fix(pool): fix race condition with small pool sizes

.github/workflows/main.yml

@@ -23,7 +23,7 @@ jobs:
         go-version: ${{ matrix.go-version }}
 
     - name: Test
-      run: make test
+      run: make test-ci
   codecov:
     name: Coverage report
     runs-on: ubuntu-latest

Makefile

@@ -1,5 +1,8 @@
 test:
-	go test -race -v -timeout 1m ./...
+	go test -race -v -timeout 15s -count=1 ./...
+
+test-ci:
+	go test -race -v -timeout 1m -count=3 ./...
 
 coverage:
 	go test -race -v -timeout 1m -coverprofile=coverage.out -covermode=atomic ./...

group_test.go

@@ -36,8 +36,9 @@ func TestResultTaskGroupWait(t *testing.T) {
 
 func TestResultTaskGroupWaitWithError(t *testing.T) {
 
-	group := NewResultPool[int](1).
-		NewGroup()
+	pool := NewResultPool[int](1)
+
+	group := pool.NewGroup()
 
 	sampleErr := errors.New("sample error")
 

internal/dispatcher/dispatcher.go

@@ -30,7 +30,6 @@ func NewDispatcher[T any](ctx context.Context, dispatchFunc func([]T), batchSize
 		bufferHasElements: make(chan struct{}, 1),
 		dispatchFunc:      dispatchFunc,
 		batchSize:         batchSize,
-		closed:            atomic.Bool{},
 	}
 
 	dispatcher.waitGroup.Add(1)
@@ -118,9 +117,6 @@ func (d *Dispatcher[T]) run(ctx context.Context) {
 
 				// Submit the next batch of values
 				d.dispatchFunc(batch[0:batchSize])
-
-				// Reset batch
-				batch = batch[:0]
 			}
 
 			if !ok || d.closed.Load() {

internal/linkedbuffer/linkedbuffer.go

@@ -17,7 +17,7 @@ type LinkedBuffer[T any] struct {
 	maxCapacity int
 	writeCount  atomic.Uint64
 	readCount   atomic.Uint64
-	mutex       sync.RWMutex
+	mutex       sync.Mutex
 }
 
 func NewLinkedBuffer[T any](initialCapacity, maxCapacity int) *LinkedBuffer[T] {
@@ -78,28 +78,25 @@ func (b *LinkedBuffer[T]) Write(values []T) {
 
 // Read reads values from the buffer and returns the number of elements read
 func (b *LinkedBuffer[T]) Read(values []T) int {
+	b.mutex.Lock()
+	defer b.mutex.Unlock()
 
 	var readBuffer *Buffer[T]
 
 	for {
-		b.mutex.RLock()
 		readBuffer = b.readBuffer
-		b.mutex.RUnlock()
 
 		// Read element
 		n, err := readBuffer.Read(values)
 
 		if err == ErrEOF {
 			// Move to next buffer
-			b.mutex.Lock()
 			if readBuffer.next == nil {
-				b.mutex.Unlock()
 				return n
 			}
 			if b.readBuffer != readBuffer.next {
 				b.readBuffer = readBuffer.next
 			}
-			b.mutex.Unlock()
 			continue
 		}
 

internal/linkedbuffer/linkedbuffer_test.go

@@ -91,3 +91,42 @@ func TestLinkedBufferLen(t *testing.T) {
 	buf.readCount.Add(1)
 	assert.Equal(t, uint64(0), buf.Len())
 }
+
+func TestLinkedBufferWithReusedBuffer(t *testing.T) {
+
+	buf := NewLinkedBuffer[int](2, 1)
+
+	values := make([]int, 1)
+
+	buf.Write([]int{1})
+	buf.Write([]int{2})
+
+	n := buf.Read(values)
+
+	assert.Equal(t, 1, n)
+	assert.Equal(t, 1, values[0])
+
+	assert.Equal(t, 1, len(values))
+	assert.Equal(t, 1, cap(values))
+
+	n = buf.Read(values)
+
+	assert.Equal(t, 1, n)
+	assert.Equal(t, 1, len(values))
+	assert.Equal(t, 2, values[0])
+
+	buf.Write([]int{3})
+	buf.Write([]int{4})
+
+	n = buf.Read(values)
+
+	assert.Equal(t, 1, n)
+	assert.Equal(t, 1, len(values))
+	assert.Equal(t, 3, values[0])
+
+	n = buf.Read(values)
+
+	assert.Equal(t, 1, n)
+	assert.Equal(t, 1, len(values))
+	assert.Equal(t, 4, values[0])
+}

pool.go

@@ -14,6 +14,8 @@
 
 var MAX_TASKS_CHAN_LENGTH = runtime.NumCPU() * 128
 
+var PERSISTENT_WORKER_COUNT = int64(runtime.NumCPU())
+
 var ErrPoolStopped = errors.New("pool stopped")
 
 var poolStoppedFuture = func() Task {
@@ -91,6 +93,7 @@
 	workerCount         atomic.Int64
 	workerWaitGroup     sync.WaitGroup
 	dispatcher          *dispatcher.Dispatcher[any]
+	dispatcherRunning   sync.Mutex
 	successfulTaskCount atomic.Uint64
 	failedTaskCount     atomic.Uint64
 }
@@ -196,15 +199,16 @@
 }
 
 func (p *pool) dispatch(incomingTasks []any) {
+	p.dispatcherRunning.Lock()
+	defer p.dispatcherRunning.Unlock()
+
 	// Submit tasks
 	for _, task := range incomingTasks {
 		p.dispatchTask(task)
 	}
 }
 
 func (p *pool) dispatchTask(task any) {
-	workerCount := int(p.workerCount.Load())
-
 	// Attempt to submit task without blocking
 	select {
 	case p.tasks <- task:
@@ -214,19 +218,13 @@
 		// 1. There are no idle workers (all spawned workers are processing a task)
 		// 2. There are no workers in the pool
 		// In either case, we should launch a new worker as long as the number of workers is less than the size of the task queue.
-		if workerCount < p.tasksLen {
-			// Launch a new worker
-			p.startWorker()
-		}
+		p.startWorker(p.tasksLen)
 		return
 	default:
 	}
 
 	// Task queue is full, launch a new worker if the number of workers is less than the maximum concurrency
-	if workerCount < p.maxConcurrency {
-		// Launch a new worker
-		p.startWorker()
-	}
+	p.startWorker(p.maxConcurrency)
 
 	// Block until task is submitted
 	select {
@@ -238,15 +236,41 @@
 	}
 }
 
-func (p *pool) startWorker() {
+func (p *pool) startWorker(limit int) {
+	if p.workerCount.Load() >= int64(limit) {
+		return
+	}
 	p.workerWaitGroup.Add(1)
-	p.workerCount.Add(1)
-	go p.worker()
+	workerNumber := p.workerCount.Add(1)
+	// Guarantee at least PERSISTENT_WORKER_COUNT workers are always running during dispatch to prevent deadlocks
+	canExitDuringDispatch := workerNumber > PERSISTENT_WORKER_COUNT
+	go p.worker(canExitDuringDispatch)
 }
 
-func (p *pool) worker() {
-	defer func() {
+func (p *pool) workerCanExit(canExitDuringDispatch bool) bool {
+	if canExitDuringDispatch {
 		p.workerCount.Add(-1)
+		return true
+	}
+
+	// Check if the dispatcher is running
+	if !p.dispatcherRunning.TryLock() {
+		// Dispatcher is running, cannot exit yet
+		return false
+	}
+	if len(p.tasks) > 0 {
+		// There are tasks in the queue, cannot exit yet
+		p.dispatcherRunning.Unlock()
+		return false
+	}
+	p.workerCount.Add(-1)
+	p.dispatcherRunning.Unlock()
+
+	return true
+}
+
+func (p *pool) worker(canExitDuringDispatch bool) {
+	defer func() {
 		p.workerWaitGroup.Done()
 	}()
 
@@ -255,17 +279,20 @@
 		select {
 		case <-p.ctx.Done():
 			// Context cancelled, exit
+			p.workerCount.Add(-1)
 			return
 		default:
 		}
 
 		select {
 		case <-p.ctx.Done():
 			// Context cancelled, exit
+			p.workerCount.Add(-1)
 			return
 		case task, ok := <-p.tasks:
 			if !ok || task == nil {
 				// Channel closed or worker killed, exit
+				p.workerCount.Add(-1)
 				return
 			}
 
@@ -276,8 +303,13 @@
 			p.updateMetrics(err)
 
 		default:
-			// No tasks left, exit
-			return
+			// No tasks left
+
+			// Check if the worker can exit
+			if p.workerCanExit(canExitDuringDispatch) {
+				return
+			}
+			continue
 		}
 	}
 }