Michael–Scott MPMC queue

A C++17 implementation of the Michael–Scott concurrent queue. Producers link a node and then advance tail; a thread that observes a lagging tail helps complete that update. Producers and consumers operate on separate, cache-line-aligned atomics.

Memory reclamation

Dequeued dummy nodes are placed on a lock-free retired list and reclaimed by the destructor. This avoids use-after-free and ABA caused by immediate address reuse, but memory consumption grows with the number of dequeues while the queue is alive. A long-lived implementation would need hazard pointers or epoch-based reclamation.

Build and test

make test
make tsan

The stress test runs four producers and four consumers over 400,000 unique integers, then verifies the item count and sum. The current implementation allocates one node with new per enqueue; its benchmark therefore measures both queue synchronization and allocation cost.