TSAN detected wrong use of the old deque-based internal queue. To avoid
unwanted/undetected mis-use the patch uses the thread-safe block_queue
data structure instead.
- the consumer does multi-staged waiting:
1. spins first across all queues in a RR fashion
2. each queue access is done with a try_lock.
3. if the try_lock fails, it increases the number of spins needed
2. if no queue had data, the consumer sleeps for 100 usec.
- no differentiation between queues, in terms of notification features
With the new design, each queue created in the multiqueue object has its own mutex.
Pushing tasks to separate queues will, therefore, not cause contention.
There will be, however, still contention between the popping thread and the pushing threads.