> Yes, I noticed this for SCSI (it's the disconnect/reselect, rather than
> command queuing, which is the killer). But most scsi controllers are 
> intelligent these days, and the complexity overhead is mostly handled by the
> controller; from a CPU point of view it's mostly a latency overhead.
> With IDE, the CPU will have to do all the work. So we'll have more latency,
> and more CPU time spent in interrupt handlers. But I think it's worth it,
> with eventually a way to disable it.

CPU time is to all intents and purposes more of interest in NetBSD than
anywhere else because NetBSD covers older, slower, platforms. The number of
instructions spent at 900MHz can be an awfully large amount.