Mojo C++ bindings: introduce MultiplexRouter and related classes.

mojo/public/cpp/bindings/lib/multiplex_router.cc

Index: mojo/public/cpp/bindings/lib/multiplex_router.cc
diff --git a/mojo/public/cpp/bindings/lib/multiplex_router.cc b/mojo/public/cpp/bindings/lib/multiplex_router.cc
new file mode 100644
index 0000000000000000000000000000000000000000..75ed6cd61409d13a7ab7d835e8f7c088eea61830
--- /dev/null
+++ b/mojo/public/cpp/bindings/lib/multiplex_router.cc
@@ -0,0 +1,492 @@
+// Copyright 2015 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "mojo/public/cpp/bindings/lib/multiplex_router.h"
+
+#include "base/bind.h"
+#include "base/memory/ref_counted.h"
+#include "base/message_loop/message_loop.h"
+#include "base/stl_util.h"
+#include "mojo/public/cpp/bindings/lib/interface_endpoint_client.h"
+
+namespace mojo {
+namespace internal {
+
+// InterfaceEndpoint stores the endpoint of an interface endpoint registered
+// with

[sky, 2015/11/19 17:16:27]

nit: fix wrapping.

[yzshen1, 2015/11/19 22:00:40]

Done.

+// the router. Always accessed under the router's lock.

[sky, 2015/11/19 17:16:27]

To help debugging could this object take the lock and assert that it is
acquired?

[yzshen1, 2015/11/19 22:00:40]

Done. I added the asserts.

It adds a little cost though, because otherwise this object doesn't need to
store a pointer to the router's lock.

+//
+// On creation, this object holds two refs to itself. Those refs are released
+// when SetClosedMayDestruct() and SetPeerClosedMayDestruct() are called.
+// Other than that, the task queue in the router may hold refs to it, too.
+class MultiplexRouter::InterfaceEndpoint
+    : public base::RefCounted<InterfaceEndpoint> {
+ public:
+  InterfaceEndpoint(MultiplexRouter* router, InterfaceId id)
+      : router_(router),
+        id_(id),
+        closed_(false),
+        peer_closed_(false),
+        client_(nullptr) {
+    AddRef();
+    AddRef();

[sky, 2015/11/19 17:16:27]

I suspect a scoped_refptr to this member would be less error prone, and it gets
rid of the double AddRefs() and is better documenting what is going on.

[yzshen1, 2015/11/19 22:00:40]

Done.

+  }
+
+  InterfaceId id() const { return id_; }
+
+  bool closed() const { return closed_; }
+  void SetClosedMayDestruct() {
+    if (closed_)
+      return;
+
+    closed_ = true;
+    Release();
+  }
+
+  bool peer_closed() const { return peer_closed_; }
+  void SetPeerClosedMayDestruct() {
+    if (peer_closed_)
+      return;
+
+    peer_closed_ = true;
+    Release();
+  }
+
+  const scoped_refptr<base::SingleThreadTaskRunner> task_runner() const {
+    return task_runner_;
+  }
+  void set_task_runner(
+      scoped_refptr<base::SingleThreadTaskRunner> task_runner) {
+    task_runner_ = task_runner.Pass();
+  }
+
+  InterfaceEndpointClient* client() const { return client_; }
+  void set_client(InterfaceEndpointClient* client) { client_ = client; }
+
+ private:
+  friend class base::RefCounted<InterfaceEndpoint>;
+
+  ~InterfaceEndpoint() {
+    DCHECK(!client_);
+    router_->OnEndpointDestructed(id_);
+  }
+
+  MultiplexRouter* const router_;
+  const InterfaceId id_;
+
+  // Whether the endpoint has been closed.
+  bool closed_;
+  // Whether the peer endpoint has been closed.
+  bool peer_closed_;
+
+  // The task runner on which |client_| can be accessed.
+  scoped_refptr<base::SingleThreadTaskRunner> task_runner_;
+  // Not owned. It is null if no client is attached to this endpoint.
+  InterfaceEndpointClient* client_;
+
+  DISALLOW_COPY_AND_ASSIGN(InterfaceEndpoint);
+};
+
+struct MultiplexRouter::Task {
+ public:
+  // Doesn't take ownership of |message| but takes its contents.
+  static Task* CreateIncomingMessageTask(Message* message) {
+    Task* task = new Task();
+    task->message.reset(new Message);
+    message->MoveTo(task->message.get());
+    return task;
+  }
+  static Task* CreateNotifyErrorTask(InterfaceEndpoint* endpoint) {
+    Task* task = new Task();
+    task->endpoint_to_notify = endpoint;
+    return task;
+  }
+
+  ~Task() {}
+
+  bool IsIncomingMessageTask() const { return !!message; }
+  bool IsNotifyErrorTask() const { return !!endpoint_to_notify; }
+
+  scoped_ptr<Message> message;
+  scoped_refptr<InterfaceEndpoint> endpoint_to_notify;
+
+ private:
+  Task() {}
+};
+
+MultiplexRouter::MultiplexRouter(bool set_interface_id_namesapce_bit,
+                                 ScopedMessagePipeHandle message_pipe,
+                                 const MojoAsyncWaiter* waiter)
+    : RefCountedDeleteOnMessageLoop(base::MessageLoop::current()
+                                        ->task_runner()),
+      set_interface_id_namespace_bit_(set_interface_id_namesapce_bit),
+      header_validator_(this),
+      connector_(message_pipe.Pass(), Connector::MULTI_THREADED_SEND, waiter),
+      control_message_handler_(this),
+      control_message_proxy_(&connector_),
+      next_interface_id_value_(1),
+      testing_mode_(false) {
+  connector_.set_incoming_receiver(&header_validator_);
+  connector_.set_connection_error_handler(
+      [this]() { OnPipeConnectionError(); });
+}
+
+MultiplexRouter::~MultiplexRouter() {
+  base::AutoLock locker(lock_);
+
+  while (!tasks_.empty()) {
+    delete tasks_.front();
+    tasks_.pop_front();
+  }
+
+  auto iter = endpoints_.begin();

[sky, 2015/11/19 17:16:27]

Use a for loop (you can still increment inside the for loop), but that way iter
is scoped to the loop.

[yzshen1, 2015/11/19 22:00:40]

Done.

+  while (iter != endpoints_.end()) {
+    InterfaceEndpoint* endpoint = iter->second;
+    ++iter;

[sky, 2015/11/19 17:16:27]

This is subtle and worth a comment.

[yzshen1, 2015/11/19 22:00:40]

Done.

+
+    DCHECK(endpoint->closed());
+    endpoint->SetPeerClosedMayDestruct();
+  }
+
+  DCHECK(endpoints_.empty());
+}
+
+void MultiplexRouter::CreateEndpointHandlePair(
+    ScopedInterfaceEndpointHandle* local_endpoint,
+    ScopedInterfaceEndpointHandle* remote_endpoint) {
+  base::AutoLock locker(lock_);
+  uint32_t id = 0;
+  do {
+    if (next_interface_id_value_ >= kInterfaceIdNamespaceMask)
+      next_interface_id_value_ = 1;
+    id = next_interface_id_value_++;
+    if (set_interface_id_namespace_bit_)
+      id |= kInterfaceIdNamespaceMask;
+  } while (ContainsKey(endpoints_, id));
+
+  endpoints_[id] = new InterfaceEndpoint(this, id);
+  *local_endpoint = ScopedInterfaceEndpointHandle(id, true, this);
+  *remote_endpoint = ScopedInterfaceEndpointHandle(id, false, this);
+}
+
+ScopedInterfaceEndpointHandle MultiplexRouter::CreateLocalEndpointHandle(
+    InterfaceId id) {
+  if (!IsValidInterfaceId(id))
+    return ScopedInterfaceEndpointHandle();
+
+  base::AutoLock locker(lock_);
+  if (ContainsKey(endpoints_, id)) {
+    // If the endpoint already exist, it is because we have received a

[sky, 2015/11/19 17:16:27]

It seems weird that you have to detect this case in CreateLocalEndpointHandle,
can't it be detected before calling this?

[yzshen1, 2015/11/19 22:00:40]

This is legitimate:
this method is a public method, typically called by the generated bindings code
when it deserializes an incoming message containing an interface endpoint. The
callsite shouldn't care (nor be able to know without races) whether a control
message PeerAssociatedEndpointClosed for that interface has been received by
this router or not.

+    // notification that the peer endpoint has closed.
+    InterfaceEndpoint* endpoint = endpoints_[id];
+    CHECK(!endpoint->closed());
+    CHECK(endpoint->peer_closed());
+  } else {
+    endpoints_[id] = new InterfaceEndpoint(this, id);
+  }
+  return ScopedInterfaceEndpointHandle(id, true, this);
+}
+
+void MultiplexRouter::CloseEndpointHandle(InterfaceId id, bool is_local) {
+  if (!IsValidInterfaceId(id))
+    return;
+
+  base::AutoLock locker(lock_);
+
+  if (!is_local) {
+    DCHECK(ContainsKey(endpoints_, id));
+    DCHECK(!IsMasterInterfaceId(id));
+
+    // We will receive a NotifyPeerEndpointClosed message from the other side.
+    control_message_proxy_.NotifyEndpointClosedBeforeSent(id);
+
+    return;
+  }
+
+  DCHECK(ContainsKey(endpoints_, id));
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  DCHECK(!endpoint->client());
+  DCHECK(!endpoint->closed());
+  endpoint->SetClosedMayDestruct();
+
+  if (!IsMasterInterfaceId(id))
+    control_message_proxy_.NotifyPeerEndpointClosed(id);
+
+  ProcessTasks(true);
+}
+
+void MultiplexRouter::AttachEndpointClient(
+    const ScopedInterfaceEndpointHandle& handle,
+    InterfaceEndpointClient* client) {
+  const InterfaceId id = handle.id();
+
+  DCHECK(IsValidInterfaceId(id));
+  DCHECK(client);
+
+  base::AutoLock locker(lock_);
+  DCHECK(ContainsKey(endpoints_, id));
+
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  DCHECK(!endpoint->client());
+  DCHECK(!endpoint->closed());
+
+  endpoint->set_task_runner(base::MessageLoop::current()->task_runner());
+  endpoint->set_client(client);
+
+  if (endpoint->peer_closed())
+    tasks_.push_back(Task::CreateNotifyErrorTask(endpoint));
+  ProcessTasks(true);
+}
+
+void MultiplexRouter::DetachEndpointClient(
+    const ScopedInterfaceEndpointHandle& handle) {
+  const InterfaceId id = handle.id();
+
+  DCHECK(IsValidInterfaceId(id));
+
+  base::AutoLock locker(lock_);
+  DCHECK(ContainsKey(endpoints_, id));
+
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  DCHECK(endpoint->client());
+  DCHECK(!endpoint->closed());
+
+  endpoint->set_task_runner(nullptr);
+  endpoint->set_client(nullptr);
+}
+
+bool MultiplexRouter::SendMessage(const ScopedInterfaceEndpointHandle& handle,
+                                  Message* message) {
+  const InterfaceId id = handle.id();
+
+  base::AutoLock locker(lock_);
+  if (!ContainsKey(endpoints_, id))
+    return false;
+
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  if (endpoint->peer_closed())
+    return false;
+
+  message->set_interface_id(id);
+  return connector_.Accept(message);
+}
+
+void MultiplexRouter::RaiseError() {
+  if (task_runner_->BelongsToCurrentThread()) {
+    connector_.RaiseError();
+  } else {
+    task_runner_->PostTask(FROM_HERE,
+                           base::Bind(&MultiplexRouter::RaiseError, this));
+  }
+}
+
+ScopedMessagePipeHandle MultiplexRouter::PassMessagePipe() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  {
+    base::AutoLock locker(lock_);
+    DCHECK(endpoints_.empty() || (endpoints_.size() == 1 &&
+                                  ContainsKey(endpoints_, kMasterInterfaceId)));
+  }
+  return connector_.PassMessagePipe();
+}
+
+void MultiplexRouter::EnableTestingMode() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock locker(lock_);
+
+  testing_mode_ = true;
+  connector_.set_enforce_errors_from_incoming_receiver(false);
+}
+
+bool MultiplexRouter::Accept(Message* message) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  scoped_refptr<MultiplexRouter> protector(this);
+  base::AutoLock locker(lock_);
+  tasks_.push_back(Task::CreateIncomingMessageTask(message));
+  ProcessTasks(false);
+
+  // Always return true. If we see errors during message processing, we will
+  // explicitly call Connector::RaiseError() to disconnect the message pipe.
+  return true;
+}
+
+bool MultiplexRouter::OnPeerAssociatedEndpointClosed(InterfaceId id) {
+  lock_.AssertAcquired();
+
+  if (IsMasterInterfaceId(id))
+    return false;
+
+  if (!ContainsKey(endpoints_, id))
+    endpoints_[id] = new InterfaceEndpoint(this, id);
+
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  DCHECK(!endpoint->peer_closed());
+
+  if (endpoint->client())
+    tasks_.push_back(Task::CreateNotifyErrorTask(endpoint));
+  endpoint->SetPeerClosedMayDestruct();
+
+  // No need to trigger a ProcessTasks() because it is already on the stack.
+
+  return true;
+}
+
+bool MultiplexRouter::OnAssociatedEndpointClosedBeforeSent(InterfaceId id) {
+  lock_.AssertAcquired();
+
+  if (IsMasterInterfaceId(id))
+    return false;
+
+  if (!ContainsKey(endpoints_, id))
+    endpoints_[id] = new InterfaceEndpoint(this, id);
+
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  DCHECK(!endpoint->closed());
+  endpoint->SetClosedMayDestruct();
+
+  control_message_proxy_.NotifyPeerEndpointClosed(id);
+
+  return true;
+}
+
+void MultiplexRouter::OnPipeConnectionError() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  scoped_refptr<MultiplexRouter> protector(this);
+  base::AutoLock locker(lock_);
+
+  auto iter = endpoints_.begin();
+  while (iter != endpoints_.end()) {
+    InterfaceEndpoint* endpoint = iter->second;
+    ++iter;
+
+    if (endpoint->client())
+      tasks_.push_back(Task::CreateNotifyErrorTask(endpoint));
+
+    endpoint->SetPeerClosedMayDestruct();
+  }
+
+  ProcessTasks(false);
+}
+
+void MultiplexRouter::ProcessTasks(bool force_async) {
+  lock_.AssertAcquired();
+
+  while (!tasks_.empty()) {
+    scoped_ptr<Task> task(tasks_.front());
+    tasks_.pop_front();
+
+    bool processed = task->IsNotifyErrorTask()
+                         ? ProcessNotifyErrorTask(task.get(), &force_async)
+                         : ProcessIncomingMessageTask(task.get(), &force_async);
+
+    if (!processed) {
+      tasks_.push_front(task.release());
+      break;
+    }
+  }
+}
+
+bool MultiplexRouter::ProcessNotifyErrorTask(Task* task, bool* force_async) {
+  lock_.AssertAcquired();
+  InterfaceEndpoint* endpoint = task->endpoint_to_notify.get();
+  if (!endpoint->client())
+    return true;
+
+  if (!endpoint->task_runner()->BelongsToCurrentThread() || *force_async) {
+    endpoint->task_runner()->PostTask(
+        FROM_HERE, base::Bind(&MultiplexRouter::LockAndCallProcessTasks, this));
+    return false;
+  }
+
+  *force_async = true;
+  InterfaceEndpointClient* client = endpoint->client();
+  {
+    base::AutoUnlock unlocker(lock_);
+    client->NotifyError();

[sky, 2015/11/19 17:16:27]

Can bad things happen between the time you unlock and call NotifyError? For
example, could DetachEndpointClient() be called?

[yzshen1, 2015/11/19 22:00:40]

|client| lives on a single thread. NotifyError() and DetachEndpointClient() are
both called on that thread. So it is not possible for DetachEndpointClient() to
be called in the middle.

I added a comment.

+  }
+  return true;
+}
+
+bool MultiplexRouter::ProcessIncomingMessageTask(Task* task,
+                                                 bool* force_async) {
+  lock_.AssertAcquired();
+  Message* message = task->message.get();
+
+  if (PipeControlMessageHandler::IsPipeControlMessage(message)) {
+    if (!control_message_handler_.Accept(message))
+      RaiseErrorInNonTestingMode();
+    return true;
+  }
+
+  InterfaceId id = message->interface_id();
+  DCHECK(IsValidInterfaceId(id));
+
+  if (!ContainsKey(endpoints_, id)) {
+    DCHECK(!IsMasterInterfaceId(id));
+
+    // Currently, it is legitimate to receive messages for an endpoint
+    // that is not registered. For example, the endpoint is transferred in
+    // a message that is discarded. Once we add support to specify all
+    // enclosing endpoints in message header, we should be able to remove
+    // this.
+    InterfaceEndpoint* endpoint = new InterfaceEndpoint(this, id);
+    endpoints_[id] = endpoint;
+    endpoint->SetClosedMayDestruct();
+
+    control_message_proxy_.NotifyPeerEndpointClosed(id);
+    return true;
+  }
+
+  InterfaceEndpoint* endpoint = endpoints_[id];
+  if (endpoint->closed())
+    return true;
+
+  if (!endpoint->client()) {
+    // We need to wait until a client is attached in order to dispatch further
+    // messages.
+    return false;
+  }
+
+  if (!endpoint->task_runner()->BelongsToCurrentThread() || *force_async) {
+    endpoint->task_runner()->PostTask(
+        FROM_HERE, base::Bind(&MultiplexRouter::LockAndCallProcessTasks, this));
+    return false;
+  }
+
+  *force_async = true;
+  InterfaceEndpointClient* client = endpoint->client();
+  scoped_ptr<Message> owned_message = task->message.Pass();
+  bool result = false;
+  {
+    base::AutoUnlock unlocker(lock_);
+    result = client->HandleIncomingMessage(owned_message.get());

[sky, 2015/11/19 17:16:27]

Similar question to above about unlocking.

[yzshen1, 2015/11/19 22:00:40]

(Please see my comment for the previous question.)

+  }
+  if (!result)
+    RaiseErrorInNonTestingMode();
+
+  return true;
+}
+
+void MultiplexRouter::LockAndCallProcessTasks() {
+  // There is no need to hold a ref to this class in this case because this is
+  // always called using base::Bind(), which holds a ref.
+  base::AutoLock locker(lock_);
+  ProcessTasks(false);
+}
+
+void MultiplexRouter::OnEndpointDestructed(InterfaceId id) {
+  lock_.AssertAcquired();
+  endpoints_.erase(id);
+}
+
+void MultiplexRouter::RaiseErrorInNonTestingMode() {
+  lock_.AssertAcquired();
+  if (!testing_mode_)
+    RaiseError();
+}
+
+}  // namespace internal
+}  // namespace mojo