Re-land: cc: Implement shared worker contexts.

cc/trees/layer_tree_host.h

Index: cc/trees/layer_tree_host.h
diff --git a/cc/trees/layer_tree_host.h b/cc/trees/layer_tree_host.h
index e17431d89c00f4b35cec785252110860e273ec28..29990f930c3725aabb473c96b0d1309caa7bf61f 100644
--- a/cc/trees/layer_tree_host.h
+++ b/cc/trees/layer_tree_host.h
@@ -434,6 +434,13 @@ class CC_EXPORT LayerTreeHost : public MutatorHostClient {
   int meta_information_sequence_number_;
   scoped_ptr<RenderingStatsInstrumentation> rendering_stats_instrumentation_;
 
+  // |current_output_surface_| can't be updated until we've successfully
+  // initialized a new output surface. |new_output_surface_| contains the
+  // new output surface that is currently being initialized. If initialization
+  // is successful then |new_output_surface_| replaces
+  // |current_output_surface_|.
+  scoped_ptr<OutputSurface> new_output_surface_;

[danakj, 2015/09/12 00:09:14]

+sievers: From what I can tell there's no good reason that LTH even owns these
after this, we could just have the embedders own the output surface entirely and
tell them when cc is done with it or something.

[reveman, 2015/09/12 12:49:57]

Sgtm but I think I prefer to save that for a follow up patch. Let me know if you
want me to add a TODO here.

[danakj, 2015/09/14 18:05:46]

Ya that's all I mean. No TODO is fine I think.

I also was thinking that if we made the RequestNewSurface pass back ownership of
the OS (ie make the LTHI pointer null at the time of loss, not at the time of
re-create-and-fail) then we could avoid having scoped_ptrs here on the LTH at
all, and just pass around ownership instead of "ok you can delete it now LTH"
more-or-less, which would be nice.

It changes when the LTHI::OS* pointer is null though, which might be a bit
tricky. Def not suggesting that here.

[no sievers, 2015/09/14 21:18:05]

Are you suggesting as a follow-up to move the ownership of the active OS back to
LTHI, but release (or return) it early on context lost so that we don't have to
track ownership of two surfaces?

Or are you saying that embedders should really 'own the output surface'
entirely, in which case we wouldn't pass ownership around? But that probably
isn't better, and we'd still need the same APIs to coordinate who is doing what
with it.

[danakj, 2015/09/14 21:22:47]

I was thinking the former. Documenting ownership/lifetime with scoped_ptr
passing is nice.

+  scoped_ptr<OutputSurface> current_output_surface_;
   bool output_surface_lost_;
 
   scoped_refptr<Layer> root_layer_;