If you put some serious intelligence on the board, you can even have the
   board do scatter/gather DMA directly into mbufs.  I'm a little leery of that,
   since you need to have a piece of "hardware" understand the layout of a 
   software structure, which might change -- [...]

Mbuf `clusters' use an optional function pointer to a `free' routine,
so you can DMA to and from the `cluster' area, and not have to teach
the board about the structure layouts.  This is also advantageous
because you can make your `cluster' whatever size is convenient.