[ca] Root rotation reconciliation loop in the CA server

ca/reconciler.go

@@ -0,0 +1,259 @@
+package ca
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"reflect"
+	"sync"
+	"time"
+
+	"github.com/cloudflare/cfssl/helpers"
+	"github.com/docker/swarmkit/api"
+	"github.com/docker/swarmkit/api/equality"
+	"github.com/docker/swarmkit/log"
+	"github.com/docker/swarmkit/manager/state/store"
+	"github.com/pkg/errors"
+)
+
+// IssuanceStateRotateMaxBatchSize is the maximum number of nodes we'll tell to rotate their certificates in any given update
+const IssuanceStateRotateMaxBatchSize = 30
+
+func hasIssuer(n *api.Node, info *IssuerInfo) bool {
+	if n.Description == nil || n.Description.TLSInfo == nil {
+		return false
+	}
+	return bytes.Equal(info.Subject, n.Description.TLSInfo.CertIssuerSubject) && bytes.Equal(info.PublicKey, n.Description.TLSInfo.CertIssuerPublicKey)
+}
+
+var errRootRotationChanged = errors.New("target root rotation has changed")
+
+// rootRotationReconciler keeps track of all the nodes in the store so that we can determine which ones need reconciliation when nodes are updated
+// or the root CA is updated.  This is meant to be used with watches on nodes and the cluster, and provides functions to be called when the
+// cluster's RootCA has changed and when a node is added, updated, or removed.
+type rootRotationReconciler struct {
+	mu                  sync.Mutex
+	clusterID           string
+	batchUpdateInterval time.Duration
+	ctx                 context.Context
+	store               *store.MemoryStore
+
+	currentRootCA    *api.RootCA
+	currentIssuer    IssuerInfo
+	unconvergedNodes map[string]*api.Node
+
+	wg     sync.WaitGroup
+	cancel func()
+}
+
+// IssuerFromAPIRootCA returns the desired issuer given an API root CA object
+func IssuerFromAPIRootCA(rootCA *api.RootCA) (*IssuerInfo, error) {
+	wantedIssuer := rootCA.CACert
+	if rootCA.RootRotation != nil {
+		wantedIssuer = rootCA.RootRotation.CACert
+	}
+	issuerCerts, err := helpers.ParseCertificatesPEM(wantedIssuer)
+	if err != nil {
+		return nil, errors.Wrap(err, "invalid certificate in cluster root CA object")
+	}
+	if len(issuerCerts) == 0 {
+		return nil, errors.New("invalid certificate in cluster root CA object")
+	}
+	return &IssuerInfo{
+		Subject:   issuerCerts[0].RawSubject,
+		PublicKey: issuerCerts[0].RawSubjectPublicKeyInfo,
+	}, nil
+}
+
+// assumption:  UpdateRootCA will never be called with a `nil` root CA because the caller will be acting in response to
+// a store update event
+func (r *rootRotationReconciler) UpdateRootCA(newRootCA *api.RootCA) {
+	issuerInfo, err := IssuerFromAPIRootCA(newRootCA)
+	if err != nil {
+		log.G(r.ctx).WithError(err).Error("unable to update process the current root CA")
+		return
+	}
+
+	var (
+		shouldStartNewLoop, waitForPrevLoop bool
+		loopCtx                             context.Context
+	)
+	r.mu.Lock()
+	defer func() {
+		r.mu.Unlock()
+		if shouldStartNewLoop {
+			if waitForPrevLoop {
+				r.wg.Wait()
+			}
+			go r.runReconcilerLoop(loopCtx, newRootCA)
+		}
+	}()
+
+	// check if the issuer has changed, first
+	if reflect.DeepEqual(&r.currentIssuer, issuerInfo) {
+		r.currentRootCA = newRootCA
+		return
+	}
+	// If the issuer has changed, iterate through all the nodes to figure out which ones need rotation
+	if newRootCA.RootRotation != nil {
+		var nodes []*api.Node
+		r.store.View(func(tx store.ReadTx) {
+			nodes, err = store.FindNodes(tx, store.ByMembership(api.NodeMembershipAccepted))
+		})
+		if err != nil {
+			log.G(r.ctx).WithError(err).Error("unable to list nodes, so unable to process the current root CA")
+			return
+		}
+
+		// from here on out, there will be no more errors that cause us to have to abandon updating the Root CA,
+		// so we can start making changes to r's fields
+		r.unconvergedNodes = make(map[string]*api.Node)
+		for _, n := range nodes {
+			if !hasIssuer(n, issuerInfo) {
+				r.unconvergedNodes[n.ID] = n
+			}
+		}
+		shouldStartNewLoop = true
+		if r.cancel != nil { // there's already a loop going, so cancel it
+			r.cancel()
+			waitForPrevLoop = true
+		}
+		loopCtx, r.cancel = context.WithCancel(r.ctx)
+	} else {
+		r.unconvergedNodes = nil
+	}
+	r.currentRootCA = newRootCA
+	r.currentIssuer = *issuerInfo
+}
+
+// assumption:  UpdateNode will never be called with a `nil` node because the caller will be acting in response to
+// a store update event
+func (r *rootRotationReconciler) UpdateNode(node *api.Node) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	// if we're not in the middle of a root rotation, or if this node does not have membership, ignore it
+	if r.currentRootCA == nil || r.currentRootCA.RootRotation == nil || node.Spec.Membership != api.NodeMembershipAccepted {
+		return
+	}
+	if hasIssuer(node, &r.currentIssuer) {
+		delete(r.unconvergedNodes, node.ID)
+	} else {
+		r.unconvergedNodes[node.ID] = node
+	}
+}
+
+// assumption:  DeleteNode will never be called with a `nil` node because the caller will be acting in response to
+// a store update event
+func (r *rootRotationReconciler) DeleteNode(node *api.Node) {
+	r.mu.Lock()
+	delete(r.unconvergedNodes, node.ID)
+	r.mu.Unlock()
+}
+
+func (r *rootRotationReconciler) runReconcilerLoop(ctx context.Context, loopRootCA *api.RootCA) {
+	r.wg.Add(1)
+	defer r.wg.Done()
+	for {
+		r.mu.Lock()
+		if len(r.unconvergedNodes) == 0 {
+			r.mu.Unlock()
+
+			err := r.store.Update(func(tx store.Tx) error {
+				return r.finishRootRotation(tx, loopRootCA)
+			})
+			if err == nil {
+				log.G(r.ctx).Info("completed root rotation")
+				return
+			}
+			log.G(r.ctx).WithError(err).Error("could not complete root rotation")
+			if err == errRootRotationChanged {
+				// if the root rotation has changed, this loop will be cancelled anyway, so may as well abort early
+				return
+			}
+		} else {
+			var toUpdate []*api.Node
+			for _, n := range r.unconvergedNodes {
+				iState := n.Certificate.Status.State
+				if iState != api.IssuanceStateRenew && iState != api.IssuanceStatePending && iState != api.IssuanceStateRotate {
+					n = n.Copy()
+					n.Certificate.Status.State = api.IssuanceStateRotate
+					toUpdate = append(toUpdate, n)
+					if len(toUpdate) >= IssuanceStateRotateMaxBatchSize {
+						break
+					}
+				}
+			}
+			r.mu.Unlock()
+
+			if err := r.batchUpdateNodes(toUpdate); err != nil {
+				log.G(r.ctx).WithError(err).Errorf("store error when trying to batch update %d nodes to request certificate rotation", len(toUpdate))
+			}
+		}
+
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(r.batchUpdateInterval):
+		}
+	}
+}
+
+// This function assumes that the expected root CA has root rotation.  This is intended to be used by
+// `reconcileNodeRootsAndCerts`, which uses the root CA from the `lastSeenClusterRootCA`, and checks
+// that it has a root rotation before calling this function.
+func (r *rootRotationReconciler) finishRootRotation(tx store.Tx, expectedRootCA *api.RootCA) error {
+	cluster := store.GetCluster(tx, r.clusterID)
+	if cluster == nil {
+		return fmt.Errorf("unable to get cluster %s", r.clusterID)
+	}
+
+	// If the RootCA object has changed (because another root rotation was started or because some other node
+	// had finished the root rotation), we cannot finish the root rotation that we were working on.
+	if !equality.RootCAEqualStable(expectedRootCA, &cluster.RootCA) {
+		return errRootRotationChanged
+	}
+
+	var signerCert []byte
+	if len(cluster.RootCA.RootRotation.CAKey) > 0 {
+		signerCert = cluster.RootCA.RootRotation.CACert
+	}
+	// we don't actually have to parse out the default node expiration from the cluster - we are just using
+	// the ca.RootCA object to generate new tokens and the digest
+	updatedRootCA, err := NewRootCA(cluster.RootCA.RootRotation.CACert, signerCert, cluster.RootCA.RootRotation.CAKey,
+		DefaultNodeCertExpiration, nil)
+	if err != nil {
+		return errors.Wrap(err, "invalid cluster root rotation object")
+	}
+	cluster.RootCA = api.RootCA{
+		CACert:     cluster.RootCA.RootRotation.CACert,
+		CAKey:      cluster.RootCA.RootRotation.CAKey,
+		CACertHash: updatedRootCA.Digest.String(),
+		JoinTokens: api.JoinTokens{
+			Worker:  GenerateJoinToken(&updatedRootCA),
+			Manager: GenerateJoinToken(&updatedRootCA),
+		},
+		LastForcedRotation: cluster.RootCA.LastForcedRotation,
+	}
+	return store.UpdateCluster(tx, cluster)
+}
+
+func (r *rootRotationReconciler) batchUpdateNodes(toUpdate []*api.Node) error {
+	if len(toUpdate) == 0 {
+		return nil
+	}
+	_, err := r.store.Batch(func(batch *store.Batch) error {
+		// Directly update the nodes rather than get + update, and ignore version errors.  Since
+		// `rootRotationReconciler` should be hooked up to all node update/delete/create events, we should have
+		// close to the latest versions of all the nodes.  If not, the node will updated later and the
+		// next batch of updates should catch it.
+		for _, n := range toUpdate {
+			if err := batch.Update(func(tx store.Tx) error {
+				return store.UpdateNode(tx, n)
+			}); err != nil && err != store.ErrSequenceConflict {
+				log.G(r.ctx).WithError(err).Errorf("unable to update node %s to request a certificate rotation", n.ID)
+			}
+		}
+		return nil
+	})
+	return err
+}