All the devices that use devwalk have n squared behavior.  It wasn't
supposed to be a problem because n isn't supposed to be large for them.
I agree that it's short sighted for something with potentially many files
like #e.  Your solution is also O(n**2), though the constant
is much smaller (no devdir inside the loop).  I'll replace the devwalk
with your search and put it on the next update wrap.

Also, on network performance, we played a bunch with TCP speed after
Dong Lin showed us how slow we were.  Last we tested our TCP stack,
we came out somewhere between Linux and FreeBSD in speed between 2
500 MHZ Pentiums on a 100meg ether.  That's partially why loopbackmedium
appeared, to see how much was stack and how much driver.  I'll run
a new set of tests in a few weeks and post when I'm done.
Since both all 3 OS's are moving targets, the validity
of the tests is transient.  Sic transit perfunctio.