Subject: Re: My dec 5000/25 went boink on me today
To: None <port-pmax@netbsd.org>
From: Toru Nishimura <locore32@gaea.ocn.ne.jp>
List: port-pmax
Date: 04/15/2004 17:09:08
Here is a technical tip.

Due to the distinctive feature of software TLB update scheme kernel stack
needs to be "protected" from TLB miss, that is, kernel stack should never
make TLB miss to use.  Software TLB update processing is likely to
produce a string of TLB miss traversing the kernel date structure to
complete.  Then to make kernel stack "TLB miss free" NetBSD kernel
hardwires the 8KB memory  with 2 pcs of TLB entry on every context switch.
If the hardwiring goes wrong (out of sync), kernel stack gets effectively
destroyed at a moment.

Toru Nishimura/ALKYL Technology