> Someone made the claim that bounce-buffers need to be supported in a MI
> fashion.  What other machines besides an ISA x86 box make use of boards
> which need to DMA more than 16MB of core? [Chris, was this you?]
> I mean the alphas and the HP9000 series (for which we don't yet have
> a port) have ISA busses in them -- or some of them do.  But are bounce
> buffers CRITICAL to their success?

Yes, at least on the Alpha.

On the Alpha, there is by default _NO_ memory mapped into the 0-16M
range on the ISA/EISA/PCI busses.  If you want memory there, you have
to use DMA mapping registers to get it there.  This is not bounce
buffering, but it's effectively equivalent in terms of the function
calls required.

_ANY_ ISA devices you have on an Alpha which use DMA _must_ be
"bounce-buffered" to work _at all_.  You'll note that the Alpha port
currently does not support floppy or sound drivers.  That is why.


> I'm sure that if I'm blowing smoke, someone will be sure to let me know,
> but before you flame me, forgive me if I don't have a firm grasp on the
> details.  It just seems to me that to go MI _right now_ for bounce buffering
> seems a bit overblown for our current needs:  It's more effort to do a
> MI version than to do a port-specific version, and there's one port which
> needs it more than the others, to the extent that NOT having it is causing
> much unnecessary suffering.  I agree with the premise that one should NOT
> have to buy all new hardware when there is only one component that imposes
> a restriction.

Actually, it's not overblown for "our current needs," if you consider
other ports besides the	i386.  And, i've explained this in detail
several times before, when justifying the need for real MI bounce
buffers.


cgd