In message <200111272233.fARMXgp39283@mrynet.com>, you write: 

-> I've discovered what appears to be a correlation between console
-> output and seemingly random panics.  When doing a trivial execution
-> which does a lot of output, such as:
-> 	# cd /usr/src
-> 	# make build >&BUILD& tail -f BUILD

Hmm, my Challenge S has been doing this for a few days (on and off)
and I've seen no problems.  This is the R4400 or the R5k?

[...]
-> If I avoid all console output on the serial console, I can have
-> a kernel running for days on the INDY (this one is an R5K).
-> If I DO perform much work which outputs to the serial console,
-> that time is cut down to hours, if not minutes.

I'd suspect the fact that the cache handling on R5k (and R4600) CPUs
in SGI machines is `odd' and has never been tested on a box with a
secondary cache.  I meant to check out the post-thorpej-mips-cache
code on my R4600 Indy but haven't had the time.

[...]
-> Nonetheless, I wanted to present this bit of information as
-> food for thought.  Perhaps it might pique someone's interest
-> in the meantime to see if they can duplicate the problem with
-> kernels (has been happening from as far back as I can 
-> remember, namely netbsd.IP22, and continues to date with a
-> current kernel).

Hmm, this is interesting.  I'll try and make sure to keep serial
output going on my Challenge to see if I see similar effects there,
but it's been quite stable (except for a mystery pmap panic that
I blew off a few weeks ago because I needed the machine to do some
other work).  Maybe I'll also try and fire up my old, slow R4600
box and see how it fares these days.

--rafal

----
Rafal Boni                                                  rafal@mediaone.net