* separate submap for zio buffers
  * different allocation strategies (best fit, first fit, next fit)
  * per-cpu caching of vm allocations (similarly to uma)
  * 2^N freelists similar to our phys pages allocator