Attempt to make Clear concurrency-safe

This is an attempt at fixing #81 without imposing a performance hit on the
cache's "normal" (get/set/fetch) activity. Calling "Clear" is now considerably
more expensive.

bucket.go

@@ -98,8 +98,10 @@ func (b *bucket[T]) deletePrefix(prefix string, deletables chan *Item[T]) int {
 	}, deletables)
 }
 
+// we expect the caller to have acquired a write lock
 func (b *bucket[T]) clear() {
-	b.Lock()
+	for _, item := range b.lookup {
+		item.promotions = -2
+	}
 	b.lookup = make(map[string]*Item[T])
-	b.Unlock()
 }

cache.go

@@ -206,6 +206,24 @@ func (c *Cache[T]) bucket(key string) *bucket[T] {
 	return c.buckets[h.Sum32()&c.bucketMask]
 }
 
+func (c *Cache[T]) halted(fn func()) {
+	c.halt()
+	defer c.unhalt()
+	fn()
+}
+
+func (c *Cache[T]) halt() {
+	for _, bucket := range c.buckets {
+		bucket.Lock()
+	}
+}
+
+func (c *Cache[T]) unhalt() {
+	for _, bucket := range c.buckets {
+		bucket.Unlock()
+	}
+}
+
 func (c *Cache[T]) worker() {
 	dropped := 0
 	cc := c.control
@@ -236,11 +254,22 @@ func (c *Cache[T]) worker() {
 				}
 				msg.done <- struct{}{}
 			case controlClear:
+				c.halted(func() {
+					promotables := c.promotables
+					for len(promotables) > 0 {
+						<-promotables
+					}
+					deletables := c.deletables
+					for len(deletables) > 0 {
+						<-deletables
+					}
+
 					for _, bucket := range c.buckets {
 						bucket.clear()
 					}
 					c.size = 0
 					c.list = NewList[*Item[T]]()
+				})
 				msg.done <- struct{}{}
 			case controlGetSize:
 				msg.res <- c.size

cache_test.go

@@ -4,6 +4,7 @@ import (
 	"math/rand"
 	"sort"
 	"strconv"
+	"sync"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -361,6 +362,40 @@ func Test_ConcurrentStop(t *testing.T) {
 	}
 }
 
+func Test_ConcurrentClearAndSet(t *testing.T) {
+	for i := 0; i < 100; i++ {
+		var stop atomic.Bool
+		var wg sync.WaitGroup
+
+		cache := New(Configure[string]())
+		r := func() {
+			for !stop.Load() {
+				cache.Set("a", "a", time.Minute)
+			}
+			wg.Done()
+		}
+		go r()
+		wg.Add(1)
+		cache.Clear()
+		stop.Store(true)
+		wg.Wait()
+		time.Sleep(time.Millisecond)
+		cache.SyncUpdates()
+
+		known := make(map[string]struct{})
+		for node := cache.list.Head; node != nil; node = node.Next {
+			known[node.Value.key] = struct{}{}
+		}
+
+		for _, bucket := range cache.buckets {
+			for key := range bucket.lookup {
+				_, exists := known[key]
+				assert.True(t, exists)
+			}
+		}
+	}
+}
+
 type SizedItem struct {
 	id int
 	s  int64

layeredbucket.go

@@ -111,9 +111,8 @@ func (b *layeredBucket[T]) forEachFunc(primary string, matches func(key string,
 	}
 }
 
+// we expect the caller to have acquired a write lock
 func (b *layeredBucket[T]) clear() {
-	b.Lock()
-	defer b.Unlock()
 	for _, bucket := range b.buckets {
 		bucket.clear()
 	}

layeredcache.go

@@ -196,6 +196,24 @@ func (c *LayeredCache[T]) bucket(key string) *layeredBucket[T] {
 	return c.buckets[h.Sum32()&c.bucketMask]
 }
 
+func (c *LayeredCache[T]) halted(fn func()) {
+	c.halt()
+	defer c.unhalt()
+	fn()
+}
+
+func (c *LayeredCache[T]) halt() {
+	for _, bucket := range c.buckets {
+		bucket.Lock()
+	}
+}
+
+func (c *LayeredCache[T]) unhalt() {
+	for _, bucket := range c.buckets {
+		bucket.Unlock()
+	}
+}
+
 func (c *LayeredCache[T]) promote(item *Item[T]) {
 	c.promotables <- item
 }
@@ -230,11 +248,22 @@ func (c *LayeredCache[T]) worker() {
 				}
 				msg.done <- struct{}{}
 			case controlClear:
+				promotables := c.promotables
+				for len(promotables) > 0 {
+					<-promotables
+				}
+				deletables := c.deletables
+				for len(deletables) > 0 {
+					<-deletables
+				}
+
+				c.halted(func() {
 					for _, bucket := range c.buckets {
 						bucket.clear()
 					}
 					c.size = 0
 					c.list = NewList[*Item[T]]()
+				})
 				msg.done <- struct{}{}
 			case controlGetSize:
 				msg.res <- c.size