GitLab

Please see https://crbug.com/658759 for more context.

The TL;DR is I broke some subtle behavior in
https://crrev.com/425358, and this fixes it. Here's what happens
in ProcessDied, in this order:

  1. We reset channel_ so that future Send() calls throw away
     messages. Note that at this point, any Channel-associated
     interfaces remain valid but are intentionally dropping
     all messages as well.

  2. We fire various notifications about the death of the process.

  3. We re-initialize channel_, allowing subsequent Send() calls
     and associated interface calls to be queued and eventually
     delivered to the new process once it's restarted.

The subtlety here is that if a RenderViewHostImpl is created at any
point during step (2) above, we must re-initialize channel_ and
start allowing messages to be queued on it. Formerly this was
accomplished by EnableSendQueue() setting is_initialized_ to
false. This CL establishes the same behavior by restoring
EnableSendQueue() to essentially have the same effect by re-
initializing the channel earlier.

This is likely not the precise behavior we want, since it seems like
this can cause messages intended for the old process to end up getting
delivered to the new process, but it is the behavior we already had in
place before https://crrev.com/425358 and we apparently rely on it.

BUG=658759

Review-Url: https://codereview.chromium.org/2446543004
Cr-Commit-Position: refs/heads/master@{#427257}