docker镜像是如何导入的？

镜像导入是由image/tarexport/load.go#tarexporter.Load()完成的

以下代码参考github.com/docker/docker版本v0.0.0-20181129155816-baab736a3649

主要是注册镜像信息以及解包镜像tar流到新root

导出和保存的区别在于

• 导出(export): 仅导出文件结构
• 保存(save): 保存镜像历史和元数据

这意味着导出将不会包含USER、EXPOSE等Dockerfile里面的命令，也就无法转移镜像到另一台机器上了

func (l *tarexporter) Load(inTar io.ReadCloser, outStream io.Writer, quiet bool) error {
	var progressOutput progress.Output
	if !quiet {
		progressOutput = streamformatter.NewJSONProgressOutput(outStream, false)
	}
	outStream = streamformatter.NewStdoutWriter(outStream)

  // 1. 创建docker-import的临时目录
	tmpDir, err := ioutil.TempDir("", "docker-import-")
	if err != nil {
		return err
	}
	defer os.RemoveAll(tmpDir)

  // 2. 解包tar流到临时目录
	if err := chrootarchive.Untar(inTar, tmpDir, nil); err != nil {
		return err
	}

  // 3. 打开manifest文件，并解析
	manifestPath, err := safePath(tmpDir, manifestFileName)
	if err != nil {
		return err
	}
	manifestFile, err := os.Open(manifestPath)
	if err != nil {
		if os.IsNotExist(err) {
			return l.legacyLoad(tmpDir, outStream, progressOutput)
		}
		return err
	}
	defer manifestFile.Close()

	var manifest []manifestItem
	if err := json.NewDecoder(manifestFile).Decode(&manifest); err != nil {
		return err
	}

	var parentLinks []parentLink
	var imageIDsStr string
	var imageRefCount int

  // 4. 从manifest中读取并解析到image
	for _, m := range manifest {
		configPath, err := safePath(tmpDir, m.Config)
		if err != nil {
			return err
		}
		config, err := ioutil.ReadFile(configPath)
		if err != nil {
			return err
		}
		img, err := image.NewFromJSON(config)
		if err != nil {
			return err
		}
		if err := checkCompatibleOS(img.OS); err != nil {
			return err
		}
		rootFS := *img.RootFS
		rootFS.DiffIDs = nil

    // 若image rootFS diffID数量与manifest中记录的层数不一致，则报错
		if expected, actual := len(m.Layers), len(img.RootFS.DiffIDs); expected != actual {
			return fmt.Errorf("invalid manifest, layers length mismatch: expected %d, got %d", expected, actual)
		}

		// On Windows, validate the platform, defaulting to windows if not present.
		os := img.OS
		if os == "" {
			os = runtime.GOOS
		}
		if runtime.GOOS == "windows" {
			if (os != "windows") && (os != "linux") {
				return fmt.Errorf("configuration for this image has an unsupported operating system: %s", os)
			}
		}

    // 5. 注册层
		for i, diffID := range img.RootFS.DiffIDs {
			layerPath, err := safePath(tmpDir, m.Layers[i])
			if err != nil {
				return err
			}
			r := rootFS
			r.Append(diffID)
			newLayer, err := l.lss[os].Get(r.ChainID())
			if err != nil {
        // 如果没有注册，那就注册layer
				newLayer, err = l.loadLayer(layerPath, rootFS, diffID.String(), os, m.LayerSources[diffID], progressOutput)
				if err != nil {
					return err
				}
			}
			defer layer.ReleaseAndLog(l.lss[os], newLayer)
      // 若manifest与缓存中layer diffID不一致，则报错
			if expected, actual := diffID, newLayer.DiffID(); expected != actual {
				return fmt.Errorf("invalid diffID for layer %d: expected %q, got %q", i, expected, actual)
			}
			rootFS.Append(diffID)
		}

    // 6. 缓存该层镜像配置
		imgID, err := l.is.Create(config)
		if err != nil {
			return err
		}
		imageIDsStr += fmt.Sprintf("Loaded image ID: %s\n", imgID)

		imageRefCount = 0
		for _, repoTag := range m.RepoTags {
			named, err := reference.ParseNormalizedNamed(repoTag)
			if err != nil {
				return err
			}
			ref, ok := named.(reference.NamedTagged)
			if !ok {
				return fmt.Errorf("invalid tag %q", repoTag)
			}
      // 设置已加载的id、reference
			l.setLoadedTag(ref, imgID.Digest(), outStream)
			outStream.Write([]byte(fmt.Sprintf("Loaded image: %s\n", reference.FamiliarString(ref))))
			imageRefCount++
		}

		parentLinks = append(parentLinks, parentLink{imgID, m.Parent})
		l.loggerImgEvent.LogImageEvent(imgID.String(), imgID.String(), "load")
	}

	for _, p := range validatedParentLinks(parentLinks) {
		if p.parentID != "" {
			if err := l.setParentID(p.id, p.parentID); err != nil {
				return err
			}
		}
	}

	if imageRefCount == 0 {
		outStream.Write([]byte(imageIDsStr))
	}

	return nil
}

Untar

主要过程是将tar流解包到新root

untar操作实际由chrootarchive/archive_unix.go untar()执行

// untar is the entry-point for docker-untar on re-exec. This is not used on
// Windows as it does not support chroot, hence no point sandboxing through
// chroot and rexec.
func untar() {
	runtime.LockOSThread()
	flag.Parse()

	var options *archive.TarOptions

	//read the options from the pipe "ExtraFiles"
	if err := json.NewDecoder(os.NewFile(3, "options")).Decode(&options); err != nil {
		fatal(err)
	}

  // Linux上的Chroot使用pivot_root，而不是Chroot。 pivot_root需要一个新根和一个旧根。旧根必须是新根的子目录，它是调用pivot_root后当前rootfs驻留的位置。New root是新rootfs设置的位置。在调用pivot_root之后，旧根会被移除，因此在新根下不再可用。这类似于libcontainer设置容器rootfs的方式
  // 在这里是以前面创建的临时目录作为新root，并在其下创建privot_root作为老root，最后切换到新root
	if err := chroot(flag.Arg(0)); err != nil {
		fatal(err)
	}

  // 将tar流解包到新root
	if err := archive.Unpack(os.Stdin, "/", options); err != nil {
		fatal(err)
	}
	// fully consume stdin in case it is zero padded
	if _, err := flush(os.Stdin); err != nil {
		fatal(err)
	}

	os.Exit(0)
}

loadLayer

注册镜像层以及加载层tar流到对应目录下

image/tarexport/load.go#tarexpoter.loadLayer()

func (l *tarexporter) loadLayer(filename string, rootFS image.RootFS, id string, os string, foreignSrc distribution.Descriptor, progressOutput progress.Output) (layer.Layer, error) {
	// We use system.OpenSequential to use sequential file access on Windows, avoiding
	// depleting the standby list. On Linux, this equates to a regular os.Open.
	rawTar, err := system.OpenSequential(filename)
	if err != nil {
		logrus.Debugf("Error reading embedded tar: %v", err)
		return nil, err
	}
	defer rawTar.Close()

	var r io.Reader
	if progressOutput != nil {
		fileInfo, err := rawTar.Stat()
		if err != nil {
			logrus.Debugf("Error statting file: %v", err)
			return nil, err
		}

		r = progress.NewProgressReader(rawTar, progressOutput, fileInfo.Size(), stringid.TruncateID(id), "Loading layer")
	} else {
		r = rawTar
	}

	inflatedLayerData, err := archive.DecompressStream(r)
	if err != nil {
		return nil, err
	}
	defer inflatedLayerData.Close()

	if ds, ok := l.lss[os].(layer.DescribableStore); ok {
		return ds.RegisterWithDescriptor(inflatedLayerData, rootFS.ChainID(), foreignSrc)
	}
  // 到这里是去注册层tar流和本层镜像的chainID
	return l.lss[os].Register(inflatedLayerData, rootFS.ChainID())
}

func (ls *layerStore) registerWithDescriptor(ts io.Reader, parent ChainID, descriptor distribution.Descriptor) (Layer, error) {
	// err is used to hold the error which will always trigger
	// cleanup of creates sources but may not be an error returned
	// to the caller (already exists).
	var err error
	var pid string
	var p *roLayer

  // 1. 从缓存中获取到给定chainID的层信息
	if string(parent) != "" {
		p = ls.get(parent)
		if p == nil {
			return nil, ErrLayerDoesNotExist
		}
		pid = p.cacheID
		// Release parent chain if error
		defer func() {
			if err != nil {
				ls.layerL.Lock()
				ls.releaseLayer(p)
				ls.layerL.Unlock()
			}
		}()
		if p.depth() >= maxLayerDepth {
			err = ErrMaxDepthExceeded
			return nil, err
		}
	}

	// 2. 创建新的只读层
	layer := &roLayer{
		parent:         p,
		cacheID:        stringid.GenerateRandomID(),
		referenceCount: 1,
		layerStore:     ls,
		references:     map[Layer]struct{}{},
		descriptor:     descriptor,
	}

  // 3. 准备文件系统(overlay2)文件目录结构
	if err = ls.driver.Create(layer.cacheID, pid, nil); err != nil {
		return nil, err
	}

	tx, err := ls.store.StartTransaction()
	if err != nil {
		return nil, err
	}

	defer func() {
		if err != nil {
			logrus.Debugf("Cleaning up layer %s: %v", layer.cacheID, err)
			if err := ls.driver.Remove(layer.cacheID); err != nil {
				logrus.Errorf("Error cleaning up cache layer %s: %v", layer.cacheID, err)
			}
			if err := tx.Cancel(); err != nil {
				logrus.Errorf("Error canceling metadata transaction %q: %s", tx.String(), err)
			}
		}
	}()

  // 4. 从给定读写层流中提取变化的内容到镜像层挂载点
	if err = ls.applyTar(tx, ts, pid, layer); err != nil {
		return nil, err
	}

  // 5. 若本层无父层，那么chainID就是自己的diffID。否则从parent和自己的diffID中生成
	if layer.parent == nil {
		layer.chainID = ChainID(layer.diffID)
	} else {
		layer.chainID = createChainIDFromParent(layer.parent.chainID, layer.diffID)
	}

  // 6. 储存层diffID、size、cacheID、descriptor、parent、os等信息
	if err = storeLayer(tx, layer); err != nil {
		return nil, err
	}

	ls.layerL.Lock()
	defer ls.layerL.Unlock()

	if existingLayer := ls.getWithoutLock(layer.chainID); existingLayer != nil {
		// Set error for cleanup, but do not return the error
		err = errors.New("layer already exists")
		return existingLayer.getReference(), nil
	}

	if err = tx.Commit(layer.chainID); err != nil {
		return nil, err
	}

	ls.layerMap[layer.chainID] = layer

	return layer.getReference(), nil
}

driver.Create

为镜像层创建diff、work、lower目录，并写入镜像层tar流lower内容到对应lower目录

func (d *Driver) create(id, parent string, opts *graphdriver.CreateOpts) (retErr error) {
	dir := d.dir(id)

  // 1. 获取当前用户在宿主机对应的userID、groupID
	rootUID, rootGID, err := idtools.GetRootUIDGID(d.uidMaps, d.gidMaps)
	if err != nil {
		return err
	}
	root := idtools.Identity{UID: rootUID, GID: rootGID}

  // 2. 为当前用户创建镜像目录
	if err := idtools.MkdirAllAndChown(path.Dir(dir), 0700, root); err != nil {
		return err
	}
	if err := idtools.MkdirAndChown(dir, 0700, root); err != nil {
		return err
	}

	defer func() {
		// Clean up on failure
		if retErr != nil {
			os.RemoveAll(dir)
		}
	}()

  // 3. 解析储存选项
	if opts != nil && len(opts.StorageOpt) > 0 {
		driver := &Driver{}
		if err := d.parseStorageOpt(opts.StorageOpt, driver); err != nil {
			return err
		}

    // 4. 设置储存配额
		if driver.options.quota.Size > 0 {
			// Set container disk quota limit
			if err := d.quotaCtl.SetQuota(dir, driver.options.quota); err != nil {
				return err
			}
		}
	}

  // 5. 创建镜像diff目录
	if err := idtools.MkdirAndChown(path.Join(dir, "diff"), 0755, root); err != nil {
		return err
	}

  // 6. 创建指向diff目录的链接
	lid := generateID(idLength)
	if err := os.Symlink(path.Join("..", id, "diff"), path.Join(d.home, linkDir, lid)); err != nil {
		return err
	}

	// 7. 将链接id写入链接文件
	if err := ioutil.WriteFile(path.Join(dir, "link"), []byte(lid), 0644); err != nil {
		return err
	}

	// 8. 父层不存在就直接返回
	if parent == "" {
		return nil
	}

  // 9. 创建镜像work目录作为overlay2内部使用
	if err := idtools.MkdirAndChown(path.Join(dir, "work"), 0700, root); err != nil {
		return err
	}

  // 10. 找到父层(也就是tar中的镜像层)lower文件，并写入到当前层lower文件中
	lower, err := d.getLower(parent)
	if err != nil {
		return err
	}
	if lower != "" {
		if err := ioutil.WriteFile(path.Join(dir, lowerFile), []byte(lower), 0666); err != nil {
			return err
		}
	}

	return nil
}

applyTar

将层tar流解包到层挂载点

func (gdw *NaiveDiffDriver) ApplyDiff(id, parent string, diff io.Reader) (size int64, err error) {
	driver := gdw.ProtoDriver

	// Mount the root filesystem so we can apply the diff/layer.
  // 返回由id引用的分层文件系统的挂载点
	layerRootFs, err := driver.Get(id, "")
	if err != nil {
		return
	}
	defer driver.Put(id)

	layerFs := layerRootFs.Path()
	options := &archive.TarOptions{UIDMaps: gdw.uidMaps,
		GIDMaps: gdw.gidMaps}
	start := time.Now().UTC()
	logrus.WithField("id", id).Debug("Start untar layer")
  // 将层tar流解包到层挂载点
	if size, err = ApplyUncompressedLayer(layerFs, diff, options); err != nil {
		return
	}
	logrus.WithField("id", id).Debugf("Untar time: %vs", time.Now().UTC().Sub(start).Seconds())

	return
}

Create

创建就是在缓存中添加镜像信息，保存配置

func (is *store) Create(config []byte) (ID, error) {
	var img Image
	err := json.Unmarshal(config, &img)
	if err != nil {
		return "", err
	}

	// Must reject any config that references diffIDs from the history
	// which aren't among the rootfs layers.
	rootFSLayers := make(map[layer.DiffID]struct{})
	for _, diffID := range img.RootFS.DiffIDs {
		rootFSLayers[diffID] = struct{}{}
	}

  // 如果记录的创建历史非空层大于rootFS层数，报错
	layerCounter := 0
	for _, h := range img.History {
		if !h.EmptyLayer {
			layerCounter++
		}
	}
	if layerCounter > len(img.RootFS.DiffIDs) {
		return "", errors.New("too many non-empty layers in History section")
	}

  // 将解析配置写入content目录
	dgst, err := is.fs.Set(config)
	if err != nil {
		return "", err
	}
	imageID := IDFromDigest(dgst)

	is.Lock()
	defer is.Unlock()

  // 若镜像已经存在镜像元数据缓存中，就直接返回
	if _, exists := is.images[imageID]; exists {
		return imageID, nil
	}

	layerID := img.RootFS.ChainID()

	var l layer.Layer
  // 获取镜像只读层，并缓存
	if layerID != "" {
		if !system.IsOSSupported(img.OperatingSystem()) {
			return "", system.ErrNotSupportedOperatingSystem
		}
		l, err = is.lss[img.OperatingSystem()].Get(layerID)
		if err != nil {
			return "", errors.Wrapf(err, "failed to get layer %s", layerID)
		}
	}

	imageMeta := &imageMeta{
		layer:    l,
		children: make(map[ID]struct{}),
	}

	is.images[imageID] = imageMeta
  // 添加reference和id缓存
	if err := is.digestSet.Add(imageID.Digest()); err != nil {
		delete(is.images, imageID)
		return "", err
	}

	return imageID, nil
}

创建容器时是如何使用image的？

1. 从缓存获取镜像配置进行校验以及合并容器配置
2. 以镜像chainID作为容器挂载层(也是读写层)的parent
3. 复制镜像目录内容到容器目录

Ref

1. https://stackoverflow.com/questions/22655867/what-is-the-difference-between-save-and-export-in-docker