Use RenderProcessHost::Shutdown in test to avoid flakiness.

content/browser/frame_host/frame_tree_browsertest.cc

Index: content/browser/frame_host/frame_tree_browsertest.cc
diff --git a/content/browser/frame_host/frame_tree_browsertest.cc b/content/browser/frame_host/frame_tree_browsertest.cc
index 9ce4131f87773a340eb7e7bcd837cd40803c5698..a1d670d185ed8298b208ed5bb20552941149e6ac 100644
--- a/content/browser/frame_host/frame_tree_browsertest.cc
+++ b/content/browser/frame_host/frame_tree_browsertest.cc
@@ -135,7 +135,7 @@ IN_PROC_BROWSER_TEST_F(FrameTreeBrowserTest, FrameTreeAfterCrash) {
   RenderProcessHostWatcher crash_observer(
       shell()->web_contents(),
       RenderProcessHostWatcher::WATCH_FOR_PROCESS_EXIT);
-  NavigateToURL(shell(), GURL(kChromeUICrashURL));
+  shell()->web_contents()->GetMainFrame()->GetProcess()->Shutdown(0, false);

[Łukasz Anforowicz, 2017/06/01 16:10:29]

It's probably a good idea to check (probably via EXPECT_TRUE rather than
ASSERT_TRUE) the return value from Shutdown (it can be failure-indicating-false
if !child_process_launcher_.get() or IsStarting().

I wonder if we should pass |true| as the second argument - I see in
base::Process::Terminate that it starts with SIGTERM, but when |wait| is true
and the process takes a long time to die, then it also does SIGKILL.  OTOH, if
the process hung, then I guess chrome://crash would also take a long time...
(but then maybe this is the source of flakiness that we are trying to avoid...).

[Charlie Reis, 2017/06/01 17:04:16]

Good idea.  But this is one of the cases that ASSERT_TRUE is good for-- the rest
of the test doesn't make sense if this kill fails, so it's good to early exit. 
The rest of the results won't have meaning otherwise.
Might as well-- we want this kill to be as effective as possible to make the
test meaningful.

   crash_observer.Wait();
 
   // The frame tree should be cleared.