Finally a clue:

ahc0: WARNING no command for scb 21 (cmdcmplt)
QOUTPOS = 23
 
Unfortunately that's the only thing in /var/log/messages, as that's when
the disks froze. Then plenty zipped up the screen, the last of which
was (fat fingered):
...

Sequencer SCB Info: 
  0 SCB_CONTROL[0xe8] SCB_SCSIID[0x27] SCB_LUN[0x0] SCB_TAG[0xff]
  1 SCB_CONTROL[0x6c] SCB_SCSIID[0x7]  SCB_LUN[0x0] SCB_TAG[0xff]
  2 SCB_CONTROL[0xe8] SCB_SCSIID[0x27] SCB_LUN[0x0] SCB_TAG[0xff] 
Pending list:
  8 SCB_CONTROL[0x68] SCB_SCSIID[0x27] SCB_LUN[0x0]
 10 SCB_CONTROL[0x68] SCB_SCSIID[0x27] SCB_LUN[0x0]
 29 SCB_CONTROL[0x68] SCB_SCSIID[0x27] SCB_LUN[0x0]
Kernel Free SCB list: 6 1 23 11 18 20 17 31 26 16 19 0 12 25 28 30 47 24 22 27 7
 14 13 2 15 5 9 21 4 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
 
<<<<<<< Dump Card State Ends >>>>>>>>
sg[0] - Addr 0x50215000 : Length 1024
ahc: Queuing a BDR SCB
ahc0: Bus Device Reset Message Sent
sd1(ahc0:0:2:0): ahc0: no longer in timeout, status = 0
ahc0: Bus Device Reset on A:2. 3 SCBs aborted



This time /var was corrupted too => softdeps was a red herring.

Looking through the archives, I had roughly the same thing around 15th April
2000 on current-users. It's the same disk & card. There the conclusion seemed
to be "disable tagged queueing", however things must have improved since then,
as I'm sure I have been using tagged queueing at least since September 2001.

BTW, it isn't just the freezing which is a problem. All I need to do is
newfs /usr/obj, make build for a bit, fsck -f /usr/obj, and fsck will report
failures. (This is the new post ufs2 fsck - no problems there - all fine with
16th April kernel & userland)

(and on i386 what is the magic key combination to get into ddb? I thought it
was ctl-esc, but that doesn't get me there, despite ddb.{onpanic,fromconsole}=1
.. hmm must install a serial console..)

Hints and Tips?

Cheers,

Patrick