> That is an oversimplification.  There are two problems with the VLB cards.
> DMA to pages above 16MB that fall into a "mirror" of the BIOS window fail
> on some revisions.  The other is that transfers that wrap a 16MB boundary
> additionally fail on other early versions.  Bouncing all memory above 16MB
> is a "sledgehammer" approach to fixing these problems.

and how many of these machines are there?  and are any new ones being built?

i'm happy to see support added to support historical systems (like VAX dilog
serial controllers that aren't quite DHV11's) if the hardware in question
was once dominant and expensive, or if someone just feels like plinking at
it and it's harmless to the rest of the system.

but to change an API design in a way that impacts the rest of the drivers,
one would have to show that the change is going to help the system grow to
be more useful to more NEW users.  i've got to say that VLB only *deserves*
a sledgehammer at this point.  let's support it, but let's not change every
other part of the system to make it possible to eke out the last erg of
performance from it.

so this fails, with me, as an attempt to justify the proposed API change.
the API change may yet be proven useful and nec'y -- but not because of VLB.

arguments of the form "but this generalization, however painful on a few
architectures, means that drivers are more portable to other architectures"
hold a lot more weight, but for the same reason.