---
id: cs-lock-free-patterns
title: Lock-Free Patterns — atomic / CAS / queue
category: Coding
status: draft
source_trust_level: B
verification_status: conceptual
created_at: 2026-05-09
updated_at: 2026-05-09
tags: [cs, concurrency, vibe-coding]
tech_stack: { language: "C++ / Rust", applicable_to: ["CS", "Backend"] }
applied_in: []
aliases: [lock-free, wait-free, CAS, compare and swap, atomic, MPSC, MPMC, ABA problem]
---

# Lock-Free Patterns

> Lock = 느린, deadlock 위험. **Atomic + CAS = lock-free**. Wait-free 가 강함.

## 📖 핵심 개념
- Atomic: 원자 operation.
- CAS: Compare-And-Swap.
- Lock-free: 1+ thread 가 progress.
- Wait-free: 매 thread 가 progress.

## 💻 코드 패턴

### Atomic counter
```rust
use std::sync::atomic::{AtomicU64, Ordering};

let counter = AtomicU64::new(0);

// Thread-safe
counter.fetch_add(1, Ordering::SeqCst);

let val = counter.load(Ordering::SeqCst);
```

### CAS (Compare-And-Swap)
```rust
let atomic = AtomicU64::new(0);

loop {
    let old = atomic.load(Ordering::SeqCst);
    let new = old + 1;
    if atomic.compare_exchange(old, new, Ordering::SeqCst, Ordering::SeqCst).is_ok() {
        break;
    }
    // 다른 thread 가 변경. Retry.
}
```

→ "if 같으면 set". Atomic.

### Memory order
```
Relaxed: 매 thread 의 자체 ordering 만.
Acquire / Release: 인접 op 의 ordering.
SeqCst: total ordering (제일 안전, 느린).
```

### TS / JavaScript SharedArrayBuffer
```ts
const sab = new SharedArrayBuffer(8);
const view = new Int32Array(sab);

// Atomic
Atomics.add(view, 0, 1);
Atomics.compareExchange(view, 0, oldVal, newVal);
Atomics.wait(view, 0, 0);
Atomics.notify(view, 0);
```

→ Worker 간 shared.

### Lock-free queue (Michael-Scott)
```rust
struct Node<T> {
    value: T,
    next: AtomicPtr<Node<T>>,
}

struct Queue<T> {
    head: AtomicPtr<Node<T>>,
    tail: AtomicPtr<Node<T>>,
}

impl<T> Queue<T> {
    fn enqueue(&self, value: T) {
        let new_node = Box::into_raw(Box::new(Node { value, next: AtomicPtr::new(null_mut()) }));
        
        loop {
            let tail = self.tail.load(SeqCst);
            let next = unsafe { (*tail).next.load(SeqCst) };
            
            if next.is_null() {
                if unsafe { (*tail).next.compare_exchange(null_mut(), new_node, SeqCst, SeqCst) }.is_ok() {
                    self.tail.compare_exchange(tail, new_node, SeqCst, SeqCst).ok();
                    return;
                }
            } else {
                self.tail.compare_exchange(tail, next, SeqCst, SeqCst).ok();
            }
        }
    }
}
```

→ Multi-producer multi-consumer.

### MPSC (single consumer)
```rust
// Crossbeam (Rust)
use crossbeam::channel::unbounded;

let (tx, rx) = unbounded();

// Multiple producer
thread::spawn(move || tx.send(1).unwrap());

// Single consumer
let v = rx.recv().unwrap();
```

→ Lock-free queue.

### Crossbeam (Rust)
```
- Unbounded / bounded channel.
- ArrayQueue / SegQueue.
- AtomicCell.
- Epoch-based GC.
```

### ABA problem
```
Thread A:
- Load X = 1.
Thread B:
- Set X = 2.
- Set X = 1.
Thread A:
- CAS X (expect 1) = success.
- → wrong (X 가 변경 됐는데 안 알아).
```

```rust
// Solution: tagged pointer (counter + value).
let old: (u64, T) = atomic.load();   // counter + value
let new = (old.0 + 1, new_value);
atomic.compare_exchange(old, new, ...);
```

→ Counter 가 increment. 같은 value 가 다른 counter.

### Memory reclamation
```
Lock-free 의 함정:
- Free 한 node 의 다른 thread 가 still 참조.

Solution:
- Hazard pointer.
- Epoch-based GC (crossbeam).
- RCU (Read-Copy-Update).
```

### Lock-free hash map
```
DashMap (Rust):
- Sharded (16 partition).
- 매 partition 의 Lock-free or fine lock.
- 큰 throughput.
```

```rust
use dashmap::DashMap;
let map: DashMap<String, u32> = DashMap::new();

map.insert("key".into(), 42);
let v = map.get("key");
```

### Atomic reference counting (Rust Arc)
```rust
use std::sync::Arc;
use std::sync::atomic::AtomicUsize;

let arc = Arc::new(MyData);
let arc2 = arc.clone();   // count++

// Drop = count--. 0 = free.
```

### Java AtomicInteger / AtomicReference
```java
AtomicInteger counter = new AtomicInteger(0);
counter.incrementAndGet();

AtomicReference<Node> head = new AtomicReference<>();
head.compareAndSet(expected, newValue);
```

### Go (sync/atomic)
```go
import "sync/atomic"

var counter int64
atomic.AddInt64(&counter, 1)

old := atomic.LoadInt64(&counter)
atomic.CompareAndSwapInt64(&counter, old, old + 1)
```

### Wait-free
```
Lock-free: 1+ thread 가 progress.
Wait-free: 매 thread 가 progress.

Wait-free = stronger. 보통 lock-free 가 충분.
```

### Software Transactional Memory (STM)
```haskell
-- Haskell
atomically $ do
  x <- readTVar var
  writeTVar var (x + 1)
```

```clojure
;; Clojure
(dosync
  (alter counter inc))
```

→ Optimistic concurrency. Conflict 시 retry.

### Use case
```
- Counter (statistics).
- Single producer queue (logging).
- Lock-free hash (cache).
- Reference count (shared ownership).
- LSM tree memtable.
- WAL (write-ahead log).
```

### vs Mutex
```
Mutex:
- Simple.
- Block 가능 (sleep).
- Deadlock 위험.

Lock-free:
- Complex.
- Non-blocking.
- 더 빠름 (high contention).
- 더 어려움.

→ 작은 critical section = mutex.
큰 contention = lock-free.
```

### Lock contention 측정
```
perf record / perf report.
- Hot mutex 발견.
- Lock-free 의 candidate.

→ Profile 가 답.
```

### Lock-free의 함정
```
- ABA problem: tagged pointer.
- Memory reclamation: hazard pointer / epoch.
- Memory ordering: SeqCst 가 안전 default.
- Algorithm 가 매우 어려움.
- Test 어려움 (race rare).

→ Library (Crossbeam / Folly) 사용. 자체 X.
```

### Folly (Facebook C++)
```cpp
#include <folly/MPMCQueue.h>
folly::MPMCQueue<int> queue(1024);
queue.write(42);
int x;
queue.read(x);
```

### LMAX Disruptor (Java)
```java
RingBuffer<Event> ringBuffer = ...;

long sequence = ringBuffer.next();
Event event = ringBuffer.get(sequence);
event.setValue(42);
ringBuffer.publish(sequence);
```

→ 매우 빠른 (millions / sec).

### Database 의 lock-free
```
MVCC: lock-free read.
LSM tree: lock-free memtable insert.
B-tree (some): lock-free split.

→ Modern DB 의 핵심.
```

### When use?
```
✓ High contention.
✓ Real-time (no block).
✓ Embedded / kernel.

✗ Low contention (mutex 충분).
✗ Logic 가 복잡.
✗ Team 가 expert X.

→ Profile 후. Library 사용.
```

### Real-world
- **Crossbeam** (Rust).
- **Folly** (C++ Facebook).
- **DashMap** (Rust hash).
- **LMAX Disruptor** (Java).
- **Java concurrent.atomic**.

## 🤔 의사결정 기준
| 작업 | 추천 |
|---|---|
| Counter | Atomic |
| Queue | Lock-free queue (Crossbeam) |
| Hash map | DashMap / ConcurrentHashMap |
| Ref count | Arc / shared_ptr |
| 큰 throughput | LMAX Disruptor |
| Database | MVCC + LSM |
| Complex sync | STM (Haskell / Clojure) |

## ❌ 안티패턴
- **자체 lock-free queue**: bug.
- **No memory order**: race.
- **No ABA fix**: corruption.
- **No memory reclamation**: leak / use-after-free.
- **Mutex on hot path**: contention.

## 🤖 LLM 활용 힌트
- Library 사용 (자체 X).
- ABA + memory reclamation 가 큰 함정.
- Profile + measure.
- MVCC + LSM 가 DB 의 lock-free.

## 🔗 관련 문서
- [[CS_LockFree_Atomic]]
- [[CS_MVCC_Concurrency]]
- [[CS_Skip_List_Persistent_Data]]