On getting more memory, I patched my sources with the bounce-buffer patch
from ftp://bseis.eis.cs.tu-bs.de/pub/hannken/isadma-patch.  The disk IO
works fine (when it happens, at any rate; I love not paging!!!).  However,
the gods of portability have chosen to frown on my for my sin :-); it turns
out that the interface to isa_dmadone was changed, and the backward
compatability macro added to isadmavar.h is wrong; instead of

#define isa_dmadone(a,s,c,f) isadma_done((c))

it should be

#define isa_dmadone(a,s,n,c) isadma_done((c))

as inspection of the original sources will indicate.  This causes bizarre
behavior on every driver that actually uses the ISA dma channels.  (I was
probably lucky that it didn't just crash my system; all I got was tens of
kilobytes of

Oct 16 23:38:55 jfwhome /netbsd: isadma_done: no request active on 256
Oct 16 23:38:55 jfwhome /netbsd: isadma_start: old request active on 1

in /var/log/messages.)

I'm going to read this stuff more carefully; if I spot any more bugs when
I get that fixed, I'll send mail again.