There are several magazines for allocations [segmented by allocation size]. 

The large magazine has a global lock. 

For the small magazine, there is a lock per logical cpu. Allocations grab the lock of the current CPU, but frees grab the lock of the CPU at time of allocation. There will be heavy lock contention if one thread allocates and another frees. 

I didn't check the code for tiny/nano magazines.

lizeb@ mentioned that tcmalloc has better performance for this use case.