device: use a waiting sync.Pool instead of a channel

Channels are FIFO which means we have guaranteed cache misses.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>

device/alignment_test.go

@@ -42,7 +42,6 @@ func TestPeerAlignment(t *testing.T) {
 	checkAlignment(t, "Peer.isRunning", unsafe.Offsetof(p.isRunning))
 }
 
-
 // TestDeviceAlignment checks that atomically-accessed fields are
 // aligned to 64-bit boundaries, as required by the atomic package.
 //

device/device.go

@@ -67,12 +67,9 @@ type Device struct {
 	}
 
 	pool struct {
-		messageBufferPool        *sync.Pool
-		messageBufferReuseChan   chan *[MaxMessageSize]byte
-		inboundElementPool       *sync.Pool
-		inboundElementReuseChan  chan *QueueInboundElement
-		outboundElementPool      *sync.Pool
-		outboundElementReuseChan chan *QueueOutboundElement
+		messageBuffers   *WaitPool
+		inboundElements  *WaitPool
+		outboundElements *WaitPool
 	}
 
 	queue struct {

device/pools.go

@@ -5,87 +5,80 @@
 
 package device
 
-import "sync"
+import (
+	"sync"
+	"sync/atomic"
+)
+
+type WaitPool struct {
+	pool  sync.Pool
+	cond  sync.Cond
+	lock  sync.Mutex
+	count uint32
+	max   uint32
+}
+
+func NewWaitPool(max uint32, new func() interface{}) *WaitPool {
+	p := &WaitPool{pool: sync.Pool{New: new}, max: max}
+	p.cond = sync.Cond{L: &p.lock}
+	return p
+}
+
+func (p *WaitPool) Get() interface{} {
+	if p.max != 0 {
+		p.lock.Lock()
+		for atomic.LoadUint32(&p.count) >= p.max {
+			p.cond.Wait()
+		}
+		atomic.AddUint32(&p.count, 1)
+		p.lock.Unlock()
+	}
+	return p.pool.Get()
+}
+
+func (p *WaitPool) Put(x interface{}) {
+	p.pool.Put(x)
+	if p.max == 0 {
+		return
+	}
+	atomic.AddUint32(&p.count, ^uint32(0))
+	p.cond.Signal()
+}
 
 func (device *Device) PopulatePools() {
-	if PreallocatedBuffersPerPool == 0 {
-		device.pool.messageBufferPool = &sync.Pool{
-			New: func() interface{} {
+	device.pool.messageBuffers = NewWaitPool(PreallocatedBuffersPerPool, func() interface{} {
 		return new([MaxMessageSize]byte)
-			},
-		}
-		device.pool.inboundElementPool = &sync.Pool{
-			New: func() interface{} {
+	})
+	device.pool.inboundElements = NewWaitPool(PreallocatedBuffersPerPool, func() interface{} {
 		return new(QueueInboundElement)
-			},
-		}
-		device.pool.outboundElementPool = &sync.Pool{
-			New: func() interface{} {
+	})
+	device.pool.outboundElements = NewWaitPool(PreallocatedBuffersPerPool, func() interface{} {
 		return new(QueueOutboundElement)
-			},
-		}
-	} else {
-		device.pool.messageBufferReuseChan = make(chan *[MaxMessageSize]byte, PreallocatedBuffersPerPool)
-		for i := 0; i < PreallocatedBuffersPerPool; i++ {
-			device.pool.messageBufferReuseChan <- new([MaxMessageSize]byte)
-		}
-		device.pool.inboundElementReuseChan = make(chan *QueueInboundElement, PreallocatedBuffersPerPool)
-		for i := 0; i < PreallocatedBuffersPerPool; i++ {
-			device.pool.inboundElementReuseChan <- new(QueueInboundElement)
-		}
-		device.pool.outboundElementReuseChan = make(chan *QueueOutboundElement, PreallocatedBuffersPerPool)
-		for i := 0; i < PreallocatedBuffersPerPool; i++ {
-			device.pool.outboundElementReuseChan <- new(QueueOutboundElement)
-		}
-	}
+	})
 }
 
 func (device *Device) GetMessageBuffer() *[MaxMessageSize]byte {
-	if PreallocatedBuffersPerPool == 0 {
-		return device.pool.messageBufferPool.Get().(*[MaxMessageSize]byte)
-	} else {
-		return <-device.pool.messageBufferReuseChan
-	}
+	return device.pool.messageBuffers.Get().(*[MaxMessageSize]byte)
 }
 
 func (device *Device) PutMessageBuffer(msg *[MaxMessageSize]byte) {
-	if PreallocatedBuffersPerPool == 0 {
-		device.pool.messageBufferPool.Put(msg)
-	} else {
-		device.pool.messageBufferReuseChan <- msg
-	}
+	device.pool.messageBuffers.Put(msg)
 }
 
 func (device *Device) GetInboundElement() *QueueInboundElement {
-	if PreallocatedBuffersPerPool == 0 {
-		return device.pool.inboundElementPool.Get().(*QueueInboundElement)
-	} else {
-		return <-device.pool.inboundElementReuseChan
-	}
+	return device.pool.inboundElements.Get().(*QueueInboundElement)
 }
 
 func (device *Device) PutInboundElement(elem *QueueInboundElement) {
 	elem.clearPointers()
-	if PreallocatedBuffersPerPool == 0 {
-		device.pool.inboundElementPool.Put(elem)
-	} else {
-		device.pool.inboundElementReuseChan <- elem
-	}
+	device.pool.inboundElements.Put(elem)
 }
 
 func (device *Device) GetOutboundElement() *QueueOutboundElement {
-	if PreallocatedBuffersPerPool == 0 {
-		return device.pool.outboundElementPool.Get().(*QueueOutboundElement)
-	} else {
-		return <-device.pool.outboundElementReuseChan
-	}
+	return device.pool.outboundElements.Get().(*QueueOutboundElement)
 }
 
 func (device *Device) PutOutboundElement(elem *QueueOutboundElement) {
 	elem.clearPointers()
-	if PreallocatedBuffersPerPool == 0 {
-		device.pool.outboundElementPool.Put(elem)
-	} else {
-		device.pool.outboundElementReuseChan <- elem
-	}
+	device.pool.outboundElements.Put(elem)
 }

device/pools_test.go

@@ -0,0 +1,60 @@
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (C) 2019-2021 WireGuard LLC. All Rights Reserved.
+ */
+
+package device
+
+import (
+	"math/rand"
+	"runtime"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestWaitPool(t *testing.T) {
+	var wg sync.WaitGroup
+	trials := int32(100000)
+	workers := runtime.NumCPU() + 2
+	if workers-4 <= 0 {
+		t.Skip("Not enough cores")
+	}
+	p := NewWaitPool(uint32(workers-4), func() interface{} { return make([]byte, 16) })
+	wg.Add(workers)
+	max := uint32(0)
+	updateMax := func() {
+		count := atomic.LoadUint32(&p.count)
+		if count > p.max {
+			t.Errorf("count (%d) > max (%d)", count, p.max)
+		}
+		for {
+			old := atomic.LoadUint32(&max)
+			if count <= old {
+				break
+			}
+			if atomic.CompareAndSwapUint32(&max, old, count) {
+				break
+			}
+		}
+	}
+	for i := 0; i < workers; i++ {
+		go func() {
+			defer wg.Done()
+			for atomic.AddInt32(&trials, -1) > 0 {
+				updateMax()
+				x := p.Get()
+				updateMax()
+				time.Sleep(time.Duration(rand.Intn(100)) * time.Microsecond)
+				updateMax()
+				p.Put(x)
+				updateMax()
+			}
+		}()
+	}
+	wg.Wait()
+	if max != p.max {
+		t.Errorf("Actual maximum count (%d) != ideal maximum count (%d)", max, p.max)
+	}
+}