archive: Optimize ChangesDirs on Linux

If we tear through a few layers of abstraction, we can get at the inodes
contained in a directory without having to stat all the files. This
allows us to eliminate identical files much earlier in the changelist
generation process.

Signed-off-by: Burke Libbey <burke@libbey.me>

pkg/archive/changes.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"fmt"
 	"io"
+	"io/ioutil"
 	"os"
 	"path/filepath"
 	"sort"
@@ -277,78 +278,23 @@ func newRootFileInfo() *FileInfo {
 	return root
 }
 
-func collectFileInfo(sourceDir string) (*FileInfo, error) {
-	root := newRootFileInfo()
-
-	err := filepath.Walk(sourceDir, func(path string, f os.FileInfo, err error) error {
-		if err != nil {
-			return err
-		}
-
-		// Rebase path
-		relPath, err := filepath.Rel(sourceDir, path)
-		if err != nil {
-			return err
-		}
-		relPath = filepath.Join("/", relPath)
-
-		if relPath == "/" {
-			return nil
-		}
-
-		parent := root.LookUp(filepath.Dir(relPath))
-		if parent == nil {
-			return fmt.Errorf("collectFileInfo: Unexpectedly no parent for %s", relPath)
-		}
-
-		info := &FileInfo{
-			name:     filepath.Base(relPath),
-			children: make(map[string]*FileInfo),
-			parent:   parent,
-		}
-
-		s, err := system.Lstat(path)
-		if err != nil {
-			return err
-		}
-		info.stat = s
-
-		info.capability, _ = system.Lgetxattr(path, "security.capability")
-
-		parent.children[info.name] = info
-
-		return nil
-	})
-	if err != nil {
-		return nil, err
-	}
-	return root, nil
-}
-
 // ChangesDirs compares two directories and generates an array of Change objects describing the changes.
 // If oldDir is "", then all files in newDir will be Add-Changes.
 func ChangesDirs(newDir, oldDir string) ([]Change, error) {
 	var (
 		oldRoot, newRoot *FileInfo
-		err1, err2       error
-		errs             = make(chan error, 2)
 	)
-	go func() {
-		if oldDir != "" {
-			oldRoot, err1 = collectFileInfo(oldDir)
-		}
-		errs <- err1
-	}()
-	go func() {
-		newRoot, err2 = collectFileInfo(newDir)
-		errs <- err2
-	}()
-
-	// block until both routines have returned
-	for i := 0; i < 2; i++ {
-		if err := <-errs; err != nil {
+	if oldDir == "" {
+		emptyDir, err := ioutil.TempDir("", "empty")
+		if err != nil {
 			return nil, err
 		}
+		defer os.Remove(emptyDir)
+		oldDir = emptyDir
+	}
+	oldRoot, newRoot, err := collectFileInfoForChanges(oldDir, newDir)
+	if err != nil {
+		return nil, err
 	}
 
 	return newRoot.Changes(oldRoot), nil

pkg/archive/changes_linux.go

@@ -0,0 +1,276 @@
+package archive
+
+import (
+	"bytes"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"syscall"
+	"unsafe"
+
+	"github.com/docker/docker/pkg/system"
+)
+
+// walker is used to implement collectFileInfoForChanges on linux. Where this
+// method in general returns the entire contents of two directory trees, we
+// optimize some FS calls out on linux. In particular, we take advantage of the
+// fact that getdents(2) returns the inode of each file in the directory being
+// walked, which, when walking two trees in parallel to generate a list of
+// changes, can be used to prune subtrees without ever having to lstat(2) them
+// directly. Eliminating stat calls in this way can save up to seconds on large
+// images.
+type walker struct {
+	dir1  string
+	dir2  string
+	root1 *FileInfo
+	root2 *FileInfo
+}
+
+// collectFileInfoForChanges returns a complete representation of the trees
+// rooted at dir1 and dir2, with one important exception: any subtree or
+// leaf where the inode and device numbers are an exact match between dir1
+// and dir2 will be pruned from the results. This method is *only* to be used
+// to generating a list of changes between the two directories, as it does not
+// reflect the full contents.
+func collectFileInfoForChanges(dir1, dir2 string) (*FileInfo, *FileInfo, error) {
+	w := &walker{
+		dir1:  dir1,
+		dir2:  dir2,
+		root1: newRootFileInfo(),
+		root2: newRootFileInfo(),
+	}
+
+	i1, err := os.Lstat(w.dir1)
+	if err != nil {
+		return nil, nil, err
+	}
+	i2, err := os.Lstat(w.dir2)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	if err := w.walk("/", i1, i2); err != nil {
+		return nil, nil, err
+	}
+
+	return w.root1, w.root2, nil
+}
+
+// Given a FileInfo, its path info, and a reference to the root of the tree
+// being constructed, register this file with the tree.
+func walkchunk(path string, fi os.FileInfo, dir string, root *FileInfo) error {
+	if fi == nil {
+		return nil
+	}
+	parent := root.LookUp(filepath.Dir(path))
+	if parent == nil {
+		return fmt.Errorf("collectFileInfoForChanges: Unexpectedly no parent for %s", path)
+	}
+	info := &FileInfo{
+		name:     filepath.Base(path),
+		children: make(map[string]*FileInfo),
+		parent:   parent,
+	}
+	cpath := filepath.Join(dir, path)
+	stat, err := system.FromStatT(fi.Sys().(*syscall.Stat_t))
+	if err != nil {
+		return err
+	}
+	info.stat = stat
+	info.capability, _ = system.Lgetxattr(cpath, "security.capability") // lgetxattr(2): fs access
+	parent.children[info.name] = info
+	return nil
+}
+
+// Walk a subtree rooted at the same path in both trees being iterated. For
+// example, /docker/overlay/1234/a/b/c/d and /docker/overlay/8888/a/b/c/d
+func (w *walker) walk(path string, i1, i2 os.FileInfo) (err error) {
+	// Register these nodes with the return trees, unless we're still at the
+	// (already-created) roots:
+	if path != "/" {
+		if err := walkchunk(path, i1, w.dir1, w.root1); err != nil {
+			return err
+		}
+		if err := walkchunk(path, i2, w.dir2, w.root2); err != nil {
+			return err
+		}
+	}
+
+	is1Dir := i1 != nil && i1.IsDir()
+	is2Dir := i2 != nil && i2.IsDir()
+
+	// If these files are both non-existent, or leaves (non-dirs), we are done.
+	if !is1Dir && !is2Dir {
+		return nil
+	}
+
+	// Fetch the names of all the files contained in both directories being walked:
+	var names1, names2 []nameIno
+	if is1Dir {
+		names1, err = readdirnames(filepath.Join(w.dir1, path)) // getdents(2): fs access
+		if err != nil {
+			return err
+		}
+	}
+	if is2Dir {
+		names2, err = readdirnames(filepath.Join(w.dir2, path)) // getdents(2): fs access
+		if err != nil {
+			return err
+		}
+	}
+
+	// We have lists of the files contained in both parallel directories, sorted
+	// in the same order. Walk them in parallel, generating a unique merged list
+	// of all items present in either or both directories.
+	var names []string
+	ix1 := 0
+	ix2 := 0
+
+	for {
+		if ix1 >= len(names1) {
+			break
+		}
+		if ix2 >= len(names2) {
+			break
+		}
+
+		ni1 := names1[ix1]
+		ni2 := names2[ix2]
+
+		switch bytes.Compare([]byte(ni1.name), []byte(ni2.name)) {
+		case -1: // ni1 < ni2 -- advance ni1
+			// we will not encounter ni1 in names2
+			names = append(names, ni1.name)
+			ix1++
+		case 0: // ni1 == ni2
+			if ni1.ino != ni2.ino {
+				names = append(names, ni1.name)
+			}
+			ix1++
+			ix2++
+		case 1: // ni1 > ni2 -- advance ni2
+			// we will not encounter ni2 in names1
+			names = append(names, ni2.name)
+			ix2++
+		}
+	}
+	for ix1 < len(names1) {
+		names = append(names, names1[ix1].name)
+		ix1++
+	}
+	for ix2 < len(names2) {
+		names = append(names, names2[ix2].name)
+		ix2++
+	}
+
+	// For each of the names present in either or both of the directories being
+	// iterated, stat the name under each root, and recurse the pair of them:
+	for _, name := range names {
+		fname := filepath.Join(path, name)
+		var cInfo1, cInfo2 os.FileInfo
+		if is1Dir {
+			cInfo1, err = os.Lstat(filepath.Join(w.dir1, fname)) // lstat(2): fs access
+			if err != nil && !os.IsNotExist(err) {
+				return err
+			}
+		}
+		if is2Dir {
+			cInfo2, err = os.Lstat(filepath.Join(w.dir2, fname)) // lstat(2): fs access
+			if err != nil && !os.IsNotExist(err) {
+				return err
+			}
+		}
+		if err = w.walk(fname, cInfo1, cInfo2); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// {name,inode} pairs used to support the early-pruning logic of the walker type
+type nameIno struct {
+	name string
+	ino  uint64
+}
+
+type nameInoSlice []nameIno
+
+func (s nameInoSlice) Len() int           { return len(s) }
+func (s nameInoSlice) Swap(i, j int)      { s[i], s[j] = s[j], s[i] }
+func (s nameInoSlice) Less(i, j int) bool { return s[i].name < s[j].name }
+
+// readdirnames is a hacked-apart version of the Go stdlib code, exposing inode
+// numbers further up the stack when reading directory contents. Unlike
+// os.Readdirnames, which returns a list of filenames, this function returns a
+// list of {filename,inode} pairs.
+func readdirnames(dirname string) (names []nameIno, err error) {
+	var (
+		size = 100
+		buf  = make([]byte, 4096)
+		nbuf int
+		bufp int
+		nb   int
+	)
+
+	f, err := os.Open(dirname)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	names = make([]nameIno, 0, size) // Empty with room to grow.
+	for {
+		// Refill the buffer if necessary
+		if bufp >= nbuf {
+			bufp = 0
+			nbuf, err = syscall.ReadDirent(int(f.Fd()), buf) // getdents on linux
+			if nbuf < 0 {
+				nbuf = 0
+			}
+			if err != nil {
+				return nil, os.NewSyscallError("readdirent", err)
+			}
+			if nbuf <= 0 {
+				break // EOF
+			}
+		}
+
+		// Drain the buffer
+		nb, names = parseDirent(buf[bufp:nbuf], names)
+		bufp += nb
+	}
+
+	sl := nameInoSlice(names)
+	sort.Sort(sl)
+	return sl, nil
+}
+
+// parseDirent is a minor modification of syscall.ParseDirent (linux version)
+// which returns {name,inode} pairs instead of just names.
+func parseDirent(buf []byte, names []nameIno) (consumed int, newnames []nameIno) {
+	origlen := len(buf)
+	for len(buf) > 0 {
+		dirent := (*syscall.Dirent)(unsafe.Pointer(&buf[0]))
+		buf = buf[dirent.Reclen:]
+		if dirent.Ino == 0 { // File absent in directory.
+			continue
+		}
+		bytes := (*[10000]byte)(unsafe.Pointer(&dirent.Name[0]))
+		var name = string(bytes[0:clen(bytes[:])])
+		if name == "." || name == ".." { // Useless names
+			continue
+		}
+		names = append(names, nameIno{name, dirent.Ino})
+	}
+	return origlen - len(buf), names
+}
+
+func clen(n []byte) int {
+	for i := 0; i < len(n); i++ {
+		if n[i] == 0 {
+			return i
+		}
+	}
+	return len(n)
+}

pkg/archive/changes_other.go

@@ -0,0 +1,84 @@
+// +build !linux
+
+package archive
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+
+	"github.com/docker/docker/pkg/system"
+)
+
+func collectFileInfoForChanges(oldDir, newDir string) (*FileInfo, *FileInfo, error) {
+	var (
+		oldRoot, newRoot *FileInfo
+		err1, err2       error
+		errs             = make(chan error, 2)
+	)
+	go func() {
+		oldRoot, err1 = collectFileInfo(oldDir)
+		errs <- err1
+	}()
+	go func() {
+		newRoot, err2 = collectFileInfo(newDir)
+		errs <- err2
+	}()
+
+	// block until both routines have returned
+	for i := 0; i < 2; i++ {
+		if err := <-errs; err != nil {
+			return nil, nil, err
+		}
+	}
+
+	return oldRoot, newRoot, nil
+}
+
+func collectFileInfo(sourceDir string) (*FileInfo, error) {
+	root := newRootFileInfo()
+
+	err := filepath.Walk(sourceDir, func(path string, f os.FileInfo, err error) error {
+		if err != nil {
+			return err
+		}
+
+		// Rebase path
+		relPath, err := filepath.Rel(sourceDir, path)
+		if err != nil {
+			return err
+		}
+		relPath = filepath.Join("/", relPath)
+
+		if relPath == "/" {
+			return nil
+		}
+
+		parent := root.LookUp(filepath.Dir(relPath))
+		if parent == nil {
+			return fmt.Errorf("collectFileInfo: Unexpectedly no parent for %s", relPath)
+		}
+
+		info := &FileInfo{
+			name:     filepath.Base(relPath),
+			children: make(map[string]*FileInfo),
+			parent:   parent,
+		}
+
+		s, err := system.Lstat(path)
+		if err != nil {
+			return err
+		}
+		info.stat = s
+
+		info.capability, _ = system.Lgetxattr(path, "security.capability")
+
+		parent.children[info.name] = info
+
+		return nil
+	})
+	if err != nil {
+		return nil, err
+	}
+	return root, nil
+}

pkg/system/stat_linux.go

@@ -14,6 +14,12 @@ func fromStatT(s *syscall.Stat_t) (*Stat_t, error) {
 		mtim: s.Mtim}, nil
 }
 
+// FromStatT exists only on linux, and loads a system.Stat_t from a
+// syscal.Stat_t.
+func FromStatT(s *syscall.Stat_t) (*Stat_t, error) {
+	return fromStatT(s)
+}
+
 // Stat takes a path to a file and returns
 // a system.Stat_t type pertaining to that file.
 //