Fix some flaky tests.

content/browser/site_per_process_browsertest.cc

Index: content/browser/site_per_process_browsertest.cc
diff --git a/content/browser/site_per_process_browsertest.cc b/content/browser/site_per_process_browsertest.cc
index 477294d4ff3293afaf7946ad66680995578bb6f8..641dac150d461576e1923a9814e96f986bb82e3b 100644
--- a/content/browser/site_per_process_browsertest.cc
+++ b/content/browser/site_per_process_browsertest.cc
@@ -8357,6 +8357,8 @@ IN_PROC_BROWSER_TEST_F(SitePerProcessBrowserTest, NavigateInUnloadHandler) {
                     "};\n"));
 
   // Navigate B's subframe to a cross-site C.
+  RenderFrameDeletedObserver deleted_observer(
+      root->child_at(0)->child_at(0)->current_frame_host());

[nasko, 2016/11/23 17:48:29]

I agree that moving this code makes sense. However, I don't follow your CL
comment about deleted_observer binding to the "c.com" RenderFrameHost. Even if
the navigation commits for c.com, there is no way for the observer to bind to
that RFH, as it will be root->child_at(0) and not
root->child_at(0)->child_at(0). What is possible is for the
root->child_at(0)->child_at(0) to be null of that happens.

[Alexander Semashko, 2016/11/23 18:38:08]

c.com is loaded in the third-level frame, because we navigate the child of
root->child_at(0).

[nasko, 2016/11/24 00:00:43]

Ah, yes!

   std::string script =
       std::string("window.document.getElementById('child-0').src = \"") +
       embedded_test_server()
@@ -8367,8 +8369,6 @@ IN_PROC_BROWSER_TEST_F(SitePerProcessBrowserTest, NavigateInUnloadHandler) {
       ExecuteScript(root->child_at(0)->current_frame_host(), script.c_str()));
 
   // Wait until B's subframe RenderFrameHost is destroyed.
-  RenderFrameDeletedObserver deleted_observer(
-      root->child_at(0)->child_at(0)->current_frame_host());
   deleted_observer.WaitUntilDeleted();
 
   // Check that C's subframe is alive and the navigation in the unload handler