Here's another failing bot: https://uberchromegw.corp.google.com/i/chromium.linux/builders/Linux%20Tests/builds/58277

I'm going to go ahead and roll this out.

This has been reverted in https://chromium.googlesource.com/chromium/src/+/15f5b5b2ef86151279bbad41a0976375b9f40977

 Issue 734862  has been merged into this issue.

Thanks for this report and for rolling back the change.

It seems the test is hanging, but flakily. I have been able to repro at least once locally.

Don't understand the problem yet, but I believe I understand the source of the flakiness: it seems the test hangs in the case where, when we initiate the second navigation, is_waiting_for_beforeunload_ack_ is still true from the first navigation.

I reverted everything but the new-test part of my change, and can still repro the intermittent hang. This suggests the behavior here is preexisting.

Still digging to determine if it's a test bug or a functional bug. RenderFrameHostImpl::DispatchBeforeUnload definitely seems to deliberately drop the dispatch operation on the floor here, but this intermediate state is weird, all the following are simultaneously true:
  - The first load has stopped already (from the perspective of content::WaitForLoadStop)
 - The omnibox state has reset to the original URL, as the pending entry has been discarded.
 - BUT we're still waiting for an beforeunload ACK to report back the "proceed=false" information (even though proceed=false came from an earlier dialog result in the browser process).

We could paper over it by waiting longer (until is_waiting_for_beforeunload_ack_ transitions back to false) but before doing that I want to prove there's not a more serious bug lurking underneath.

The hang seems to go away if I change my implementation of RunBeforeUnloadDialog to run the cancellation callback immediately, instead of doing a PostTask. This makes sense, since every PostTask operation provides an opportunity for nondeterminism. I don't actually see a bug here, just my own confusion, so this is the approach I'll go with.