services/ui: Allow injecting the IO thread task runner for gpu channel.

services/ui/public/cpp/gpu_service.h

Index: services/ui/public/cpp/gpu_service.h
diff --git a/services/ui/public/cpp/gpu_service.h b/services/ui/public/cpp/gpu_service.h
index a1155ae0bdc3745714270da2455db5d2c75720fc..f7af2a56e221545481eeccceb2dca6e30936faf8 100644
--- a/services/ui/public/cpp/gpu_service.h
+++ b/services/ui/public/cpp/gpu_service.h
@@ -35,9 +35,14 @@ class GpuService : public gpu::GpuChannelHostFactory,
   }
 
   // The GpuService has to be initialized in the main thread before establishing
-  // the gpu channel.
+  // the gpu channel. The IO thread task-runner needed for the gpu channel can
+  // be set using SetIOTaskRunner(). If no task-runner is set, then a new thread
+  // is created for IO instead.
   static std::unique_ptr<GpuService> Initialize(shell::Connector* connector);
 
+  void SetIOThreadTaskRunner(
+      scoped_refptr<base::SingleThreadTaskRunner> task_runner);

[Fady Samuel, 2016/08/18 11:00:56]

Can this not be passed in Initialize to avoid giving the impression that the
IOThreadTaskRunner can change at any time?

Also optional nit: GpuService::Initialize sounds weird because it actually
creates the object rather than initialize an already constructed object.

It would be nice to call this GpuService::Create, but optional I guess

[sadrul, 2016/08/18 12:56:53]

It's not straight forward, because for the chrome browser, the GpuService is
created by the views::WindowManagerConnection object. So we would have to plumb
the task-runner through that to GpuService. This is cleaner.
SetIOThreadTaskRunner() already has a DCHECK() to make sure the task runner
cannot be changed once set. I have added some messaging to the DCHECK() to
clarify that as well.
I have disliked Initialize() myself. Renamed to Create().

[Fady Samuel, 2016/08/18 13:05:39]

LGTM but I still feel like this is brittle because we could accidentally create
an IOThread if someone calls GetIOThreadTaskRunner() too early before
SetIOThreadTaskRunner is called

+
   // gpu::GpuChannelEstablishFactory:
   void EstablishGpuChannel(
       const gpu::GpuChannelEstablishedCallback& callback) override;
@@ -65,9 +70,10 @@ class GpuService : public gpu::GpuChannelHostFactory,
       size_t size) override;
 
   scoped_refptr<base::SingleThreadTaskRunner> main_task_runner_;
+  scoped_refptr<base::SingleThreadTaskRunner> io_task_runner_;
   shell::Connector* connector_;
   base::WaitableEvent shutdown_event_;
-  base::Thread io_thread_;
+  std::unique_ptr<base::Thread> io_thread_;
   std::unique_ptr<MojoGpuMemoryBufferManager> gpu_memory_buffer_manager_;
 
   // Lock for |gpu_channel_|, |establish_callbacks_| & |is_establishing_|.