gpu: Replace GpuMsg_CollectGraphicsInfo with mojom API.

content/browser/gpu/gpu_data_manager_impl_private.cc

Index: content/browser/gpu/gpu_data_manager_impl_private.cc
diff --git a/content/browser/gpu/gpu_data_manager_impl_private.cc b/content/browser/gpu/gpu_data_manager_impl_private.cc
index 319a130a4c0c68108af5018bf44776b414c81fdb..586b90d56c23ccf4906a4420e24038ba03adcf3a 100644
--- a/content/browser/gpu/gpu_data_manager_impl_private.cc
+++ b/content/browser/gpu/gpu_data_manager_impl_private.cc
@@ -418,13 +418,21 @@ void GpuDataManagerImplPrivate::RequestCompleteGpuInfoIfNeeded() {
 
   complete_gpu_info_already_requested_ = true;
 
-  GpuProcessHost::SendOnIO(
+  GpuProcessHost::CallOnIO(
 #if defined(OS_WIN)
       GpuProcessHost::GPU_PROCESS_KIND_UNSANDBOXED,
 #else
       GpuProcessHost::GPU_PROCESS_KIND_SANDBOXED,
 #endif
-      true /* force_create */, new GpuMsg_CollectGraphicsInfo());
+      true /* force_create */, base::Bind([](GpuProcessHost* host) {
+        if (!host)
+          return;
+        host->gpu_service()->RequestCompleteGpuInfo(
+            base::Bind([](const gpu::GPUInfo& gpu_info) {
+              TRACE_EVENT0("test_gpu", "OnGraphicsInfoCollected");
+              GpuDataManagerImpl::GetInstance()->UpdateGpuInfo(gpu_info);

[piman, 2017/03/20 21:02:36]

Should we hop back to the UI thread, though?
I know there's a lock around accesses to GpuDataManagerImplPrivate, but 
GpuDataManagerImplPrivate::UpdateGpuInfo calls back into things (e.g. 
GetContentClient()->SetGpuInfo ) which I'm not sure is safe to do from the IO
thread.

[sadrul, 2017/03/20 21:21:44]

Ah, that's a good point! I have moved this to hop onto the UI thread before
updating GPUInfo.

Would it make sense to add some DCHECKs into GpuDataManager[Impl[Private]] to
make sure such functions are called from the right thread (where applicable) to
avoid accidental regressions?

[piman, 2017/03/20 21:33:54]

I think it would be useful. GpuDataManagerImplPrivate is a bit of a mess and not
necessarily well behaved everywhere. It "looks" thread-safe for the most part,
but I think it's not in several places (UpdateGpuInfo being one).

+            }));
+      }));
 }
 
 bool GpuDataManagerImplPrivate::IsEssentialGpuInfoAvailable() const {