Find the source code for computing size of a docker image
Before digging too deep, you may find it useful to understand how Linux implements the overlay filesystem. I include a bit on this the first exercise of my intro presentation's build section. The demo notes include each of the commands I'm running and it gives you an idea of how layers are merged, and what happens when you add/modify/delete from a layer.
This is implementation dependent, based on your host OS and the graph driver being used. I'm taking the example of a Linux OS and Overlay2 since that's the most common use case.
It starts by looking at the image layer storage size:
// GetContainerLayerSize returns the real size & virtual size of the container.
func (i *ImageService) GetContainerLayerSize(containerID string) (int64, int64) {
var (
sizeRw, sizeRootfs int64
err error
)
// Safe to index by runtime.GOOS as Unix hosts don't support multiple
// container operating systems.
rwlayer, err := i.layerStores[runtime.GOOS].GetRWLayer(containerID)
if err != nil {
logrus.Errorf("Failed to compute size of container rootfs %v: %v", containerID, err)
return sizeRw, sizeRootfs
}
defer i.layerStores[runtime.GOOS].ReleaseRWLayer(rwlayer)
sizeRw, err = rwlayer.Size()
if err != nil {
logrus.Errorf("Driver %s couldn't return diff size of container %s: %s",
i.layerStores[runtime.GOOS].DriverName(), containerID, err)
// FIXME: GetSize should return an error. Not changing it now in case
// there is a side-effect.
sizeRw = -1
}
if parent := rwlayer.Parent(); parent != nil {
sizeRootfs, err = parent.Size()
if err != nil {
sizeRootfs = -1
} else if sizeRw != -1 {
sizeRootfs += sizeRw
}
}
return sizeRw, sizeRootfs
}
In there is a call to layerStores
which itself is a mapping to layer.Store:
// ImageServiceConfig is the configuration used to create a new ImageService
type ImageServiceConfig struct {
ContainerStore containerStore
DistributionMetadataStore metadata.Store
EventsService *daemonevents.Events
ImageStore image.Store
LayerStores map[string]layer.Store
MaxConcurrentDownloads int
MaxConcurrentUploads int
MaxDownloadAttempts int
ReferenceStore dockerreference.Store
RegistryService registry.Service
TrustKey libtrust.PrivateKey
}
Digging into the layer.Store
implementation for GetRWLayer
, there is the following definition:
func (ls *layerStore) GetRWLayer(id string) (RWLayer, error) {
ls.locker.Lock(id)
defer ls.locker.Unlock(id)
ls.mountL.Lock()
mount := ls.mounts[id]
ls.mountL.Unlock()
if mount == nil {
return nil, ErrMountDoesNotExist
}
return mount.getReference(), nil
}
Following that to find the Size
implementation for the mount reference, there is this function that gets into the specific graph driver:
func (ml *mountedLayer) Size() (int64, error) {
return ml.layerStore.driver.DiffSize(ml.mountID, ml.cacheParent())
}
Looking at the overlay2 graph driver to find the DiffSize function:
func (d *Driver) DiffSize(id, parent string) (size int64, err error) {
if useNaiveDiff(d.home) || !d.isParent(id, parent) {
return d.naiveDiff.DiffSize(id, parent)
}
return directory.Size(context.TODO(), d.getDiffPath(id))
}
That is calling naiveDiff
which implements Size in the graphDriver package:
func (gdw *NaiveDiffDriver) DiffSize(id, parent string) (size int64, err error) {
driver := gdw.ProtoDriver
changes, err := gdw.Changes(id, parent)
if err != nil {
return
}
layerFs, err := driver.Get(id, "")
if err != nil {
return
}
defer driver.Put(id)
return archive.ChangesSize(layerFs.Path(), changes), nil
}
Following archive.ChangeSize
we can see this implementation:
// ChangesSize calculates the size in bytes of the provided changes, based on newDir.
func ChangesSize(newDir string, changes []Change) int64 {
var (
size int64
sf = make(map[uint64]struct{})
)
for _, change := range changes {
if change.Kind == ChangeModify || change.Kind == ChangeAdd {
file := filepath.Join(newDir, change.Path)
fileInfo, err := os.Lstat(file)
if err != nil {
logrus.Errorf("Can not stat %q: %s", file, err)
continue
}
if fileInfo != nil && !fileInfo.IsDir() {
if hasHardlinks(fileInfo) {
inode := getIno(fileInfo)
if _, ok := sf[inode]; !ok {
size += fileInfo.Size()
sf[inode] = struct{}{}
}
} else {
size += fileInfo.Size()
}
}
}
}
return size
}
At which point we are using os.Lstat
to return a struct that includes Size
on each entry that is an add or modify to each directory. Note that this is one of several possible paths the code takes, but I believe it's one of the more common ones for this scenario.