- * Lock-free RCU queue using reference counting. Enqueue and dequeue operations
- * hold a RCU read lock to deal with cmpxchg ABA problem. This implementation
- * keeps a dummy head node to ensure we can always update the queue locklessly.
- * Given that this is a queue, the dummy head node must always advance as we
- * dequeue entries. Therefore, we keep a reference count on each entry we are
- * dequeueing, so they can be kept as dummy head node until the next dequeue, at
- * which point their reference count will be decremented.
+ * Lock-free RCU queue. Enqueue and dequeue operations hold a RCU read
+ * lock to deal with cmpxchg ABA problem. This queue is *not* circular:
+ * head points to the oldest node, tail points to the newest node.
+ * A dummy node is kept to ensure enqueue and dequeue can always proceed
+ * concurrently. Keeping a separate head and tail helps with large
+ * queues: enqueue and dequeue can proceed concurrently without
+ * wrestling for exclusive access to the same variables.
+ *
+ * Dequeue retry if it detects that it would be dequeueing the last node
+ * (it means a dummy node dequeue-requeue is in progress). This ensures
+ * that there is always at least one node in the queue.
+ *
+ * In the dequeue operation, we internally reallocate the dummy node
+ * upon dequeue/requeue and use call_rcu to free the old one after a
+ * grace period.