This is under 3.1, a rebuilt kernel but built from absolutely stock
source and the stock GENERIC config file.

On my WorkPad, running diskless, I did a bulid of the world.  (This
took about a week....)  Today, to put it on a CF card, I first created
a file (on the NFS server) containing a dd image of the CF card.  Then
I did "vnconfig vnd0 /wd0.image" and did a newfs of vnd0a, then mounted
it on /mnt.  Semeed to work.  So I started to do an install of the
build, into /mnt (env USETOOLS=yes TOOLDIR=... ./build.sh -D ... -O ...
-U install=/mnt).

This started, but after no more than a few minutes the machine crashed:

trap: TLB miss (load or instr. fetch) in kernel mode
status=0xf803, cause=0x8, epc=0x800af274, vaddr=0x0
pid=882 cmd=vnd0 usp=0x0 ksp=0xc54a1d40
Stopped in pid 882.1 (vnd0) at  netbsd:nfs_asynco:0x5cc:       lw      v0,28(v0)
db> 

This happened once before; that time, I had the ffs mounted -o async,
so I tried this time without that.  The first time, when I told ddb to
"reboot", it failed; a second "reboot" reset the machine and I had to
come back in through wince.  This time, "reboot" worked; kloader
managed to load /netbsd OK and it rebooted nicely.

What's wrong here?  Does layering vnd atop nfs just not work, or is
there something more to it?  The whole file exists (it doesn't have any
holes), though that shouldn't matter over NFS AIUI.  Clearly NFS per se
is not respnosible, or it wouldn't've stayed up for a week doing the
build of the world and then crashed within minutes of my doing this;
there must be some bad interaction going on somewhere.

Oh, and, fsck_ffs on the FFS partition on the vnd doesn't cause any
problem (or at least it hasn't yet, though I've tried it only about
three times - but it's run to normal completion).

I'm going to retry with -o sync on the FFS mount (slow is better than
crashing, I figure).

/~\ The ASCII				der Mouse
\ / Ribbon Campaign
 X  Against HTML	       mouse@rodents.montreal.qc.ca
/ \ Email!	     7D C8 61 52 5D E7 2D 39  4E F1 31 3E E8 B3 27 4B