> Basically, his idea amounts to a complete redesign of the network
> stack architecture, moving most of the code into userland and
> eliminating the two-stage transfer to the user.

Protocol stack should be run in its own 'CPU scheduling' domain
whose discipline is to-be-fair-and-efficient between netifs.  dmr's
STREAM is a pretty ascent idea, but few understand it was driven
by own scheduler independent from process scheduling.  Personally
I think it's a bad idea to move protocol stack to user space.  It'd be
rather better to have the intermediate domain between kernel and
user.  It's subject to research how really-stackable protocol stuff
should be implemented in that context.

Toru Nishimura/ALKYL Technology