Create a new RenderFrameHost per child frame when --site-per-process is enabled.

content/browser/renderer_host/frame_tree.h

Index: content/browser/renderer_host/frame_tree.h
diff --git a/content/browser/renderer_host/frame_tree.h b/content/browser/renderer_host/frame_tree.h
new file mode 100644
index 0000000000000000000000000000000000000000..ad7a97e04fd769f3c6941017214e0d7360a93061
--- /dev/null
+++ b/content/browser/renderer_host/frame_tree.h
@@ -0,0 +1,84 @@
+// Copyright 2013 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef CONTENT_BROWSER_RENDERER_HOST_FRAME_TREE_H_
+#define CONTENT_BROWSER_RENDERER_HOST_FRAME_TREE_H_
+
+#include <string>
+
+#include "base/callback.h"
+#include "base/memory/scoped_ptr.h"
+#include "content/browser/renderer_host/frame_tree_node.h"
+#include "content/common/content_export.h"
+
+namespace content {
+
+class FrameTreeNode;
+class RenderProcessHost;
+
+// Represents the frame tree for a page. Used to manage RenderFrameHost during

[Charlie Reis, 2013/09/11 22:25:04]

FrameTree doesn't manage RFHs; FrameTreeNode does.

[awong, 2013/09/21 01:19:56]

Done.

+// navigation. Eventually each node in this class will optionally contain a

[Charlie Reis, 2013/09/11 22:25:04]

This seems hard to understand unless you know our plans to migrate BrowserPlugin
to this.  I'd either leave it out or have a separate TODO comment at the end of
the paragraph saying we're going to migrate BrowserPlugin to RenderFrameHost,
storing its NavigationController on a lower FrameTreeNode.

[awong, 2013/09/21 01:19:56]

Moved to TODO

+// NavigationController. The main frame node is guaranteed to have a

[Charlie Reis, 2013/09/11 22:25:04]

frame node -> FrameTreeNode (here and below)
(Too easy to get lost here unless we're precise in the names.)

[awong, 2013/09/21 01:19:56]

Yep...rewrote the comment.

+// NavigationController. Because of this, the main frame node cannot change
+// during navigation.  Instead, we reset its children and navigation related
+// state variables.
+class CONTENT_EXPORT FrameTree {
+ public:
+  explicit FrameTree();

[nasko, 2013/09/11 22:26:24]

Isn't explicit needed only for one argument constructors?

[awong, 2013/09/21 01:19:56]

Done.

+  ~FrameTree();
+
+  // Indiates the that RenderFrameHost for the main frame needs to be changed.

[nasko, 2013/09/11 22:26:24]

nit: Indicates

[awong, 2013/09/21 01:19:56]

Done.

+  // This effectively resets the frame tree but retains the root node so that
+  // navigation state may be kept on it between process swaps.
+  scoped_ptr<RenderFrameHostImpl> SwapMainFrame(

[Charlie Reis, 2013/09/11 22:25:04]

I'm not sure I understand why this is an operation on FrameTree.  Any
FrameTreeNode will be able to swap its RenderFrameHost, so it feels like it
belongs as SwapRenderFrameHost on FrameTreeNode.

nit: Indicates

[awong, 2013/09/21 01:19:56]

I think the name didn't match the semantics.  Also, I completely redid the
ownership semantics for the RenderFrameHost of the mainframe which should change
this function's behavior.

+      scoped_ptr<RenderFrameHostImpl> render_frame_host);
+
+  // Returns the FrameTreeNode with the given |frame_id|.
+  FrameTreeNode* FindByID(int64 frame_id);
+
+  // Executes |on_node| on each node in the frame tree.  If |on_node| returns
+  // false, terminates the loop without continuing.
+  void ForEach(const base::Callback<bool(FrameTreeNode*)>& on_node) const;
+
+  // After the FrameTree is created, or after SwapMainFrame() has been called,
+  // the root node does not yet have a frame id. This is allocated by the
+  // renderer and is published to the browser process on the first navigation.
+  // These two functions are used to set the root node's frame id.
+  //
+  // TODO(ajwong): Remove these once RenderFrameHost's routing id replaces
+  // frame_id.
+  bool IsFirstNavigation() const;
+  void OnFirstNavigation(int main_frame_id);

[Charlie Reis, 2013/09/11 22:25:04]

Which kind of frame ID are we talking about here, and why is it only relevant on
the first navigation?

(If we're talking about the current renderer-allocated frame_id from the
WebFrameImpl, it changes after every cross-process navigation, so
OnFirstNavigation seems misleading to me.)

[awong, 2013/09/21 01:19:56]

It's the render-allocated frame_id. It's only relevant on the initial navigation
after the tree has been created, or where the mainframe has been reset. In these
situations, we know we have a frame, but the id may yet to have been
allocated...these should be removed later, but there are some trickiness with
timing until we move everything over to routing_id (and there are some
trickinesses with WebWorkers and HelperPlugins that make it hard to move to
routing_id).

+
+  // Frame tree manipulation routines.
+  void AddFrame(int render_frame_host_id, int64 parent_frame_id,
+                int64 frame_id, const std::string& frame_name);
+  void RemoveFrame(int64 parent_frame_id, int64 frame_id);
+  void SetFrameUrl(int64 frame_id, const GURL& url);
+
+  // Convenience accessor for the main frame's RenderFrameHostImpl.
+  RenderFrameHostImpl* GetMainFrame() const;
+
+  // Allows a client to listen for frame removal.
+  void SetFrameRemoveListener(
+      const base::Callback<void(int64)>& on_frame_removed);
+
+  FrameTreeNode* GetRootForTesting() { return root_.get(); }
+
+ private:
+  scoped_ptr<FrameTreeNode> CreateNode(int64 frame_id,
+                                       const std::string& frame_name,
+                                       int render_frame_host_id,
+                                       RenderProcessHost* render_process_host);
+
+  scoped_ptr<FrameTreeNode> root_;
+
+  base::Callback<void(int64)> on_frame_removed_;
+
+  DISALLOW_COPY_AND_ASSIGN(FrameTree);
+};
+
+}  // namespace content
+
+#endif  // CONTENT_BROWSER_RENDERER_HOST_FRAME_TREE_H_