Merge pull request #9819 from miminar/remote-layer-federation

Merged by openshift-bot

images/dockerregistry/config.yml

@@ -23,5 +23,6 @@ middleware:
         pullthrough: true
         enforcequota: false
         projectcachettl: 1m
+        blobrepositorycachettl: 10m
   storage:
     - name: openshift

pkg/dockerregistry/server/auth.go

@@ -40,26 +40,36 @@ const (
 	TokenRealmKey = "token-realm"
 )
 
+// RegistryClient encapsulates getting access to the OpenShift API.
+type RegistryClient interface {
+	// Clients return the authenticated client to use with the server.
+	Clients() (client.Interface, kclient.Interface, error)
+	// SafeClientConfig returns a client config without authentication info.
+	SafeClientConfig() restclient.Config
+}
+
 // DefaultRegistryClient is exposed for testing the registry with fake client.
 var DefaultRegistryClient = NewRegistryClient(clientcmd.NewConfig().BindToFile())
 
-// RegistryClient encapsulates getting access to the OpenShift API.
-type RegistryClient struct {
+// registryClient implements RegistryClient
+type registryClient struct {
 	config *clientcmd.Config
 }
 
+var _ RegistryClient = &registryClient{}
+
 // NewRegistryClient creates a registry client.
-func NewRegistryClient(config *clientcmd.Config) *RegistryClient {
-	return &RegistryClient{config: config}
+func NewRegistryClient(config *clientcmd.Config) RegistryClient {
+	return &registryClient{config: config}
 }
 
 // Client returns the authenticated client to use with the server.
-func (r *RegistryClient) Clients() (client.Interface, kclient.Interface, error) {
+func (r *registryClient) Clients() (client.Interface, kclient.Interface, error) {
 	return r.config.Clients()
 }
 
 // SafeClientConfig returns a client config without authentication info.
-func (r *RegistryClient) SafeClientConfig() restclient.Config {
+func (r *registryClient) SafeClientConfig() restclient.Config {
 	return clientcmd.AnonymousClientConfig(r.config.OpenShiftConfig())
 }
 
@@ -327,22 +337,6 @@ func (ac *AccessController) Authorized(ctx context.Context, accessRecords ...reg
 	return WithUserClient(ctx, osClient), nil
 }
 
-func getNamespaceName(resourceName string) (string, string, error) {
-	repoParts := strings.SplitN(resourceName, "/", 2)
-	if len(repoParts) != 2 {
-		return "", "", ErrNamespaceRequired
-	}
-	ns := repoParts[0]
-	if len(ns) == 0 {
-		return "", "", ErrNamespaceRequired
-	}
-	name := repoParts[1]
-	if len(name) == 0 {
-		return "", "", ErrNamespaceRequired
-	}
-	return ns, name, nil
-}
-
 func getOpenShiftAPIToken(ctx context.Context, req *http.Request) (string, error) {
 	token := ""
 

pkg/dockerregistry/server/blobdescriptorservice.go

@@ -1,13 +1,28 @@
 package server
 
 import (
+	"fmt"
+	"sort"
+	"time"
+
 	"github.com/docker/distribution"
 	"github.com/docker/distribution/context"
 	"github.com/docker/distribution/digest"
 	"github.com/docker/distribution/registry/middleware/registry"
 	"github.com/docker/distribution/registry/storage"
+
+	kerrors "k8s.io/kubernetes/pkg/api/errors"
+
+	imageapi "github.com/openshift/origin/pkg/image/api"
 )
 
+// ByGeneration allows for sorting tag events from latest to oldest.
+type ByGeneration []*imageapi.TagEvent
+
+func (b ByGeneration) Less(i, j int) bool { return b[i].Generation > b[j].Generation }
+func (b ByGeneration) Len() int           { return len(b) }
+func (b ByGeneration) Swap(i, j int)      { b[i], b[j] = b[j], b[i] }
+
 func init() {
 	middleware.RegisterOptions(storage.BlobDescriptorServiceFactory(&blobDescriptorServiceFactory{}))
 }
@@ -25,6 +40,168 @@ type blobDescriptorService struct {
 	distribution.BlobDescriptorService
 }
 
+// Stat returns a a blob descriptor if the given blob is either linked in repository or is referenced in
+// corresponding image stream. This method is invoked from inside of upstream's linkedBlobStore. It expects
+// a proper repository object to be set on given context by upper openshift middleware wrappers.
 func (bs *blobDescriptorService) Stat(ctx context.Context, dgst digest.Digest) (distribution.Descriptor, error) {
-	return dockerRegistry.BlobStatter().Stat(ctx, dgst)
+	repo, found := RepositoryFrom(ctx)
+	if !found || repo == nil {
+		err := fmt.Errorf("failed to retrieve repository from context")
+		context.GetLogger(ctx).Error(err)
+		return distribution.Descriptor{}, err
+	}
+
+	// if there is a repo layer link, return its descriptor
+	desc, err := bs.BlobDescriptorService.Stat(ctx, dgst)
+	if err == nil {
+		// and remember the association
+		repo.cachedLayers.RememberDigest(dgst, repo.blobrepositorycachettl, imageapi.DockerImageReference{
+			Namespace: repo.namespace,
+			Name:      repo.name,
+		}.Exact())
+		return desc, nil
+	}
+
+	context.GetLogger(ctx).Debugf("could not stat layer link %q in repository %q: %v", dgst.String(), repo.Named().Name(), err)
+
+	// verify the blob is stored locally
+	desc, err = dockerRegistry.BlobStatter().Stat(ctx, dgst)
+	if err != nil {
+		return desc, err
+	}
+
+	// ensure it's referenced inside of corresponding image stream
+	if imageStreamHasBlob(repo, dgst) {
+		return desc, nil
+	}
+
+	return distribution.Descriptor{}, distribution.ErrBlobUnknown
+}
+
+func (bs *blobDescriptorService) Clear(ctx context.Context, dgst digest.Digest) error {
+	repo, found := RepositoryFrom(ctx)
+	if !found || repo == nil {
+		err := fmt.Errorf("failed to retrieve repository from context")
+		context.GetLogger(ctx).Error(err)
+		return err
+	}
+
+	repo.cachedLayers.ForgetDigest(dgst, imageapi.DockerImageReference{
+		Namespace: repo.namespace,
+		Name:      repo.name,
+	}.Exact())
+	return bs.BlobDescriptorService.Clear(ctx, dgst)
+}
+
+// imageStreamHasBlob returns true if the given blob digest is referenced in image stream corresponding to
+// given repository. If not found locally, image stream's images will be iterated and fetched from newest to
+// oldest until found. Each processed image will update local cache of blobs.
+func imageStreamHasBlob(r *repository, dgst digest.Digest) bool {
+	repoCacheName := imageapi.DockerImageReference{Namespace: r.namespace, Name: r.name}.Exact()
+	if r.cachedLayers.RepositoryHasBlob(repoCacheName, dgst) {
+		context.GetLogger(r.ctx).Debugf("found cached blob %q in repository %s", dgst.String(), r.Named().Name())
+		return true
+	}
+
+	context.GetLogger(r.ctx).Debugf("verifying presence of blob %q in image stream %s/%s", dgst.String(), r.namespace, r.name)
+	started := time.Now()
+	logFound := func(found bool) bool {
+		elapsed := time.Now().Sub(started)
+		if found {
+			context.GetLogger(r.ctx).Debugf("verified presence of blob %q in image stream %s/%s after %s", dgst.String(), r.namespace, r.name, elapsed.String())
+		} else {
+			context.GetLogger(r.ctx).Debugf("detected absence of blob %q in image stream %s/%s after %s", dgst.String(), r.namespace, r.name, elapsed.String())
+		}
+		return found
+	}
+
+	// verify directly with etcd
+	is, err := r.getImageStream()
+	if err != nil {
+		context.GetLogger(r.ctx).Errorf("failed to get image stream: %v", err)
+		return logFound(false)
+	}
+
+	tagEvents := []*imageapi.TagEvent{}
+	event2Name := make(map[*imageapi.TagEvent]string)
+	for name, eventList := range is.Status.Tags {
+		for i := range eventList.Items {
+			event := &eventList.Items[i]
+			tagEvents = append(tagEvents, event)
+			event2Name[event] = name
+		}
+	}
+	// search from youngest to oldest
+	sort.Sort(ByGeneration(tagEvents))
+
+	processedImages := map[string]struct{}{}
+
+	for _, tagEvent := range tagEvents {
+		if _, processed := processedImages[tagEvent.Image]; processed {
+			continue
+		}
+		if imageHasBlob(r, repoCacheName, tagEvent.Image, dgst.String(), !r.pullthrough) {
+			tagName := event2Name[tagEvent]
+			context.GetLogger(r.ctx).Debugf("blob found under istag %s/%s:%s in image %s", r.namespace, r.name, tagName, tagEvent.Image)
+			return logFound(true)
+		}
+		processedImages[tagEvent.Image] = struct{}{}
+	}
+
+	context.GetLogger(r.ctx).Warnf("blob %q exists locally but is not referenced in repository %s/%s", dgst.String(), r.namespace, r.name)
+
+	return logFound(false)
+}
+
+// imageHasBlob returns true if the image identified by imageName refers to the given blob. The image is
+// fetched. If requireManaged is true and the image is not managed (it refers to remote registry), the image
+// will not be processed. Fetched image will update local cache of blobs -> repositories with (blobDigest,
+// cacheName) pairs.
+func imageHasBlob(
+	r *repository,
+	cacheName,
+	imageName,
+	blobDigest string,
+	requireManaged bool,
+) bool {
+	context.GetLogger(r.ctx).Debugf("getting image %s", imageName)
+	image, err := r.getImage(digest.Digest(imageName))
+	if err != nil {
+		if kerrors.IsNotFound(err) {
+			context.GetLogger(r.ctx).Debugf("image %q not found: imageName")
+		} else {
+			context.GetLogger(r.ctx).Errorf("failed to get image: %v", err)
+		}
+		return false
+	}
+
+	// in case of pullthrough disabled, client won't be able to download a blob belonging to not managed image
+	// (image stored in external registry), thus don't consider them as candidates
+	if managed := image.Annotations[imageapi.ManagedByOpenShiftAnnotation]; requireManaged && managed != "true" {
+		context.GetLogger(r.ctx).Debugf("skipping not managed image")
+		return false
+	}
+
+	if len(image.DockerImageLayers) == 0 {
+		if len(image.DockerImageManifestMediaType) > 0 {
+			// If the media type is set, we can safely assume that the best effort to fill the image layers
+			// has already been done. There are none.
+			return false
+		}
+		err = imageapi.ImageWithMetadata(image)
+		if err != nil {
+			context.GetLogger(r.ctx).Errorf("failed to get metadata for image %s: %v", imageName, err)
+			return false
+		}
+	}
+
+	for _, layer := range image.DockerImageLayers {
+		if layer.Name == blobDigest {
+			// remember all the layers of matching image
+			r.rememberLayersOfImage(image, cacheName)
+			return true
+		}
+	}
+
+	return false
 }