package archive import ( "archive/tar" "errors" "io" "io/ioutil" "os" "path/filepath" "strings" "github.com/Sirupsen/logrus" "github.com/docker/docker/pkg/system" ) // Errors used or returned by this file. var ( ErrNotDirectory = errors.New("not a directory") ErrDirNotExists = errors.New("no such directory") ErrCannotCopyDir = errors.New("cannot copy directory") ErrInvalidCopySource = errors.New("invalid copy source content") ) // PreserveTrailingDotOrSeparator returns the given cleaned path (after // processing using any utility functions from the path or filepath stdlib // packages) and appends a trailing `/.` or `/` if its corresponding original // path (from before being processed by utility functions from the path or // filepath stdlib packages) ends with a trailing `/.` or `/`. If the cleaned // path already ends in a `.` path segment, then another is not added. If the // clean path already ends in a path separator, then another is not added. func PreserveTrailingDotOrSeparator(cleanedPath, originalPath string) string { // Ensure paths are in platform semantics cleanedPath = normalizePath(cleanedPath) originalPath = normalizePath(originalPath) if !specifiesCurrentDir(cleanedPath) && specifiesCurrentDir(originalPath) { if !hasTrailingPathSeparator(cleanedPath) { // Add a separator if it doesn't already end with one (a cleaned // path would only end in a separator if it is the root). cleanedPath += string(filepath.Separator) } cleanedPath += "." } if !hasTrailingPathSeparator(cleanedPath) && hasTrailingPathSeparator(originalPath) { cleanedPath += string(filepath.Separator) } return cleanedPath } // assertsDirectory returns whether the given path is // asserted to be a directory, i.e., the path ends with // a trailing '/' or `/.`, assuming a path separator of `/`. func assertsDirectory(path string) bool { return hasTrailingPathSeparator(path) || specifiesCurrentDir(path) } // hasTrailingPathSeparator returns whether the given // path ends with the system's path separator character. func hasTrailingPathSeparator(path string) bool { return len(path) > 0 && os.IsPathSeparator(path[len(path)-1]) } // specifiesCurrentDir returns whether the given path specifies // a "current directory", i.e., the last path segment is `.`. func specifiesCurrentDir(path string) bool { return filepath.Base(path) == "." } // SplitPathDirEntry splits the given path between its directory name and its // basename by first cleaning the path but preserves a trailing "." if the // original path specified the current directory. func SplitPathDirEntry(path string) (dir, base string) { cleanedPath := filepath.Clean(normalizePath(path)) if specifiesCurrentDir(path) { cleanedPath += string(filepath.Separator) + "." } return filepath.Dir(cleanedPath), filepath.Base(cleanedPath) } // TarResource archives the resource described by the given CopyInfo to a Tar // archive. A non-nil error is returned if sourcePath does not exist or is // asserted to be a directory but exists as another type of file. // // This function acts as a convenient wrapper around TarWithOptions, which // requires a directory as the source path. TarResource accepts either a // directory or a file path and correctly sets the Tar options. func TarResource(sourceInfo CopyInfo) (content io.ReadCloser, err error) { return TarResourceRebase(sourceInfo.Path, sourceInfo.RebaseName) } // TarResourceRebase is like TarResource but renames the first path element of // items in the resulting tar archive to match the given rebaseName if not "". func TarResourceRebase(sourcePath, rebaseName string) (content io.ReadCloser, err error) { sourcePath = normalizePath(sourcePath) if _, err = os.Lstat(sourcePath); err != nil { // Catches the case where the source does not exist or is not a // directory if asserted to be a directory, as this also causes an // error. return } // Separate the source path between its directory and // the entry in that directory which we are archiving. sourceDir, sourceBase := SplitPathDirEntry(sourcePath) filter := []string{sourceBase} logrus.Debugf("copying %q from %q", sourceBase, sourceDir) return TarWithOptions(sourceDir, &TarOptions{ Compression: Uncompressed, IncludeFiles: filter, IncludeSourceDir: true, RebaseNames: map[string]string{ sourceBase: rebaseName, }, }) } // CopyInfo holds basic info about the source // or destination path of a copy operation. type CopyInfo struct { Path string Exists bool IsDir bool RebaseName string } // CopyInfoSourcePath stats the given path to create a CopyInfo // struct representing that resource for the source of an archive copy // operation. The given path should be an absolute local path. A source path // has all symlinks evaluated that appear before the last path separator ("/" // on Unix). As it is to be a copy source, the path must exist. func CopyInfoSourcePath(path string, followLink bool) (CopyInfo, error) { // normalize the file path and then evaluate the symbol link // we will use the target file instead of the symbol link if // followLink is set path = normalizePath(path) resolvedPath, rebaseName, err := ResolveHostSourcePath(path, followLink) if err != nil { return CopyInfo{}, err } stat, err := os.Lstat(resolvedPath) if err != nil { return CopyInfo{}, err } return CopyInfo{ Path: resolvedPath, Exists: true, IsDir: stat.IsDir(), RebaseName: rebaseName, }, nil } // CopyInfoDestinationPath stats the given path to create a CopyInfo // struct representing that resource for the destination of an archive copy // operation. The given path should be an absolute local path. func CopyInfoDestinationPath(path string) (info CopyInfo, err error) { maxSymlinkIter := 10 // filepath.EvalSymlinks uses 255, but 10 already seems like a lot. path = normalizePath(path) originalPath := path stat, err := os.Lstat(path) if err == nil && stat.Mode()&os.ModeSymlink == 0 { // The path exists and is not a symlink. return CopyInfo{ Path: path, Exists: true, IsDir: stat.IsDir(), }, nil } // While the path is a symlink. for n := 0; err == nil && stat.Mode()&os.ModeSymlink != 0; n++ { if n > maxSymlinkIter { // Don't follow symlinks more than this arbitrary number of times. return CopyInfo{}, errors.New("too many symlinks in " + originalPath) } // The path is a symbolic link. We need to evaluate it so that the // destination of the copy operation is the link target and not the // link itself. This is notably different than CopyInfoSourcePath which // only evaluates symlinks before the last appearing path separator. // Also note that it is okay if the last path element is a broken // symlink as the copy operation should create the target. var linkTarget string linkTarget, err = os.Readlink(path) if err != nil { return CopyInfo{}, err } if !system.IsAbs(linkTarget) { // Join with the parent directory. dstParent, _ := SplitPathDirEntry(path) linkTarget = filepath.Join(dstParent, linkTarget) } path = linkTarget stat, err = os.Lstat(path) } if err != nil { // It's okay if the destination path doesn't exist. We can still // continue the copy operation if the parent directory exists. if !os.IsNotExist(err) { return CopyInfo{}, err } // Ensure destination parent dir exists. dstParent, _ := SplitPathDirEntry(path) parentDirStat, err := os.Lstat(dstParent) if err != nil { return CopyInfo{}, err } if !parentDirStat.IsDir() { return CopyInfo{}, ErrNotDirectory } return CopyInfo{Path: path}, nil } // The path exists after resolving symlinks. return CopyInfo{ Path: path, Exists: true, IsDir: stat.IsDir(), }, nil } // PrepareArchiveCopy prepares the given srcContent archive, which should // contain the archived resource described by srcInfo, to the destination // described by dstInfo. Returns the possibly modified content archive along // with the path to the destination directory which it should be extracted to. func PrepareArchiveCopy(srcContent io.Reader, srcInfo, dstInfo CopyInfo) (dstDir string, content io.ReadCloser, err error) { // Ensure in platform semantics srcInfo.Path = normalizePath(srcInfo.Path) dstInfo.Path = normalizePath(dstInfo.Path) // Separate the destination path between its directory and base // components in case the source archive contents need to be rebased. dstDir, dstBase := SplitPathDirEntry(dstInfo.Path) _, srcBase := SplitPathDirEntry(srcInfo.Path) switch { case dstInfo.Exists && dstInfo.IsDir: // The destination exists as a directory. No alteration // to srcContent is needed as its contents can be // simply extracted to the destination directory. return dstInfo.Path, ioutil.NopCloser(srcContent), nil case dstInfo.Exists && srcInfo.IsDir: // The destination exists as some type of file and the source // content is a directory. This is an error condition since // you cannot copy a directory to an existing file location. return "", nil, ErrCannotCopyDir case dstInfo.Exists: // The destination exists as some type of file and the source content // is also a file. The source content entry will have to be renamed to // have a basename which matches the destination path's basename. if len(srcInfo.RebaseName) != 0 { srcBase = srcInfo.RebaseName } return dstDir, RebaseArchiveEntries(srcContent, srcBase, dstBase), nil case srcInfo.IsDir: // The destination does not exist and the source content is an archive // of a directory. The archive should be extracted to the parent of // the destination path instead, and when it is, the directory that is // created as a result should take the name of the destination path. // The source content entries will have to be renamed to have a // basename which matches the destination path's basename. if len(srcInfo.RebaseName) != 0 { srcBase = srcInfo.RebaseName } return dstDir, RebaseArchiveEntries(srcContent, srcBase, dstBase), nil case assertsDirectory(dstInfo.Path): // The destination does not exist and is asserted to be created as a // directory, but the source content is not a directory. This is an // error condition since you cannot create a directory from a file // source. return "", nil, ErrDirNotExists default: // The last remaining case is when the destination does not exist, is // not asserted to be a directory, and the source content is not an // archive of a directory. It this case, the destination file will need // to be created when the archive is extracted and the source content // entry will have to be renamed to have a basename which matches the // destination path's basename. if len(srcInfo.RebaseName) != 0 { srcBase = srcInfo.RebaseName } return dstDir, RebaseArchiveEntries(srcContent, srcBase, dstBase), nil } } // RebaseArchiveEntries rewrites the given srcContent archive replacing // an occurrence of oldBase with newBase at the beginning of entry names. func RebaseArchiveEntries(srcContent io.Reader, oldBase, newBase string) io.ReadCloser { if oldBase == string(os.PathSeparator) { // If oldBase specifies the root directory, use an empty string as // oldBase instead so that newBase doesn't replace the path separator // that all paths will start with. oldBase = "" } rebased, w := io.Pipe() go func() { srcTar := tar.NewReader(srcContent) rebasedTar := tar.NewWriter(w) for { hdr, err := srcTar.Next() if err == io.EOF { // Signals end of archive. rebasedTar.Close() w.Close() return } if err != nil { w.CloseWithError(err) return } hdr.Name = strings.Replace(hdr.Name, oldBase, newBase, 1) if hdr.Typeflag == tar.TypeLink { hdr.Linkname = strings.Replace(hdr.Linkname, oldBase, newBase, 1) } if err = rebasedTar.WriteHeader(hdr); err != nil { w.CloseWithError(err) return } if _, err = io.Copy(rebasedTar, srcTar); err != nil { w.CloseWithError(err) return } } }() return rebased } // CopyResource performs an archive copy from the given source path to the // given destination path. The source path MUST exist and the destination // path's parent directory must exist. func CopyResource(srcPath, dstPath string, followLink bool) error { var ( srcInfo CopyInfo err error ) // Ensure in platform semantics srcPath = normalizePath(srcPath) dstPath = normalizePath(dstPath) // Clean the source and destination paths. srcPath = PreserveTrailingDotOrSeparator(filepath.Clean(srcPath), srcPath) dstPath = PreserveTrailingDotOrSeparator(filepath.Clean(dstPath), dstPath) if srcInfo, err = CopyInfoSourcePath(srcPath, followLink); err != nil { return err } content, err := TarResource(srcInfo) if err != nil { return err } defer content.Close() return CopyTo(content, srcInfo, dstPath) } // CopyTo handles extracting the given content whose // entries should be sourced from srcInfo to dstPath. func CopyTo(content io.Reader, srcInfo CopyInfo, dstPath string) error { // The destination path need not exist, but CopyInfoDestinationPath will // ensure that at least the parent directory exists. dstInfo, err := CopyInfoDestinationPath(normalizePath(dstPath)) if err != nil { return err } dstDir, copyArchive, err := PrepareArchiveCopy(content, srcInfo, dstInfo) if err != nil { return err } defer copyArchive.Close() options := &TarOptions{ NoLchown: true, NoOverwriteDirNonDir: true, } return Untar(copyArchive, dstDir, options) } // ResolveHostSourcePath decides real path need to be copied with parameters such as // whether to follow symbol link or not, if followLink is true, resolvedPath will return // link target of any symbol link file, else it will only resolve symlink of directory // but return symbol link file itself without resolving. func ResolveHostSourcePath(path string, followLink bool) (resolvedPath, rebaseName string, err error) { if followLink { resolvedPath, err = filepath.EvalSymlinks(path) if err != nil { return } resolvedPath, rebaseName = GetRebaseName(path, resolvedPath) } else { dirPath, basePath := filepath.Split(path) // if not follow symbol link, then resolve symbol link of parent dir var resolvedDirPath string resolvedDirPath, err = filepath.EvalSymlinks(dirPath) if err != nil { return } // resolvedDirPath will have been cleaned (no trailing path separators) so // we can manually join it with the base path element. resolvedPath = resolvedDirPath + string(filepath.Separator) + basePath if hasTrailingPathSeparator(path) && filepath.Base(path) != filepath.Base(resolvedPath) { rebaseName = filepath.Base(path) } } return resolvedPath, rebaseName, nil } // GetRebaseName normalizes and compares path and resolvedPath, // return completed resolved path and rebased file name func GetRebaseName(path, resolvedPath string) (string, string) { // linkTarget will have been cleaned (no trailing path separators and dot) so // we can manually join it with them var rebaseName string if specifiesCurrentDir(path) && !specifiesCurrentDir(resolvedPath) { resolvedPath += string(filepath.Separator) + "." } if hasTrailingPathSeparator(path) && !hasTrailingPathSeparator(resolvedPath) { resolvedPath += string(filepath.Separator) } if filepath.Base(path) != filepath.Base(resolvedPath) { // In the case where the path had a trailing separator and a symlink // evaluation has changed the last path component, we will need to // rebase the name in the archive that is being copied to match the // originally requested name. rebaseName = filepath.Base(path) } return resolvedPath, rebaseName }