core: make subchannel creation timing restriction stricter (#7790)

Throw for subchannel creation if the channel is being shutting down and the delayed transport is terminated (aka, all retry calls has been finished). This enforces load balancer implementations to avoid creating subchannels after being shut down.

core/src/main/java/io/grpc/internal/ManagedChannelImpl.java

@@ -1391,12 +1391,8 @@ private class LbHelperImpl extends LoadBalancer.Helper {
     @Override
     public AbstractSubchannel createSubchannel(CreateSubchannelArgs args) {
       syncContext.throwIfNotInThisSynchronizationContext();
-      return createSubchannelInternal(args);
-    }
-
-    private SubchannelImpl createSubchannelInternal(CreateSubchannelArgs args) {
-      // TODO(ejona): can we be even stricter? Like loadBalancer == null?
-      checkState(!terminated, "Channel is terminated");
+      // No new subchannel should be created after load balancer has been shutdown.
+      checkState(!terminating, "Channel is being terminated");
       return new SubchannelImpl(args, this);
     }
 
@@ -1823,18 +1819,8 @@ private final class SubchannelImpl extends AbstractSubchannel {
     private void internalStart(final SubchannelStateListener listener) {
       checkState(!started, "already started");
       checkState(!shutdown, "already shutdown");
+      checkState(!terminating, "Channel is being terminated");
       started = true;
-      // TODO(zhangkun): possibly remove the volatile of terminating when this whole method is
-      // required to be called from syncContext
-      if (terminating) {
-        syncContext.execute(new Runnable() {
-            @Override
-            public void run() {
-              listener.onSubchannelState(ConnectivityStateInfo.forNonError(SHUTDOWN));
-            }
-          });
-        return;
-      }
       final class ManagedInternalSubchannelCallback extends InternalSubchannel.Callback {
         // All callbacks are run in syncContext
         @Override