The page on concurrency in map/reduce (http://docs.mongodb.org/manual/core/map-reduce-concurrency/) should make it clear that the write lock taken during the reduce phase is actually a global lock, rather than just a database-level lock. Also, while in non-atomic mode the lock is released and re-acquired for every single document that is written, the lock is pre-emptive in that case and thus has essentially the same effect as in atomic mode, rendering the server inaccessible for any other operation.

(Compare SERVER-13552 where this issue is tracked.)