views/mus: Fix some flaky crashes during test teardown.

ui/views/mus/surface_binding.cc

Index: ui/views/mus/surface_binding.cc
diff --git a/ui/views/mus/surface_binding.cc b/ui/views/mus/surface_binding.cc
index 4161811c44992bfa72440ed0a1279b1b02903f3f..9aca4af0d6a7e6488099f917aadc35deebbe6b66 100644
--- a/ui/views/mus/surface_binding.cc
+++ b/ui/views/mus/surface_binding.cc
@@ -98,6 +98,8 @@ std::unique_ptr<cc::OutputSurface>
 SurfaceBinding::PerConnectionState::CreateOutputSurface(
     mus::Window* window,
     mus::mojom::SurfaceType surface_type) {
+  if (gpu_.encountered_error())
+    return nullptr;
   // TODO(sky): figure out lifetime here. Do I need to worry about the return
   // value outliving this?
   mus::mojom::CommandBufferPtr cb;
@@ -127,6 +129,7 @@ SurfaceBinding::PerConnectionState::~PerConnectionState() {
 
 void SurfaceBinding::PerConnectionState::Init() {
   connector_->ConnectToInterface("mojo:mus", &gpu_);
+  gpu_.set_connection_error_handler([]{});

[sadrul, 2016/05/18 20:00:19]

rockot@: This seems to be necessary for the change above to work (i.e. without
this, |gpu_.encountered_error()| doesn't seem to see any errors). I assume this
is unexpected? Any ideas/suggestions as to why this may be making a difference?

[sky, 2016/05/18 20:10:32]

This seems like a bug in the bindings code.

[Ken Rockot(use gerrit already), 2016/05/18 20:11:08]

Because we lazily set up the internal proxy in InterfacePtr, and
set_connection_error_handler is one of the calls which forces proxy setup.
+yzshen@ +amistry@ I remember that we made this change for a reason, but I don't
remember why. Why doesn't Bind also configure the proxy?

[sadrul, 2016/05/19 20:14:17]

OK. Sounds like this is not entirely unexpected, then. Can I add a TODO here and
refer to some bug (does one already exist?) so this doesn't get lost?

[sky, 2016/05/19 20:59:42]

Yes, go for it.

 }
 
 // SurfaceBinding --------------------------------------------------------------