I found several drawbacks in the MD portion of NetBSD/mips VM.

- current code base looks trying vast efforts to sync. the content of
TLB with 'pmap' changes at any moment.  However, in most cases, the
target PTE is not 'hot', hot == (resides in TLB && active), searching
and updating such entries are not only just verbose, but even might
hurt other useful entries. 

- some useful VM notions of RO, COW or defered modification attributes
could be implemented and manipulated by D bit of TLB entryhi register.
The D bit has many overlapped implications by the software managed MMU
architecture.  I can not figure out how the current code base handles
COW so far, but the right solution would be heavily coupled with
TLBmod exception handler.  Current code base is very fragile to be
'corrected' and I've adandoned to 'improve' it.  trap() routine and
pmap.c are to be reworked (keeping multiprocessor mips64 in mind,
crossing fingers...)

Tohru Nishimura
Nara Institute of Science and Technology