content/gpu: Set up a mojom connection between browser and gpu.

content/gpu/gpu_child_thread.cc

Index: content/gpu/gpu_child_thread.cc
diff --git a/content/gpu/gpu_child_thread.cc b/content/gpu/gpu_child_thread.cc
index 637b4802f198153f6e398699538b81439ebdf867..726dffc1b8878c16e5f9e9be9eaa85eab8527ac0 100644
--- a/content/gpu/gpu_child_thread.cc
+++ b/content/gpu/gpu_child_thread.cc
@@ -203,12 +203,11 @@ void GpuChildThread::Init(const base::Time& process_start_time) {
   if (!in_browser_process_)
     media::SetMediaClientAndroid(GetContentClient()->GetMediaClientAndroid());
 #endif
-  // We don't want to process any incoming interface requests until
-  // OnInitialize() is invoked.

[jbauman, 2017/01/11 01:49:22]

Just checking, but is the reason we can remove this that mojo messages are now
processed in order with the GpuMsg_Initialize IPC message?

[sadrul, 2017/01/11 01:55:46]

The Pause/Resume was added in https://crrev.com/402996, to make sure
OnInitialize() is called before any of the mojo requests are processed.
https://crrev.com/406761 then moved the mojo registrations in OnInitialize(),
which means the mojo requests could not be processed before OnInitialize() is
called anyway. So I think that makes it safe to remove Pause/Remove now.

This is also necessary because OnInitialize() itself will be replaced by a mojom
call (see more details in https://codereview.chromium.org/2612623002/), which
means this needs to allow processing mojo requests from the start.

I have added rockot@ to the reviewer list to make sure this reasoning is OK, and
there isn't a better way for doing this.

[Ken Rockot(use gerrit already), 2017/01/11 18:17:00]

This rationale seems kind of circular to me. I don't think it's safe to remove
this behavior. r406761 was able to move interface registration to OnInitialize
precisely *because* we pause the registry binding here. If we don't pause it,
interface requests can be dispatched (and thus those interfaces can be bound and
receive messages) before OnInitialize() runs.

-  GetInterfaceRegistry()->PauseBinding();
 
   if (GetContentClient()->gpu())  // NULL in tests.
     GetContentClient()->gpu()->Initialize(this);
+  channel()->AddAssociatedInterface(base::Bind(
+      &GpuChildThread::CreateGpuMainService, base::Unretained(this)));
 }
 
 void GpuChildThread::OnFieldTrialGroupFinalized(const std::string& trial_name,
@@ -216,6 +215,11 @@ void GpuChildThread::OnFieldTrialGroupFinalized(const std::string& trial_name,
   Send(new GpuHostMsg_FieldTrialActivated(trial_name));
 }
 
+void GpuChildThread::CreateGpuMainService(
+    ui::mojom::GpuMainAssociatedRequest request) {
+  // TODO(sad): Implement.
+}
+
 bool GpuChildThread::Send(IPC::Message* msg) {
   // The GPU process must never send a synchronous IPC message to the browser
   // process. This could result in deadlock.
@@ -369,8 +373,6 @@ void GpuChildThread::OnInitialize(const gpu::GpuPreferences& gpu_preferences) {
     GetContentClient()->gpu()->ConsumeInterfacesFromBrowser(
         GetRemoteInterfaces());
   }
-
-  GetInterfaceRegistry()->ResumeBinding();
 }
 
 void GpuChildThread::OnFinalize() {