grpc: support channel idleness

balancer_conn_wrappers.go

@@ -46,36 +46,53 @@ import (
 // It uses the gracefulswitch.Balancer internally to ensure that balancer
 // switches happen in a graceful manner.
 type ccBalancerWrapper struct {
-	cc *ClientConn
+	// The following fields are initialized when the wrapper is created and are
+	// read-only afterwards, and therefore can be accessed without a mutex.
+	cc   *ClientConn
+	opts balancer.BuildOptions
 
 	// Outgoing (gRPC --> balancer) calls are guaranteed to execute in a
-	// mutually exclusive manner as they are scheduled on the
-	// CallbackSerializer. Fields accessed *only* in serializer callbacks, can
-	// therefore be accessed without a mutex.
-	serializer       *grpcsync.CallbackSerializer
-	serializerCancel context.CancelFunc
-	balancer         *gracefulswitch.Balancer
-	curBalancerName  string
+	// mutually exclusive manner as they are scheduled in the serializer. Fields
+	// accessed *only* in these serializer callbacks, can therefore be accessed
+	// without a mutex.
+	balancer        *gracefulswitch.Balancer
+	curBalancerName string
+
+	// During the window of time when the channel is entring idle and the
+	// underlying balancer is being shut down, keeping track of whether the
+	// channel is in idle mode to ensure that calls from the underlying balancer
+	// are not forwarded to grpc.
+
+	// mu guards access to the below fields. The serializer and its cancel
+	// function need to be mutex protected because they are overwritten when the
+	// wrapper exits idle mode.
+	mu               sync.Mutex
+	serializer       *grpcsync.CallbackSerializer // To serialize all outoing calls.
+	serializerCancel context.CancelFunc           // To close the seralizer at close/enterIdle time.
+	isIdleOrClosed   bool
 }
 
 // newCCBalancerWrapper creates a new balancer wrapper. The underlying balancer
 // is not created until the switchTo() method is invoked.
 func newCCBalancerWrapper(cc *ClientConn, bopts balancer.BuildOptions) *ccBalancerWrapper {
-	ctx, cancel := context.WithCancel(context.Background())
 	ccb := &ccBalancerWrapper{
-		cc:               cc,
-		serializer:       grpcsync.NewCallbackSerializer(ctx),
-		serializerCancel: cancel,
+		cc:   cc,
+		opts: bopts,
 	}
+
 	ccb.balancer = gracefulswitch.NewBalancer(ccb, bopts)
+	ctx, cancel := context.WithCancel(context.Background())
+	ccb.serializer = grpcsync.NewCallbackSerializer(ctx)
+	ccb.serializerCancel = cancel
 	return ccb
 }
 
 // updateClientConnState is invoked by grpc to push a ClientConnState update to
 // the underlying balancer.
 func (ccb *ccBalancerWrapper) updateClientConnState(ccs *balancer.ClientConnState) error {
+	ccb.mu.Lock()
 	errCh := make(chan error, 1)
-	ccb.serializer.Schedule(func(_ context.Context) {
+	ok := ccb.serializer.Schedule(func(_ context.Context) {
 		// If the addresses specified in the update contain addresses of type
 		// "grpclb" and the selected LB policy is not "grpclb", these addresses
 		// will be filtered out and ccs will be modified with the updated
@@ -92,16 +109,19 @@ func (ccb *ccBalancerWrapper) updateClientConnState(ccs *balancer.ClientConnStat
 		}
 		errCh <- ccb.balancer.UpdateClientConnState(*ccs)
 	})
-
-	// If the balancer wrapper is closed when waiting for this state update to
-	// be handled, the callback serializer will be closed as well, and we can
-	// rely on its Done channel to ensure that we don't block here forever.
-	select {
-	case err := <-errCh:
-		return err
-	case <-ccb.serializer.Done:
-		return nil
+	if !ok {
+		// If we are unable to schedule a function with the serializer, it
+		// indicates that it has been closed. A serializer is only closed when
+		// the wrapper is closed or is in idle.
+		ccb.mu.Unlock()
+		return fmt.Errorf("grpc: cannot send state update to a closed or idle balancer")
 	}
+	ccb.mu.Unlock()
+
+	// We get here only if the above call to Schedule succeeds, in which case it
+	// is guaranteed that the scheduled function will run. Therefore it is safe
+	// to block on this channel.
+	return <-errCh
 }
 
 // updateSubConnState is invoked by grpc to push a subConn state update to the
@@ -120,21 +140,19 @@ func (ccb *ccBalancerWrapper) updateSubConnState(sc balancer.SubConn, s connecti
 	if sc == nil {
 		return
 	}
+	ccb.mu.Lock()
 	ccb.serializer.Schedule(func(_ context.Context) {
 		ccb.balancer.UpdateSubConnState(sc, balancer.SubConnState{ConnectivityState: s, ConnectionError: err})
 	})
-}
-
-func (ccb *ccBalancerWrapper) exitIdle() {
-	ccb.serializer.Schedule(func(_ context.Context) {
-		ccb.balancer.ExitIdle()
-	})
+	ccb.mu.Unlock()
 }
 
 func (ccb *ccBalancerWrapper) resolverError(err error) {
+	ccb.mu.Lock()
 	ccb.serializer.Schedule(func(_ context.Context) {
 		ccb.balancer.ResolverError(err)
 	})
+	ccb.mu.Unlock()
 }
 
 // switchTo is invoked by grpc to instruct the balancer wrapper to switch to the
@@ -148,41 +166,142 @@ func (ccb *ccBalancerWrapper) resolverError(err error) {
 // the ccBalancerWrapper keeps track of the current LB policy name, and skips
 // the graceful balancer switching process if the name does not change.
 func (ccb *ccBalancerWrapper) switchTo(name string) {
+	ccb.mu.Lock()
 	ccb.serializer.Schedule(func(_ context.Context) {
 		// TODO: Other languages use case-sensitive balancer registries. We should
 		// switch as well. See: https://github.com/grpc/grpc-go/issues/5288.
 		if strings.EqualFold(ccb.curBalancerName, name) {
 			return
 		}
+		ccb.buildLoadBalancingPolicy(name)
+	})
+	ccb.mu.Unlock()
+}
 
-		// Use the default LB policy, pick_first, if no LB policy with name is
-		// found in the registry.
-		builder := balancer.Get(name)
-		if builder == nil {
-			channelz.Warningf(logger, ccb.cc.channelzID, "Channel switches to new LB policy %q, since the specified LB policy %q was not registered", PickFirstBalancerName, name)
-			builder = newPickfirstBuilder()
-		} else {
-			channelz.Infof(logger, ccb.cc.channelzID, "Channel switches to new LB policy %q", name)
-		}
+// buildLoadBalancingPolicy performs the following:
+//   - retrieve a balancer builder for the given name. Use the default LB
+//     policy, pick_first, if no LB policy with name is found in the registry.
+//   - instruct the gracefulswitch balancer to switch to the above builder. This
+//     will actually build the new balancer.
+//   - update the `curBalancerName` field
+//
+// Must be called from a serializer callback.
+func (ccb *ccBalancerWrapper) buildLoadBalancingPolicy(name string) {
+	builder := balancer.Get(name)
+	if builder == nil {
+		channelz.Warningf(logger, ccb.cc.channelzID, "Channel switches to new LB policy %q, since the specified LB policy %q was not registered", PickFirstBalancerName, name)
+		builder = newPickfirstBuilder()
+	} else {
+		channelz.Infof(logger, ccb.cc.channelzID, "Channel switches to new LB policy %q", name)
+	}
+
+	if err := ccb.balancer.SwitchTo(builder); err != nil {
+		channelz.Errorf(logger, ccb.cc.channelzID, "Channel failed to build new LB policy %q: %v", name, err)
+		return
+	}
+	ccb.curBalancerName = builder.Name()
+}
+
+func (ccb *ccBalancerWrapper) close() {
+	done := make(chan struct{})
+	ccb.mu.Lock()
+	if ok := ccb.serializer.Schedule(ccb.handleCloseAndEnterIdle(done)); !ok {
+		// This indicates that the wrapper is already closed or is in idle mode.
+		// Nothing needs to be done here.
+		ccb.mu.Unlock()
+		return
+	}
+	ccb.mu.Unlock()
+	<-done
+}
+
+// enterIdleMode is invoked by grpc when the channel enters idle mode upon
+// expiry of idle_timeout. This call blocks until the balancer is closed.
+func (ccb *ccBalancerWrapper) enterIdleMode() {
+	channelz.Info(logger, ccb.cc.channelzID, "ccBalancerWrapper: entering idle mode")
+	ccb.close()
+}
+
+func (ccb *ccBalancerWrapper) handleCloseAndEnterIdle(done chan struct{}) func(context.Context) {
+	channelz.Info(logger, ccb.cc.channelzID, "easwars: in handleCloseAndEnterIdle")
+	return func(context.Context) {
+		channelz.Info(logger, ccb.cc.channelzID, "easwars: in close func in serializer")
+		ccb.mu.Lock()
+		// Close the serializer to ensure that no more calls from gRPC are sent
+		// to the balancer.
+		ccb.serializerCancel()
+		ccb.isIdleOrClosed = true
+		ccb.mu.Unlock()
+
+		ccb.balancer.Close()
+		close(done)
+	}
+}
+
+// exitIdleMode is invoked by grpc when the channel exits idle mode either
+// because of an RPC or because of an invocation of the Connect() API. This
+// recreates the balancer that was closed previously when entering idle mode.
+//
+// If the channel is not in idle mode, we know for a fact that we are here as a
+// result of the user calling the Connect() method on the ClientConn. In this
+// case, we can simply forward the call to the underlying balancer, instructing
+// it to reconnect to the backends.
+func (ccb *ccBalancerWrapper) exitIdleMode() {
+	channelz.Info(logger, ccb.cc.channelzID, "ccBalancerWrapper: exiting idle mode")
+
+	ccb.mu.Lock()
+	if ccb.isIdleOrClosed {
+		ctx, cancel := context.WithCancel(context.Background())
+		ccb.serializer = grpcsync.NewCallbackSerializer(ctx)
+		ccb.serializerCancel = cancel
+	}
 
-		if err := ccb.balancer.SwitchTo(builder); err != nil {
-			channelz.Errorf(logger, ccb.cc.channelzID, "Channel failed to build new LB policy %q: %v", name, err)
+	done := make(chan struct{})
+	ccb.serializer.Schedule(func(_ context.Context) {
+		defer close(done)
+
+		ccb.mu.Lock()
+		defer ccb.mu.Unlock()
+
+		if !ccb.isIdleOrClosed {
+			ccb.balancer.ExitIdle()
 			return
 		}
-		ccb.curBalancerName = builder.Name()
+
+		// Gracefulswitch balancer does not support a switchTo operation after
+		// being closed. Hence we need to create a new one here.
+		ccb.balancer = gracefulswitch.NewBalancer(ccb, ccb.opts)
+		ccb.buildLoadBalancingPolicy(ccb.curBalancerName)
+		ccb.isIdleOrClosed = false
 	})
+	ccb.mu.Unlock()
+
+	<-done
 }
 
-func (ccb *ccBalancerWrapper) close() {
-	// Close the serializer to ensure that no more calls from gRPC are sent to
-	// the balancer. We don't have to worry about suppressing calls from a
-	// closed balancer because these are handled by the ClientConn (balancer
-	// wrapper is only ever closed when the ClientConn is closed).
-	ccb.serializerCancel()
-	<-ccb.serializer.Done
-	ccb.balancer.Close()
+func (ccb *ccBalancerWrapper) isIdleOrClosedLocked() bool {
+	ccb.mu.Lock()
+	defer ccb.mu.Unlock()
+	return ccb.isIdleOrClosed
 }
 
+// We cannot schedule incoming calls on the serailzer becasue outoing call updateClientConnState is blocked on the seralizer. When this call is forwarded
+// to the actual balancer, it might call NewSubConn inline which also needs to block.
+// But holding the lock for the duration of the call is fine because:
+// a: the lock is held on the outoing side only to queue callbacks in the serializer
+// b: holding the lock for the whole duration is necessary here to guarantee
+// that balancer close/enterIdle does not happen after we check for
+// `isInactive`, but before we do that actual operation required for the
+// incoming call.
+//
+// holding the lock for the whole duration does not work:
+// - when a balancer switch happens, the old balancer is closed, wherein it might remove subConns
+// - removeSubConn is handled with the lock and as part of it, and UpdateSubConnState is pushed by grpc
+// - updateSubConnState tries to grab the lock to add a callback to the serializer, but its deadlocks
+//
+// graceful switch balancer drops calls from the balancer once close is called.
+// So, we might not have to check for `isInactiveLocked()` here.
+// No, we need it since nothing stops them from calling any methods on the balancer.ClientConn interface inline from their Close(), like removeAddrConn etc
 func (ccb *ccBalancerWrapper) NewSubConn(addrs []resolver.Address, opts balancer.NewSubConnOptions) (balancer.SubConn, error) {
 	if len(addrs) <= 0 {
 		return nil, fmt.Errorf("grpc: cannot create SubConn with empty address list")
@@ -200,6 +319,12 @@ func (ccb *ccBalancerWrapper) NewSubConn(addrs []resolver.Address, opts balancer
 }
 
 func (ccb *ccBalancerWrapper) RemoveSubConn(sc balancer.SubConn) {
+	/*
+		if ccb.isIdleOrClosedLocked() {
+			return
+		}
+	*/
+
 	acbw, ok := sc.(*acBalancerWrapper)
 	if !ok {
 		return
@@ -208,6 +333,10 @@ func (ccb *ccBalancerWrapper) RemoveSubConn(sc balancer.SubConn) {
 }
 
 func (ccb *ccBalancerWrapper) UpdateAddresses(sc balancer.SubConn, addrs []resolver.Address) {
+	if ccb.isIdleOrClosedLocked() {
+		return
+	}
+
 	acbw, ok := sc.(*acBalancerWrapper)
 	if !ok {
 		return
@@ -216,6 +345,10 @@ func (ccb *ccBalancerWrapper) UpdateAddresses(sc balancer.SubConn, addrs []resol
 }
 
 func (ccb *ccBalancerWrapper) UpdateState(s balancer.State) {
+	if ccb.isIdleOrClosedLocked() {
+		return
+	}
+
 	// Update picker before updating state.  Even though the ordering here does
 	// not matter, it can lead to multiple calls of Pick in the common start-up
 	// case where we wait for ready and then perform an RPC.  If the picker is
@@ -226,6 +359,10 @@ func (ccb *ccBalancerWrapper) UpdateState(s balancer.State) {
 }
 
 func (ccb *ccBalancerWrapper) ResolveNow(o resolver.ResolveNowOptions) {
+	if ccb.isIdleOrClosedLocked() {
+		return
+	}
+
 	ccb.cc.resolveNow(o)
 }
 

call.go

@@ -27,6 +27,11 @@ import (
 //
 // All errors returned by Invoke are compatible with the status package.
 func (cc *ClientConn) Invoke(ctx context.Context, method string, args, reply interface{}, opts ...CallOption) error {
+	if err := cc.idlenessMgr.onCallBegin(); err != nil {
+		return err
+	}
+	defer cc.idlenessMgr.onCallEnd()
+
 	// allow interceptor to see all applicable call options, which means those
 	// configured as defaults from dial option as well as per-call options
 	opts = combine(cc.dopts.callOptions, opts)