GitLab

On uniprocessor, the main thread reached Engine::Destroy before the
worker threads had even started, so they did not see the shutdown.

On SMP, the worker threads were started (and blocked), so it was OK.

This new code also has the advantage of shutting down quicker
if there are pending callback requests in the queue.

Change-Id: Iff8cca042ebffe48ce7cc5d798e7b3b9911fb64b