OK, it's looking like a scsi dma problem.  I've enabled debug messages in
the asc driver.  On one session, telnetted in, I'm running "newfs /dev/rsd0a".
On the console, here's where it goes bye-bye...

...
asc_startcmd: asc0 target 0 cmd a len 8192                                      
asc_intr: status 80 ss c4 ir 18 cond 3:18                                       
asc_dma_out: buflen 8192, len 8192                                              
asc_intr: status 93 ss c4 ir 10 cond 4:310                                      
asc_last_dma_out: buflen 8192 dmalen 8192 tc 0 fifo 0                           
asc_intr: status 97 ss c4 ir 8 cond 7:708                                       
asc_intr: status 90 ss c4 ir 20 cond 8:20                                       
asc_end: asc0 target 0 cmd a err 0 resid 0                                      
asc_startcmd: asc0 target 0 cmd a len 8192                                      
asc_intr: status 80 ss c4 ir 18 cond 3:18                                       
asc_dma_out: buflen 8192, len 8192                                              
asc_intr: status 93 ss c4 ir 10 cond 4:310                                      
asc_last_dma_out: buflen 8192 dmalen 8192 tc 0 fifo 0                           
asc_intr: status 97 ss c4 ir 8 cond 7:708                                       
asc_intr: status 90 ss c4 ir 20 cond 8:20                                       
asc_end: asc0 target 0 cmd a err 0 resid 0                                      
asc_startcmd: asc0 target 0 cmd a len 8192                                      
asc_intr: status 80 ss c4 ir 18 cond 3:18                                       
asc_dma_out: buflen 8192, len 8192                                              

...and that's it.  There's no subsequent intr, last_dma_out, intr, intr, and
end.  Basically, the final dma operation was never acknowledged.  The Magnum
isn't hung, but scsi operations effectively are.
Has anyone seen this before?  Any hints?  This seems to happen with any
sustained scsi i/o.  Small amounts of data, like writing disklabels, etc.,
work fine.  There must be some dma threshold possibly getting screwed up?

Cheers,
-Mark