AsyncLocal和Async原理解读

  • 2019 年 10 月 3 日
  • 筆記

AsyncLocal 的实现很简单,将AsyncLocal实例和当前线程的值以键值对的形式保存在Thread.CurrentThread.ExecutionContext.m_localValues.中。由于使用[ThreadStatic] 修饰了 Thread.CurrentThread属性对应的字段,所以实现了多个线程之间各自维护不同的一份数据。同时,在每一次修改AsyncLocal.Value的时候,都新建了ExecutionContextIAsyncLocalValueMap对象并赋值给当前的线程。

以下为AsyncLocal的测试代码

class AsyncLocalTests : Singleton<AsyncLocalTests>,ITestMethod  {      private readonly AsyncLocal<int> asyncLocalVariable = new AsyncLocal<int>();        public async Task MethodAsync()      {          asyncLocalVariable.Value = 88;          await Task.Run(() =>          {              Console.WriteLine($"进入 Task,值:{asyncLocalVariable.Value};线程Id:{Thread.CurrentThread.ManagedThreadId};ExecutionContext:Hashcode:{Thread.CurrentThread.ExecutionContext.GetHashCode()}");              asyncLocalVariable.Value = 888;          });          Console.WriteLine($"await Task 后,值:{asyncLocalVariable.Value};线程Id:{Thread.CurrentThread.ManagedThreadId};ExecutionContext:Hashcode:{Thread.CurrentThread.ExecutionContext.GetHashCode()}");      }        public void RunTest()      {          asyncLocalVariable.Value = 1;          Console.WriteLine($"初始值:{asyncLocalVariable.Value};线程Id:{Thread.CurrentThread.ManagedThreadId};ExecutionContext:Hashcode:{Thread.CurrentThread.ExecutionContext.GetHashCode()}");          MethodAsync();          Thread.Sleep(1000);          Console.WriteLine($"async方法后,值:{asyncLocalVariable.Value};线程Id:{Thread.CurrentThread.ManagedThreadId};ExecutionContext:Hashcode:{Thread.CurrentThread.ExecutionContext.GetHashCode()}");          Console.ReadKey();      }  }

测试结果

image

从上面的测试结果我们看出:

  1. MethodAsync() 异步方法前后,AsyncLocal.Value的值相同
  2. await MethodAsync() 异步方法前后,AsyncLocal.Value的值相同
  3. await Task.Run() 前后代码块中,AsyncLocal.Value的值相同

由于AsyncLocal.Value 是从 Thread.CurrentThread.ExecutionContext 获取实际的值,那么理解ExecutionContextasync、Task、Thread 的中流动就十分重要。先说结论,并简单描述一下原因:

1、进入Task.Run()前后,ExecutionContext相同(处于不同线程)

原因:这是因为 Task.Run()Thread.Start() 会捕获当前线程的 ExecutionContext 传递给工作线程,并且在工作线程修改 AsyncLocal.Value 的值, 不会影响原线程的ExecutionContext 。因为每次修改 AsyncLocal.Value 的值,都会新建 ExecutionContext 实例并保存到工作线程

2、 MethodAsync() 前后代码块的 ExecutionContext 相同(不使用await)

原因:在状态机第一次执行前后会备份、恢复ExecutionContext(线程并没有进行切换)

3、await Task.Run() 前后代码块的 ExecutionContext 相同(处于不同线程)

原因:我们知道await Task.Run()肯定位于一个异步方法中,该异步方法会被编译成一个状态机,通过状态的切换,将await前后的代码分成了两步来执行。第一次执行由当前线程执行,在开启新 Task 后、当前线程返回之前,会保存当前线程的 ExecutionContext,供状态机第二次执行使用(工作线程)。从第一点我们知道新建Task实例的时候会捕获一次ExecutionContext给工作线程,碰到await返回之前会捕获一次ExecutionContext给状态机,这两次捕获的实际上是同一个对象。

4、 await MethodAsync() 前后代码块的 ExecutionContext 相同(处于不同线程)

原因:在状态机第一次执行(当前线程)的时候,会捕获当前线程的 ExecutionContext,供状态机第二次执行使用(工作线程)。

2. async关键字对ExecutionContext的影响

async关键字实际上是编译器的语法糖,可以通过Dnspy 反编译查看去除语法糖的原始代码。
Dnspy配置如下,去除勾选"反编译异步方法(async/await)"

image

原代码:

public async Task MethodAsyncWithAwait()  {      asyncLocalVariable.Value = 88;      await Task.Run(() =>      {          asyncLocalVariable.Value = 888;      });      asyncLocalVariable.Value = 8888;  }

去除async/await 语法糖代码:

异步方法代码:

可以看到编译器生成了一个实现 IAsyncStateMachine 接口的异步状态机,并生成了一个私有方法,保存了Task.Run()中的的代码块。异步方法中实际上做了以下四个步骤:

  1. 实例化异步状态机, 将状态置为 -1
  2. 创建 AsyncTaskMethodBuilder
  3. 通过 AsyncTaskMethodBuilder.Sratr(ref IAsyncStateMachine)启动状态机
  4. 返回 Task
public Task MethodAsyncWithAwait()  {      AsyncLocalTests.<MethodAsyncWithAwait>d__1 <MethodAsyncWithAwait>d__ = new AsyncLocalTests.<MethodAsyncWithAwait>d__1();      <MethodAsyncWithAwait>d__.<>4__this = this;      <MethodAsyncWithAwait>d__.<>t__builder = AsyncTaskMethodBuilder.Create();      <MethodAsyncWithAwait>d__.<>1__state = -1;      AsyncTaskMethodBuilder <>t__builder = <MethodAsyncWithAwait>d__.<>t__builder;      <>t__builder.Start<AsyncLocalTests.<MethodAsyncWithAwait>d__1>(ref <MethodAsyncWithAwait>d__);      return <MethodAsyncWithAwait>d__.<>t__builder.Task;  }

状态机启动代码:

可以看到在执行 stateMachine.MoveNext() 之前备份了当前线程的 _executionContext_synchronizationContext,并且在 finally 代码块中恢复了备份的数据。
这样也就解释了:在不使用 await 等待异步方法的情况下,虽然在原线程修改了AsyncLocal.Value的值,但是离开async方法后,我们获取的还是原来的值。值得注意的是,这里的备份恢复针对的都是当前线程,而不涉及到工作线程。

public void Start<[Nullable(0)] TStateMachine>(ref TStateMachine stateMachine) where TStateMachine : IAsyncStateMachine  {      AsyncMethodBuilderCore.Start<TStateMachine>(ref stateMachine);  }  internal static class AsyncMethodBuilderCore  {  [DebuggerStepThrough]  public static void Start<TStateMachine>(ref TStateMachine stateMachine) where TStateMachine : IAsyncStateMachine  {      if (stateMachine == null)      {          ThrowHelper.ThrowArgumentNullException(ExceptionArgument.stateMachine);      }      Thread currentThread = Thread.CurrentThread;      Thread thread = currentThread;      ExecutionContext executionContext = currentThread._executionContext;      ExecutionContext executionContext2 = executionContext;      SynchronizationContext synchronizationContext = currentThread._synchronizationContext;      try      {          stateMachine.MoveNext();      }      finally      {          SynchronizationContext synchronizationContext2 = synchronizationContext;          Thread thread2 = thread;          if (synchronizationContext2 != thread2._synchronizationContext)          {              thread2._synchronizationContext = synchronizationContext2;          }          ExecutionContext executionContext3 = executionContext2;          ExecutionContext executionContext4 = thread2._executionContext;          if (executionContext3 != executionContext4)          {              ExecutionContext.RestoreChangedContextToThread(thread2, executionContext3, executionContext4);          }      }  }

状态机代码:

实际上编译器将标记为 async 的方法分成了两部分,一部分是 await 之前的代码(包括新建并启动启动Task部分),另一部分是 await之后的代码。通过状态的改变,这两部分代码分两次执行。如果没有使用await修饰异步方法,那么该状态机没有 else代码块, 只会执行一次stateMachine.MoveNext()。可以看到在第一次执行stateMachine.MoveNext() 之后,当前线程就直接返回了,然后一层层的返回到最外层。这也是为什么说碰到await之后,当前线程就直接返回,当然最内层的返回是在开启新Task之后。

[CompilerGenerated]  private void <MethodAsyncWithAwait>b__1_0()  {      this.asyncLocalVariable.Value = 888;  }    [CompilerGenerated]  private sealed class <MethodAsyncWithAwait>d__1 : IAsyncStateMachine  {      void IAsyncStateMachine.MoveNext()      {          int num = this.<>1__state;          try          {              TaskAwaiter awaiter;              if (num != 0)              {                  this.<>4__this.asyncLocalVariable.Value = 88;                  awaiter = Task.Run(new Action(this.<>4__this.<MethodAsyncWithAwait>b__1_0)).GetAwaiter();                  if (!awaiter.IsCompleted)                  {                      this.<>1__state = 0;                      this.<>u__1 = awaiter;                      AsyncLocalTests.<MethodAsyncWithAwait>d__1 <MethodAsyncWithAwait>d__ = this;                      this.<>t__builder.AwaitUnsafeOnCompleted<TaskAwaiter, AsyncLocalTests.<MethodAsyncWithAwait>d__1>(ref awaiter, ref <MethodAsyncWithAwait>d__);                      return;                  }              }              else              {                  awaiter = this.<>u__1;                  this.<>u__1 = default(TaskAwaiter);                  this.<>1__state = -1;              }              awaiter.GetResult();              this.<>4__this.asyncLocalVariable.Value = 8888;          }          catch (Exception exception)          {              this.<>1__state = -2;              this.<>t__builder.SetException(exception);              return;          }          this.<>1__state = -2;          this.<>t__builder.SetResult();      }      [DebuggerHidden]      void IAsyncStateMachine.SetStateMachine(IAsyncStateMachine stateMachine)      {      }        public int <>1__state;        public AsyncTaskMethodBuilder <>t__builder;        public AsyncLocalTests <>4__this;        private TaskAwaiter <>u__1;  }

我们可以看到在第一次执行stateMachine.MoveNext()的时候,会通过当前线程执行 await 之前的代码块,并通过Task.Run()启用工作线程去完成任务。IAsyncStateMachine.MoveNext() 里面有三句代码比较重要,这里先大概描述一下作用:

1、 Task.Run():

新建Task实例、捕获当前线程的ExecutionContext 保存到Task实例中、启动新任务

2、 AsyncTaskMethodBuilder.AwaitUnsafeOnCompleted(ref awaiter, ref stateMachine):

将状态机和当前线程的上下文包装成 AsyncStateMachineBox:Task 对象,保存到在 Task(第一步新建的 Task实例).m_continuationObject 字段中,最后将其传递到 stateMachine.AsyncTaskMethodBuilder.Task属性中, 这样外层状态机可以通过MethodAsync().GetAwaiter().m_task获取到内层状态机的AsyncStateMachineBox:Task 对象,同样的外层状态机再次执行本步骤,将自身的 AsyncStateMachineBox:Task 对象赋值给内层 AsyncStateMachineBox:Task 对象的 m_continuationObject字段,这样的话,就构建了一个单向链表,该链表保存了每一层异步方法的stateMachineExecutionContext

3、 this.<>t__builder.SetResult():
设置异步方法的结果,并检查 Task.m_continuationObject 是否为空,不为空的情况下,执行外层状态机的第二次 stateMachine.MoveNext()。最内层 Task.m_continuationObject的执行会在Task完成之后调用,接下来通过SetResult()一层层调用了外层状态机的第二次 stateMachine.MoveNext()

3. AwaitUnsafeOnCompleted() 代码分析

以下只放出了简化的代码。这里首先构建 IAsyncStateMachineBox实例,并将其赋值给 m_task 供外层状态机使用。IAsyncStateMachineBox实例保存了本层的状态机,并捕获了当前线程的ExecutionContextawaiter.m_task 是调用内层异步方法返回的 Task实例。最后在 TaskAwaiter.UnsafeOnCompletedInternal() 方法中,将构建的IAsyncStateMachineBox实例保存到 awaiter(内层).m_task.m_continuationObject字段中,使得内层状态机指向本层状态机。因为每一层状态机都会调用AwaitUnsafeOnCompleted 方法,所以一层层构建了 await 后的所有回调,并且每一层回调的 ExecutionContext 都不同。

public void AwaitUnsafeOnCompleted<TAwaiter, TStateMachine>(ref TAwaiter awaiter, ref TStateMachine stateMachine)      where TAwaiter : ICriticalNotifyCompletion where TStateMachine : IAsyncStateMachine  {      IAsyncStateMachineBox box = GetStateMachineBox(ref stateMachine);        if ((null != (object)default(TAwaiter)!) && (awaiter is ITaskAwaiter))      {          ref TaskAwaiter ta = ref Unsafe.As<TAwaiter, TaskAwaiter>(ref awaiter);          TaskAwaiter.UnsafeOnCompletedInternal(ta.m_task, box, continueOnCapturedContext: true);      }  }
private IAsyncStateMachineBox GetStateMachineBox<TStateMachine>(ref TStateMachine stateMachine) where TStateMachine : IAsyncStateMachine  {      ExecutionContext? currentContext = ExecutionContext.Capture();        AsyncStateMachineBox<TStateMachine> box = AsyncMethodBuilderCore.TrackAsyncMethodCompletion ?          CreateDebugFinalizableAsyncStateMachineBox<TStateMachine>() :          new AsyncStateMachineBox<TStateMachine>();      m_task = box;      box.StateMachine = stateMachine;      box.Context = currentContext;        return box;  }

4. SetResult() 代码分析

该方法最后调用了 Task.TrySetResult() 方法,读取了Task.m_continuationObject 来获取外一层的回调(包括状态机、ExecutionContext),并通过FinishContinuations()执行外层状态机的第二次执行。

internal bool TrySetResult([AllowNull] TResult result)  {      bool result2 = false;      if (base.AtomicStateUpdate(67108864, 90177536))      {          this.m_result = result;          Interlocked.Exchange(ref this.m_stateFlags, this.m_stateFlags | 16777216);          Task.ContingentProperties contingentProperties = this.m_contingentProperties;          if (contingentProperties != null)          {              base.NotifyParentIfPotentiallyAttachedTask();              contingentProperties.SetCompleted();          }          base.FinishContinuations();          result2 = true;      }      return result2;  }

FinishContinuations() 方法接下来的调用在 Task.Run() 部分会有介绍。

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

5. Task.Run() 代码分析

实际上在Task.Run()内部也是先新建一个Task 实例,然后通过Task.ScheduleAndStart()方法来调度并启动任务。两者的区别在于传入的Optionsscheduler 是不相同的。

var task1 = Task.Run(() => { });  //默认配置无法更改: InternalTaskOptions.QueuedByRuntime, TaskCreationOptions.DenyChildAttach, TaskScheduler.Default  //t.ScheduleAndStart(false);  var task2 = new Task(() => { }, TaskCreationOptions.LongRunning);  //默认配置可以修改: InternalTaskOptions.None, TaskCreationOptions.None, scheduler:null  task2.Start();  //可以传入scheduler  //默认使用TaskScheduler.Current: 先取[ThreadStatic]Task.InternalCurrent,如果为空取 TaskScheduler.Default  //t.ScheduleAndStart(true);

通过在Task的构造函数中调用ExecutionContext.Capture() 方法来保存当前线程的ExecutionContextTask实例中,这样的话,只要将到Task实例作为参数传入到工作线程中,工作线程就可以获取到ExecutionContext

internal Task(Delegate action, object state, Task parent, CancellationToken cancellationToken, TaskCreationOptions creationOptions, InternalTaskOptions internalOptions, TaskScheduler scheduler)  {      if (action == null)      {          ThrowHelper.ThrowArgumentNullException(ExceptionArgument.action);      }      if (parent != null && (creationOptions & TaskCreationOptions.AttachedToParent) != TaskCreationOptions.None)      {          this.EnsureContingentPropertiesInitializedUnsafe().m_parent = parent;      }      this.TaskConstructorCore(action, state, cancellationToken, creationOptions, internalOptions, scheduler);      this.CapturedContext = ExecutionContext.Capture();  }

Task.ScheduleAndStart()方法:

Task的调度分两种情况:1、配置了TaskCreationOptions.LongRunningTask实例直接新建一个后台 Thread,并将Task实例作为启动参数来启动工作线程;2、对于没有配置TaskCreationOptions.LongRunningTask实例,将其加入ThreadPool的线程池,由线程池来调度运行

internal sealed class ThreadPoolTaskScheduler : TaskScheduler  {      protected internal override void QueueTask(Task task)      {          TaskCreationOptions options = task.Options;          if ((options & TaskCreationOptions.LongRunning) != 0)          {              // Run LongRunning tasks on their own dedicated thread.              Thread thread = new Thread(s_longRunningThreadWork);              thread.IsBackground = true; // Keep this thread from blocking process shutdown              thread.Start(task);          }          else          {              // Normal handling for non-LongRunning tasks.              bool preferLocal = ((options & TaskCreationOptions.PreferFairness) == 0);              ThreadPool.UnsafeQueueUserWorkItemInternal(task, preferLocal);          }      }  }

虽然已经在Task的构造函数中,捕获了ExecutionContext,但是对于直接新建的Thread实例,启动的时候同样也需要捕获当前线程的ExecutionContext

public void Start()  {      this.StartupSetApartmentStateInternal();      if (this._delegate != null)      {          ThreadHelper threadHelper = (ThreadHelper)this._delegate.Target;          ExecutionContext executionContextHelper = ExecutionContext.Capture();          threadHelper.SetExecutionContextHelper(executionContextHelper);      }      this.StartInternal();  }

线程池调度Task、IThreadPoolWorkItem逻辑:

从线程池取出工作线程,工作线程调用Dispatch()方法,对于IThreadPoolWorkItem,直接执行IThreadPoolWorkItem.Execute() 方法,所以线程池处理IThreadPoolWorkItem是不涉及到上下文切换的。对于 Task ,将ExecutionContext赋值给工作线程,调用委托,然后清除工作线程的上下文,最后调用Finish(true) 来执行任务完成的回调方法。调用链路很长,这里直接跳到RunContinuations()方法。

调用链路:

    // 最内层 async 状态机 执行await Task.Run()之后的代码块      //=>Task.Finish(true);      //=>FinishStageTwo();      //===>FinishStageThree();      //=====>FinishContinuations();      //=======>RunContinuations(continuationObject);      //=========>AwaitTaskContinuation.RunOrScheduleAction(asyncStateMachineBox, flag); 状态机的回调执行这个      //==========>box.ExecuteFromThreadPool(threadPoolThread); 或者 box.MoveNext();

我们知道,异步方法的最内层肯定有一个 await Task。 正是Task.Finish(true)这个方法调用了最内层状态机,去执行第二次stateMachine.MoveNext()方法,并且在MoveNext()方法中都会调用SetResult() 方法,从而触发外层状态机的第二次stateMachine.MoveNext()执行,就这样一层层的调用完成了所有的层次的回调。可以看到,工作线程在执行 await Task.Run()/MethodAsnc() 后代码块时,传入的是在 AwaitUnsafeOnCompleted() 方法中捕获的 ExecutionContext
s_callback字段保存了状态机的MoveNext()方法。

private class AsyncStateMachineBox<TStateMachine> :Task<TResult>,IAsyncStateMachineBox  where TStateMachine : IAsyncStateMachine  {      private static readonly ContextCallback s_callback = ExecutionContextCallback;        private static void ExecutionContextCallback(object? s)      {          Unsafe.As<AsyncStateMachineBox<TStateMachine>>(s).StateMachine!.MoveNext();      }        public TStateMachine StateMachine = default;        public ExecutionContext? Context;        internal sealed override void ExecuteFromThreadPool(Thread threadPoolThread) => MoveNext(threadPoolThread);          internal sealed override void ExecuteFromThreadPool(Thread threadPoolThread) => MoveNext(threadPoolThread);        public void MoveNext() => MoveNext(threadPoolThread: null);        private void MoveNext(Thread? threadPoolThread)      {          bool loggingOn = AsyncCausalityTracer.LoggingOn;          if (loggingOn)          {              AsyncCausalityTracer.TraceSynchronousWorkStart(this, CausalitySynchronousWork.Execution);          }            ExecutionContext? context = Context;          if (context == null)          {              Debug.Assert(StateMachine != null);              StateMachine.MoveNext();          }          else          {              if (threadPoolThread is null)              {                  ExecutionContext.RunInternal(context, s_callback, this);              }              else              {                  ExecutionContext.RunFromThreadPoolDispatchLoop(threadPoolThread, context, s_callback, this);              }          }            if (IsCompleted)          {              if (System.Threading.Tasks.Task.s_asyncDebuggingEnabled)              {                  System.Threading.Tasks.Task.RemoveFromActiveTasks(this);              }              StateMachine = default;              Context = default;          }            if (loggingOn)          {              AsyncCausalityTracer.TraceSynchronousWorkCompletion(CausalitySynchronousWork.Execution);          }      }  }

最后在 RunInternalRunFromThreadPoolDispatchLoop 中,都会使用在AwaitUnsafeOnCompleted()方法里面捕获的 ExecutionContext,这也就解释了为什么在 await Task.Run()/MethodAsync() 前后的代码块中,ExecutionContext始终相同。需要注意的一点是,不管在Task 任务执行之后,还是 await 回调执行之后,都会把工作线程的上下文清空。

internal static void RunFromThreadPoolDispatchLoop(Thread threadPoolThread, ExecutionContext executionContext, ContextCallback callback, object state)  {      if (executionContext != null && !executionContext.m_isDefault)      {          ExecutionContext.RestoreChangedContextToThread(threadPoolThread, executionContext, null);      }      ExceptionDispatchInfo exceptionDispatchInfo = null;      try      {          callback(state);      }      catch (Exception source)      {          exceptionDispatchInfo = ExceptionDispatchInfo.Capture(source);      }      ExecutionContext executionContext2 = threadPoolThread._executionContext;      threadPoolThread._synchronizationContext = null;      if (executionContext2 != null)      {          ExecutionContext.RestoreChangedContextToThread(threadPoolThread, null, executionContext2);      }      if (exceptionDispatchInfo != null)      {          exceptionDispatchInfo.Throw();      }  }