you can look at it both ways.  i like the splice() idea.  i want to,
at the very basic level, join two fd's and pass the data between them.
that's a common, useful operation.


but you are right about UVM -- we should be able to provide zero-copy
IO.  chuck has some test-data on a 'null' socket protocol using the
page loan and/or transfer (i forget, it has a long time ago i read this
chapter :) mechanisms that, once you got beyond the page-sized data
manipulations, the loan/xfer stuff *really* shows.  i know that thorpej
was interested in looking at this but he's a busy guy :)  it would be
really neat if someone started using the cooler UVM features around
the kernel now that mach is gone :-)



BTW:  the splice() thread has been hashed out here before;  please check
the archives.