1 files changed, 31 insertions, 26 deletions

diff --git a/device/pools.go b/device/pools.go
index 02a5d6a..2c18f41 100644
--- a/device/pools.go
+++ b/device/pools.go
@@ -1,20 +1,19 @@
 /* SPDX-License-Identifier: MIT
  *
- * Copyright (C) 2017-2023 WireGuard LLC. All Rights Reserved.
+ * Copyright (C) 2017-2025 WireGuard LLC. All Rights Reserved.
  */
 
 package device
 
 import (
 	"sync"
-	"sync/atomic"
 )
 
 type WaitPool struct {
 	pool  sync.Pool
 	cond  sync.Cond
 	lock  sync.Mutex
-	count atomic.Uint32
+	count uint32 // Get calls not yet Put back
 	max   uint32
 }
 
@@ -27,10 +26,10 @@ func NewWaitPool(max uint32, new func() any) *WaitPool {
 func (p *WaitPool) Get() any {
 	if p.max != 0 {
 		p.lock.Lock()
-		for p.count.Load() >= p.max {
+		for p.count >= p.max {
 			p.cond.Wait()
 		}
-		p.count.Add(1)
+		p.count++
 		p.lock.Unlock()
 	}
 	return p.pool.Get()
@@ -41,18 +40,20 @@ func (p *WaitPool) Put(x any) {
 	if p.max == 0 {
 		return
 	}
-	p.count.Add(^uint32(0))
+	p.lock.Lock()
+	defer p.lock.Unlock()
+	p.count--
 	p.cond.Signal()
 }
 
 func (device *Device) PopulatePools() {
-	device.pool.outboundElementsSlice = NewWaitPool(PreallocatedBuffersPerPool, func() any {
-		s := make([]*QueueOutboundElement, 0, device.BatchSize())
-		return &s
-	})
-	device.pool.inboundElementsSlice = NewWaitPool(PreallocatedBuffersPerPool, func() any {
+	device.pool.inboundElementsContainer = NewWaitPool(PreallocatedBuffersPerPool, func() any {
 		s := make([]*QueueInboundElement, 0, device.BatchSize())
-		return &s
+		return &QueueInboundElementsContainer{elems: s}
+	})
+	device.pool.outboundElementsContainer = NewWaitPool(PreallocatedBuffersPerPool, func() any {
+		s := make([]*QueueOutboundElement, 0, device.BatchSize())
+		return &QueueOutboundElementsContainer{elems: s}
 	})
 	device.pool.messageBuffers = NewWaitPool(PreallocatedBuffersPerPool, func() any {
 		return new([MaxMessageSize]byte)
@@ -65,28 +66,32 @@ func (device *Device) PopulatePools() {
 	})
 }
 
-func (device *Device) GetOutboundElementsSlice() *[]*QueueOutboundElement {
-	return device.pool.outboundElementsSlice.Get().(*[]*QueueOutboundElement)
+func (device *Device) GetInboundElementsContainer() *QueueInboundElementsContainer {
+	c := device.pool.inboundElementsContainer.Get().(*QueueInboundElementsContainer)
+	c.Mutex = sync.Mutex{}
+	return c
 }
 
-func (device *Device) PutOutboundElementsSlice(s *[]*QueueOutboundElement) {
-	for i := range *s {
-		(*s)[i] = nil
+func (device *Device) PutInboundElementsContainer(c *QueueInboundElementsContainer) {
+	for i := range c.elems {
+		c.elems[i] = nil
 	}
-	*s = (*s)[:0]
-	device.pool.outboundElementsSlice.Put(s)
+	c.elems = c.elems[:0]
+	device.pool.inboundElementsContainer.Put(c)
 }
 
-func (device *Device) GetInboundElementsSlice() *[]*QueueInboundElement {
-	return device.pool.inboundElementsSlice.Get().(*[]*QueueInboundElement)
+func (device *Device) GetOutboundElementsContainer() *QueueOutboundElementsContainer {
+	c := device.pool.outboundElementsContainer.Get().(*QueueOutboundElementsContainer)
+	c.Mutex = sync.Mutex{}
+	return c
 }
 
-func (device *Device) PutInboundElementsSlice(s *[]*QueueInboundElement) {
-	for i := range *s {
-		(*s)[i] = nil
+func (device *Device) PutOutboundElementsContainer(c *QueueOutboundElementsContainer) {
+	for i := range c.elems {
+		c.elems[i] = nil
 	}
-	*s = (*s)[:0]
-	device.pool.inboundElementsSlice.Put(s)
+	c.elems = c.elems[:0]
+	device.pool.outboundElementsContainer.Put(c)
 }
 
 func (device *Device) GetMessageBuffer() *[MaxMessageSize]byte {