rlp: use atomic.Value for type cache (#22902)

All encoding/decoding operations read the type cache to find the
writer/decoder function responsible for a type. When analyzing CPU
profiles of geth during sync, I found that the use of sync.RWMutex in
cache lookups appears in the profiles. It seems we are running into
CPU cache contention problems when package rlp is heavily used
on all CPU cores during sync.

This change makes it use atomic.Value + a writer lock instead of
sync.RWMutex. In the common case where the typeinfo entry is present in
the cache, we simply fetch the map and lookup the type.

rlp/decode.go

@@ -245,7 +245,7 @@ func makeListDecoder(typ reflect.Type, tag tags) (decoder, error) {
 		}
 		return decodeByteSlice, nil
 	}
-	etypeinfo := cachedTypeInfo1(etype, tags{})
+	etypeinfo := theTC.infoWhileGenerating(etype, tags{})
 	if etypeinfo.decoderErr != nil {
 		return nil, etypeinfo.decoderErr
 	}
@@ -424,7 +424,7 @@ func zeroFields(structval reflect.Value, fields []field) {
 // makePtrDecoder creates a decoder that decodes into the pointer's element type.
 func makePtrDecoder(typ reflect.Type, tag tags) (decoder, error) {
 	etype := typ.Elem()
-	etypeinfo := cachedTypeInfo1(etype, tags{})
+	etypeinfo := theTC.infoWhileGenerating(etype, tags{})
 	switch {
 	case etypeinfo.decoderErr != nil:
 		return nil, etypeinfo.decoderErr

rlp/encode.go

@@ -517,7 +517,7 @@ func writeInterface(val reflect.Value, w *encbuf) error {
 }
 
 func makeSliceWriter(typ reflect.Type, ts tags) (writer, error) {
-	etypeinfo := cachedTypeInfo1(typ.Elem(), tags{})
+	etypeinfo := theTC.infoWhileGenerating(typ.Elem(), tags{})
 	if etypeinfo.writerErr != nil {
 		return nil, etypeinfo.writerErr
 	}
@@ -585,7 +585,7 @@ func makeStructWriter(typ reflect.Type) (writer, error) {
 }
 
 func makePtrWriter(typ reflect.Type, ts tags) (writer, error) {
-	etypeinfo := cachedTypeInfo1(typ.Elem(), tags{})
+	etypeinfo := theTC.infoWhileGenerating(typ.Elem(), tags{})
 	if etypeinfo.writerErr != nil {
 		return nil, etypeinfo.writerErr
 	}

rlp/encode_test.go

@@ -23,6 +23,7 @@ import (
 	"io"
 	"io/ioutil"
 	"math/big"
+	"runtime"
 	"sync"
 	"testing"
 
@@ -480,3 +481,35 @@ func BenchmarkEncodeBigInts(b *testing.B) {
 		}
 	}
 }
+
+func BenchmarkEncodeConcurrentInterface(b *testing.B) {
+	type struct1 struct {
+		A string
+		B *big.Int
+		C [20]byte
+	}
+	value := []interface{}{
+		uint(999),
+		&struct1{A: "hello", B: big.NewInt(0xFFFFFFFF)},
+		[10]byte{1, 2, 3, 4, 5, 6},
+		[]string{"yeah", "yeah", "yeah"},
+	}
+
+	var wg sync.WaitGroup
+	for cpu := 0; cpu < runtime.NumCPU(); cpu++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+
+			var buffer bytes.Buffer
+			for i := 0; i < b.N; i++ {
+				buffer.Reset()
+				err := Encode(&buffer, value)
+				if err != nil {
+					panic(err)
+				}
+			}
+		}()
+	}
+	wg.Wait()
+}

rlp/typecache.go

@@ -21,13 +21,10 @@ import (
 	"reflect"
 	"strings"
 	"sync"
+	"sync/atomic"
 )
 
-var (
+// typeinfo is an entry in the type cache.
-	typeCacheMutex sync.RWMutex
-	typeCache      = make(map[typekey]*typeinfo)
-)
-
 type typeinfo struct {
 	decoder    decoder
 	decoderErr error // error from makeDecoder
@@ -65,41 +62,76 @@ type decoder func(*Stream, reflect.Value) error
 
 type writer func(reflect.Value, *encbuf) error
 
+var theTC = newTypeCache()
+
+type typeCache struct {
+	cur atomic.Value
+
+	// This lock synchronizes writers.
+	mu   sync.Mutex
+	next map[typekey]*typeinfo
+}
+
+func newTypeCache() *typeCache {
+	c := new(typeCache)
+	c.cur.Store(make(map[typekey]*typeinfo))
+	return c
+}
+
 func cachedDecoder(typ reflect.Type) (decoder, error) {
-	info := cachedTypeInfo(typ, tags{})
+	info := theTC.info(typ)
 	return info.decoder, info.decoderErr
 }
 
 func cachedWriter(typ reflect.Type) (writer, error) {
-	info := cachedTypeInfo(typ, tags{})
+	info := theTC.info(typ)
 	return info.writer, info.writerErr
 }
 
-func cachedTypeInfo(typ reflect.Type, tags tags) *typeinfo {
+func (c *typeCache) info(typ reflect.Type) *typeinfo {
-	typeCacheMutex.RLock()
+	key := typekey{Type: typ}
-	info := typeCache[typekey{typ, tags}]
+	if info := c.cur.Load().(map[typekey]*typeinfo)[key]; info != nil {
-	typeCacheMutex.RUnlock()
-	if info != nil {
 		return info
 	}
-	// not in the cache, need to generate info for this type.
+
-	typeCacheMutex.Lock()
+	// Not in the cache, need to generate info for this type.
-	defer typeCacheMutex.Unlock()
+	return c.generate(typ, tags{})
-	return cachedTypeInfo1(typ, tags)
+}
+
+func (c *typeCache) generate(typ reflect.Type, tags tags) *typeinfo {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	cur := c.cur.Load().(map[typekey]*typeinfo)
+	if info := cur[typekey{typ, tags}]; info != nil {
+		return info
+	}
+
+	// Copy cur to next.
+	c.next = make(map[typekey]*typeinfo, len(cur)+1)
+	for k, v := range cur {
+		c.next[k] = v
+	}
+
+	// Generate.
+	info := c.infoWhileGenerating(typ, tags)
+
+	// next -> cur
+	c.cur.Store(c.next)
+	c.next = nil
+	return info
 }
 
-func cachedTypeInfo1(typ reflect.Type, tags tags) *typeinfo {
+func (c *typeCache) infoWhileGenerating(typ reflect.Type, tags tags) *typeinfo {
 	key := typekey{typ, tags}
-	info := typeCache[key]
+	if info := c.next[key]; info != nil {
-	if info != nil {
-		// another goroutine got the write lock first
 		return info
 	}
-	// put a dummy value into the cache before generating.
+	// Put a dummy value into the cache before generating.
-	// if the generator tries to lookup itself, it will get
+	// If the generator tries to lookup itself, it will get
 	// the dummy value and won't call itself recursively.
-	info = new(typeinfo)
+	info := new(typeinfo)
-	typeCache[key] = info
+	c.next[key] = info
 	info.generate(typ, tags)
 	return info
 }
@@ -133,7 +165,7 @@ func structFields(typ reflect.Type) (fields []field, err error) {
 			} else if anyOptional {
 				return nil, fmt.Errorf(`rlp: struct field %v.%s needs "optional" tag`, typ, f.Name)
 			}
-			info := cachedTypeInfo1(f.Type, tags)
+			info := theTC.infoWhileGenerating(f.Type, tags)
 			fields = append(fields, field{i, info, tags.optional})
 		}
 	}