[why flag?]

I'd do both.

>Um, how about "so you can enable it for one device and not another"? :-)

good reason. but it ocnflicts with 

>How about "flags are patchable"?  :-)

Not very patchable in practice. You have to find the right ioconf
entry, which is nontrivial unless you have the config file or you can
reverse-enigneer the parent array.  Ugh.  And if the multiple devices
are wildcarded and on the same parent, you're still scrooed.

But even a novice can patch a A kernel variable, and the initial value
can be set by an option.

If the common case is (a) frobbing a precompiled kenerel, probably a
[GENERIC one, and (b) all IDE controllers on the same kind of parent
bus -- all PCI or all ISA -- then a kernel variable to turn on DMA for
IDE controllers which arent specifically recognised seems better.


For the wizards who need complete control in their config files, I'd
use two flag bits to force either no DMA or DMA, irrespective of the
whether the device is recognized as "dma supported" or the "default"
variable.