    static int Work()
    {
        Console.WriteLine("In Task.Run");
        return 1;
    }

    static Task TestAsync()
    {
        var stateMachine = new StateMachine()
        {
            _builder = AsyncTaskMethodBuilder.Create(),
            _state = -1
        };
        stateMachine._builder.Start(ref stateMachine);
        return stateMachine._builder.Task;
    }

    static void Main()
    {
        _ = TestAsync();
        Console.WriteLine("End");
        Console.ReadKey();
    }

    class StateMachine : IAsyncStateMachine
    {
        public int _state;
        public AsyncTaskMethodBuilder _builder;
        private TaskAwaiter<int> _awaiter;

        void IAsyncStateMachine.MoveNext()
        {
            int num = _state;
            try
            {
                TaskAwaiter<int> awaiter;
                if (num != 0)
                {
                    Console.WriteLine("Before Task.Run");
                    awaiter = Task.Run(Work).GetAwaiter();
                    if (!awaiter.IsCompleted)
                    {
                        _state = 0;
                        _awaiter = awaiter;
                        StateMachine stateMachine = this;
                        _builder.AwaitUnsafeOnCompleted(ref awaiter, ref stateMachine);
                        return;
                    }
                }
                else
                {
                    awaiter = _awaiter;
                    _awaiter = default;
                    _state = -1;
                }
                awaiter.GetResult();
                Console.WriteLine("After Task.Run");
            }
            catch (Exception exception)
            {
                _state = -2;
                _builder.SetException(exception);
                return;
            }
            _state = -2;
            _builder.SetResult();
        }

        void IAsyncStateMachine.SetStateMachine(IAsyncStateMachine stateMachine) { }
    }
}

• 编译后的代码经过我的整理,命名简化了,更容易理解

状态机实现

• 我们看到实际是生成了一个隐藏的状态机类 StateMachine

• 把状态机的初始状态 _state 设置 -1

• stateMachine._builder.Start(ref stateMachine); 启动状态机,内部实际调用的就是状态机的 MoveNext 方法

• Task.Run 创建一个任务, 把委托放在 Task.m_action 字段,丢到线程池,等待调度

• 任务在线程池内被调度完成后,是怎么回到这个状态机继续执行后续代码的呢?
  _builder.AwaitUnsafeOnCompleted(ref awaiter, ref stateMachine); 就是关键了, 跟下去,到了如下的代码:

  if (!this.AddTaskContinuation(stateMachineBox, false))
  {
      ThreadPool.UnsafeQueueUserWorkItemInternal(stateMachineBox, true);
  }
  bool AddTaskContinuation(object tc, bool addBeforeOthers)
  {
      return !this.IsCompleted && ((this.m_continuationObject == null && Interlocked.CompareExchange(ref this.m_continuationObject, tc, null) == null) || this.AddTaskContinuationComplex(tc, addBeforeOthers));
  }

  • 这里很清楚的看到,尝试把状态机对象(实际是状态机的包装类),赋值到 Task.m_continuationObject, 如果操作失败,则把状态机对象丢进线程池等待调度,这里为什么这么实现,看一下线程池是怎么执行的就清楚了

线程池实现

• .NET6 的线程池实现,实际是放到了 PortableThreadPool, 具体调试步骤我就不放了,直接说结果就是, 线程池线程从任务队列中拿到任务后都执行了 DispatchWorkItem 方法

static void DispatchWorkItem(object workItem, Thread currentThread)
{
    Task task = workItem as Task;
    if (task != null)
    {
        task.ExecuteFromThreadPool(currentThread);
        return;
    }
    Unsafe.As<IThreadPoolWorkItem>(workItem).Execute();
}
virtual void ExecuteFromThreadPool(Thread threadPoolThread)
{
    this.ExecuteEntryUnsafe(threadPoolThread);
}

• 我们看到, 线程池队列中的任务都是 object 类型的, 这里进行了类型判断, 如果是 Task , 直接执行 task.ExecuteFromThreadPool, 更有意思的这个方法是个虚方法,后面说明

• ExecuteFromThreadPool 继续追下去,我们来到了这里,代码做了简化

  private void ExecuteWithThreadLocal(ref Task currentTaskSlot, Thread threadPoolThread = null)
  {
      this.InnerInvoke();
      this.Finish(true);
  }
  
  virtual void InnerInvoke()
  {
      Action action = this.m_action as Action;
      if (action != null)
      {
          action();
          return;
      }
  }

• 很明显 this.InnerInvoke 就是执行了最开始 Task.Run(Work) 封装的委托了, 在 m_action 字段

• this.Finish(true); 跟下去会发现会调用 FinishStageTwo 设置任务的完成状态,异常等, 继续调用 FinishStageThree 就来了重点: FinishContinuations 这个方法就是衔接后续回调的核心

  internal void FinishContinuations()
  {
      object obj = Interlocked.Exchange(ref this.m_continuationObject, Task.s_taskCompletionSentinel);
      if (obj != null)
      {
          this.RunContinuations(obj);
      }
  }

• 还记得状态机实现么, Task.m_continuationObject 字段实际存储的就是状态机的包装类,这里线程池线程也会判断这个字段有值的话,就直接使用它执行后续代码了

  void RunContinuations(object continuationObject)
  {
      var asyncStateMachineBox = continuationObject as IAsyncStateMachineBox;
      if (asyncStateMachineBox != null)
      {
          AwaitTaskContinuation.RunOrScheduleAction(asyncStateMachineBox, flag2);
          return;
      }
  }
  
  static void RunOrScheduleAction(IAsyncStateMachineBox box, bool allowInlining)
  {
      if (allowInlining && AwaitTaskContinuation.IsValidLocationForInlining)
      {
          box.MoveNext();
          return;
      }
  }

总结

1. Task.Run 创建 Task, 把委托放在 m_action 字段, 把 Task 压入线程池队列,等待调度
2. _builder.AwaitUnsafeOnCompleted(ref awaiter, ref stateMachine); 尝试把状态机对象放在 Task.m_continuationObject 字段上,等待线程池线程调度完成任务后使用(用来执行后续),若操作失败,直接把状态机对象压入线程池队列,等待调度
3. 线程池线程调度任务完成后,会判断 Task.m_continuationObject 有值,直接执行它的 MoveNext

备注

1. 状态机实现中,尝试修改 Task.m_continuationObject,可能会失败,
   就会直接把状态机对象压入线程池, 但是线程池调度,不都是判断是不是 Task 类型么, 其实状态机的包装类是 Task 的子类,哈哈,是不是明白了

   class AsyncStateMachineBox<TStateMachine> : Task<TResult>, IAsyncStateMachineBox where TStateMachine : IAsyncStateMachine
   
   static void DispatchWorkItem(object workItem, Thread currentThread)
   {
       Task task = workItem as Task;
       if (task != null)
       {
           task.ExecuteFromThreadPool(currentThread);
           return;
       }
       Unsafe.As<IThreadPoolWorkItem>(workItem).Execute();
   }

• 还有就是状态机包装类,重写了 Task.ExecuteFromThreadPool,所以线程池调用 task.ExecuteFromThreadPool 就是直接调用了状态机的 MoveNext 了, Soga ^_^

  override void ExecuteFromThreadPool(Thread threadPoolThread)
  {
      this.MoveNext(threadPoolThread);
  }

参考链接

• 关于线程池和异步的更深刻的原理,大家可以参考下面的文章

概述 .NET 6 ThreadPool 实现: https://www.cnblogs.com/eventhorizon/p/15316955.html

.NET Task 揭秘（2）：Task 的回调执行与 await: https://www.cnblogs.com/eventhorizon/p/15912383.html