Goals

We'd like to have Huge Tracts of Buffer Cache 


When not using memory for other things.

Challenges
Some machines don't have lots of memory. (We run lots of RASH)
The kernel memory map (kva) may be limited on some architectures. 
We need to know when memory is needed for other things
Lots of subsystem interdependence (VFS, UVM, name cache, drivers, etc.)
UVM and ilk have been hacked on for years, surely there aren't bugs.