While looking at something unrelated, I noticed that pmap_prefer() is
rather simplistic -- that is, it always tries to match
offset%cache_size with address%cache_size.

The problem I see with this is that shared objects will always get
mapped at an address that causes them to start at the beginning of the
cache.  Thus, the beginning of the cache will tend to get hit much
more heavily than the rest.

Would it not make more sense to rotate this, perhaps based on the
inode number (which is persisent) or even the object pointer in core
(which is at least persistent over the lifetime of multiple mappings)?