I've made couple of test now that my 3c529 is usable, and observed
following things (on 386DX):

* the  mbuf preallocation code indeed helps on the machine - I gain about
  extra 80KB/s (or ~9.5%) for ftp test. It doesn't improve netio ratings,
  though. Thus, this should be kept.
  It's interesting to note the mbuf preallocation doesn't help as much
  as Herb reported. So it seems the mbuf allocation in -current
  is (much) faster than it used to be.
* Moving the mbuf data space allocation out of the splsched()-guarded
  section improved the performance for non-ftp test slighly - netio reported
  like 1-2% improvement. Also, nfs boot (to singleuser) was much faster
  with this change, my estimate is about 2-3 _times_ (not measured).
  There doesn't seem to be any change in speed for ftp test, though.
  Since this looks like a good change, this will go in (this would even
  solve an ancient PR - kern/2478)
* RX Early code doesn't help performance any; actually, it slighly hurts
  (like 0.1%). Since such optimization is targetted on 'slow' systems
  and it doesn't help, it won't go in.

Record of test results is available at
	ftp://ftp.NetBSD.org/pub/NetBSD/misc/jdolecek/ep_perf.txt

The current version of patch is available at
	ftp://ftp.NetBSD.org/pub/NetBSD/misc/jdolecek/ep-2.diff
It's against most current version of sys/dev/ic/elink3.c (rev. 1.92).

If you have appropriate hw, please test the change.  Unless there
would be some problems found, that change will go into tree later
this week.

Jaromir
-- 
Jaromir Dolecek <jdolecek@NetBSD.org>      http://www.ics.muni.cz/~dolecek/
NetBSD - just plain best OS! -=*=- Got spare MCA cards or docs? Hand me them!