From a423e54ba43515af045c794529a98ced8b1bb45b Mon Sep 17 00:00:00 2001 From: Burke Libbey Date: Tue, 14 Apr 2015 15:28:54 -0400 Subject: [PATCH] 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 --- archive/changes.go | 74 ++--------- archive/changes_linux.go | 276 +++++++++++++++++++++++++++++++++++++++ archive/changes_other.go | 84 ++++++++++++ system/stat_linux.go | 6 + 4 files changed, 376 insertions(+), 64 deletions(-) create mode 100644 archive/changes_linux.go create mode 100644 archive/changes_other.go diff --git a/archive/changes.go b/archive/changes.go index 06fad8e..30eee60 100644 --- a/archive/changes.go +++ b/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 diff --git a/archive/changes_linux.go b/archive/changes_linux.go new file mode 100644 index 0000000..b1d34b3 --- /dev/null +++ b/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) +} diff --git a/archive/changes_other.go b/archive/changes_other.go new file mode 100644 index 0000000..e2d9c23 --- /dev/null +++ b/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 +} diff --git a/system/stat_linux.go b/system/stat_linux.go index 3899b3e..80262d9 100644 --- a/system/stat_linux.go +++ b/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. //