[performance] update storage backend and make use of seek syscall when available (#2924)

* update to use go-storage/ instead of go-store/v2/storage/

* pull in latest version from codeberg

* remove test output 😇

* add code comments

* set the exclusive bit when creating new files in disk config

* bump to actual release version

* bump to v0.1.1 (tis a simple no-logic change)

* update readme

* only use a temporary read seeker when decoding video if required (should only be S3 now)

* use fastcopy library to use memory pooled buffers when calling TempFileSeeker()

* update to use seek call in serveFileRange()

vendor/codeberg.org/gruf/go-storage/block.archived

@@ -0,0 +1,887 @@
+package storage
+
+import (
+	"bytes"
+	"context"
+	"crypto/sha256"
+	"fmt"
+	"io"
+	"io/fs"
+	"os"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"codeberg.org/gruf/go-byteutil"
+	"codeberg.org/gruf/go-errors/v2"
+	"codeberg.org/gruf/go-fastcopy"
+	"codeberg.org/gruf/go-hashenc"
+	"codeberg.org/gruf/go-iotools"
+	"codeberg.org/gruf/go-pools"
+	"codeberg.org/gruf/go-store/v2/util"
+)
+
+var (
+	nodePathPrefix  = "node/"
+	blockPathPrefix = "block/"
+)
+
+// DefaultBlockConfig is the default BlockStorage configuration.
+var DefaultBlockConfig = &BlockConfig{
+	BlockSize:    1024 * 16,
+	WriteBufSize: 4096,
+	Overwrite:    false,
+	Compression:  NoCompression(),
+}
+
+// BlockConfig defines options to be used when opening a BlockStorage.
+type BlockConfig struct {
+	// BlockSize is the chunking size to use when splitting and storing blocks of data.
+	BlockSize int
+
+	// ReadBufSize is the buffer size to use when reading node files.
+	ReadBufSize int
+
+	// WriteBufSize is the buffer size to use when writing file streams.
+	WriteBufSize int
+
+	// Overwrite allows overwriting values of stored keys in the storage.
+	Overwrite bool
+
+	// Compression is the Compressor to use when reading / writing files,
+	// default is no compression.
+	Compression Compressor
+}
+
+// getBlockConfig returns a valid BlockConfig for supplied ptr.
+func getBlockConfig(cfg *BlockConfig) BlockConfig {
+	// If nil, use default
+	if cfg == nil {
+		cfg = DefaultBlockConfig
+	}
+
+	// Assume nil compress == none
+	if cfg.Compression == nil {
+		cfg.Compression = NoCompression()
+	}
+
+	// Assume 0 chunk size == use default
+	if cfg.BlockSize <= 0 {
+		cfg.BlockSize = DefaultBlockConfig.BlockSize
+	}
+
+	// Assume 0 buf size == use default
+	if cfg.WriteBufSize <= 0 {
+		cfg.WriteBufSize = DefaultDiskConfig.WriteBufSize
+	}
+
+	// Return owned config copy
+	return BlockConfig{
+		BlockSize:    cfg.BlockSize,
+		WriteBufSize: cfg.WriteBufSize,
+		Overwrite:    cfg.Overwrite,
+		Compression:  cfg.Compression,
+	}
+}
+
+// BlockStorage is a Storage implementation that stores input data as chunks on
+// a filesystem. Each value is chunked into blocks of configured size and these
+// blocks are stored with name equal to their base64-encoded SHA256 hash-sum. A
+// "node" file is finally created containing an array of hashes contained within
+// this value.
+type BlockStorage struct {
+	path      string            // path is the root path of this store
+	blockPath string            // blockPath is the joined root path + block path prefix
+	nodePath  string            // nodePath is the joined root path + node path prefix
+	config    BlockConfig       // cfg is the supplied configuration for this store
+	hashPool  sync.Pool         // hashPool is this store's hashEncoder pool
+	bufpool   pools.BufferPool  // bufpool is this store's bytes.Buffer pool
+	cppool    fastcopy.CopyPool // cppool is the prepared io copier with buffer pool
+	lock      *Lock             // lock is the opened lockfile for this storage instance
+
+	// NOTE:
+	// BlockStorage does not need to lock each of the underlying block files
+	// as the filename itself directly relates to the contents. If there happens
+	// to be an overwrite, it will just be of the same data since the filename is
+	// the hash of the data.
+}
+
+// OpenBlock opens a BlockStorage instance for given folder path and configuration.
+func OpenBlock(path string, cfg *BlockConfig) (*BlockStorage, error) {
+	// Acquire path builder
+	pb := util.GetPathBuilder()
+	defer util.PutPathBuilder(pb)
+
+	// Clean provided path, ensure ends in '/' (should
+	// be dir, this helps with file path trimming later)
+	path = pb.Clean(path) + "/"
+
+	// Get checked config
+	config := getBlockConfig(cfg)
+
+	// Attempt to open path
+	file, err := os.OpenFile(path, defaultFileROFlags, defaultDirPerms)
+	if err != nil {
+		// If not a not-exist error, return
+		if !os.IsNotExist(err) {
+			return nil, err
+		}
+
+		// Attempt to make store path dirs
+		err = os.MkdirAll(path, defaultDirPerms)
+		if err != nil {
+			return nil, err
+		}
+
+		// Reopen dir now it's been created
+		file, err = os.OpenFile(path, defaultFileROFlags, defaultDirPerms)
+		if err != nil {
+			return nil, err
+		}
+	}
+	defer file.Close()
+
+	// Double check this is a dir (NOT a file!)
+	stat, err := file.Stat()
+	if err != nil {
+		return nil, err
+	} else if !stat.IsDir() {
+		return nil, new_error("path is file")
+	}
+
+	// Open and acquire storage lock for path
+	lock, err := OpenLock(pb.Join(path, LockFile))
+	if err != nil {
+		return nil, err
+	}
+
+	// Figure out the largest size for bufpool slices
+	bufSz := encodedHashLen
+	if bufSz < config.BlockSize {
+		bufSz = config.BlockSize
+	}
+	if bufSz < config.WriteBufSize {
+		bufSz = config.WriteBufSize
+	}
+
+	// Prepare BlockStorage
+	st := &BlockStorage{
+		path:      path,
+		blockPath: pb.Join(path, blockPathPrefix),
+		nodePath:  pb.Join(path, nodePathPrefix),
+		config:    config,
+		hashPool: sync.Pool{
+			New: func() interface{} {
+				return newHashEncoder()
+			},
+		},
+		bufpool: pools.NewBufferPool(bufSz),
+		lock:    lock,
+	}
+
+	// Set copypool buffer size
+	st.cppool.Buffer(config.ReadBufSize)
+
+	return st, nil
+}
+
+// Clean implements storage.Clean().
+func (st *BlockStorage) Clean(ctx context.Context) error {
+	// Check if open
+	if st.lock.Closed() {
+		return ErrClosed
+	}
+
+	// Check context still valid
+	if err := ctx.Err(); err != nil {
+		return err
+	}
+
+	// Acquire path builder
+	pb := util.GetPathBuilder()
+	defer util.PutPathBuilder(pb)
+
+	nodes := map[string]*node{}
+
+	// Walk nodes dir for entries
+	err := walkDir(pb, st.nodePath, func(npath string, fsentry fs.DirEntry) error {
+		// Only deal with regular files
+		if !fsentry.Type().IsRegular() {
+			return nil
+		}
+
+		// Get joined node path name
+		npath = pb.Join(npath, fsentry.Name())
+
+		// Attempt to open RO file
+		file, err := open(npath, defaultFileROFlags)
+		if err != nil {
+			return err
+		}
+		defer file.Close()
+
+		// Alloc new Node + acquire hash buffer for writes
+		hbuf := st.bufpool.Get()
+		defer st.bufpool.Put(hbuf)
+		hbuf.Guarantee(encodedHashLen)
+		node := node{}
+
+		// Write file contents to node
+		_, err = io.CopyBuffer(
+			&nodeWriter{
+				node: &node,
+				buf:  hbuf,
+			},
+			file,
+			nil,
+		)
+		if err != nil {
+			return err
+		}
+
+		// Append to nodes slice
+		nodes[fsentry.Name()] = &node
+		return nil
+	})
+
+	// Handle errors (though nodePath may not have been created yet)
+	if err != nil && !os.IsNotExist(err) {
+		return err
+	}
+
+	// Walk blocks dir for entries
+	err = walkDir(pb, st.blockPath, func(bpath string, fsentry fs.DirEntry) error {
+		// Only deal with regular files
+		if !fsentry.Type().IsRegular() {
+			return nil
+		}
+
+		inUse := false
+		for key, node := range nodes {
+			if node.removeHash(fsentry.Name()) {
+				if len(node.hashes) < 1 {
+					// This node contained hash, and after removal is now empty.
+					// Remove this node from our tracked nodes slice
+					delete(nodes, key)
+				}
+				inUse = true
+			}
+		}
+
+		// Block hash is used by node
+		if inUse {
+			return nil
+		}
+
+		// Get joined block path name
+		bpath = pb.Join(bpath, fsentry.Name())
+
+		// Remove this unused block path
+		return os.Remove(bpath)
+	})
+
+	// Handle errors (though blockPath may not have been created yet)
+	if err != nil && !os.IsNotExist(err) {
+		return err
+	}
+
+	// If there are nodes left at this point, they are corrupt
+	// (i.e. they're referencing block hashes that don't exist)
+	if len(nodes) > 0 {
+		nodeKeys := []string{}
+		for key := range nodes {
+			nodeKeys = append(nodeKeys, key)
+		}
+		return fmt.Errorf("store/storage: corrupted nodes: %v", nodeKeys)
+	}
+
+	return nil
+}
+
+// ReadBytes implements Storage.ReadBytes().
+func (st *BlockStorage) ReadBytes(ctx context.Context, key string) ([]byte, error) {
+	// Get stream reader for key
+	rc, err := st.ReadStream(ctx, key)
+	if err != nil {
+		return nil, err
+	}
+	defer rc.Close()
+
+	// Read all bytes and return
+	return io.ReadAll(rc)
+}
+
+// ReadStream implements Storage.ReadStream().
+func (st *BlockStorage) ReadStream(ctx context.Context, key string) (io.ReadCloser, error) {
+	// Get node file path for key
+	npath, err := st.nodePathForKey(key)
+	if err != nil {
+		return nil, err
+	}
+
+	// Check if open
+	if st.lock.Closed() {
+		return nil, ErrClosed
+	}
+
+	// Check context still valid
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+
+	// Attempt to open RO file
+	file, err := open(npath, defaultFileROFlags)
+	if err != nil {
+		return nil, errSwapNotFound(err)
+	}
+	defer file.Close()
+
+	// Acquire hash buffer for writes
+	hbuf := st.bufpool.Get()
+	defer st.bufpool.Put(hbuf)
+
+	var node node
+
+	// Write file contents to node
+	_, err = st.cppool.Copy(
+		&nodeWriter{
+			node: &node,
+			buf:  hbuf,
+		},
+		file,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	// Prepare block reader and return
+	return iotools.NopReadCloser(&blockReader{
+		storage: st,
+		node:    &node,
+	}), nil
+}
+
+// readBlock reads the block with hash (key) from the filesystem.
+func (st *BlockStorage) readBlock(key string) ([]byte, error) {
+	// Get block file path for key
+	bpath := st.blockPathForKey(key)
+
+	// Attempt to open RO file
+	file, err := open(bpath, defaultFileROFlags)
+	if err != nil {
+		return nil, wrap(new_error("corrupted node"), err)
+	}
+	defer file.Close()
+
+	// Wrap the file in a compressor
+	cFile, err := st.config.Compression.Reader(file)
+	if err != nil {
+		return nil, wrap(new_error("corrupted node"), err)
+	}
+	defer cFile.Close()
+
+	// Read the entire file
+	return io.ReadAll(cFile)
+}
+
+// WriteBytes implements Storage.WriteBytes().
+func (st *BlockStorage) WriteBytes(ctx context.Context, key string, value []byte) (int, error) {
+	n, err := st.WriteStream(ctx, key, bytes.NewReader(value))
+	return int(n), err
+}
+
+// WriteStream implements Storage.WriteStream().
+func (st *BlockStorage) WriteStream(ctx context.Context, key string, r io.Reader) (int64, error) {
+	// Get node file path for key
+	npath, err := st.nodePathForKey(key)
+	if err != nil {
+		return 0, err
+	}
+
+	// Check if open
+	if st.lock.Closed() {
+		return 0, ErrClosed
+	}
+
+	// Check context still valid
+	if err := ctx.Err(); err != nil {
+		return 0, err
+	}
+
+	// Check if this exists
+	ok, err := stat(key)
+	if err != nil {
+		return 0, err
+	}
+
+	// Check if we allow overwrites
+	if ok && !st.config.Overwrite {
+		return 0, ErrAlreadyExists
+	}
+
+	// Ensure nodes dir (and any leading up to) exists
+	err = os.MkdirAll(st.nodePath, defaultDirPerms)
+	if err != nil {
+		return 0, err
+	}
+
+	// Ensure blocks dir (and any leading up to) exists
+	err = os.MkdirAll(st.blockPath, defaultDirPerms)
+	if err != nil {
+		return 0, err
+	}
+
+	var node node
+	var total atomic.Int64
+
+	// Acquire HashEncoder
+	hc := st.hashPool.Get().(*hashEncoder)
+	defer st.hashPool.Put(hc)
+
+	// Create new waitgroup and OnceError for
+	// goroutine error tracking and propagating
+	wg := sync.WaitGroup{}
+	onceErr := errors.OnceError{}
+
+loop:
+	for !onceErr.IsSet() {
+		// Fetch new buffer for this loop
+		buf := st.bufpool.Get()
+		buf.Grow(st.config.BlockSize)
+
+		// Read next chunk
+		n, err := io.ReadFull(r, buf.B)
+		switch err {
+		case nil, io.ErrUnexpectedEOF:
+			// do nothing
+		case io.EOF:
+			st.bufpool.Put(buf)
+			break loop
+		default:
+			st.bufpool.Put(buf)
+			return 0, err
+		}
+
+		// Hash the encoded data
+		sum := hc.EncodeSum(buf.B)
+
+		// Append to the node's hashes
+		node.hashes = append(node.hashes, sum)
+
+		// If already on disk, skip
+		has, err := st.statBlock(sum)
+		if err != nil {
+			st.bufpool.Put(buf)
+			return 0, err
+		} else if has {
+			st.bufpool.Put(buf)
+			continue loop
+		}
+
+		// Check if reached EOF
+		atEOF := (n < buf.Len())
+
+		wg.Add(1)
+		go func() {
+			// Perform writes in goroutine
+
+			defer func() {
+				// Defer release +
+				// signal we're done
+				st.bufpool.Put(buf)
+				wg.Done()
+			}()
+
+			// Write block to store at hash
+			n, err := st.writeBlock(sum, buf.B[:n])
+			if err != nil {
+				onceErr.Store(err)
+				return
+			}
+
+			// Increment total.
+			total.Add(int64(n))
+		}()
+
+		// Break at end
+		if atEOF {
+			break loop
+		}
+	}
+
+	// Wait, check errors
+	wg.Wait()
+	if onceErr.IsSet() {
+		return 0, onceErr.Load()
+	}
+
+	// If no hashes created, return
+	if len(node.hashes) < 1 {
+		return 0, new_error("no hashes written")
+	}
+
+	// Prepare to swap error if need-be
+	errSwap := errSwapNoop
+
+	// Build file RW flags
+	// NOTE: we performed an initial check for
+	//       this before writing blocks, but if
+	//       the utilizer of this storage didn't
+	//       correctly mutex protect this key then
+	//       someone may have beaten us to the
+	//       punch at writing the node file.
+	flags := defaultFileRWFlags
+	if !st.config.Overwrite {
+		flags |= syscall.O_EXCL
+
+		// Catch + replace err exist
+		errSwap = errSwapExist
+	}
+
+	// Attempt to open RW file
+	file, err := open(npath, flags)
+	if err != nil {
+		return 0, errSwap(err)
+	}
+	defer file.Close()
+
+	// Acquire write buffer
+	buf := st.bufpool.Get()
+	defer st.bufpool.Put(buf)
+	buf.Grow(st.config.WriteBufSize)
+
+	// Finally, write data to file
+	_, err = io.CopyBuffer(file, &nodeReader{node: node}, buf.B)
+	return total.Load(), err
+}
+
+// writeBlock writes the block with hash and supplied value to the filesystem.
+func (st *BlockStorage) writeBlock(hash string, value []byte) (int, error) {
+	// Get block file path for key
+	bpath := st.blockPathForKey(hash)
+
+	// Attempt to open RW file
+	file, err := open(bpath, defaultFileRWFlags)
+	if err != nil {
+		if err == syscall.EEXIST {
+			err = nil /* race issue describe in struct NOTE */
+		}
+		return 0, err
+	}
+	defer file.Close()
+
+	// Wrap the file in a compressor
+	cFile, err := st.config.Compression.Writer(file)
+	if err != nil {
+		return 0, err
+	}
+	defer cFile.Close()
+
+	// Write value to file
+	return cFile.Write(value)
+}
+
+// statBlock checks for existence of supplied block hash.
+func (st *BlockStorage) statBlock(hash string) (bool, error) {
+	return stat(st.blockPathForKey(hash))
+}
+
+// Stat implements Storage.Stat()
+func (st *BlockStorage) Stat(ctx context.Context, key string) (bool, error) {
+	// Get node file path for key
+	kpath, err := st.nodePathForKey(key)
+	if err != nil {
+		return false, err
+	}
+
+	// Check if open
+	if st.lock.Closed() {
+		return false, ErrClosed
+	}
+
+	// Check context still valid
+	if err := ctx.Err(); err != nil {
+		return false, err
+	}
+
+	// Check for file on disk
+	return stat(kpath)
+}
+
+// Remove implements Storage.Remove().
+func (st *BlockStorage) Remove(ctx context.Context, key string) error {
+	// Get node file path for key
+	kpath, err := st.nodePathForKey(key)
+	if err != nil {
+		return err
+	}
+
+	// Check if open
+	if st.lock.Closed() {
+		return ErrClosed
+	}
+
+	// Check context still valid
+	if err := ctx.Err(); err != nil {
+		return err
+	}
+
+	// Remove at path (we know this is file)
+	if err := unlink(kpath); err != nil {
+		return errSwapNotFound(err)
+	}
+
+	return nil
+}
+
+// Close implements Storage.Close().
+func (st *BlockStorage) Close() error {
+	return st.lock.Close()
+}
+
+// WalkKeys implements Storage.WalkKeys().
+func (st *BlockStorage) WalkKeys(ctx context.Context, opts WalkKeysOptions) error {
+	// Check if open
+	if st.lock.Closed() {
+		return ErrClosed
+	}
+
+	// Check context still valid
+	if err := ctx.Err(); err != nil {
+		return err
+	}
+
+	// Acquire path builder
+	pb := util.GetPathBuilder()
+	defer util.PutPathBuilder(pb)
+
+	// Walk dir for entries
+	return walkDir(pb, st.nodePath, func(npath string, fsentry fs.DirEntry) error {
+		if !fsentry.Type().IsRegular() {
+			// Only deal with regular files
+			return nil
+		}
+
+		// Perform provided walk function
+		return opts.WalkFn(ctx, Entry{
+			Key:  fsentry.Name(),
+			Size: -1,
+		})
+	})
+}
+
+// nodePathForKey calculates the node file path for supplied key.
+func (st *BlockStorage) nodePathForKey(key string) (string, error) {
+	// Path separators are illegal, as directory paths
+	if strings.Contains(key, "/") || key == "." || key == ".." {
+		return "", ErrInvalidKey
+	}
+
+	// Acquire path builder
+	pb := util.GetPathBuilder()
+	defer util.PutPathBuilder(pb)
+
+	// Return joined + cleaned node-path
+	return pb.Join(st.nodePath, key), nil
+}
+
+// blockPathForKey calculates the block file path for supplied hash.
+func (st *BlockStorage) blockPathForKey(hash string) string {
+	pb := util.GetPathBuilder()
+	defer util.PutPathBuilder(pb)
+	return pb.Join(st.blockPath, hash)
+}
+
+// hashSeparator is the separating byte between block hashes.
+const hashSeparator = byte('\n')
+
+// node represents the contents of a node file in storage.
+type node struct {
+	hashes []string
+}
+
+// removeHash attempts to remove supplied block hash from the node's hash array.
+func (n *node) removeHash(hash string) bool {
+	for i := 0; i < len(n.hashes); {
+		if n.hashes[i] == hash {
+			// Drop this hash from slice
+			n.hashes = append(n.hashes[:i], n.hashes[i+1:]...)
+			return true
+		}
+
+		// Continue iter
+		i++
+	}
+	return false
+}
+
+// nodeReader is an io.Reader implementation for the node file representation,
+// which is useful when calculated node file is being written to the store.
+type nodeReader struct {
+	node node
+	idx  int
+	last int
+}
+
+func (r *nodeReader) Read(b []byte) (int, error) {
+	n := 0
+
+	// '-1' means we missed writing
+	// hash separator on last iteration
+	if r.last == -1 {
+		b[n] = hashSeparator
+		n++
+		r.last = 0
+	}
+
+	for r.idx < len(r.node.hashes) {
+		hash := r.node.hashes[r.idx]
+
+		// Copy into buffer + update read count
+		m := copy(b[n:], hash[r.last:])
+		n += m
+
+		// If incomplete copy, return here
+		if m < len(hash)-r.last {
+			r.last = m
+			return n, nil
+		}
+
+		// Check we can write last separator
+		if n == len(b) {
+			r.last = -1
+			return n, nil
+		}
+
+		// Write separator, iter, reset
+		b[n] = hashSeparator
+		n++
+		r.idx++
+		r.last = 0
+	}
+
+	// We reached end of hashes
+	return n, io.EOF
+}
+
+// nodeWriter is an io.Writer implementation for the node file representation,
+// which is useful when calculated node file is being read from the store.
+type nodeWriter struct {
+	node *node
+	buf  *byteutil.Buffer
+}
+
+func (w *nodeWriter) Write(b []byte) (int, error) {
+	n := 0
+
+	for {
+		// Find next hash separator position
+		idx := bytes.IndexByte(b[n:], hashSeparator)
+		if idx == -1 {
+			// Check we shouldn't be expecting it
+			if w.buf.Len() > encodedHashLen {
+				return n, new_error("invalid node")
+			}
+
+			// Write all contents to buffer
+			w.buf.Write(b[n:])
+			return len(b), nil
+		}
+
+		// Found hash separator, write
+		// current buf contents to Node hashes
+		w.buf.Write(b[n : n+idx])
+		n += idx + 1
+		if w.buf.Len() != encodedHashLen {
+			return n, new_error("invalid node")
+		}
+
+		// Append to hashes & reset
+		w.node.hashes = append(w.node.hashes, w.buf.String())
+		w.buf.Reset()
+	}
+}
+
+// blockReader is an io.Reader implementation for the combined, linked block
+// data contained with a node file. Basically, this allows reading value data
+// from the store for a given node file.
+type blockReader struct {
+	storage *BlockStorage
+	node    *node
+	buf     []byte
+	prev    int
+}
+
+func (r *blockReader) Read(b []byte) (int, error) {
+	n := 0
+
+	// Data left in buf, copy as much as we
+	// can into supplied read buffer
+	if r.prev < len(r.buf)-1 {
+		n += copy(b, r.buf[r.prev:])
+		r.prev += n
+		if n >= len(b) {
+			return n, nil
+		}
+	}
+
+	for {
+		// Check we have any hashes left
+		if len(r.node.hashes) < 1 {
+			return n, io.EOF
+		}
+
+		// Get next key from slice
+		key := r.node.hashes[0]
+		r.node.hashes = r.node.hashes[1:]
+
+		// Attempt to fetch next batch of data
+		var err error
+		r.buf, err = r.storage.readBlock(key)
+		if err != nil {
+			return n, err
+		}
+		r.prev = 0
+
+		// Copy as much as can from new buffer
+		m := copy(b[n:], r.buf)
+		r.prev += m
+		n += m
+
+		// If we hit end of supplied buf, return
+		if n >= len(b) {
+			return n, nil
+		}
+	}
+}
+
+var (
+	// base64Encoding is our base64 encoding object.
+	base64Encoding = hashenc.Base64()
+
+	// encodedHashLen is the once-calculated encoded hash-sum length
+	encodedHashLen = base64Encoding.EncodedLen(
+		sha256.New().Size(),
+	)
+)
+
+// hashEncoder is a HashEncoder with built-in encode buffer.
+type hashEncoder struct {
+	henc hashenc.HashEncoder
+	ebuf []byte
+}
+
+// newHashEncoder returns a new hashEncoder instance.
+func newHashEncoder() *hashEncoder {
+	return &hashEncoder{
+		henc: hashenc.New(sha256.New(), base64Encoding),
+		ebuf: make([]byte, encodedHashLen),
+	}
+}
+
+// EncodeSum encodes the src data and returns resulting bytes, only valid until next call to EncodeSum().
+func (henc *hashEncoder) EncodeSum(src []byte) string {
+	henc.henc.EncodeSum(henc.ebuf, src)
+	return string(henc.ebuf)
+}