title	TODO
order	0

TODO

EVENTING as a failure method
speed differential if combining cache lines (actual impact of false sharing)?
generic nmath functions so 32-bit size_t case is cared for
no safety checking or locking on init/deinit - unsure of the best approach here; maybe a strenuous warning to the caller not to shoot themselves in the foot?
Python bindings
C++ extensions?
non-contention cost of operations (reserving and releasing buffer blocks one by one)
contention-ONLY cost (no operation on underlying memory)
example of stack allocation
example of underlying file access
example of returning data to producers
example of using zero-copy I/O (split nmem from nonlibc?)
man pages
return both 'res' and 'pos' by value in registers??

Benchmark against other implementations