NoState Prefetch: nostate prefetch browser tests.

chrome/browser/prerender/prerender_final_status.h

Index: chrome/browser/prerender/prerender_final_status.h
diff --git a/chrome/browser/prerender/prerender_final_status.h b/chrome/browser/prerender/prerender_final_status.h
index 36d38b7571077b13add8218cbc13c8e20980fd0e..78036b8967a6041df43323e25d89d52ab5719b61 100644
--- a/chrome/browser/prerender/prerender_final_status.h
+++ b/chrome/browser/prerender/prerender_final_status.h
@@ -67,6 +67,7 @@ enum FinalStatus {
   FINAL_STATUS_PRERENDERING_DISABLED = 52,
   FINAL_STATUS_CELLULAR_NETWORK = 53,
   FINAL_STATUS_BLOCK_THIRD_PARTY_COOKIES = 54,
+  FINAL_STATUS_TESTING_IGNORE = 55,

[pasko, 2016/10/14 12:49:13]

It is unfortunate that we need this (not too serious, but makes test code harder
to understand). Is this because the FinalStatus can be different depending on
racy conditions? Wondering whether it can be avoided, after all,
prereneder_browsertests managed to avoid mispredicting FinalStatus in tests.

[mattcary, 2016/10/14 12:54:45]

Actually, see TestPrerenderContents::RenderProcessGone, where we see that app
terminating or renderer crashed can both occur for prerender_browsertests.

For nostate prefetch, it's even harder, because the prerender process may or may
not have been killed by the time all the requests are done (or the requests
didn't even occur, as is the case in the tests needing this).

Hopefully your lifetime changes will plug a lot of these cases and we can remove
it.

I thought about setting the constant to some high value, so that we can remove
it from the enum if it turns out to be necessary without polluting the range (we
know this can never occur in a histogram. Thoughts?

[pasko, 2016/10/14 13:28:05]

Can you explain a little more on why is it killed? Is it flaky because we did
not allow for enough RAM in the renderer? I thought the current behavior should
keep the PrerenderContents alive until the test wraps up (and a long timeout
that we disable?).
I am OK with a high number, yes, since reusing values in histogram-related enums
is not allowed. Also potentially augmenting, if there is still a reason to
survive-or-kill undeterministically, tiny nit: the value could be named as
PREFETCH_ALIVE_OR_GONE_FOR_TESTING to keep it targeted at a specific problem and
suggest not reusing it for anything else in the meantime.

[mattcary, 2016/10/14 13:41:52]

The race in question for the original prerender_browsertests is between
app_terminating and renderer_crashed because, I presume, of the race between the
renderprocesses getting killed and the prerender manager getting destroyed.

The race that inspired my odious final status is between things like the
renderer getting killed and safe browsing returning, or the renderer getting
destroyed and the unsupport-scheme final status that I believe comes from the
error page one gets redirected to when trying to load a nonexisting page.
As explained above, I don't think that value is actually an accurate description
of what's going on.

   FINAL_STATUS_MAX,
 };