Cleanup GpuMemoryBuffer

gpu/command_buffer/client/gpu_memory_buffer_manager.h

Index: gpu/command_buffer/client/gpu_memory_buffer_manager.h
diff --git a/gpu/command_buffer/client/gpu_memory_buffer_manager.h b/gpu/command_buffer/client/gpu_memory_buffer_manager.h
index cc422bfcba78c5fb75253a82bf1a4e227ef501bb..f3ca3177b0fd50e739b86b716a5a5ec25f4233a9 100644
--- a/gpu/command_buffer/client/gpu_memory_buffer_manager.h
+++ b/gpu/command_buffer/client/gpu_memory_buffer_manager.h
@@ -27,6 +27,14 @@ class GPU_EXPORT GpuMemoryBufferManager {
   virtual gfx::GpuMemoryBuffer* GpuMemoryBufferFromClientBuffer(
       ClientBuffer buffer) = 0;
 
+  // Returns whether a sync point is required for a specific buffer type during
+  // desruction.
+  virtual bool NeedsDestructionSyncPoint(gfx::GpuMemoryBufferType type) = 0;

[reveman, 2014/10/31 11:39:36]

Let's avoid this and just use the switch statement in
GpuChannelHost::ShareGpuMemoryBufferToGpuProcess instead.

+
+  // Associates destruction sync point with |buffer|.
+  virtual void RecordDestructionSyncPoint(ClientBuffer buffer,

[reveman, 2014/10/31 11:39:36]

We shouldn't be using ClientBuffer except for C API.

+                                          uint32 sync_point) = 0;

[reveman, 2014/10/31 11:39:36]

Can we make this:

virtual void SetGpuMemoryBufferDestructionSyncPointCallback(
    gfx::GpuMemoryBufferType type,
    const gfx::GpuMemoryBufferId& id,
    const InsertSyncPointCallback& insert_sync_point_callback) = 0;

and keep these callbacks in a map in the manager instead of adding
GpuMemoryBufferImpl API. The map might seem unnecessary but we'll need it to
handle process removal anyhow (in follow up patch).

The InsertSyncPointCallback can either return a sync point directly or take a
SyncPointInserted callback depending on if you need to make a thread hop or not.
I don't think you have to make a thread hop if you just make sure to track the
route_id as part of the InsertSyncPointCallback.

The manager's buffer deletion handler would use this callback if set when
deleting a buffer.

This would allow us to handle the sync point decision as part of the existing
ShareGpuMemoryBufferToGpuProcess logic and the need for a destruction
insertsyncpoint callback can simply be decided based on the existing switch
statement in that function. It also moves the sync point creation to destruction
time, which I think is a bit better as it's not going to affect the "initial
frame up on screen" time.

+
  protected:
   virtual ~GpuMemoryBufferManager();
 };