> During a warm boot, the chips' DMA engine might be still active.
> The BIOS can switch off all the bus master bits before the
> OS is loaded again, without doing a hard reset, to save the
> fresh OS brom being corrupted.

That's a good point.

I think the right approach is for the BIOS to clear that bit on all 
devices and then for the drivers to turn it on, but only after they 
have setup things so the device won't do anything dumb. 

The code I looked at all turned the enable on very early in the initialization 
routine - clearly too early to avoid this problem.  (The window would 
be pretty small, but I've seen lots of troubles sneak through tiny 
cracks.) 



> It was N slots behind a PPB.  (a DEC PPB, a 21050 iirc).  At least one
> of them was indicated by the manual and a sticker as "dma not
> supported in this slot."

I haven't been tracking PPBs for a while.  I think the 21050 (first 
one from DEC) had 7 Request-Grant pairs.  The next one had only 4.  
That let them use a much cheaper package.  (It covers the quad Tulip 
case.) 

Unless the BIOS makes a copy of that sticker available it doesn't 
do any good at the level we are discussing.  If somebody drops a 
card in there that tries to do DMA it won't work - with or without 
the master-enable flag being set.  And the symptoms will be pretty 
mysterious.  (Maybe less so if this discussion gets added to some 
web page.)