Report a lane enqueue or SynchronizationContext.Post from a non-simulation thread as a deterministic failure

オープン
#76 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
43/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
csharp

調査の方向性

Start by reading SimulationSchedulerLane.Enqueue, SimulationSynchronizationContext.Post and Send, and SimulationScheduler.RegisterCore; the issue identifies these as candidate points for detecting off-thread operation registration. Trace how RunStep surfaces failures and how the driver reports stalls or completion. Done means an off-thread post fails deterministically with its escaping stack, including when it arrives late.

索引モデルが issue の本文から書いたものです。

説明

Summary

Work posted into a simulation from a thread outside it is accepted silently today. SimulationSchedulerLane.Enqueue (and so SimulationSynchronizationContext.Post) succeeds on a real thread-pool thread. Its SingleThreadedGuard detects only concurrent entry, so the post becomes an operation that no seed decides.

The run then goes wrong in one of two ways:

  • It diverges silently. The recorded schedule replays only up to the point where the operation appeared.
  • It stalls spuriously. If the post lands after the driver's RunStep returned false, TryAdvanceVirtualTime declines, because something is now runnable. The driver then sees nothing runnable and no timer pending, while DescribeLiveness() reports runnable 1 or ExternallyCompletable.

The existing affinity check doesn't catch this. The escaped thread carries no simulation runtime context, so no controlled API reports a SimulationThreadAffinityException.

This was found in ReubenBond/clog#360 and fixed there on the consumer side in ReubenBond/clog#361.

Minimal reproduction

The escape needs unrewritten code (for example a simulation test assembly, which the build doesn't rewrite) that awaits a BCL task combinator over a task that runs its continuations asynchronously:

// In the (unrewritten) test assembly, inside a scenario started on a lane:
var answered = new TaskCompletionSource(TaskCreationOptions.RunContinuationsAsynchronously);
var never = new TaskCompletionSource();
var any = Task.WhenAny(answered.Task, never.Task);   // the BCL's combinator, not ControlledTask.WhenAny
bool? inside = null;
_ = any.ContinueWith(_ => inside = scheduler.IsSimulationThread,
    CancellationToken.None, TaskContinuationOptions.ExecuteSynchronously, TaskScheduler.Default);

answered.SetResult();
await any;          // the continuation is posted to the lane's SynchronizationContext from a .NET TP Worker
never.SetResult();
// inside == false, every run.

Here is why the BCL escapes. Task.RunContinuations doesn't run an ITaskCompletionAction whose InvokeMayRunArbitraryCode is true inline when the completing task has RunContinuationsAsynchronously. It queues it to the thread pool instead, and that includes the TwoTaskWhenAnyPromise. The promise then completes any on the pool, and the await continuation is posted to the lane from that thread.

A diagnostic build logged the stack of every operation registered while !IsSimulationThread. Each one was:

SimulationScheduler.RegisterCore
SimulationScheduler.Schedule / ScheduleAfter
SimulationSchedulerLane.Schedule / Enqueue(ScheduledItem)
SimulationSynchronizationContext.Post
System.Threading.Tasks.AwaitTaskContinuation.RunCallback
System.Threading.Tasks.Task.RunContinuations
System.Threading.Tasks.Task`1.TrySetResult
System.Threading.Tasks.Task.TwoTaskWhenAnyPromise`1.Invoke
System.Threading.ThreadPoolWorkQueue.Dispatch          (".NET TP Worker")

Proposed hardening

When a lane enqueue, and so SimulationSynchronizationContext.Post/Send, comes from a thread that isn't one of the scheduler's operation threads and doesn't hold its control scope (!IsSimulationThread):

  1. Record it as a thread-affinity violation, as ReportThreadAffinityViolation does for controlled APIs.
  2. Have the driving thread's next RunStep surface it, so the run fails with the escaping stack rather than diverging or stalling.

A second check would make the failure deterministic even when the escaped thread posts late. When the driver reports a stall or finishes, it could check whether any operation was registered off the simulation thread since the last step.

Possibly related: RegisterCore (and ScheduleAfter's timer path) is the common point where any operation is created, so one check there would cover lanes, timers, and direct Schedule calls alike.

Consumer-side fix, for reference

ReubenBond/clog#361 bans the BCL's Task.WhenAll, Task.WhenAny and Task.WhenEach in simulation test projects, and routes them through a rewritten helper that binds to ControlledTask's combinators. The hardening proposed here would turn any similar escape (Task.Yield or an await with no synchronization context, a raw ThreadPool call, and so on) into an immediate, attributable failure.

主要言語
C#
スター
0
フォーク
0
平均マージ
2時間 32分
マージ済み PR(30日)
23

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

ReubenBond/Clockwork のほかの issue

ReubenBond/Clockwork の issue をすべて見る

似ている issue

C# の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。