GitLab

If a renderer process crashes, then we retain the associated
RenderProcessHostImpl object (to retain the old render_process_host_id
that may have been stored elsewhere in the system).  Before this CL this
had an undesired interaction with the ability of RenderProcessHostImpl
to queue IPCs between the call to Init and the time when a renderer
process is actually spawned (disclaimer - in practice the queueing has a
slightly/undesirably longer timeline).

Before this CL, RenderProcessHostImpl::ProcessDied(...) would call
EnableSendQueue().  This meant that an unbounded number of IPC messages
may be accumulated in
ChannelAssociatedGroupController::outgoing_messages_ (until the crashed
process is restarted).

After this CL, RenderProcessHostImpl::ProcessDied(...) no longer calls
EnableSendQueue().
- Short-term: subsystems that require queueing of IPCs should
  call RPH::EnableSendQueue() themselves (e.g. like RenderViewHostImpl's
  constructor does).
- Long-term: we should discard IPCs until RenderProcessHostImpl::Init
  is called - queueing should only take place in the short time between
  calling RenderProcessHostImpl::Init and the time the renderer process
  is actually launched.

There is a small risk that this CL will introduce crashes similar to the
ones seen in https://crbug.com/658759.

Bug: 813045
Change-Id: I838bf7ea443633ba0f2314d2de5c83891ec9ac5f
Reviewed-on: https://chromium-review.googlesource.com/1073621Reviewed-by: Nasko Oskov <nasko@chromium.org>
Reviewed-by: Ken Rockot <rockot@chromium.org>
Commit-Queue: Łukasz Anforowicz <lukasza@chromium.org>
Cr-Commit-Position: refs/heads/master@{#562491}