Merge e759d2a^

32 files changed, 812 insertions, 861 deletions

diff --git a/include/grpc++/alarm.h b/include/grpc++/alarm.h
index 2d88d868e5..b43425e224 100644
--- a/include/grpc++/alarm.h
+++ b/include/grpc++/alarm.h
@@ -92,7 +92,7 @@ class Alarm : private GrpcLibraryCodegen {
   }
 
  private:
-  class AlarmEntry : public CompletionQueueTag {
+  class AlarmEntry : public internal::CompletionQueueTag {
    public:
     AlarmEntry(void* tag) : tag_(tag) {}
     void Set(void* tag) { tag_ = tag; }
diff --git a/include/grpc++/channel.h b/include/grpc++/channel.h
index c50091d6ac..e9fb5a5d09 100644
--- a/include/grpc++/channel.h
+++ b/include/grpc++/channel.h
@@ -32,7 +32,7 @@ struct grpc_channel;
 namespace grpc {
 /// Channels represent a connection to an endpoint. Created by \a CreateChannel.
 class Channel final : public ChannelInterface,
-                      public CallHook,
+                      public internal::CallHook,
                       public std::enable_shared_from_this<Channel>,
                       private GrpcLibraryCodegen {
  public:
@@ -51,18 +51,16 @@ class Channel final : public ChannelInterface,
 
  private:
   template <class InputMessage, class OutputMessage>
-  friend Status BlockingUnaryCall(ChannelInterface* channel,
-                                  const RpcMethod& method,
-                                  ClientContext* context,
-                                  const InputMessage& request,
-                                  OutputMessage* result);
+  friend class internal::BlockingUnaryCallImpl;
   friend std::shared_ptr<Channel> CreateChannelInternal(
       const grpc::string& host, grpc_channel* c_channel);
   Channel(const grpc::string& host, grpc_channel* c_channel);
 
-  Call CreateCall(const RpcMethod& method, ClientContext* context,
-                  CompletionQueue* cq) override;
-  void PerformOpsOnCall(CallOpSetInterface* ops, Call* call) override;
+  internal::Call CreateCall(const internal::RpcMethod& method,
+                            ClientContext* context,
+                            CompletionQueue* cq) override;
+  void PerformOpsOnCall(internal::CallOpSetInterface* ops,
+                        internal::Call* call) override;
   void* RegisterMethod(const char* method) override;
 
   void NotifyOnStateChangeImpl(grpc_connectivity_state last_observed,
diff --git a/include/grpc++/impl/codegen/async_stream.h b/include/grpc++/impl/codegen/async_stream.h
index e60572fc93..4476033463 100644
--- a/include/grpc++/impl/codegen/async_stream.h
+++ b/include/grpc++/impl/codegen/async_stream.h
@@ -30,6 +30,7 @@ namespace grpc {
 
 class CompletionQueue;
 
+namespace internal {
 /// Common interface for all client side asynchronous streaming.
 class ClientAsyncStreamingInterface {
  public:
@@ -151,15 +152,16 @@ class AsyncWriterInterface {
   }
 };
 
+}  // namespace internal
+
 template <class R>
-class ClientAsyncReaderInterface : public ClientAsyncStreamingInterface,
-                                   public AsyncReaderInterface<R> {};
+class ClientAsyncReaderInterface
+    : public internal::ClientAsyncStreamingInterface,
+      public internal::AsyncReaderInterface<R> {};
 
-/// Async client-side API for doing server-streaming RPCs,
-/// where the incoming message stream coming from the server has
-/// messages of type \a R.
+namespace internal {
 template <class R>
-class ClientAsyncReader final : public ClientAsyncReaderInterface<R> {
+class ClientAsyncReaderFactory {
  public:
   /// Create a stream object.
   /// Write the first request out if \a start is set.
@@ -169,16 +171,25 @@ class ClientAsyncReader final : public ClientAsyncReaderInterface<R> {
   /// Note that \a context will be used to fill in custom initial metadata
   /// used to send to the server when starting the call.
   template <class W>
-  static ClientAsyncReader* Create(ChannelInterface* channel,
-                                   CompletionQueue* cq, const RpcMethod& method,
-                                   ClientContext* context, const W& request,
-                                   bool start, void* tag) {
-    Call call = channel->CreateCall(method, context, cq);
+  static ClientAsyncReader<R>* Create(ChannelInterface* channel,
+                                      CompletionQueue* cq,
+                                      const ::grpc::internal::RpcMethod& method,
+                                      ClientContext* context, const W& request,
+                                      bool start, void* tag) {
+    ::grpc::internal::Call call = channel->CreateCall(method, context, cq);
     return new (g_core_codegen_interface->grpc_call_arena_alloc(
-        call.call(), sizeof(ClientAsyncReader)))
-        ClientAsyncReader(call, context, request, start, tag);
+        call.call(), sizeof(ClientAsyncReader<R>)))
+        ClientAsyncReader<R>(call, context, request, start, tag);
   }
+};
+}  // namespace internal
 
+/// Async client-side API for doing server-streaming RPCs,
+/// where the incoming message stream coming from the server has
+/// messages of type \a R.
+template <class R>
+class ClientAsyncReader final : public ClientAsyncReaderInterface<R> {
+ public:
   // always allocated against a call arena, no memory free required
   static void operator delete(void* ptr, std::size_t size) {
     assert(size == sizeof(ClientAsyncReader));
@@ -233,9 +244,10 @@ class ClientAsyncReader final : public ClientAsyncReaderInterface<R> {
   }
 
  private:
+  friend class internal::ClientAsyncReaderFactory<R>;
   template <class W>
-  ClientAsyncReader(Call call, ClientContext* context, const W& request,
-                    bool start, void* tag)
+  ClientAsyncReader(::grpc::internal::Call call, ClientContext* context,
+                    const W& request, bool start, void* tag)
       : context_(context), call_(call), started_(start) {
     // TODO(ctiller): don't assert
     GPR_CODEGEN_ASSERT(init_ops_.SendMessage(request).ok());
@@ -255,19 +267,27 @@ class ClientAsyncReader final : public ClientAsyncReaderInterface<R> {
   }
 
   ClientContext* context_;
-  Call call_;
+  ::grpc::internal::Call call_;
   bool started_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage, CallOpClientSendClose>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpClientSendClose>
       init_ops_;
-  CallOpSet<CallOpRecvInitialMetadata> meta_ops_;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpRecvMessage<R>> read_ops_;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpClientRecvStatus> finish_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+      meta_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpRecvMessage<R>>
+      read_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpClientRecvStatus>
+      finish_ops_;
 };
 
 /// Common interface for client side asynchronous writing.
 template <class W>
-class ClientAsyncWriterInterface : public ClientAsyncStreamingInterface,
-                                   public AsyncWriterInterface<W> {
+class ClientAsyncWriterInterface
+    : public internal::ClientAsyncStreamingInterface,
+      public internal::AsyncWriterInterface<W> {
  public:
   /// Signal the client is done with the writes (half-close the client stream).
   /// Thread-safe with respect to \a AsyncReaderInterface::Read
@@ -276,11 +296,9 @@ class ClientAsyncWriterInterface : public ClientAsyncStreamingInterface,
   virtual void WritesDone(void* tag) = 0;
 };
 
-/// Async API on the client side for doing client-streaming RPCs,
-/// where the outgoing message stream going to the server contains
-/// messages of type \a W.
+namespace internal {
 template <class W>
-class ClientAsyncWriter final : public ClientAsyncWriterInterface<W> {
+class ClientAsyncWriterFactory {
  public:
   /// Create a stream object.
   /// Start the RPC if \a start is set
@@ -294,16 +312,25 @@ class ClientAsyncWriter final : public ClientAsyncWriterInterface<W> {
   /// message from the server upon a successful call to the \a Finish
   /// method of this instance.
   template <class R>
-  static ClientAsyncWriter* Create(ChannelInterface* channel,
-                                   CompletionQueue* cq, const RpcMethod& method,
-                                   ClientContext* context, R* response,
-                                   bool start, void* tag) {
-    Call call = channel->CreateCall(method, context, cq);
+  static ClientAsyncWriter<W>* Create(ChannelInterface* channel,
+                                      CompletionQueue* cq,
+                                      const ::grpc::internal::RpcMethod& method,
+                                      ClientContext* context, R* response,
+                                      bool start, void* tag) {
+    ::grpc::internal::Call call = channel->CreateCall(method, context, cq);
     return new (g_core_codegen_interface->grpc_call_arena_alloc(
-        call.call(), sizeof(ClientAsyncWriter)))
-        ClientAsyncWriter(call, context, response, start, tag);
+        call.call(), sizeof(ClientAsyncWriter<W>)))
+        ClientAsyncWriter<W>(call, context, response, start, tag);
   }
+};
+}  // namespace internal
 
+/// Async API on the client side for doing client-streaming RPCs,
+/// where the outgoing message stream going to the server contains
+/// messages of type \a W.
+template <class W>
+class ClientAsyncWriter final : public ClientAsyncWriterInterface<W> {
+ public:
   // always allocated against a call arena, no memory free required
   static void operator delete(void* ptr, std::size_t size) {
     assert(size == sizeof(ClientAsyncWriter));
@@ -376,9 +403,10 @@ class ClientAsyncWriter final : public ClientAsyncWriterInterface<W> {
   }
 
  private:
+  friend class internal::ClientAsyncWriterFactory<W>;
   template <class R>
-  ClientAsyncWriter(Call call, ClientContext* context, R* response, bool start,
-                    void* tag)
+  ClientAsyncWriter(::grpc::internal::Call call, ClientContext* context,
+                    R* response, bool start, void* tag)
       : context_(context), call_(call), started_(start) {
     finish_ops_.RecvMessage(response);
     finish_ops_.AllowNoMessage();
@@ -401,13 +429,17 @@ class ClientAsyncWriter final : public ClientAsyncWriterInterface<W> {
   }
 
   ClientContext* context_;
-  Call call_;
+  ::grpc::internal::Call call_;
   bool started_;
-  CallOpSet<CallOpRecvInitialMetadata> meta_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage, CallOpClientSendClose>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+      meta_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpClientSendClose>
       write_ops_;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpGenericRecvMessage,
-            CallOpClientRecvStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpGenericRecvMessage,
+                              ::grpc::internal::CallOpClientRecvStatus>
       finish_ops_;
 };
 
@@ -415,9 +447,10 @@ class ClientAsyncWriter final : public ClientAsyncWriterInterface<W> {
 /// where the client-to-server message stream has messages of type \a W,
 /// and the server-to-client message stream has messages of type \a R.
 template <class W, class R>
-class ClientAsyncReaderWriterInterface : public ClientAsyncStreamingInterface,
-                                         public AsyncWriterInterface<W>,
-                                         public AsyncReaderInterface<R> {
+class ClientAsyncReaderWriterInterface
+    : public internal::ClientAsyncStreamingInterface,
+      public internal::AsyncWriterInterface<W>,
+      public internal::AsyncReaderInterface<R> {
  public:
   /// Signal the client is done with the writes (half-close the client stream).
   /// Thread-safe with respect to \a AsyncReaderInterface::Read
@@ -426,13 +459,9 @@ class ClientAsyncReaderWriterInterface : public ClientAsyncStreamingInterface,
   virtual void WritesDone(void* tag) = 0;
 };
 
-/// Async client-side interface for bi-directional streaming,
-/// where the outgoing message stream going to the server
-/// has messages of type \a W,  and the incoming message stream coming
-/// from the server has messages of type \a R.
+namespace internal {
 template <class W, class R>
-class ClientAsyncReaderWriter final
-    : public ClientAsyncReaderWriterInterface<W, R> {
+class ClientAsyncReaderWriterFactory {
  public:
   /// Create a stream object.
   /// Start the RPC request if \a start is set.
@@ -441,18 +470,27 @@ class ClientAsyncReaderWriter final
   /// nullptr and the actual call must be initiated by StartCall
   /// Note that \a context will be used to fill in custom initial metadata
   /// used to send to the server when starting the call.
-  static ClientAsyncReaderWriter* Create(ChannelInterface* channel,
-                                         CompletionQueue* cq,
-                                         const RpcMethod& method,
-                                         ClientContext* context, bool start,
-                                         void* tag) {
-    Call call = channel->CreateCall(method, context, cq);
+  static ClientAsyncReaderWriter<W, R>* Create(
+      ChannelInterface* channel, CompletionQueue* cq,
+      const ::grpc::internal::RpcMethod& method, ClientContext* context,
+      bool start, void* tag) {
+    ::grpc::internal::Call call = channel->CreateCall(method, context, cq);
 
     return new (g_core_codegen_interface->grpc_call_arena_alloc(
-        call.call(), sizeof(ClientAsyncReaderWriter)))
-        ClientAsyncReaderWriter(call, context, start, tag);
+        call.call(), sizeof(ClientAsyncReaderWriter<W, R>)))
+        ClientAsyncReaderWriter<W, R>(call, context, start, tag);
   }
+};
+}  // namespace internal
 
+/// Async client-side interface for bi-directional streaming,
+/// where the outgoing message stream going to the server
+/// has messages of type \a W,  and the incoming message stream coming
+/// from the server has messages of type \a R.
+template <class W, class R>
+class ClientAsyncReaderWriter final
+    : public ClientAsyncReaderWriterInterface<W, R> {
+ public:
   // always allocated against a call arena, no memory free required
   static void operator delete(void* ptr, std::size_t size) {
     assert(size == sizeof(ClientAsyncReaderWriter));
@@ -532,8 +570,9 @@ class ClientAsyncReaderWriter final
   }
 
  private:
-  ClientAsyncReaderWriter(Call call, ClientContext* context, bool start,
-                          void* tag)
+  friend class internal::ClientAsyncReaderWriterFactory<W, R>;
+  ClientAsyncReaderWriter(::grpc::internal::Call call, ClientContext* context,
+                          bool start, void* tag)
       : context_(context), call_(call), started_(start) {
     if (start) {
       StartCallInternal(tag);
@@ -554,18 +593,26 @@ class ClientAsyncReaderWriter final
   }
 
   ClientContext* context_;
-  Call call_;
+  ::grpc::internal::Call call_;
   bool started_;
-  CallOpSet<CallOpRecvInitialMetadata> meta_ops_;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpRecvMessage<R>> read_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage, CallOpClientSendClose>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+      meta_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpRecvMessage<R>>
+      read_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpClientSendClose>
       write_ops_;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpClientRecvStatus> finish_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpClientRecvStatus>
+      finish_ops_;
 };
 
 template <class W, class R>
-class ServerAsyncReaderInterface : public ServerAsyncStreamingInterface,
-                                   public AsyncReaderInterface<R> {
+class ServerAsyncReaderInterface
+    : public internal::ServerAsyncStreamingInterface,
+      public internal::AsyncReaderInterface<R> {
  public:
   /// Indicate that the stream is to be finished with a certain status code
   /// and also send out \a msg response to the client.
@@ -692,20 +739,23 @@ class ServerAsyncReader final : public ServerAsyncReaderInterface<W, R> {
   }
 
  private:
-  void BindCall(Call* call) override { call_ = *call; }
+  void BindCall(::grpc::internal::Call* call) override { call_ = *call; }
 
-  Call call_;
+  ::grpc::internal::Call call_;
   ServerContext* ctx_;
-  CallOpSet<CallOpSendInitialMetadata> meta_ops_;
-  CallOpSet<CallOpRecvMessage<R>> read_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-            CallOpServerSendStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata>
+      meta_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvMessage<R>> read_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpServerSendStatus>
       finish_ops_;
 };
 
 template <class W>
-class ServerAsyncWriterInterface : public ServerAsyncStreamingInterface,
-                                   public AsyncWriterInterface<W> {
+class ServerAsyncWriterInterface
+    : public internal::ServerAsyncStreamingInterface,
+      public internal::AsyncWriterInterface<W> {
  public:
   /// Indicate that the stream is to be finished with a certain status code.
   /// Request notification for when the server has sent the appropriate
@@ -823,7 +873,7 @@ class ServerAsyncWriter final : public ServerAsyncWriterInterface<W> {
   }
 
  private:
-  void BindCall(Call* call) override { call_ = *call; }
+  void BindCall(::grpc::internal::Call* call) override { call_ = *call; }
 
   template <class T>
   void EnsureInitialMetadataSent(T* ops) {
@@ -837,20 +887,25 @@ class ServerAsyncWriter final : public ServerAsyncWriterInterface<W> {
     }
   }
 
-  Call call_;
+  ::grpc::internal::Call call_;
   ServerContext* ctx_;
-  CallOpSet<CallOpSendInitialMetadata> meta_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-            CallOpServerSendStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata>
+      meta_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpServerSendStatus>
       write_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpServerSendStatus> finish_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpServerSendStatus>
+      finish_ops_;
 };
 
 /// Server-side interface for asynchronous bi-directional streaming.
 template <class W, class R>
-class ServerAsyncReaderWriterInterface : public ServerAsyncStreamingInterface,
-                                         public AsyncWriterInterface<W>,
-                                         public AsyncReaderInterface<R> {
+class ServerAsyncReaderWriterInterface
+    : public internal::ServerAsyncStreamingInterface,
+      public internal::AsyncWriterInterface<W>,
+      public internal::AsyncReaderInterface<R> {
  public:
   /// Indicate that the stream is to be finished with a certain status code.
   /// Request notification for when the server has sent the appropriate
@@ -980,7 +1035,7 @@ class ServerAsyncReaderWriter final
  private:
   friend class ::grpc::Server;
 
-  void BindCall(Call* call) override { call_ = *call; }
+  void BindCall(::grpc::internal::Call* call) override { call_ = *call; }
 
   template <class T>
   void EnsureInitialMetadataSent(T* ops) {
@@ -994,14 +1049,18 @@ class ServerAsyncReaderWriter final
     }
   }
 
-  Call call_;
+  ::grpc::internal::Call call_;
   ServerContext* ctx_;
-  CallOpSet<CallOpSendInitialMetadata> meta_ops_;
-  CallOpSet<CallOpRecvMessage<R>> read_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-            CallOpServerSendStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata>
+      meta_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvMessage<R>> read_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpServerSendStatus>
       write_ops_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpServerSendStatus> finish_ops_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpServerSendStatus>
+      finish_ops_;
 };
 
 }  // namespace grpc
diff --git a/include/grpc++/impl/codegen/async_unary_call.h b/include/grpc++/impl/codegen/async_unary_call.h
index e472f04f56..6d51c78d5c 100644
--- a/include/grpc++/impl/codegen/async_unary_call.h
+++ b/include/grpc++/impl/codegen/async_unary_call.h
@@ -69,11 +69,9 @@ class ClientAsyncResponseReaderInterface {
   virtual void Finish(R* msg, Status* status, void* tag) = 0;
 };
 
-/// Async API for client-side unary RPCs, where the message response
-/// received from the server is of type \a R.
+namespace internal {
 template <class R>
-class ClientAsyncResponseReader final
-    : public ClientAsyncResponseReaderInterface<R> {
+class ClientAsyncResponseReaderFactory {
  public:
   /// Start a call and write the request out if \a start is set.
   /// \a tag will be notified on \a cq when the call has been started (i.e.
@@ -82,17 +80,24 @@ class ClientAsyncResponseReader final
   /// Note that \a context will be used to fill in custom initial metadata
   /// used to send to the server when starting the call.
   template <class W>
-  static ClientAsyncResponseReader* Create(ChannelInterface* channel,
-                                           CompletionQueue* cq,
-                                           const RpcMethod& method,
-                                           ClientContext* context,
-                                           const W& request, bool start) {
-    Call call = channel->CreateCall(method, context, cq);
+  static ClientAsyncResponseReader<R>* Create(
+      ChannelInterface* channel, CompletionQueue* cq,
+      const ::grpc::internal::RpcMethod& method, ClientContext* context,
+      const W& request, bool start) {
+    ::grpc::internal::Call call = channel->CreateCall(method, context, cq);
     return new (g_core_codegen_interface->grpc_call_arena_alloc(
-        call.call(), sizeof(ClientAsyncResponseReader)))
-        ClientAsyncResponseReader(call, context, request, start);
+        call.call(), sizeof(ClientAsyncResponseReader<R>)))
+        ClientAsyncResponseReader<R>(call, context, request, start);
   }
+};
+}  // namespace internal
 
+/// Async API for client-side unary RPCs, where the message response
+/// received from the server is of type \a R.
+template <class R>
+class ClientAsyncResponseReader final
+    : public ClientAsyncResponseReaderInterface<R> {
+ public:
   // always allocated against a call arena, no memory free required
   static void operator delete(void* ptr, std::size_t size) {
     assert(size == sizeof(ClientAsyncResponseReader));
@@ -137,13 +142,14 @@ class ClientAsyncResponseReader final
   }
 
  private:
+  friend class internal::ClientAsyncResponseReaderFactory<R>;
   ClientContext* const context_;
-  Call call_;
+  ::grpc::internal::Call call_;
   bool started_;
 
   template <class W>
-  ClientAsyncResponseReader(Call call, ClientContext* context, const W& request,
-                            bool start)
+  ClientAsyncResponseReader(::grpc::internal::Call call, ClientContext* context,
+                            const W& request, bool start)
       : context_(context), call_(call), started_(start) {
     // Bind the metadata at time of StartCallInternal but set up the rest here
     // TODO(ctiller): don't assert
@@ -162,19 +168,23 @@ class ClientAsyncResponseReader final
   static void* operator new(std::size_t size);
   static void* operator new(std::size_t size, void* p) { return p; }
 
-  SneakyCallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-                  CallOpClientSendClose>
+  ::grpc::internal::SneakyCallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                                    ::grpc::internal::CallOpSendMessage,
+                                    ::grpc::internal::CallOpClientSendClose>
       init_buf;
-  CallOpSet<CallOpRecvInitialMetadata> meta_buf;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpRecvMessage<R>,
-            CallOpClientRecvStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+      meta_buf;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpRecvMessage<R>,
+                              ::grpc::internal::CallOpClientRecvStatus>
       finish_buf;
 };
 
 /// Async server-side API for handling unary calls, where the single
 /// response message sent to the client is of type \a W.
 template <class W>
-class ServerAsyncResponseWriter final : public ServerAsyncStreamingInterface {
+class ServerAsyncResponseWriter final
+    : public internal::ServerAsyncStreamingInterface {
  public:
   explicit ServerAsyncResponseWriter(ServerContext* ctx)
       : call_(nullptr, nullptr, nullptr), ctx_(ctx) {}
@@ -262,13 +272,15 @@ class ServerAsyncResponseWriter final : public ServerAsyncStreamingInterface {
   }
 
  private:
-  void BindCall(Call* call) override { call_ = *call; }
+  void BindCall(::grpc::internal::Call* call) override { call_ = *call; }
 
-  Call call_;
+  ::grpc::internal::Call call_;
   ServerContext* ctx_;
-  CallOpSet<CallOpSendInitialMetadata> meta_buf_;
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-            CallOpServerSendStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata>
+      meta_buf_;
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                              ::grpc::internal::CallOpSendMessage,
+                              ::grpc::internal::CallOpServerSendStatus>
       finish_buf_;
 };
 
diff --git a/include/grpc++/impl/codegen/byte_buffer.h b/include/grpc++/impl/codegen/byte_buffer.h
index 57d731be18..fe73ce7a83 100644
--- a/include/grpc++/impl/codegen/byte_buffer.h
+++ b/include/grpc++/impl/codegen/byte_buffer.h
@@ -31,18 +31,19 @@
 
 namespace grpc {
 
+namespace internal {
+class CallOpSendMessage;
 template <class R>
 class CallOpRecvMessage;
+class CallOpGenericRecvMessage;
 class MethodHandler;
 template <class ServiceType, class RequestType, class ResponseType>
 class RpcMethodHandler;
 template <class ServiceType, class RequestType, class ResponseType>
 class ServerStreamingHandler;
-namespace CallOpGenericRecvMessageHelper {
 template <class R>
 class DeserializeFuncType;
-}  // namespace CallOpGenericRecvMessageHelper
-
+}  // namespace internal
 /// A sequence of bytes.
 class ByteBuffer final {
  public:
@@ -97,17 +98,17 @@ class ByteBuffer final {
 
  private:
   friend class SerializationTraits<ByteBuffer, void>;
-  friend class CallOpSendMessage;
+  friend class internal::CallOpSendMessage;
   template <class R>
-  friend class CallOpRecvMessage;
-  friend class CallOpGenericRecvMessage;
-  friend class MethodHandler;
+  friend class internal::CallOpRecvMessage;
+  friend class internal::CallOpGenericRecvMessage;
+  friend class internal::MethodHandler;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class RpcMethodHandler;
+  friend class internal::RpcMethodHandler;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class ServerStreamingHandler;
+  friend class internal::ServerStreamingHandler;
   template <class R>
-  friend class CallOpGenericRecvMessageHelper::DeserializeFuncType;
+  friend class internal::DeserializeFuncType;
 
   grpc_byte_buffer* buffer_;
 
diff --git a/include/grpc++/impl/codegen/call.h b/include/grpc++/impl/codegen/call.h
index 0e7ad0b60a..63caf90d39 100644
--- a/include/grpc++/impl/codegen/call.h
+++ b/include/grpc++/impl/codegen/call.h
@@ -43,11 +43,13 @@
 namespace grpc {
 
 class ByteBuffer;
-class Call;
-class CallHook;
 class CompletionQueue;
 extern CoreCodegenInterface* g_core_codegen_interface;
 
+namespace internal {
+class Call;
+class CallHook;
+
 const char kBinaryErrorDetailsKey[] = "grpc-status-details-bin";
 
 // TODO(yangg) if the map is changed before we send, the pointers will be a
@@ -75,6 +77,7 @@ inline grpc_metadata* FillMetadataArray(
   }
   return metadata_array;
 }
+}  // namespace internal
 
 /// Per-message write options.
 class WriteOptions {
@@ -163,7 +166,7 @@ class WriteOptions {
 
   /// Clears flag indicating that this is the last message in a stream,
   /// disabling coalescing.
-  inline WriteOptions& clear_last_messsage() {
+  inline WriteOptions& clear_last_message() {
     last_message_ = false;
     return *this;
   }
@@ -199,6 +202,7 @@ class WriteOptions {
   bool last_message_;
 };
 
+namespace internal {
 /// Default argument for CallOpSet. I is unused by the class, but can be
 /// used for generating multiple names for the same thing.
 template <int I>
@@ -301,7 +305,11 @@ template <class M>
 Status CallOpSendMessage::SendMessage(const M& message, WriteOptions options) {
   write_options_ = options;
   bool own_buf;
-  Status result = SerializationTraits<M>::Serialize(
+  // TODO(vjpai): Remove the void below when possible
+  // The void in the template parameter below should not be needed
+  // (since it should be implicit) but is needed due to an observed
+  // difference in behavior between clang and gcc for certain internal users
+  Status result = SerializationTraits<M, void>::Serialize(
       message, send_buf_.bbuf_ptr(), &own_buf);
   if (!own_buf) {
     send_buf_.Duplicate();
@@ -366,7 +374,6 @@ class CallOpRecvMessage {
   bool allow_not_getting_message_;
 };
 
-namespace CallOpGenericRecvMessageHelper {
 class DeserializeFunc {
  public:
   virtual Status Deserialize(ByteBuffer* buf) = 0;
@@ -386,7 +393,6 @@ class DeserializeFuncType final : public DeserializeFunc {
  private:
   R* message_;  // Not a managed pointer because management is external to this
 };
-}  // namespace CallOpGenericRecvMessageHelper
 
 class CallOpGenericRecvMessage {
  public:
@@ -397,8 +403,7 @@ class CallOpGenericRecvMessage {
   void RecvMessage(R* message) {
     // Use an explicit base class pointer to avoid resolution error in the
     // following unique_ptr::reset for some old implementations.
-    CallOpGenericRecvMessageHelper::DeserializeFunc* func =
-        new CallOpGenericRecvMessageHelper::DeserializeFuncType<R>(message);
+    DeserializeFunc* func = new DeserializeFuncType<R>(message);
     deserialize_.reset(func);
   }
 
@@ -438,7 +443,7 @@ class CallOpGenericRecvMessage {
   }
 
  private:
-  std::unique_ptr<CallOpGenericRecvMessageHelper::DeserializeFunc> deserialize_;
+  std::unique_ptr<DeserializeFunc> deserialize_;
   ByteBuffer recv_buf_;
   bool allow_not_getting_message_;
 };
@@ -693,7 +698,7 @@ class Call final {
   grpc_call* call_;
   int max_receive_message_size_;
 };
-
+}  // namespace internal
 }  // namespace grpc
 
 #endif  // GRPCXX_IMPL_CODEGEN_CALL_H
diff --git a/include/grpc++/impl/codegen/call_hook.h b/include/grpc++/impl/codegen/call_hook.h
index d026cc8b58..44e9de220e 100644
--- a/include/grpc++/impl/codegen/call_hook.h
+++ b/include/grpc++/impl/codegen/call_hook.h
@@ -21,6 +21,7 @@
 
 namespace grpc {
 
+namespace internal {
 class CallOpSetInterface;
 class Call;
 
@@ -31,6 +32,7 @@ class CallHook {
   virtual ~CallHook() {}
   virtual void PerformOpsOnCall(CallOpSetInterface* ops, Call* call) = 0;
 };
+}  // namespace internal
 
 }  // namespace grpc
 
diff --git a/include/grpc++/impl/codegen/channel_interface.h b/include/grpc++/impl/codegen/channel_interface.h
index 1b7590bf0c..769f853974 100644
--- a/include/grpc++/impl/codegen/channel_interface.h
+++ b/include/grpc++/impl/codegen/channel_interface.h
@@ -24,10 +24,8 @@
 #include <grpc/impl/codegen/connectivity_state.h>
 
 namespace grpc {
-class Call;
+class ChannelInterface;
 class ClientContext;
-class RpcMethod;
-class CallOpSetInterface;
 class CompletionQueue;
 
 template <class R>
@@ -36,14 +34,22 @@ template <class W>
 class ClientWriter;
 template <class W, class R>
 class ClientReaderWriter;
+
+namespace internal {
+class Call;
+class CallOpSetInterface;
+class RpcMethod;
+template <class InputMessage, class OutputMessage>
+class BlockingUnaryCallImpl;
 template <class R>
-class ClientAsyncReader;
+class ClientAsyncReaderFactory;
 template <class W>
-class ClientAsyncWriter;
+class ClientAsyncWriterFactory;
 template <class W, class R>
-class ClientAsyncReaderWriter;
+class ClientAsyncReaderWriterFactory;
 template <class R>
-class ClientAsyncResponseReader;
+class ClientAsyncResponseReaderFactory;
+}  // namespace internal
 
 /// Codegen interface for \a grpc::Channel.
 class ChannelInterface {
@@ -88,23 +94,21 @@ class ChannelInterface {
   template <class W, class R>
   friend class ::grpc::ClientReaderWriter;
   template <class R>
-  friend class ::grpc::ClientAsyncReader;
+  friend class ::grpc::internal::ClientAsyncReaderFactory;
   template <class W>
-  friend class ::grpc::ClientAsyncWriter;
+  friend class ::grpc::internal::ClientAsyncWriterFactory;
   template <class W, class R>
-  friend class ::grpc::ClientAsyncReaderWriter;
+  friend class ::grpc::internal::ClientAsyncReaderWriterFactory;
   template <class R>
-  friend class ::grpc::ClientAsyncResponseReader;
+  friend class ::grpc::internal::ClientAsyncResponseReaderFactory;
   template <class InputMessage, class OutputMessage>
-  friend Status BlockingUnaryCall(ChannelInterface* channel,
-                                  const RpcMethod& method,
-                                  ClientContext* context,
-                                  const InputMessage& request,
-                                  OutputMessage* result);
-  friend class ::grpc::RpcMethod;
-  virtual Call CreateCall(const RpcMethod& method, ClientContext* context,
-                          CompletionQueue* cq) = 0;
-  virtual void PerformOpsOnCall(CallOpSetInterface* ops, Call* call) = 0;
+  friend class ::grpc::internal::BlockingUnaryCallImpl;
+  friend class ::grpc::internal::RpcMethod;
+  virtual internal::Call CreateCall(const internal::RpcMethod& method,
+                                    ClientContext* context,
+                                    CompletionQueue* cq) = 0;
+  virtual void PerformOpsOnCall(internal::CallOpSetInterface* ops,
+                                internal::Call* call) = 0;
   virtual void* RegisterMethod(const char* method) = 0;
   virtual void NotifyOnStateChangeImpl(grpc_connectivity_state last_observed,
                                        gpr_timespec deadline,
@@ -112,7 +116,6 @@ class ChannelInterface {
   virtual bool WaitForStateChangeImpl(grpc_connectivity_state last_observed,
                                       gpr_timespec deadline) = 0;
 };
-
 }  // namespace grpc
 
 #endif  // GRPCXX_IMPL_CODEGEN_CHANNEL_INTERFACE_H
diff --git a/include/grpc++/impl/codegen/client_context.h b/include/grpc++/impl/codegen/client_context.h
index 6d7e13bbf2..22b581cbc5 100644
--- a/include/grpc++/impl/codegen/client_context.h
+++ b/include/grpc++/impl/codegen/client_context.h
@@ -60,7 +60,16 @@ class Channel;
 class ChannelInterface;
 class CompletionQueue;
 class CallCredentials;
+class ClientContext;
+
+namespace internal {
 class RpcMethod;
+class CallOpClientRecvStatus;
+class CallOpRecvInitialMetadata;
+template <class InputMessage, class OutputMessage>
+class BlockingUnaryCallImpl;
+}  // namespace internal
+
 template <class R>
 class ClientReader;
 template <class W>
@@ -345,8 +354,8 @@ class ClientContext {
   ClientContext& operator=(const ClientContext&);
 
   friend class ::grpc::testing::InteropClientContextInspector;
-  friend class CallOpClientRecvStatus;
-  friend class CallOpRecvInitialMetadata;
+  friend class ::grpc::internal::CallOpClientRecvStatus;
+  friend class ::grpc::internal::CallOpRecvInitialMetadata;
   friend class Channel;
   template <class R>
   friend class ::grpc::ClientReader;
@@ -363,11 +372,7 @@ class ClientContext {
   template <class R>
   friend class ::grpc::ClientAsyncResponseReader;
   template <class InputMessage, class OutputMessage>
-  friend Status BlockingUnaryCall(ChannelInterface* channel,
-                                  const RpcMethod& method,
-                                  ClientContext* context,
-                                  const InputMessage& request,
-                                  OutputMessage* result);
+  friend class ::grpc::internal::BlockingUnaryCallImpl;
 
   grpc_call* call() const { return call_; }
   void set_call(grpc_call* call, const std::shared_ptr<Channel>& channel);
@@ -399,8 +404,8 @@ class ClientContext {
   mutable std::shared_ptr<const AuthContext> auth_context_;
   struct census_context* census_context_;
   std::multimap<grpc::string, grpc::string> send_initial_metadata_;
-  MetadataMap recv_initial_metadata_;
-  MetadataMap trailing_metadata_;
+  internal::MetadataMap recv_initial_metadata_;
+  internal::MetadataMap trailing_metadata_;
 
   grpc_call* propagate_from_call_;
   PropagationOptions propagation_options_;
diff --git a/include/grpc++/impl/codegen/client_unary_call.h b/include/grpc++/impl/codegen/client_unary_call.h
index 7c540fade9..170c562cf3 100644
--- a/include/grpc++/impl/codegen/client_unary_call.h
+++ b/include/grpc++/impl/codegen/client_unary_call.h
@@ -30,43 +30,60 @@ namespace grpc {
 class Channel;
 class ClientContext;
 class CompletionQueue;
-class RpcMethod;
 
+namespace internal {
+class RpcMethod;
 /// Wrapper that performs a blocking unary call
 template <class InputMessage, class OutputMessage>
 Status BlockingUnaryCall(ChannelInterface* channel, const RpcMethod& method,
                          ClientContext* context, const InputMessage& request,
                          OutputMessage* result) {
-  CompletionQueue cq(grpc_completion_queue_attributes{
-      GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
-      GRPC_CQ_DEFAULT_POLLING});  // Pluckable completion queue
-  Call call(channel->CreateCall(method, context, &cq));
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-            CallOpRecvInitialMetadata, CallOpRecvMessage<OutputMessage>,
-            CallOpClientSendClose, CallOpClientRecvStatus>
-      ops;
-  Status status = ops.SendMessage(request);
-  if (!status.ok()) {
-    return status;
-  }
-  ops.SendInitialMetadata(context->send_initial_metadata_,
-                          context->initial_metadata_flags());
-  ops.RecvInitialMetadata(context);
-  ops.RecvMessage(result);
-  ops.ClientSendClose();
-  ops.ClientRecvStatus(context, &status);
-  call.PerformOps(&ops);
-  if (cq.Pluck(&ops)) {
-    if (!ops.got_message && status.ok()) {
-      return Status(StatusCode::UNIMPLEMENTED,
-                    "No message returned for unary request");
+  return BlockingUnaryCallImpl<InputMessage, OutputMessage>(
+             channel, method, context, request, result)
+      .status();
+};
+
+template <class InputMessage, class OutputMessage>
+class BlockingUnaryCallImpl {
+ public:
+  BlockingUnaryCallImpl(ChannelInterface* channel, const RpcMethod& method,
+                        ClientContext* context, const InputMessage& request,
+                        OutputMessage* result) {
+    CompletionQueue cq(grpc_completion_queue_attributes{
+        GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
+        GRPC_CQ_DEFAULT_POLLING});  // Pluckable completion queue
+    Call call(channel->CreateCall(method, context, &cq));
+    CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
+              CallOpRecvInitialMetadata, CallOpRecvMessage<OutputMessage>,
+              CallOpClientSendClose, CallOpClientRecvStatus>
+        ops;
+    status_ = ops.SendMessage(request);
+    if (!status_.ok()) {
+      return;
+    }
+    ops.SendInitialMetadata(context->send_initial_metadata_,
+                            context->initial_metadata_flags());
+    ops.RecvInitialMetadata(context);
+    ops.RecvMessage(result);
+    ops.ClientSendClose();
+    ops.ClientRecvStatus(context, &status_);
+    call.PerformOps(&ops);
+    if (cq.Pluck(&ops)) {
+      if (!ops.got_message && status_.ok()) {
+        status_ = Status(StatusCode::UNIMPLEMENTED,
+                         "No message returned for unary request");
+      }
+    } else {
+      GPR_CODEGEN_ASSERT(!status_.ok());
     }
-  } else {
-    GPR_CODEGEN_ASSERT(!status.ok());
   }
-  return status;
-}
+  Status status() { return status_; }
+
+ private:
+  Status status_;
+};
 
+}  // namespace internal
 }  // namespace grpc
 
 #endif  // GRPCXX_IMPL_CODEGEN_CLIENT_UNARY_CALL_H
diff --git a/include/grpc++/impl/codegen/completion_queue.h b/include/grpc++/impl/codegen/completion_queue.h
index ca757e2a9c..14de30a62e 100644
--- a/include/grpc++/impl/codegen/completion_queue.h
+++ b/include/grpc++/impl/codegen/completion_queue.h
@@ -56,7 +56,19 @@ class ServerWriter;
 namespace internal {
 template <class W, class R>
 class ServerReaderWriterBody;
-}
+}  // namespace internal
+
+class Channel;
+class ChannelInterface;
+class ClientContext;
+class CompletionQueue;
+class Server;
+class ServerBuilder;
+class ServerContext;
+
+namespace internal {
+class CompletionQueueTag;
+class RpcMethod;
 template <class ServiceType, class RequestType, class ResponseType>
 class RpcMethodHandler;
 template <class ServiceType, class RequestType, class ResponseType>
@@ -66,16 +78,11 @@ class ServerStreamingHandler;
 template <class ServiceType, class RequestType, class ResponseType>
 class BidiStreamingHandler;
 class UnknownMethodHandler;
-
-class Channel;
-class ChannelInterface;
-class ClientContext;
-class CompletionQueueTag;
-class CompletionQueue;
-class RpcMethod;
-class Server;
-class ServerBuilder;
-class ServerContext;
+template <class Streamer, bool WriteNeeded>
+class TemplatedBidiStreamingHandler;
+template <class InputMessage, class OutputMessage>
+class BlockingUnaryCallImpl;
+}  // namespace internal
 
 extern CoreCodegenInterface* g_core_codegen_interface;
 
@@ -109,6 +116,30 @@ class CompletionQueue : private GrpcLibraryCodegen {
     TIMEOUT     ///< deadline was reached.
   };
 
+  /// EXPERIMENTAL
+  /// First executes \a F, then reads from the queue, blocking up to
+  /// \a deadline (or the queue's shutdown).
+  /// Both \a tag and \a ok are updated upon success (if an event is available
+  /// within the \a deadline).  A \a tag points to an arbitrary location usually
+  /// employed to uniquely identify an event.
+  ///
+  /// \param F[in] Function to execute before calling AsyncNext on this queue.
+  /// \param tag[out] Upon sucess, updated to point to the event's tag.
+  /// \param ok[out] Upon sucess, true if read a regular event, false otherwise.
+  /// \param deadline[in] How long to block in wait for an event.
+  ///
+  /// \return The type of event read.
+  template <typename T, typename F>
+  NextStatus DoThenAsyncNext(F&& f, void** tag, bool* ok, const T& deadline) {
+    CompletionQueueTLSCache cache = CompletionQueueTLSCache(this);
+    f();
+    if (cache.Flush(tag, ok)) {
+      return GOT_EVENT;
+    } else {
+      return AsyncNext(tag, ok, deadline);
+    }
+  }
+
   /// Read from the queue, blocking up to \a deadline (or the queue's shutdown).
   /// Both \a tag and \a ok are updated upon success (if an event is available
   /// within the \a deadline).  A \a tag points to an arbitrary location usually
@@ -196,28 +227,39 @@ class CompletionQueue : private GrpcLibraryCodegen {
   template <class W, class R>
   friend class ::grpc::internal::ServerReaderWriterBody;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class RpcMethodHandler;
+  friend class ::grpc::internal::RpcMethodHandler;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class ClientStreamingHandler;
+  friend class ::grpc::internal::ClientStreamingHandler;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class ServerStreamingHandler;
+  friend class ::grpc::internal::ServerStreamingHandler;
   template <class Streamer, bool WriteNeeded>
-  friend class TemplatedBidiStreamingHandler;
-  friend class UnknownMethodHandler;
+  friend class ::grpc::internal::TemplatedBidiStreamingHandler;
+  friend class ::grpc::internal::UnknownMethodHandler;
   friend class ::grpc::Server;
   friend class ::grpc::ServerContext;
   template <class InputMessage, class OutputMessage>
-  friend Status BlockingUnaryCall(ChannelInterface* channel,
-                                  const RpcMethod& method,
-                                  ClientContext* context,
-                                  const InputMessage& request,
-                                  OutputMessage* result);
+  friend class ::grpc::internal::BlockingUnaryCallImpl;
+
+  /// EXPERIMENTAL
+  /// Creates a Thread Local cache to store the first event
+  /// On this completion queue queued from this thread.  Once
+  /// initialized, it must be flushed on the same thread.
+  class CompletionQueueTLSCache {
+   public:
+    CompletionQueueTLSCache(CompletionQueue* cq);
+    ~CompletionQueueTLSCache();
+    bool Flush(void** tag, bool* ok);
+
+   private:
+    CompletionQueue* cq_;
+    bool flushed_;
+  };
 
   NextStatus AsyncNextInternal(void** tag, bool* ok, gpr_timespec deadline);
 
   /// Wraps \a grpc_completion_queue_pluck.
   /// \warning Must not be mixed with calls to \a Next.
-  bool Pluck(CompletionQueueTag* tag) {
+  bool Pluck(internal::CompletionQueueTag* tag) {
     auto deadline =
         g_core_codegen_interface->gpr_inf_future(GPR_CLOCK_REALTIME);
     auto ev = g_core_codegen_interface->grpc_completion_queue_pluck(
@@ -238,7 +280,7 @@ class CompletionQueue : private GrpcLibraryCodegen {
   /// implementation to simple call the other TryPluck function with a zero
   /// timeout. i.e:
   ///      TryPluck(tag, gpr_time_0(GPR_CLOCK_REALTIME))
-  void TryPluck(CompletionQueueTag* tag) {
+  void TryPluck(internal::CompletionQueueTag* tag) {
     auto deadline = g_core_codegen_interface->gpr_time_0(GPR_CLOCK_REALTIME);
     auto ev = g_core_codegen_interface->grpc_completion_queue_pluck(
         cq_, tag, deadline, nullptr);
@@ -254,7 +296,7 @@ class CompletionQueue : private GrpcLibraryCodegen {
   ///
   /// This exects tag->FinalizeResult (if called) to return 'false' i.e expects
   /// that the tag is internal not something that is returned to the user.
-  void TryPluck(CompletionQueueTag* tag, gpr_timespec deadline) {
+  void TryPluck(internal::CompletionQueueTag* tag, gpr_timespec deadline) {
     auto ev = g_core_codegen_interface->grpc_completion_queue_pluck(
         cq_, tag, deadline, nullptr);
     if (ev.type == GRPC_QUEUE_TIMEOUT || ev.type == GRPC_QUEUE_SHUTDOWN) {
diff --git a/include/grpc++/impl/codegen/completion_queue_tag.h b/include/grpc++/impl/codegen/completion_queue_tag.h
index 4d7d3a98dd..cb16bcf9ff 100644
--- a/include/grpc++/impl/codegen/completion_queue_tag.h
+++ b/include/grpc++/impl/codegen/completion_queue_tag.h
@@ -21,6 +21,7 @@
 
 namespace grpc {
 
+namespace internal {
 /// An interface allowing implementors to process and filter event tags.
 class CompletionQueueTag {
  public:
@@ -31,6 +32,7 @@ class CompletionQueueTag {
   /// queue
   virtual bool FinalizeResult(void** tag, bool* status) = 0;
 };
+}  // namespace internal
 
 }  // namespace grpc
 
diff --git a/include/grpc++/impl/codegen/metadata_map.h b/include/grpc++/impl/codegen/metadata_map.h
index b73985967d..fd4750efdd 100644
--- a/include/grpc++/impl/codegen/metadata_map.h
+++ b/include/grpc++/impl/codegen/metadata_map.h
@@ -23,6 +23,7 @@
 
 namespace grpc {
 
+namespace internal {
 class MetadataMap {
  public:
   MetadataMap() { memset(&arr_, 0, sizeof(arr_)); }
@@ -50,6 +51,7 @@ class MetadataMap {
   grpc_metadata_array arr_;
   std::multimap<grpc::string_ref, grpc::string_ref> map_;
 };
+}  // namespace internal
 
 }  // namespace grpc
 
diff --git a/include/grpc++/impl/codegen/method_handler_impl.h b/include/grpc++/impl/codegen/method_handler_impl.h
index e14cb0e926..c0af4ca130 100644
--- a/include/grpc++/impl/codegen/method_handler_impl.h
+++ b/include/grpc++/impl/codegen/method_handler_impl.h
@@ -26,6 +26,7 @@
 
 namespace grpc {
 
+namespace internal {
 /// A wrapper class of an application provided rpc method handler.
 template <class ServiceType, class RequestType, class ResponseType>
 class RpcMethodHandler : public MethodHandler {
@@ -266,6 +267,7 @@ class UnknownMethodHandler : public MethodHandler {
   }
 };
 
+}  // namespace internal
 }  // namespace grpc
 
 #endif  // GRPCXX_IMPL_CODEGEN_METHOD_HANDLER_IMPL_H
diff --git a/include/grpc++/impl/codegen/rpc_method.h b/include/grpc++/impl/codegen/rpc_method.h
index ac13ac56c7..54e52364ef 100644
--- a/include/grpc++/impl/codegen/rpc_method.h
+++ b/include/grpc++/impl/codegen/rpc_method.h
@@ -24,7 +24,7 @@
 #include <grpc++/impl/codegen/channel_interface.h>
 
 namespace grpc {
-
+namespace internal {
 /// Descriptor of an RPC method
 class RpcMethod {
  public:
@@ -55,6 +55,7 @@ class RpcMethod {
   void* const channel_tag_;
 };
 
+}  // namespace internal
 }  // namespace grpc
 
 #endif  // GRPCXX_IMPL_CODEGEN_RPC_METHOD_H
diff --git a/include/grpc++/impl/codegen/rpc_service_method.h b/include/grpc++/impl/codegen/rpc_service_method.h
index d356012ad6..5ba11e8559 100644
--- a/include/grpc++/impl/codegen/rpc_service_method.h
+++ b/include/grpc++/impl/codegen/rpc_service_method.h
@@ -32,8 +32,8 @@
 
 namespace grpc {
 class ServerContext;
-class StreamContextInterface;
 
+namespace internal {
 /// Base class for running an RPC handler.
 class MethodHandler {
  public:
@@ -71,6 +71,7 @@ class RpcServiceMethod : public RpcMethod {
   void* server_tag_;
   std::unique_ptr<MethodHandler> handler_;
 };
+}  // namespace internal
 
 }  // namespace grpc
 
diff --git a/include/grpc++/impl/codegen/server_context.h b/include/grpc++/impl/codegen/server_context.h
index b5e37fd12b..a2d6967bf8 100644
--- a/include/grpc++/impl/codegen/server_context.h
+++ b/include/grpc++/impl/codegen/server_context.h
@@ -55,7 +55,6 @@ class ServerWriter;
 namespace internal {
 template <class W, class R>
 class ServerReaderWriterBody;
-}
 template <class ServiceType, class RequestType, class ResponseType>
 class RpcMethodHandler;
 template <class ServiceType, class RequestType, class ResponseType>
@@ -65,9 +64,11 @@ class ServerStreamingHandler;
 template <class ServiceType, class RequestType, class ResponseType>
 class BidiStreamingHandler;
 class UnknownMethodHandler;
-
+template <class Streamer, bool WriteNeeded>
+class TemplatedBidiStreamingHandler;
 class Call;
-class CallOpBuffer;
+}  // namespace internal
+
 class CompletionQueue;
 class Server;
 class ServerInterface;
@@ -247,14 +248,14 @@ class ServerContext {
   template <class W, class R>
   friend class ::grpc::internal::ServerReaderWriterBody;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class RpcMethodHandler;
+  friend class ::grpc::internal::RpcMethodHandler;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class ClientStreamingHandler;
+  friend class ::grpc::internal::ClientStreamingHandler;
   template <class ServiceType, class RequestType, class ResponseType>
-  friend class ServerStreamingHandler;
+  friend class ::grpc::internal::ServerStreamingHandler;
   template <class Streamer, bool WriteNeeded>
-  friend class TemplatedBidiStreamingHandler;
-  friend class UnknownMethodHandler;
+  friend class ::grpc::internal::TemplatedBidiStreamingHandler;
+  friend class ::grpc::internal::UnknownMethodHandler;
   friend class ::grpc::ClientContext;
 
   /// Prevent copying.
@@ -263,9 +264,9 @@ class ServerContext {
 
   class CompletionOp;
 
-  void BeginCompletionOp(Call* call);
+  void BeginCompletionOp(internal::Call* call);
   /// Return the tag queued by BeginCompletionOp()
-  CompletionQueueTag* GetCompletionOpTag();
+  internal::CompletionQueueTag* GetCompletionOpTag();
 
   ServerContext(gpr_timespec deadline, grpc_metadata_array* arr);
 
@@ -282,7 +283,7 @@ class ServerContext {
   CompletionQueue* cq_;
   bool sent_initial_metadata_;
   mutable std::shared_ptr<const AuthContext> auth_context_;
-  MetadataMap client_metadata_;
+  internal::MetadataMap client_metadata_;
   std::multimap<grpc::string, grpc::string> initial_metadata_;
   std::multimap<grpc::string, grpc::string> trailing_metadata_;
 
@@ -290,7 +291,9 @@ class ServerContext {
   grpc_compression_level compression_level_;
   grpc_compression_algorithm compression_algorithm_;
 
-  CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage> pending_ops_;
+  internal::CallOpSet<internal::CallOpSendInitialMetadata,
+                      internal::CallOpSendMessage>
+      pending_ops_;
   bool has_pending_ops_;
 };
 
diff --git a/include/grpc++/impl/codegen/server_interface.h b/include/grpc++/impl/codegen/server_interface.h
index 55937f19d7..3bcf4c87e7 100644
--- a/include/grpc++/impl/codegen/server_interface.h
+++ b/include/grpc++/impl/codegen/server_interface.h
@@ -30,20 +30,21 @@ namespace grpc {
 class AsyncGenericService;
 class Channel;
 class GenericServerContext;
-class RpcService;
-class ServerAsyncStreamingInterface;
 class ServerCompletionQueue;
 class ServerContext;
 class ServerCredentials;
 class Service;
-class ThreadPoolInterface;
 
 extern CoreCodegenInterface* g_core_codegen_interface;
 
 /// Models a gRPC server.
 ///
 /// Servers are configured and started via \a grpc::ServerBuilder.
-class ServerInterface : public CallHook {
+namespace internal {
+class ServerAsyncStreamingInterface;
+}  // namespace internal
+
+class ServerInterface : public internal::CallHook {
  public:
   virtual ~ServerInterface() {}
 
@@ -78,7 +79,7 @@ class ServerInterface : public CallHook {
   virtual void Wait() = 0;
 
  protected:
-  friend class Service;
+  friend class ::grpc::Service;
 
   /// Register a service. This call does not take ownership of the service.
   /// The service must exist for the lifetime of the Server instance.
@@ -116,12 +117,13 @@ class ServerInterface : public CallHook {
 
   virtual grpc_server* server() = 0;
 
-  virtual void PerformOpsOnCall(CallOpSetInterface* ops, Call* call) = 0;
+  virtual void PerformOpsOnCall(internal::CallOpSetInterface* ops,
+                                internal::Call* call) = 0;
 
-  class BaseAsyncRequest : public CompletionQueueTag {
+  class BaseAsyncRequest : public internal::CompletionQueueTag {
    public:
     BaseAsyncRequest(ServerInterface* server, ServerContext* context,
-                     ServerAsyncStreamingInterface* stream,
+                     internal::ServerAsyncStreamingInterface* stream,
                      CompletionQueue* call_cq, void* tag,
                      bool delete_on_finalize);
     virtual ~BaseAsyncRequest();
@@ -131,7 +133,7 @@ class ServerInterface : public CallHook {
    protected:
     ServerInterface* const server_;
     ServerContext* const context_;
-    ServerAsyncStreamingInterface* const stream_;
+    internal::ServerAsyncStreamingInterface* const stream_;
     CompletionQueue* const call_cq_;
     void* const tag_;
     const bool delete_on_finalize_;
@@ -141,7 +143,7 @@ class ServerInterface : public CallHook {
   class RegisteredAsyncRequest : public BaseAsyncRequest {
    public:
     RegisteredAsyncRequest(ServerInterface* server, ServerContext* context,
-                           ServerAsyncStreamingInterface* stream,
+                           internal::ServerAsyncStreamingInterface* stream,
                            CompletionQueue* call_cq, void* tag);
 
     // uses BaseAsyncRequest::FinalizeResult
@@ -155,7 +157,7 @@ class ServerInterface : public CallHook {
    public:
     NoPayloadAsyncRequest(void* registered_method, ServerInterface* server,
                           ServerContext* context,
-                          ServerAsyncStreamingInterface* stream,
+                          internal::ServerAsyncStreamingInterface* stream,
                           CompletionQueue* call_cq,
                           ServerCompletionQueue* notification_cq, void* tag)
         : RegisteredAsyncRequest(server, context, stream, call_cq, tag) {
@@ -170,7 +172,7 @@ class ServerInterface : public CallHook {
    public:
     PayloadAsyncRequest(void* registered_method, ServerInterface* server,
                         ServerContext* context,
-                        ServerAsyncStreamingInterface* stream,
+                        internal::ServerAsyncStreamingInterface* stream,
                         CompletionQueue* call_cq,
                         ServerCompletionQueue* notification_cq, void* tag,
                         Message* request)
@@ -212,7 +214,7 @@ class ServerInterface : public CallHook {
     void* const registered_method_;
     ServerInterface* const server_;
     ServerContext* const context_;
-    ServerAsyncStreamingInterface* const stream_;
+    internal::ServerAsyncStreamingInterface* const stream_;
     CompletionQueue* const call_cq_;
     ServerCompletionQueue* const notification_cq_;
     void* const tag_;
@@ -223,7 +225,7 @@ class ServerInterface : public CallHook {
   class GenericAsyncRequest : public BaseAsyncRequest {
    public:
     GenericAsyncRequest(ServerInterface* server, GenericServerContext* context,
-                        ServerAsyncStreamingInterface* stream,
+                        internal::ServerAsyncStreamingInterface* stream,
                         CompletionQueue* call_cq,
                         ServerCompletionQueue* notification_cq, void* tag,
                         bool delete_on_finalize);
@@ -235,8 +237,9 @@ class ServerInterface : public CallHook {
   };
 
   template <class Message>
-  void RequestAsyncCall(RpcServiceMethod* method, ServerContext* context,
-                        ServerAsyncStreamingInterface* stream,
+  void RequestAsyncCall(internal::RpcServiceMethod* method,
+                        ServerContext* context,
+                        internal::ServerAsyncStreamingInterface* stream,
                         CompletionQueue* call_cq,
                         ServerCompletionQueue* notification_cq, void* tag,
                         Message* message) {
@@ -246,8 +249,9 @@ class ServerInterface : public CallHook {
                                      message);
   }
 
-  void RequestAsyncCall(RpcServiceMethod* method, ServerContext* context,
-                        ServerAsyncStreamingInterface* stream,
+  void RequestAsyncCall(internal::RpcServiceMethod* method,
+                        ServerContext* context,
+                        internal::ServerAsyncStreamingInterface* stream,
                         CompletionQueue* call_cq,
                         ServerCompletionQueue* notification_cq, void* tag) {
     GPR_CODEGEN_ASSERT(method);
@@ -256,7 +260,7 @@ class ServerInterface : public CallHook {
   }
 
   void RequestAsyncGenericCall(GenericServerContext* context,
-                               ServerAsyncStreamingInterface* stream,
+                               internal::ServerAsyncStreamingInterface* stream,
                                CompletionQueue* call_cq,
                                ServerCompletionQueue* notification_cq,
                                void* tag) {
diff --git a/include/grpc++/impl/codegen/service_type.h b/include/grpc++/impl/codegen/service_type.h
index 2dc4ea0ea6..71c3d99d5c 100644
--- a/include/grpc++/impl/codegen/service_type.h
+++ b/include/grpc++/impl/codegen/service_type.h
@@ -28,13 +28,14 @@
 
 namespace grpc {
 
-class Call;
 class CompletionQueue;
 class Server;
 class ServerInterface;
 class ServerCompletionQueue;
 class ServerContext;
 
+namespace internal {
+class Call;
 class ServerAsyncStreamingInterface {
  public:
   virtual ~ServerAsyncStreamingInterface() {}
@@ -48,9 +49,10 @@ class ServerAsyncStreamingInterface {
   virtual void SendInitialMetadata(void* tag) = 0;
 
  private:
-  friend class ServerInterface;
+  friend class ::grpc::ServerInterface;
   virtual void BindCall(Call* call) = 0;
 };
+}  // namespace internal
 
 /// Desriptor of an RPC service and its various RPC methods
 class Service {
@@ -88,40 +90,38 @@ class Service {
  protected:
   template <class Message>
   void RequestAsyncUnary(int index, ServerContext* context, Message* request,
-                         ServerAsyncStreamingInterface* stream,
+                         internal::ServerAsyncStreamingInterface* stream,
                          CompletionQueue* call_cq,
                          ServerCompletionQueue* notification_cq, void* tag) {
     server_->RequestAsyncCall(methods_[index].get(), context, stream, call_cq,
                               notification_cq, tag, request);
   }
-  void RequestAsyncClientStreaming(int index, ServerContext* context,
-                                   ServerAsyncStreamingInterface* stream,
-                                   CompletionQueue* call_cq,
-                                   ServerCompletionQueue* notification_cq,
-                                   void* tag) {
+  void RequestAsyncClientStreaming(
+      int index, ServerContext* context,
+      internal::ServerAsyncStreamingInterface* stream, CompletionQueue* call_cq,
+      ServerCompletionQueue* notification_cq, void* tag) {
     server_->RequestAsyncCall(methods_[index].get(), context, stream, call_cq,
                               notification_cq, tag);
   }
   template <class Message>
-  void RequestAsyncServerStreaming(int index, ServerContext* context,
-                                   Message* request,
-                                   ServerAsyncStreamingInterface* stream,
-                                   CompletionQueue* call_cq,
-                                   ServerCompletionQueue* notification_cq,
-                                   void* tag) {
+  void RequestAsyncServerStreaming(
+      int index, ServerContext* context, Message* request,
+      internal::ServerAsyncStreamingInterface* stream, CompletionQueue* call_cq,
+      ServerCompletionQueue* notification_cq, void* tag) {
     server_->RequestAsyncCall(methods_[index].get(), context, stream, call_cq,
                               notification_cq, tag, request);
   }
-  void RequestAsyncBidiStreaming(int index, ServerContext* context,
-                                 ServerAsyncStreamingInterface* stream,
-                                 CompletionQueue* call_cq,
-                                 ServerCompletionQueue* notification_cq,
-                                 void* tag) {
+  void RequestAsyncBidiStreaming(
+      int index, ServerContext* context,
+      internal::ServerAsyncStreamingInterface* stream, CompletionQueue* call_cq,
+      ServerCompletionQueue* notification_cq, void* tag) {
     server_->RequestAsyncCall(methods_[index].get(), context, stream, call_cq,
                               notification_cq, tag);
   }
 
-  void AddMethod(RpcServiceMethod* method) { methods_.emplace_back(method); }
+  void AddMethod(internal::RpcServiceMethod* method) {
+    methods_.emplace_back(method);
+  }
 
   void MarkMethodAsync(int index) {
     GPR_CODEGEN_ASSERT(
@@ -139,7 +139,7 @@ class Service {
     methods_[index].reset();
   }
 
-  void MarkMethodStreamed(int index, MethodHandler* streamed_method) {
+  void MarkMethodStreamed(int index, internal::MethodHandler* streamed_method) {
     GPR_CODEGEN_ASSERT(methods_[index] && methods_[index]->handler() &&
                        "Cannot mark an async or generic method Streamed");
     methods_[index]->SetHandler(streamed_method);
@@ -148,14 +148,14 @@ class Service {
     // case of BIDI_STREAMING that has 1 read and 1 write, in that order,
     // and split server-side streaming is BIDI_STREAMING with 1 read and
     // any number of writes, in that order.
-    methods_[index]->SetMethodType(::grpc::RpcMethod::BIDI_STREAMING);
+    methods_[index]->SetMethodType(internal::RpcMethod::BIDI_STREAMING);
   }
 
  private:
   friend class Server;
   friend class ServerInterface;
   ServerInterface* server_;
-  std::vector<std::unique_ptr<RpcServiceMethod>> methods_;
+  std::vector<std::unique_ptr<internal::RpcServiceMethod>> methods_;
 };
 
 }  // namespace grpc
diff --git a/include/grpc++/impl/codegen/sync_stream.h b/include/grpc++/impl/codegen/sync_stream.h
index c1784f1820..a6dd26fb00 100644
--- a/include/grpc++/impl/codegen/sync_stream.h
+++ b/include/grpc++/impl/codegen/sync_stream.h
@@ -30,6 +30,7 @@
 
 namespace grpc {
 
+namespace internal {
 /// Common interface for all synchronous client side streaming.
 class ClientStreamingInterface {
  public:
@@ -141,10 +142,12 @@ class WriterInterface {
   }
 };
 
+}  // namespace internal
+
 /// Client-side interface for streaming reads of message of type \a R.
 template <class R>
-class ClientReaderInterface : public ClientStreamingInterface,
-                              public ReaderInterface<R> {
+class ClientReaderInterface : public internal::ClientStreamingInterface,
+                              public internal::ReaderInterface<R> {
  public:
   /// Block to wait for initial metadata from server. The received metadata
   /// can only be accessed after this call returns. Should only be called before
@@ -153,35 +156,25 @@ class ClientReaderInterface : public ClientStreamingInterface,
   virtual void WaitForInitialMetadata() = 0;
 };
 
+namespace internal {
+template <class R>
+class ClientReaderFactory {
+ public:
+  template <class W>
+  static ClientReader<R>* Create(ChannelInterface* channel,
+                                 const ::grpc::internal::RpcMethod& method,
+                                 ClientContext* context, const W& request) {
+    return new ClientReader<R>(channel, method, context, request);
+  }
+};
+}  // namespace internal
+
 /// Synchronous (blocking) client-side API for doing server-streaming RPCs,
 /// where the stream of messages coming from the server has messages
 /// of type \a R.
 template <class R>
 class ClientReader final : public ClientReaderInterface<R> {
  public:
-  /// Block to create a stream and write the initial metadata and \a request
-  /// out. Note that \a context will be used to fill in custom initial
-  /// metadata used to send to the server when starting the call.
-  template <class W>
-  ClientReader(ChannelInterface* channel, const RpcMethod& method,
-               ClientContext* context, const W& request)
-      : context_(context),
-        cq_(grpc_completion_queue_attributes{
-            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
-            GRPC_CQ_DEFAULT_POLLING}),  // Pluckable cq
-        call_(channel->CreateCall(method, context, &cq_)) {
-    CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-              CallOpClientSendClose>
-        ops;
-    ops.SendInitialMetadata(context->send_initial_metadata_,
-                            context->initial_metadata_flags());
-    // TODO(ctiller): don't assert
-    GPR_CODEGEN_ASSERT(ops.SendMessage(request).ok());
-    ops.ClientSendClose();
-    call_.PerformOps(&ops);
-    cq_.Pluck(&ops);
-  }
-
   /// See the \a ClientStreamingInterface.WaitForInitialMetadata method for
   /// semantics.
   ///
@@ -192,7 +185,8 @@ class ClientReader final : public ClientReaderInterface<R> {
   void WaitForInitialMetadata() override {
     GPR_CODEGEN_ASSERT(!context_->initial_metadata_received_);
 
-    CallOpSet<CallOpRecvInitialMetadata> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+        ops;
     ops.RecvInitialMetadata(context_);
     call_.PerformOps(&ops);
     cq_.Pluck(&ops);  /// status ignored
@@ -209,7 +203,9 @@ class ClientReader final : public ClientReaderInterface<R> {
   ///   already received (if initial metadata is received, it can be then
   ///   accessed through the \a ClientContext associated with this call).
   bool Read(R* msg) override {
-    CallOpSet<CallOpRecvInitialMetadata, CallOpRecvMessage<R>> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                                ::grpc::internal::CallOpRecvMessage<R>>
+        ops;
     if (!context_->initial_metadata_received_) {
       ops.RecvInitialMetadata(context_);
     }
@@ -224,7 +220,7 @@ class ClientReader final : public ClientReaderInterface<R> {
   ///   The \a ClientContext associated with this call is updated with
   ///   possible metadata received from the server.
   Status Finish() override {
-    CallOpSet<CallOpClientRecvStatus> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpClientRecvStatus> ops;
     Status status;
     ops.ClientRecvStatus(context_, &status);
     call_.PerformOps(&ops);
@@ -233,15 +229,41 @@ class ClientReader final : public ClientReaderInterface<R> {
   }
 
  private:
+  friend class internal::ClientReaderFactory<R>;
   ClientContext* context_;
   CompletionQueue cq_;
-  Call call_;
+  ::grpc::internal::Call call_;
+
+  /// Block to create a stream and write the initial metadata and \a request
+  /// out. Note that \a context will be used to fill in custom initial
+  /// metadata used to send to the server when starting the call.
+  template <class W>
+  ClientReader(::grpc::ChannelInterface* channel,
+               const ::grpc::internal::RpcMethod& method,
+               ClientContext* context, const W& request)
+      : context_(context),
+        cq_(grpc_completion_queue_attributes{
+            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
+            GRPC_CQ_DEFAULT_POLLING}),  // Pluckable cq
+        call_(channel->CreateCall(method, context, &cq_)) {
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                                ::grpc::internal::CallOpSendMessage,
+                                ::grpc::internal::CallOpClientSendClose>
+        ops;
+    ops.SendInitialMetadata(context->send_initial_metadata_,
+                            context->initial_metadata_flags());
+    // TODO(ctiller): don't assert
+    GPR_CODEGEN_ASSERT(ops.SendMessage(request).ok());
+    ops.ClientSendClose();
+    call_.PerformOps(&ops);
+    cq_.Pluck(&ops);
+  }
 };
 
 /// Client-side interface for streaming writes of message type \a W.
 template <class W>
-class ClientWriterInterface : public ClientStreamingInterface,
-                              public WriterInterface<W> {
+class ClientWriterInterface : public internal::ClientStreamingInterface,
+                              public internal::WriterInterface<W> {
  public:
   /// Half close writing from the client. (signal that the stream of messages
   /// coming from the client is complete).
@@ -252,37 +274,25 @@ class ClientWriterInterface : public ClientStreamingInterface,
   virtual bool WritesDone() = 0;
 };
 
+namespace internal {
+template <class W>
+class ClientWriterFactory {
+ public:
+  template <class R>
+  static ClientWriter<W>* Create(::grpc::ChannelInterface* channel,
+                                 const ::grpc::internal::RpcMethod& method,
+                                 ClientContext* context, R* response) {
+    return new ClientWriter<W>(channel, method, context, response);
+  }
+};
+}  // namespace internal
+
 /// Synchronous (blocking) client-side API for doing client-streaming RPCs,
 /// where the outgoing message stream coming from the client has messages of
 /// type \a W.
 template <class W>
 class ClientWriter : public ClientWriterInterface<W> {
  public:
-  /// Block to create a stream (i.e. send request headers and other initial
-  /// metadata to the server). Note that \a context will be used to fill
-  /// in custom initial metadata. \a response will be filled in with the
-  /// single expected response message from the server upon a successful
-  /// call to the \a Finish method of this instance.
-  template <class R>
-  ClientWriter(ChannelInterface* channel, const RpcMethod& method,
-               ClientContext* context, R* response)
-      : context_(context),
-        cq_(grpc_completion_queue_attributes{
-            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
-            GRPC_CQ_DEFAULT_POLLING}),  // Pluckable cq
-        call_(channel->CreateCall(method, context, &cq_)) {
-    finish_ops_.RecvMessage(response);
-    finish_ops_.AllowNoMessage();
-
-    if (!context_->initial_metadata_corked_) {
-      CallOpSet<CallOpSendInitialMetadata> ops;
-      ops.SendInitialMetadata(context->send_initial_metadata_,
-                              context->initial_metadata_flags());
-      call_.PerformOps(&ops);
-      cq_.Pluck(&ops);
-    }
-  }
-
   /// See the \a ClientStreamingInterface.WaitForInitialMetadata method for
   /// semantics.
   ///
@@ -292,7 +302,8 @@ class ClientWriter : public ClientWriterInterface<W> {
   void WaitForInitialMetadata() {
     GPR_CODEGEN_ASSERT(!context_->initial_metadata_received_);
 
-    CallOpSet<CallOpRecvInitialMetadata> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+        ops;
     ops.RecvInitialMetadata(context_);
     call_.PerformOps(&ops);
     cq_.Pluck(&ops);  // status ignored
@@ -304,10 +315,11 @@ class ClientWriter : public ClientWriterInterface<W> {
   /// Side effect:
   ///   Also sends initial metadata if not already sent (using the
   ///   \a ClientContext associated with this call).
-  using WriterInterface<W>::Write;
+  using ::grpc::internal::WriterInterface<W>::Write;
   bool Write(const W& msg, WriteOptions options) override {
-    CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-              CallOpClientSendClose>
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                                ::grpc::internal::CallOpSendMessage,
+                                ::grpc::internal::CallOpClientSendClose>
         ops;
 
     if (options.is_last_message()) {
@@ -328,7 +340,7 @@ class ClientWriter : public ClientWriterInterface<W> {
   }
 
   bool WritesDone() override {
-    CallOpSet<CallOpClientSendClose> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpClientSendClose> ops;
     ops.ClientSendClose();
     call_.PerformOps(&ops);
     return cq_.Pluck(&ops);
@@ -352,21 +364,51 @@ class ClientWriter : public ClientWriterInterface<W> {
   }
 
  private:
+  friend class internal::ClientWriterFactory<W>;
+
+  /// Block to create a stream (i.e. send request headers and other initial
+  /// metadata to the server). Note that \a context will be used to fill
+  /// in custom initial metadata. \a response will be filled in with the
+  /// single expected response message from the server upon a successful
+  /// call to the \a Finish method of this instance.
+  template <class R>
+  ClientWriter(ChannelInterface* channel,
+               const ::grpc::internal::RpcMethod& method,
+               ClientContext* context, R* response)
+      : context_(context),
+        cq_(grpc_completion_queue_attributes{
+            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
+            GRPC_CQ_DEFAULT_POLLING}),  // Pluckable cq
+        call_(channel->CreateCall(method, context, &cq_)) {
+    finish_ops_.RecvMessage(response);
+    finish_ops_.AllowNoMessage();
+
+    if (!context_->initial_metadata_corked_) {
+      ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata>
+          ops;
+      ops.SendInitialMetadata(context->send_initial_metadata_,
+                              context->initial_metadata_flags());
+      call_.PerformOps(&ops);
+      cq_.Pluck(&ops);
+    }
+  }
+
   ClientContext* context_;
-  CallOpSet<CallOpRecvInitialMetadata, CallOpGenericRecvMessage,
-            CallOpClientRecvStatus>
+  ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                              ::grpc::internal::CallOpGenericRecvMessage,
+                              ::grpc::internal::CallOpClientRecvStatus>
       finish_ops_;
   CompletionQueue cq_;
-  Call call_;
+  ::grpc::internal::Call call_;
 };
 
 /// Client-side interface for bi-directional streaming with
 /// client-to-server stream messages of type \a W and
 /// server-to-client stream messages of type \a R.
 template <class W, class R>
-class ClientReaderWriterInterface : public ClientStreamingInterface,
-                                    public WriterInterface<W>,
-                                    public ReaderInterface<R> {
+class ClientReaderWriterInterface : public internal::ClientStreamingInterface,
+                                    public internal::WriterInterface<W>,
+                                    public internal::ReaderInterface<R> {
  public:
   /// Block to wait for initial metadata from server. The received metadata
   /// can only be accessed after this call returns. Should only be called before
@@ -375,7 +417,7 @@ class ClientReaderWriterInterface : public ClientStreamingInterface,
   virtual void WaitForInitialMetadata() = 0;
 
   /// Half close writing from the client. (signal that the stream of messages
-  /// coming from the client is complete).
+  /// coming from the clinet is complete).
   /// Blocks until currently-pending writes are completed.
   /// Thread-safe with respect to \a ReaderInterface::Read
   ///
@@ -383,6 +425,18 @@ class ClientReaderWriterInterface : public ClientStreamingInterface,
   virtual bool WritesDone() = 0;
 };
 
+namespace internal {
+template <class W, class R>
+class ClientReaderWriterFactory {
+ public:
+  static ClientReaderWriter<W, R>* Create(
+      ::grpc::ChannelInterface* channel,
+      const ::grpc::internal::RpcMethod& method, ClientContext* context) {
+    return new ClientReaderWriter<W, R>(channel, method, context);
+  }
+};
+}  // namespace internal
+
 /// Synchronous (blocking) client-side API for bi-directional streaming RPCs,
 /// where the outgoing message stream coming from the client has messages of
 /// type \a W, and the incoming messages stream coming from the server has
@@ -390,25 +444,6 @@ class ClientReaderWriterInterface : public ClientStreamingInterface,
 template <class W, class R>
 class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
  public:
-  /// Block to create a stream and write the initial metadata and \a request
-  /// out. Note that \a context will be used to fill in custom initial metadata
-  /// used to send to the server when starting the call.
-  ClientReaderWriter(ChannelInterface* channel, const RpcMethod& method,
-                     ClientContext* context)
-      : context_(context),
-        cq_(grpc_completion_queue_attributes{
-            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
-            GRPC_CQ_DEFAULT_POLLING}),  // Pluckable cq
-        call_(channel->CreateCall(method, context, &cq_)) {
-    if (!context_->initial_metadata_corked_) {
-      CallOpSet<CallOpSendInitialMetadata> ops;
-      ops.SendInitialMetadata(context->send_initial_metadata_,
-                              context->initial_metadata_flags());
-      call_.PerformOps(&ops);
-      cq_.Pluck(&ops);
-    }
-  }
-
   /// Block waiting to read initial metadata from the server.
   /// This call is optional, but if it is used, it cannot be used concurrently
   /// with or after the \a Finish method.
@@ -418,7 +453,8 @@ class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
   void WaitForInitialMetadata() override {
     GPR_CODEGEN_ASSERT(!context_->initial_metadata_received_);
 
-    CallOpSet<CallOpRecvInitialMetadata> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata>
+        ops;
     ops.RecvInitialMetadata(context_);
     call_.PerformOps(&ops);
     cq_.Pluck(&ops);  // status ignored
@@ -434,7 +470,9 @@ class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
   ///   Also receives initial metadata if not already received (updates the \a
   ///   ClientContext associated with this call in that case).
   bool Read(R* msg) override {
-    CallOpSet<CallOpRecvInitialMetadata, CallOpRecvMessage<R>> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                                ::grpc::internal::CallOpRecvMessage<R>>
+        ops;
     if (!context_->initial_metadata_received_) {
       ops.RecvInitialMetadata(context_);
     }
@@ -448,10 +486,11 @@ class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
   /// Side effect:
   ///   Also sends initial metadata if not already sent (using the
   ///   \a ClientContext associated with this call to fill in values).
-  using WriterInterface<W>::Write;
+  using ::grpc::internal::WriterInterface<W>::Write;
   bool Write(const W& msg, WriteOptions options) override {
-    CallOpSet<CallOpSendInitialMetadata, CallOpSendMessage,
-              CallOpClientSendClose>
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata,
+                                ::grpc::internal::CallOpSendMessage,
+                                ::grpc::internal::CallOpClientSendClose>
         ops;
 
     if (options.is_last_message()) {
@@ -472,7 +511,7 @@ class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
   }
 
   bool WritesDone() override {
-    CallOpSet<CallOpClientSendClose> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpClientSendClose> ops;
     ops.ClientSendClose();
     call_.PerformOps(&ops);
     return cq_.Pluck(&ops);
@@ -484,7 +523,9 @@ class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
   ///   - the \a ClientContext associated with this call is updated with
   ///     possible trailing metadata sent from the server.
   Status Finish() override {
-    CallOpSet<CallOpRecvInitialMetadata, CallOpClientRecvStatus> ops;
+    ::grpc::internal::CallOpSet<::grpc::internal::CallOpRecvInitialMetadata,
+                                ::grpc::internal::CallOpClientRecvStatus>
+        ops;
     if (!context_->initial_metadata_received_) {
       ops.RecvInitialMetadata(context_);
     }
@@ -496,15 +537,38 @@ class ClientReaderWriter final : public ClientReaderWriterInterface<W, R> {
   }
 
  private:
+  friend class internal::ClientReaderWriterFactory<W, R>;
+
   ClientContext* context_;
   CompletionQueue cq_;
-  Call call_;
+  ::grpc::internal::Call call_;
+
+  /// Block to create a stream and write the initial metadata and \a request
+  /// out. Note that \a context will be used to fill in custom initial metadata
+  /// used to send to the server when starting the call.
+  ClientReaderWriter(::grpc::ChannelInterface* channel,
+                     const ::grpc::internal::RpcMethod& method,
+                     ClientContext* context)
+      : context_(context),
+        cq_(grpc_completion_queue_attributes{
+            GRPC_CQ_CURRENT_VERSION, GRPC_CQ_PLUCK,
+            GRPC_CQ_DEFAULT_POLLING}),  // Pluckable cq
+        call_(channel->CreateCall(method, context, &cq_)) {
+    if (!context_->initial_metadata_corked_) {
+      ::grpc::internal::CallOpSet<::grpc::internal::CallOpSendInitialMetadata>
+          ops;
+      ops.SendInitialMetadata(context->send_initial_metadata_,
+                              context->initial_metadata_flags());
+      call_.PerformOps(&ops);
+      cq_.Pluck(&ops);
+    }
+  }
 };
 
 /// Server-side interface for streaming reads of message of type \a R.
 template <class R>
-class ServerReaderInterface : public ServerStreamingInterface,
-                              public ReaderInterface<R> {};
+class ServerReaderInterface : public internal::ServerStreamingInterface,
+                              public internal::ReaderInterface<R> {};
 
 /// Synchronous (blocking) server-side API for doing client-streaming RPCs,
 /// where the incoming message stream coming from the client has messages of
@@ -512,15 +576,13 @@ class ServerReaderInterface : public ServerStreamingInterface,
 template <class R>
 class ServerReader final : public ServerReaderInterface<R> {
  public:
-  ServerReader(Call* call, ServerContext* ctx) : call_(call), ctx_(ctx) {}
-
   /// See the \a ServerStreamingInterface.SendInitialMetadata method
   /// for semantics. Note that initial metadata will be affected by the
   /// \a ServerContext associated with this call.
   void SendInitialMetadata() override {
     GPR_CODEGEN_ASSERT(!ctx_->sent_initial_metadata_);
 
-    CallOpSet<CallOpSendInitialMetadata> ops;
+    internal::CallOpSet<internal::CallOpSendInitialMetadata> ops;
     ops.SendInitialMetadata(ctx_->initial_metadata_,
                             ctx_->initial_metadata_flags());
     if (ctx_->compression_level_set()) {
@@ -537,21 +599,27 @@ class ServerReader final : public ServerReaderInterface<R> {
   }
 
   bool Read(R* msg) override {
-    CallOpSet<CallOpRecvMessage<R>> ops;
+    internal::CallOpSet<internal::CallOpRecvMessage<R>> ops;
     ops.RecvMessage(msg);
     call_->PerformOps(&ops);
     return call_->cq()->Pluck(&ops) && ops.got_message;
   }
 
  private:
-  Call* const call_;
+  internal::Call* const call_;
   ServerContext* const ctx_;
+
+  template <class ServiceType, class RequestType, class ResponseType>
+  friend class internal::ClientStreamingHandler;
+
+  ServerReader(internal::Call* call, ServerContext* ctx)
+      : call_(call), ctx_(ctx) {}
 };
 
 /// Server-side interface for streaming writes of message of type \a W.
 template <class W>
-class ServerWriterInterface : public ServerStreamingInterface,
-                              public WriterInterface<W> {};
+class ServerWriterInterface : public internal::ServerStreamingInterface,
+                              public internal::WriterInterface<W> {};
 
 /// Synchronous (blocking) server-side API for doing for doing a
 /// server-streaming RPCs, where the outgoing message stream coming from the
@@ -559,8 +627,6 @@ class ServerWriterInterface : public ServerStreamingInterface,
 template <class W>
 class ServerWriter final : public ServerWriterInterface<W> {
  public:
-  ServerWriter(Call* call, ServerContext* ctx) : call_(call), ctx_(ctx) {}
-
   /// See the \a ServerStreamingInterface.SendInitialMetadata method
   /// for semantics.
   /// Note that initial metadata will be affected by the
@@ -568,7 +634,7 @@ class ServerWriter final : public ServerWriterInterface<W> {
   void SendInitialMetadata() override {
     GPR_CODEGEN_ASSERT(!ctx_->sent_initial_metadata_);
 
-    CallOpSet<CallOpSendInitialMetadata> ops;
+    internal::CallOpSet<internal::CallOpSendInitialMetadata> ops;
     ops.SendInitialMetadata(ctx_->initial_metadata_,
                             ctx_->initial_metadata_flags());
     if (ctx_->compression_level_set()) {
@@ -584,11 +650,12 @@ class ServerWriter final : public ServerWriterInterface<W> {
   /// Side effect:
   ///   Also sends initial metadata if not already sent (using the
   ///   \a ClientContext associated with this call to fill in values).
-  using WriterInterface<W>::Write;
+  using internal::WriterInterface<W>::Write;
   bool Write(const W& msg, WriteOptions options) override {
     if (options.is_last_message()) {
       options.set_buffer_hint();
     }
+
     if (!ctx_->pending_ops_.SendMessage(msg, options).ok()) {
       return false;
     }
@@ -613,15 +680,21 @@ class ServerWriter final : public ServerWriterInterface<W> {
   }
 
  private:
-  Call* const call_;
+  internal::Call* const call_;
   ServerContext* const ctx_;
+
+  template <class ServiceType, class RequestType, class ResponseType>
+  friend class internal::ServerStreamingHandler;
+
+  ServerWriter(internal::Call* call, ServerContext* ctx)
+      : call_(call), ctx_(ctx) {}
 };
 
 /// Server-side interface for bi-directional streaming.
 template <class W, class R>
-class ServerReaderWriterInterface : public ServerStreamingInterface,
-                                    public WriterInterface<W>,
-                                    public ReaderInterface<R> {};
+class ServerReaderWriterInterface : public internal::ServerStreamingInterface,
+                                    public internal::WriterInterface<W>,
+                                    public internal::ReaderInterface<R> {};
 
 /// Actual implementation of bi-directional streaming
 namespace internal {
@@ -688,6 +761,7 @@ class ServerReaderWriterBody final {
   Call* const call_;
   ServerContext* const ctx_;
 };
+
 }  // namespace internal
 
 /// Synchronous (blocking) server-side API for a bidirectional
@@ -697,8 +771,6 @@ class ServerReaderWriterBody final {
 template <class W, class R>
 class ServerReaderWriter final : public ServerReaderWriterInterface<W, R> {
  public:
-  ServerReaderWriter(Call* call, ServerContext* ctx) : body_(call, ctx) {}
-
   /// See the \a ServerStreamingInterface.SendInitialMetadata method
   /// for semantics. Note that initial metadata will be affected by the
   /// \a ServerContext associated with this call.
@@ -715,13 +787,18 @@ class ServerReaderWriter final : public ServerReaderWriterInterface<W, R> {
   /// Side effect:
   ///   Also sends initial metadata if not already sent (using the \a
   ///   ServerContext associated with this call).
-  using WriterInterface<W>::Write;
+  using internal::WriterInterface<W>::Write;
   bool Write(const W& msg, WriteOptions options) override {
     return body_.Write(msg, options);
   }
 
  private:
   internal::ServerReaderWriterBody<W, R> body_;
+
+  friend class internal::TemplatedBidiStreamingHandler<ServerReaderWriter<W, R>,
+                                                       false>;
+  ServerReaderWriter(internal::Call* call, ServerContext* ctx)
+      : body_(call, ctx) {}
 };
 
 /// A class to represent a flow-controlled unary call. This is something
@@ -736,9 +813,6 @@ template <class RequestType, class ResponseType>
 class ServerUnaryStreamer final
     : public ServerReaderWriterInterface<ResponseType, RequestType> {
  public:
-  ServerUnaryStreamer(Call* call, ServerContext* ctx)
-      : body_(call, ctx), read_done_(false), write_done_(false) {}
-
   /// Block to send initial metadata to client.
   /// Implicit input parameter:
   ///    - the \a ServerContext associated with this call will be used for
@@ -775,7 +849,7 @@ class ServerUnaryStreamer final
   /// \param options The WriteOptions affecting the write operation.
   ///
   /// \return \a true on success, \a false when the stream has been closed.
-  using WriterInterface<ResponseType>::Write;
+  using internal::WriterInterface<ResponseType>::Write;
   bool Write(const ResponseType& response, WriteOptions options) override {
     if (write_done_ || !read_done_) {
       return false;
@@ -788,6 +862,11 @@ class ServerUnaryStreamer final
   internal::ServerReaderWriterBody<ResponseType, RequestType> body_;
   bool read_done_;
   bool write_done_;
+
+  friend class internal::TemplatedBidiStreamingHandler<
+      ServerUnaryStreamer<RequestType, ResponseType>, true>;
+  ServerUnaryStreamer(internal::Call* call, ServerContext* ctx)
+      : body_(call, ctx), read_done_(false), write_done_(false) {}
 };
 
 /// A class to represent a flow-controlled server-side streaming call.
@@ -799,9 +878,6 @@ template <class RequestType, class ResponseType>
 class ServerSplitStreamer final
     : public ServerReaderWriterInterface<ResponseType, RequestType> {
  public:
-  ServerSplitStreamer(Call* call, ServerContext* ctx)
-      : body_(call, ctx), read_done_(false) {}
-
   /// Block to send initial metadata to client.
   /// Implicit input parameter:
   ///    - the \a ServerContext associated with this call will be used for
@@ -838,7 +914,7 @@ class ServerSplitStreamer final
   /// \param options The WriteOptions affecting the write operation.
   ///
   /// \return \a true on success, \a false when the stream has been closed.
-  using WriterInterface<ResponseType>::Write;
+  using internal::WriterInterface<ResponseType>::Write;
   bool Write(const ResponseType& response, WriteOptions options) override {
     return read_done_ && body_.Write(response, options);
   }
@@ -846,6 +922,11 @@ class ServerSplitStreamer final
  private:
   internal::ServerReaderWriterBody<ResponseType, RequestType> body_;
   bool read_done_;
+
+  friend class internal::TemplatedBidiStreamingHandler<
+      ServerSplitStreamer<RequestType, ResponseType>, false>;
+  ServerSplitStreamer(internal::Call* call, ServerContext* ctx)
+      : body_(call, ctx), read_done_(false) {}
 };
 
 }  // namespace grpc
diff --git a/include/grpc++/impl/codegen/time.h b/include/grpc++/impl/codegen/time.h
index 589deb4f03..d464d6ea13 100644
--- a/include/grpc++/impl/codegen/time.h
+++ b/include/grpc++/impl/codegen/time.h
@@ -19,6 +19,8 @@
 #ifndef GRPCXX_IMPL_CODEGEN_TIME_H
 #define GRPCXX_IMPL_CODEGEN_TIME_H
 
+#include <chrono>
+
 #include <grpc++/impl/codegen/config.h>
 #include <grpc/impl/codegen/grpc_types.h>
 
@@ -59,10 +61,6 @@ class TimePoint<gpr_timespec> {
 
 }  // namespace grpc
 
-#include <chrono>
-
-#include <grpc/impl/codegen/grpc_types.h>
-
 namespace grpc {
 
 // from and to should be absolute time.
diff --git a/include/grpc++/server.h b/include/grpc++/server.h
index 0a3aae8241..01c4a60d21 100644
--- a/include/grpc++/server.h
+++ b/include/grpc++/server.h
@@ -175,7 +175,8 @@ class Server final : public ServerInterface, private GrpcLibraryCodegen {
   /// \param num_cqs How many completion queues does \a cqs hold.
   void Start(ServerCompletionQueue** cqs, size_t num_cqs) override;
 
-  void PerformOpsOnCall(CallOpSetInterface* ops, Call* call) override;
+  void PerformOpsOnCall(internal::CallOpSetInterface* ops,
+                        internal::Call* call) override;
 
   void ShutdownInternal(gpr_timespec deadline) override;
 
diff --git a/include/grpc++/server_builder.h b/include/grpc++/server_builder.h
index bbf45b3e74..0888bef0d9 100644
--- a/include/grpc++/server_builder.h
+++ b/include/grpc++/server_builder.h
@@ -40,7 +40,6 @@ namespace grpc {
 class AsyncGenericService;
 class ResourceQuota;
 class CompletionQueue;
-class RpcService;
 class Server;
 class ServerCompletionQueue;
 class ServerCredentials;
@@ -140,7 +139,7 @@ class ServerBuilder {
   /// please use IPv6 any, i.e., [::]:<port>, which also accepts IPv4
   /// connections.  Valid values include dns:///localhost:1234, /
   /// 192.168.1.1:31416, dns:///[::1]:27182, etc.).
-  /// \params creds The credentials associated with the server.
+  /// \param creds The credentials associated with the server.
   /// \param selected_port[out] If not `nullptr`, gets populated with the port
   /// number bound to the \a grpc::Server for the corresponding endpoint after
   /// it is successfully bound, 0 otherwise.
diff --git a/include/grpc/census.h b/include/grpc/census.h
index de8e7a6617..2258af8898 100644
--- a/include/grpc/census.h
+++ b/include/grpc/census.h
@@ -16,10 +16,6 @@
  *
  */
 
-/** RPC-internal Census API's. These are designed to be generic enough that
- * they can (ultimately) be used in many different RPC systems (with differing
- * implementations). */
-
 #ifndef GRPC_CENSUS_H
 #define GRPC_CENSUS_H
 
@@ -29,439 +25,12 @@
 extern "C" {
 #endif
 
-/** Identify census features that can be enabled via census_initialize(). */
-enum census_features {
-  CENSUS_FEATURE_NONE = 0,    /** Do not enable census. */
-  CENSUS_FEATURE_TRACING = 1, /** Enable census tracing. */
-  CENSUS_FEATURE_STATS = 2,   /** Enable Census stats collection. */
-  CENSUS_FEATURE_CPU = 4,     /** Enable Census CPU usage collection. */
-  CENSUS_FEATURE_ALL =
-      CENSUS_FEATURE_TRACING | CENSUS_FEATURE_STATS | CENSUS_FEATURE_CPU
-};
-
-/** Shutdown and startup census subsystem. The 'features' argument should be
- * the OR (|) of census_features values. If census fails to initialize, then
- * census_initialize() will return -1, otherwise the set of enabled features
- * (which may be smaller than that provided in the `features` argument, see
- * census_supported()) is returned. It is an error to call census_initialize()
- * more than once (without an intervening census_shutdown()). These functions
- * are not thread-safe. */
-CENSUSAPI int census_initialize(int features);
-CENSUSAPI void census_shutdown(void);
-
-/** Return the features supported by the current census implementation (not all
- * features will be available on all platforms). */
-CENSUSAPI int census_supported(void);
-
-/** Return the census features currently enabled. */
-CENSUSAPI int census_enabled(void);
-
 /**
   A Census Context is a handle used by Census to represent the current tracing
   and stats collection information. Contexts should be propagated across RPC's
-  (this is the responsibility of the local RPC system). A context is typically
-  used as the first argument to most census functions. Conceptually, they
-  should be thought of as specific to a single RPC/thread. The user visible
-  context representation is that of a collection of key:value string pairs,
-  each of which is termed a 'tag'; these form the basis against which Census
-  metrics will be recorded. Keys are unique within a context. */
+  (this is the responsibility of the local RPC system). */
 typedef struct census_context census_context;
 
-/** A tag is a key:value pair. Both keys and values are nil-terminated strings,
-   containing printable ASCII characters (decimal 32-126). Keys must be at
-   least one character in length. Both keys and values can have at most
-   CENSUS_MAX_TAG_KB_LEN characters (including the terminating nil). The
-   maximum number of tags that can be propagated is
-   CENSUS_MAX_PROPAGATED_TAGS. Users should also remember that some systems
-   may have limits on, e.g., the number of bytes that can be transmitted as
-   metadata, and that larger tags means more memory consumed and time in
-   processing. */
-typedef struct {
-  const char *key;
-  const char *value;
-  uint8_t flags;
-} census_tag;
-
-/** Maximum length of a tag's key or value. */
-#define CENSUS_MAX_TAG_KV_LEN 255
-/** Maximum number of propagatable tags. */
-#define CENSUS_MAX_PROPAGATED_TAGS 255
-
-/** Tag flags. */
-#define CENSUS_TAG_PROPAGATE 1 /** Tag should be propagated over RPC */
-#define CENSUS_TAG_STATS 2    /** Tag will be used for statistics aggregation */
-#define CENSUS_TAG_RESERVED 4 /** Reserved for internal use. */
-/** Flag values 4,8,16,32,64,128 are reserved for future/internal use. Clients
-   should not use or rely on their values. */
-
-#define CENSUS_TAG_IS_PROPAGATED(flags) (flags & CENSUS_TAG_PROPAGATE)
-#define CENSUS_TAG_IS_STATS(flags) (flags & CENSUS_TAG_STATS)
-
-/** An instance of this structure is kept by every context, and records the
-   basic information associated with the creation of that context. */
-typedef struct {
-  int n_propagated_tags; /** number of propagated tags */
-  int n_local_tags;      /** number of non-propagated (local) tags */
-  int n_deleted_tags;    /** number of tags that were deleted */
-  int n_added_tags;      /** number of tags that were added */
-  int n_modified_tags;   /** number of tags that were modified */
-  int n_invalid_tags;    /** number of tags with bad keys or values (e.g.
-                            longer than CENSUS_MAX_TAG_KV_LEN) */
-  int n_ignored_tags;    /** number of tags ignored because of
-                            CENSUS_MAX_PROPAGATED_TAGS limit. */
-} census_context_status;
-
-/** Create a new context, adding and removing tags from an existing context.
-   This will copy all tags from the 'tags' input, so it is recommended
-   to add as many tags in a single operation as is practical for the client.
-   @param base Base context to build upon. Can be NULL.
-   @param tags A set of tags to be added/changed/deleted. Tags with keys that
-   are in 'tags', but not 'base', are added to the context. Keys that are in
-   both 'tags' and 'base' will have their value/flags modified. Tags with keys
-   in both, but with NULL values, will be deleted from the context. Tags with
-   invalid (too long or short) keys or values will be ignored.
-   If adding a tag will result in more than CENSUS_MAX_PROPAGATED_TAGS in either
-   binary or non-binary tags, they will be ignored, as will deletions of
-   tags that don't exist.
-   @param ntags number of tags in 'tags'
-   @param status If not NULL, will return a pointer to a census_context_status
-   structure containing information about the new context and status of the
-   tags used in its creation.
-   @return A new, valid census_context.
-*/
-CENSUSAPI census_context *census_context_create(
-    const census_context *base, const census_tag *tags, int ntags,
-    census_context_status const **status);
-
-/** Destroy a context. Once this function has been called, the context cannot
-   be reused. */
-CENSUSAPI void census_context_destroy(census_context *context);
-
-/** Get a pointer to the original status from the context creation. */
-CENSUSAPI const census_context_status *census_context_get_status(
-    const census_context *context);
-
-/** Structure used for iterating over the tags in a context. API clients should
-   not use or reference internal fields - neither their contents or
-   presence/absence are guaranteed. */
-typedef struct {
-  const census_context *context;
-  int base;
-  int index;
-  char *kvm;
-} census_context_iterator;
-
-/** Initialize a census_tag_iterator. Must be called before first use. */
-CENSUSAPI void census_context_initialize_iterator(
-    const census_context *context, census_context_iterator *iterator);
-
-/** Get the contents of the "next" tag in the context. If there are no more
-   tags, returns 0 (and 'tag' contents will be unchanged), otherwise returns 1.
-   */
-CENSUSAPI int census_context_next_tag(census_context_iterator *iterator,
-                                      census_tag *tag);
-
-/** Get a context tag by key. Returns 0 if the key is not present. */
-CENSUSAPI int census_context_get_tag(const census_context *context,
-                                     const char *key, census_tag *tag);
-
-/** Tag set encode/decode functionality. These functions are intended
-   for use by RPC systems only, for purposes of transmitting/receiving contexts.
-   */
-
-/** Encode a context into a buffer.
-   @param context context to be encoded
-   @param buffer buffer into which the context will be encoded.
-   @param buf_size number of available bytes in buffer.
-   @return The number of buffer bytes consumed for the encoded context, or
-           zero if the buffer was of insufficient size. */
-CENSUSAPI size_t census_context_encode(const census_context *context,
-                                       char *buffer, size_t buf_size);
-
-/** Decode context buffer encoded with census_context_encode(). Returns NULL
-   if there is an error in parsing either buffer. */
-CENSUSAPI census_context *census_context_decode(const char *buffer,
-                                                size_t size);
-
-/** Distributed traces can have a number of options. */
-enum census_trace_mask_values {
-  CENSUS_TRACE_MASK_NONE = 0,      /** Default, empty flags */
-  CENSUS_TRACE_MASK_IS_SAMPLED = 1 /** RPC tracing enabled for this context. */
-};
-
-/** Get the current trace mask associated with this context. The value returned
-    will be the logical OR of census_trace_mask_values values. */
-CENSUSAPI int census_trace_mask(const census_context *context);
-
-/** Set the trace mask associated with a context. */
-CENSUSAPI void census_set_trace_mask(int trace_mask);
-
-/** The concept of "operation" is a fundamental concept for Census. In an RPC
-   system, an operation typically represents a single RPC, or a significant
-   sub-part thereof (e.g. a single logical "read" RPC to a distributed storage
-   system might do several other actions in parallel, from looking up metadata
-   indices to making requests of other services - each of these could be a
-   sub-operation with the larger RPC operation). Census uses operations for the
-   following:
-
-   CPU accounting: If enabled, census will measure the thread CPU time
-   consumed between operation start and end times.
-
-   Active operations: Census will maintain information on all currently
-   active operations.
-
-   Distributed tracing: Each operation serves as a logical trace span.
-
-   Stats collection: Stats are broken down by operation (e.g. latency
-   breakdown for each unique RPC path).
-
-   The following functions serve to delineate the start and stop points for
-   each logical operation. */
-
-/**
-  This structure represents a timestamp as used by census to record the time
-  at which an operation begins.
-*/
-typedef struct {
-  /** Use gpr_timespec for default implementation. High performance
-   * implementations should use a cycle-counter based timestamp. */
-  gpr_timespec ts;
-} census_timestamp;
-
-/**
-  Mark the beginning of an RPC operation. The information required to call the
-  functions to record the start of RPC operations (both client and server) may
-  not be callable at the true start time of the operation, due to information
-  not being available (e.g. the census context data will not be available in a
-  server RPC until at least initial metadata has been processed). To ensure
-  correct CPU accounting and latency recording, RPC systems can call this
-  function to get the timestamp of operation beginning. This can later be used
-  as an argument to census_start_{client,server}_rpc_op(). NB: for correct
-  CPU accounting, the system must guarantee that the same thread is used
-  for all request processing after this function is called.
-
-  @return A timestamp representing the operation start time.
-*/
-CENSUSAPI census_timestamp census_start_rpc_op_timestamp(void);
-
-/**
-  Represent functions to map RPC name ID to service/method names. Census
-  breaks down all RPC stats by service and method names. We leave the
-  definition and format of these to the RPC system. For efficiency purposes,
-  we encode these as a single 64 bit identifier, and allow the RPC system to
-  provide a structure for functions that can convert these to service and
-  method strings.
-
-  TODO(aveitch): Instead of providing this as an argument to the rpc_start_op()
-  functions, maybe it should be set once at census initialization.
-*/
-typedef struct {
-  const char *(*get_rpc_service_name)(int64_t id);
-  const char *(*get_rpc_method_name)(int64_t id);
-} census_rpc_name_info;
-
-/**
-   Start a client rpc operation. This function should be called as early in the
-   client RPC path as possible. This function will create a new context. If
-   the context argument is non-null, then the new context will inherit all
-   its properties, with the following changes:
-   - create a new operation ID for the new context, marking it as a child of
-     the previous operation.
-   - use the new RPC path and peer information for tracing and stats
-     collection purposes, rather than those from the original context
-
-   If the context argument is NULL, then a new root context is created. This
-   is particularly important for tracing purposes (the trace spans generated
-   will be unassociated with any other trace spans, except those
-   downstream). The trace_mask will be used for tracing operations associated
-   with the new context.
-
-   In some RPC systems (e.g. where load balancing is used), peer information
-   may not be available at the time the operation starts. In this case, use a
-   NULL value for peer, and set it later using the
-   census_set_rpc_client_peer() function.
-
-   @param context The parent context. Can be NULL.
-   @param rpc_name_id The rpc name identifier to be associated with this RPC.
-   @param rpc_name_info Used to decode rpc_name_id.
-   @param peer RPC peer. If not available at the time, NULL can be used,
-               and a later census_set_rpc_client_peer() call made.
-   @param trace_mask An OR of census_trace_mask_values values. Only used in
-                     the creation of a new root context (context == NULL).
-   @param start_time A timestamp returned from census_start_rpc_op_timestamp().
-                     Can be NULL. Used to set the true time the operation
-                     begins.
-
-   @return A new census context.
- */
-CENSUSAPI census_context *census_start_client_rpc_op(
-    const census_context *context, int64_t rpc_name_id,
-    const census_rpc_name_info *rpc_name_info, const char *peer, int trace_mask,
-    const census_timestamp *start_time);
-
-/**
-  Add peer information to a context representing a client RPC operation.
-*/
-CENSUSAPI void census_set_rpc_client_peer(census_context *context,
-                                          const char *peer);
-
-/**
-   Start a server RPC operation. Returns a new context to be used in future
-   census calls. If buffer is non-NULL, then the buffer contents should
-   represent the client context, as generated by census_context_serialize().
-   If buffer is NULL, a new root context is created.
-
-   @param buffer Buffer containing bytes output from census_context_serialize().
-   @param rpc_name_id The rpc name identifier to be associated with this RPC.
-   @param rpc_name_info Used to decode rpc_name_id.
-   @param peer RPC peer.
-   @param trace_mask An OR of census_trace_mask_values values. Only used in
-                     the creation of a new root context (buffer == NULL).
-   @param start_time A timestamp returned from census_start_rpc_op_timestamp().
-                     Can be NULL. Used to set the true time the operation
-                     begins.
-
-   @return A new census context.
- */
-CENSUSAPI census_context *census_start_server_rpc_op(
-    const char *buffer, int64_t rpc_name_id,
-    const census_rpc_name_info *rpc_name_info, const char *peer, int trace_mask,
-    census_timestamp *start_time);
-
-/**
-   Start a new, non-RPC operation. In general, this function works very
-   similarly to census_start_client_rpc_op, with the primary difference being
-   the replacement of host/path information with the more generic family/name
-   tags. If the context argument is non-null, then the new context will
-   inherit all its properties, with the following changes:
-   - create a new operation ID for the new context, marking it as a child of
-     the previous operation.
-   - use the family and name information for tracing and stats collection
-     purposes, rather than those from the original context
-
-   If the context argument is NULL, then a new root context is created. This
-   is particularly important for tracing purposes (the trace spans generated
-   will be unassociated with any other trace spans, except those
-   downstream). The trace_mask will be used for tracing
-   operations associated with the new context.
-
-   @param context The base context. Can be NULL.
-   @param family Family name to associate with the trace
-   @param name Name within family to associate with traces/stats
-   @param trace_mask An OR of census_trace_mask_values values. Only used if
-                     context is NULL.
-
-   @return A new census context.
- */
-CENSUSAPI census_context *census_start_op(census_context *context,
-                                          const char *family, const char *name,
-                                          int trace_mask);
-
-/**
-  End an operation started by any of the census_start_*_op*() calls. The
-  context used in this call will no longer be valid once this function
-  completes.
-
-  @param context Context associated with operation which is ending.
-  @param status status associated with the operation. Not interpreted by
-                census.
-*/
-CENSUSAPI void census_end_op(census_context *context, int status);
-
-#define CENSUS_TRACE_RECORD_START_OP ((uint32_t)0)
-#define CENSUS_TRACE_RECORD_END_OP ((uint32_t)1)
-
-/** Insert a trace record into the trace stream. The record consists of an
-    arbitrary size buffer, the size of which is provided in 'n'.
-    @param context Trace context
-    @param type User-defined type to associate with trace entry.
-    @param buffer Pointer to buffer to use
-    @param n Number of bytes in buffer
-*/
-CENSUSAPI void census_trace_print(census_context *context, uint32_t type,
-                                  const char *buffer, size_t n);
-
-/** Trace record. */
-typedef struct {
-  census_timestamp timestamp; /** Time of record creation */
-  uint64_t trace_id;          /** Trace ID associated with record */
-  uint64_t op_id;             /** Operation ID associated with record */
-  uint32_t type;              /** Type (as used in census_trace_print() */
-  const char *buffer;         /** Buffer (from census_trace_print() */
-  size_t buf_size;            /** Number of bytes inside buffer */
-} census_trace_record;
-
-/** Start a scan of existing trace records. While a scan is ongoing, addition
-    of new trace records will be blocked if the underlying trace buffers
-    fill up, so trace processing systems should endeavor to complete
-    reading as soon as possible.
-  @param consume if non-zero, indicates that reading records also "consumes"
-         the previously read record - i.e. releases space in the trace log
-         while scanning is ongoing.
-  @returns 0 on success, non-zero on failure (e.g. if a scan is already ongoing)
-*/
-CENSUSAPI int census_trace_scan_start(int consume);
-
-/** Get a trace record. The data pointed to by the trace buffer is guaranteed
-    stable until the next census_get_trace_record() call (if the consume
-    argument to census_trace_scan_start was non-zero) or census_trace_scan_end()
-    is called (otherwise).
-  @param trace_record structure that will be filled in with oldest trace record.
-  @returns -1 if an error occurred (e.g. no previous call to
-           census_trace_scan_start()), 0 if there is no more trace data (and
-           trace_record will not be modified) or 1 otherwise.
-*/
-CENSUSAPI int census_get_trace_record(census_trace_record *trace_record);
-
-/** End a scan previously started by census_trace_scan_start() */
-CENSUSAPI void census_trace_scan_end();
-
-/** Core stats collection API's. The following concepts are used:
-   * Resource: Users record measurements for a single resource. Examples
-     include RPC latency, CPU seconds consumed, and bytes transmitted.
-   * Aggregation: An aggregation of a set of measurements. Census supports the
-       following aggregation types:
-       * Distribution - statistical distribution information, used for
-         recording average, standard deviation etc. Can include a histogram.
-       * Interval - a count of events that happen in a rolling time window.
-   * View: A view is a combination of a Resource, a set of tag keys and an
-     Aggregation. When a measurement for a Resource matches the View tags, it is
-     recorded (for each unique set of tag values) using the Aggregation type.
-     Each resource can have an arbitrary number of views by which it will be
-     broken down.
-
-  Census uses protos to define each of the above, and output results. This
-  ensures unification across the different language and runtime
-  implementations. The proto definitions can be found in src/proto/census.
-*/
-
-/** Define a new resource. `resource_pb` should contain an encoded Resource
-   protobuf, `resource_pb_size` being the size of the buffer. Returns a -ve
-   value on error, or a positive (>= 0) resource id (for use in
-   census_delete_resource() and census_record_values()). In order to be valid, a
-   resource must have a name, and at least one numerator in its unit type. The
-   resource name must be unique, and an error will be returned if it is not. */
-CENSUSAPI int32_t census_define_resource(const uint8_t *resource_pb,
-                                         size_t resource_pb_size);
-
-/** Delete a resource created by census_define_resource(). */
-CENSUSAPI void census_delete_resource(int32_t resource_id);
-
-/** Determine the id of a resource, given its name. returns -1 if the resource
-   does not exist. */
-CENSUSAPI int32_t census_resource_id(const char *name);
-
-/** A single value to be recorded comprises two parts: an ID for the particular
- * resource and the value to be recorded against it. */
-typedef struct {
-  int32_t resource_id;
-  double value;
-} census_value;
-
-/** Record new usage values against the given context. */
-CENSUSAPI void census_record_values(census_context *context,
-                                    census_value *values, size_t nvalues);
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/include/grpc/grpc.h b/include/grpc/grpc.h
index 1de289fba4..6df3b8086e 100644
--- a/include/grpc/grpc.h
+++ b/include/grpc/grpc.h
@@ -143,6 +143,23 @@ GRPCAPI void grpc_completion_queue_shutdown(grpc_completion_queue *cq);
     drained and no threads are executing grpc_completion_queue_next */
 GRPCAPI void grpc_completion_queue_destroy(grpc_completion_queue *cq);
 
+/*********** EXPERIMENTAL API ************/
+/** Initializes a thread local cache for \a cq.
+  * grpc_flush_cq_tls_cache() MUST be called on the same thread,
+  * with the same cq.
+  */
+GRPCAPI void grpc_completion_queue_thread_local_cache_init(
+    grpc_completion_queue *cq);
+
+/*********** EXPERIMENTAL API ************/
+/** Flushes the thread local cache for \a cq.
+  * Returns 1 if there was contents in the cache.  If there was an event
+  * in \a cq tls cache, its tag is placed in tag, and ok is set to the
+  * event success.
+  */
+GRPCAPI int grpc_completion_queue_thread_local_cache_flush(
+    grpc_completion_queue *cq, void **tag, int *ok);
+
 /** Create a completion queue alarm instance */
 GRPCAPI grpc_alarm *grpc_alarm_create(void *reserved);
 
diff --git a/include/grpc/grpc_security.h b/include/grpc/grpc_security.h
index 95b1447935..758aaf5b6c 100644
--- a/include/grpc/grpc_security.h
+++ b/include/grpc/grpc_security.h
@@ -316,6 +316,43 @@ typedef struct grpc_server_credentials grpc_server_credentials;
    */
 GRPCAPI void grpc_server_credentials_release(grpc_server_credentials *creds);
 
+/** Server certificate config object holds the server's public certificates and
+   associated private keys, as well as any CA certificates needed for client
+   certificate validation (if applicable). Create using
+   grpc_ssl_server_certificate_config_create(). */
+typedef struct grpc_ssl_server_certificate_config
+    grpc_ssl_server_certificate_config;
+
+/** Creates a grpc_ssl_server_certificate_config object.
+   - pem_roots_cert is the NULL-terminated string containing the PEM encoding of
+     the client root certificates. This parameter may be NULL if the server does
+     not want the client to be authenticated with SSL.
+   - pem_key_cert_pairs is an array private key / certificate chains of the
+     server. This parameter cannot be NULL.
+   - num_key_cert_pairs indicates the number of items in the private_key_files
+     and cert_chain_files parameters. It must be at least 1.
+   - It is the caller's responsibility to free this object via
+     grpc_ssl_server_certificate_config_destroy(). */
+GRPCAPI grpc_ssl_server_certificate_config *
+grpc_ssl_server_certificate_config_create(
+    const char *pem_root_certs,
+    const grpc_ssl_pem_key_cert_pair *pem_key_cert_pairs,
+    size_t num_key_cert_pairs);
+
+/** Destroys a grpc_ssl_server_certificate_config object. */
+GRPCAPI void grpc_ssl_server_certificate_config_destroy(
+    grpc_ssl_server_certificate_config *config);
+
+/** Callback to retrieve updated SSL server certificates, private keys, and
+   trusted CAs (for client authentication).
+    - user_data parameter, if not NULL, contains opaque data to be used by the
+      callback.
+    - Use grpc_ssl_server_certificate_config_create to create the config.
+    - The caller assumes ownership of the config. */
+typedef grpc_ssl_certificate_config_reload_status (
+    *grpc_ssl_server_certificate_config_callback)(
+    void *user_data, grpc_ssl_server_certificate_config **config);
+
 /** Deprecated in favor of grpc_ssl_server_credentials_create_ex.
    Creates an SSL server_credentials object.
    - pem_roots_cert is the NULL-terminated string containing the PEM encoding of
@@ -332,7 +369,8 @@ GRPCAPI grpc_server_credentials *grpc_ssl_server_credentials_create(
     const char *pem_root_certs, grpc_ssl_pem_key_cert_pair *pem_key_cert_pairs,
     size_t num_key_cert_pairs, int force_client_auth, void *reserved);
 
-/** Same as grpc_ssl_server_credentials_create method except uses
+/** Deprecated in favor of grpc_ssl_server_credentials_create_with_options.
+   Same as grpc_ssl_server_credentials_create method except uses
    grpc_ssl_client_certificate_request_type enum to support more ways to
    authenticate client cerificates.*/
 GRPCAPI grpc_server_credentials *grpc_ssl_server_credentials_create_ex(
@@ -341,6 +379,40 @@ GRPCAPI grpc_server_credentials *grpc_ssl_server_credentials_create_ex(
     grpc_ssl_client_certificate_request_type client_certificate_request,
     void *reserved);
 
+typedef struct grpc_ssl_server_credentials_options
+    grpc_ssl_server_credentials_options;
+
+/** Creates an options object using a certificate config. Use this method when
+   the certificates and keys of the SSL server will not change during the
+   server's lifetime.
+   - Takes ownership of the certificate_config parameter. */
+GRPCAPI grpc_ssl_server_credentials_options *
+grpc_ssl_server_credentials_create_options_using_config(
+    grpc_ssl_client_certificate_request_type client_certificate_request,
+    grpc_ssl_server_certificate_config *certificate_config);
+
+/** Creates an options object using a certificate config fetcher. Use this
+   method to reload the certificates and keys of the SSL server without
+   interrupting the operation of the server. Initial certificate config will be
+   fetched during server initialization.
+   - user_data parameter, if not NULL, contains opaque data which will be passed
+     to the fetcher (see definition of
+     grpc_ssl_server_certificate_config_callback). */
+GRPCAPI grpc_ssl_server_credentials_options *
+grpc_ssl_server_credentials_create_options_using_config_fetcher(
+    grpc_ssl_client_certificate_request_type client_certificate_request,
+    grpc_ssl_server_certificate_config_callback cb, void *user_data);
+
+/** Destroys a grpc_ssl_server_credentials_options object. */
+GRPCAPI void grpc_ssl_server_credentials_options_destroy(
+    grpc_ssl_server_credentials_options *options);
+
+/** Creates an SSL server_credentials object using the provided options struct.
+    - Takes ownership of the options parameter. */
+GRPCAPI grpc_server_credentials *
+grpc_ssl_server_credentials_create_with_options(
+    grpc_ssl_server_credentials_options *options);
+
 /** --- Server-side secure ports. --- */
 
 /** Add a HTTP2 over an encrypted link over tcp listener.
diff --git a/include/grpc/grpc_security_constants.h b/include/grpc/grpc_security_constants.h
index fde300dfb1..60e167eb88 100644
--- a/include/grpc/grpc_security_constants.h
+++ b/include/grpc/grpc_security_constants.h
@@ -48,6 +48,13 @@ typedef enum {
   GRPC_SSL_ROOTS_OVERRIDE_FAIL
 } grpc_ssl_roots_override_result;
 
+/** Callback results for dynamically loading a SSL certificate config. */
+typedef enum {
+  GRPC_SSL_CERTIFICATE_CONFIG_RELOAD_UNCHANGED,
+  GRPC_SSL_CERTIFICATE_CONFIG_RELOAD_NEW,
+  GRPC_SSL_CERTIFICATE_CONFIG_RELOAD_FAIL
+} grpc_ssl_certificate_config_reload_status;
+
 typedef enum {
   /** Server does not request client certificate. A client can present a self
      signed or signed certificates if it wishes to do so and they would be
diff --git a/include/grpc/impl/codegen/connectivity_state.h b/include/grpc/impl/codegen/connectivity_state.h
index 545b4fdbcc..b70dbef356 100644
--- a/include/grpc/impl/codegen/connectivity_state.h
+++ b/include/grpc/impl/codegen/connectivity_state.h
@@ -25,8 +25,6 @@ extern "C" {
 
 /** Connectivity state of a channel. */
 typedef enum {
-  /** channel has just been initialized */
-  GRPC_CHANNEL_INIT = -1,
   /** channel is idle */
   GRPC_CHANNEL_IDLE,
   /** channel is connecting */
diff --git a/include/grpc/impl/codegen/port_platform.h b/include/grpc/impl/codegen/port_platform.h
index baea4bcf52..fb4bfc3162 100644
--- a/include/grpc/impl/codegen/port_platform.h
+++ b/include/grpc/impl/codegen/port_platform.h
@@ -241,6 +241,29 @@
 #else /* _LP64 */
 #define GPR_ARCH_32 1
 #endif /* _LP64 */
+#elif defined(__OpenBSD__)
+#define GPR_PLATFORM_STRING "openbsd"
+#ifndef _BSD_SOURCE
+#define _BSD_SOURCE
+#endif
+#define GPR_OPENBSD 1
+#define GPR_CPU_POSIX 1
+#define GPR_GCC_ATOMIC 1
+#define GPR_GCC_TLS 1
+#define GPR_POSIX_LOG 1
+#define GPR_POSIX_ENV 1
+#define GPR_POSIX_TMPFILE 1
+#define GPR_POSIX_STRING 1
+#define GPR_POSIX_SUBPROCESS 1
+#define GPR_POSIX_SYNC 1
+#define GPR_POSIX_TIME 1
+#define GPR_GETPID_IN_UNISTD_H 1
+#define GPR_SUPPORT_CHANNELS_FROM_FD 1
+#ifdef _LP64
+#define GPR_ARCH_64 1
+#else /* _LP64 */
+#define GPR_ARCH_32 1
+#endif /* _LP64 */
 #elif defined(__native_client__)
 #define GPR_PLATFORM_STRING "nacl"
 #ifndef _BSD_SOURCE
diff --git a/include/grpc/module.modulemap b/include/grpc/module.modulemap
index 1ac1804084..65de47f938 100644
--- a/include/grpc/module.modulemap
+++ b/include/grpc/module.modulemap
@@ -4,7 +4,6 @@ framework module grpc {
 
   header "support/alloc.h"
   header "support/atm.h"
-  header "support/atm_gcc_atomic.h"
   header "support/avl.h"
   header "support/cmdline.h"
   header "support/cpu.h"
@@ -16,23 +15,17 @@ framework module grpc {
   header "support/string_util.h"
   header "support/subprocess.h"
   header "support/sync.h"
-  header "support/sync_custom.h"
   header "support/sync_generic.h"
-  header "support/sync_posix.h"
   header "support/thd.h"
   header "support/time.h"
   header "support/tls.h"
-  header "support/tls_pthread.h"
   header "support/useful.h"
   header "impl/codegen/atm.h"
-  header "impl/codegen/atm_gcc_atomic.h"
   header "impl/codegen/gpr_slice.h"
   header "impl/codegen/gpr_types.h"
   header "impl/codegen/port_platform.h"
   header "impl/codegen/sync.h"
-  header "impl/codegen/sync_custom.h"
   header "impl/codegen/sync_generic.h"
-  header "impl/codegen/sync_posix.h"
   header "impl/codegen/byte_buffer.h"
   header "impl/codegen/byte_buffer_reader.h"
   header "impl/codegen/compression_types.h"
@@ -43,14 +36,11 @@ framework module grpc {
   header "impl/codegen/slice.h"
   header "impl/codegen/status.h"
   header "impl/codegen/atm.h"
-  header "impl/codegen/atm_gcc_atomic.h"
   header "impl/codegen/gpr_slice.h"
   header "impl/codegen/gpr_types.h"
   header "impl/codegen/port_platform.h"
   header "impl/codegen/sync.h"
-  header "impl/codegen/sync_custom.h"
   header "impl/codegen/sync_generic.h"
-  header "impl/codegen/sync_posix.h"
   header "grpc_security.h"
   header "byte_buffer.h"
   header "byte_buffer_reader.h"
@@ -66,6 +56,28 @@ framework module grpc {
   header "support/workaround_list.h"
   header "census.h"
 
+  textual header "support/atm_gcc_atomic.h"
+  textual header "support/atm_gcc_sync.h"
+  textual header "support/atm_windows.h"
+  textual header "support/sync_custom.h"
+  textual header "support/sync_posix.h"
+  textual header "support/sync_windows.h"
+  textual header "support/tls_gcc.h"
+  textual header "support/tls_msvc.h"
+  textual header "support/tls_pthread.h"
+  textual header "impl/codegen/atm_gcc_atomic.h"
+  textual header "impl/codegen/atm_gcc_sync.h"
+  textual header "impl/codegen/atm_windows.h"
+  textual header "impl/codegen/sync_custom.h"
+  textual header "impl/codegen/sync_posix.h"
+  textual header "impl/codegen/sync_windows.h"
+  textual header "impl/codegen/atm_gcc_atomic.h"
+  textual header "impl/codegen/atm_gcc_sync.h"
+  textual header "impl/codegen/atm_windows.h"
+  textual header "impl/codegen/sync_custom.h"
+  textual header "impl/codegen/sync_posix.h"
+  textual header "impl/codegen/sync_windows.h"
+
   export *
   module * { export * }
 }
diff --git a/include/grpc/slice.h b/include/grpc/slice.h
index 3f3cff1408..2203e3887b 100644
--- a/include/grpc/slice.h
+++ b/include/grpc/slice.h
@@ -137,7 +137,6 @@ GPRAPI int grpc_slice_eq(grpc_slice a, grpc_slice b);
    versions of the API. */
 GPRAPI int grpc_slice_cmp(grpc_slice a, grpc_slice b);
 GPRAPI int grpc_slice_str_cmp(grpc_slice a, const char *b);
-GPRAPI int grpc_slice_buf_cmp(grpc_slice a, const void *b, size_t blen);
 
 /** return non-zero if the first blen bytes of a are equal to b */
 GPRAPI int grpc_slice_buf_start_eq(grpc_slice a, const void *b, size_t blen);
diff --git a/include/grpc/support/sync.h b/include/grpc/support/sync.h
index fe8a59a5d6..ddb85808c7 100644
--- a/include/grpc/support/sync.h
+++ b/include/grpc/support/sync.h
@@ -274,7 +274,23 @@ GPRAPI intptr_t gpr_stats_read(const gpr_stats_counter *c);
 #endif /* 0 */
 
 #ifdef __cplusplus
-}
+}  // extern "C"
+
+namespace grpc_core {
+
+class mu_guard {
+ public:
+  mu_guard(gpr_mu *mu) : mu_(mu) { gpr_mu_lock(mu); }
+  ~mu_guard() { gpr_mu_unlock(mu_); }
+
+  mu_guard(const mu_guard &) = delete;
+  mu_guard &operator=(const mu_guard &) = delete;
+
+ private:
+  gpr_mu *const mu_;
+};
+
+}  // namespace grpc_core
 #endif
 
 #endif /* GRPC_SUPPORT_SYNC_H */