Fix #420

(without re-introducing #394)

420 was a regression introduced by our fix for 394; if we Return()
before we fulfill(), this can result in out of order message delivery,
since incoming calls can be made directly on clients in the result
before all queued calls are drained.

Previously, if we fulfilled() before Return(), we would get a data race
as fillPayloadCapTable modified the message while pipelined calls read
it. Having split Return(error) into PrepareReturn(error) and Return(),
it is now safe to move just the latter after the call to fulfill().

server/server.go

@@ -217,12 +217,12 @@ func (srv *Server) handleCall(ctx context.Context, c *Call) {
 
 	c.recv.ReleaseArgs()
 	c.recv.Returner.PrepareReturn(err)
-	c.recv.Returner.Return()
 	if err == nil {
 		c.aq.fulfill(c.results)
 	} else {
 		c.aq.reject(err)
 	}
+	c.recv.Returner.Return()
 	c.recv.Returner.ReleaseResults()
 }