DEVBOOK

Deadlocks

10 min read

EditReport

A deadlock is a closed waiting cycle: every participant holds something another participant needs, so none can move. In .NET this can happen with ordinary locks or when synchronous code blocks an async continuation. The process may stay healthy enough to answer a liveness probe while the affected work has stopped completely.

The Four Conditions

Classical resource or lock deadlocks are characterized by all four Coffman conditions simultaneously:

  1. Mutual exclusion means a resource cannot be shared, as with a lock or monitor.
  2. Hold and wait means one path keeps a resource while waiting for another.
  3. No preemption means the owner must release the resource. Another thread cannot simply take it.
  4. Circular wait closes the cycle: A waits for B while B waits for A.

Such a classical resource deadlock needs all four at once. Consistent lock ordering is usually the cleanest prevention because it removes circular wait.

Classic Lock Deadlock

private static readonly object LockA = new();
private static readonly object LockB = new();
 
public void First()
{
    lock (LockA)
    {
        Thread.Sleep(10); // simulate work while holding A
        lock (LockB)      // then acquire B
        {
            // critical section using A + B
        }
    }
}
 
public void Second()
{
    lock (LockB)
    {
        Thread.Sleep(10); // simulate work while holding B
        lock (LockA)      // then acquire A — reverse order!
        {
            // critical section using B + A
        }
    }
}

The cycle forms in five steps:

  1. Thread T1 enters First, acquires LockA.
  2. Thread T2 enters Second, acquires LockB.
  3. T1 tries to acquire LockB — blocked (owned by T2).
  4. T2 tries to acquire LockA — blocked (owned by T1).
  5. Neither thread can continue. The wait graph now contains a cycle.

Async Deadlock (Sync-Over-Async)

Async code can create the same shape without two explicit locks. The common case blocks a thread that owns a SynchronizationContext, while the awaited continuation is queued back to that context.

// In a UI event handler or legacy ASP.NET action:
public void OnLoad()
{
    // DEADLOCK: blocks the UI thread, which the continuation needs to resume
    var data = LoadDataAsync().Result;
    Display(data);
}
 
private async Task<string> LoadDataAsync()
{
    // Default await captures the SynchronizationContext (UI thread).
    // The continuation needs the UI thread to resume — but it's blocked by .Result.
    return await _http.GetStringAsync("https://api.example.com/data");
}

The dependency is small but fatal:

  • .Result blocks the UI/context thread.
  • The await inside LoadDataAsync captured the SynchronizationContext and needs that same thread to resume.
  • Neither can proceed.

Prevention

One lock order. Every code path acquires the same set of locks in the same order. Circular wait then has no way to form.

// Both methods acquire in the same order: LockA → LockB
public void First()
{
    lock (LockA) { lock (LockB) { /* ... */ } }
}
 
public void Second()
{
    lock (LockA) { lock (LockB) { /* ... */ } }
}

Bounded waits. Monitor.TryEnter gives the caller a chance to abandon the attempt, release anything already held, and retry later. The timeout does not preempt another owner. It prevents the caller from waiting forever.

bool acquired = Monitor.TryEnter(LockA, TimeSpan.FromMilliseconds(500));
if (!acquired)
{
    // Log, retry, or throw — don't wait forever
    throw new TimeoutException("Could not acquire LockA within 500ms");
}
try { /* critical section */ }
finally { Monitor.Exit(LockA); }

Async call chains stay async. Awaiting a task yields the thread instead of holding it while the continuation waits to run.

// Correct: await all the way up
public async Task OnLoadAsync()
{
    var data = await LoadDataAsync();
    Display(data);
}

ConfigureAwait(false) can avoid context capture inside library code, but it is not a general bridge from synchronous code to async code. A synchronous boundary should usually be moved outward or given a separate synchronous implementation.

Lock scope stays small. I/O and blocking calls stay outside the critical section. The lock protects the state change, not the work used to compute it.

// Bad: I/O inside lock
lock (_lock) { var result = _http.GetStringAsync(url).Result; }
 
// Good: I/O outside lock
var result = await _http.GetStringAsync(url);
lock (_lock) { _cache[key] = result; }

Thread-pool Starvation Can Look the Same

ASP.NET Core does not need a SynchronizationContext for sync-over-async to cause an outage. Each request blocked on .Result or .Wait() occupies a pool thread. Under load, continuations compete for the same depleted pool and throughput can collapse before thread injection catches up.

// Under concurrency this exhausts the ThreadPool and the app appears "deadlocked"
public IActionResult Get() => Ok(_service.LoadAsync().Result); // never block — await instead

The service looks deadlocked, but the mechanism is resource starvation rather than a closed wait cycle. Keeping the request path async releases pool threads while I/O is pending. Thread-pool behavior is covered in ThreadPool.

Related Failure Modes

  • Livelock keeps threads active but unproductive because each reacts to the other. Two identical Monitor.TryEnter retry loops can collide indefinitely. Jitter breaks that symmetry.
  • Lock convoy queues many threads behind one hot lock. Progress continues, but latency rises and throughput falls as the queue grows.

Pitfalls

There may be no exception. A frozen UI or request timeout may be the first signal. A process dump reveals threads blocked on monitors or synchronous task waits.

A lock cannot span await. The compiler rejects the construct because monitor ownership belongs to a thread, and an async continuation may resume elsewhere. SemaphoreSlim supplies async-compatible waiting.

private readonly SemaphoreSlim _gate = new(1, 1);
 
public async Task UpdateAsync()
{
    await _gate.WaitAsync();
    try { /* critical section */ }
    finally { _gate.Release(); }
}

WARNING

SemaphoreSlim is not reentrant. Unlike Monitor/lock (and Mutex), it has no thread affinity or recursion count. A method holding the only permit self-deadlocks if it calls another method that waits on the same semaphore. The permit belongs at one clear boundary in the call chain.

Public lock objects hide dependencies. lock(this), lock(typeof(X)), and an interned string expose the same monitor to unrelated code. A private readonly gate keeps lock ordering inside the component.

Library calls may take internal locks. Calling unknown code while holding an application lock creates an ordering dependency that is difficult to see or control.

Database deadlocks belong to the database lock manager. Two transactions can touch rows or indexes in opposite orders, after which SQL Server chooses a victim and reports error 1205. The usual response combines consistent access order and short transactions with a bounded retry for the victim. CLR locks do not repair this layer. Database lock behavior is covered in Database Locks.

Questions

References

Graph View

Backlinks