rhealey@kas.helios.mn.org (Rob Healey) wrote:

   > > you wrote in <95Jul30.182102edt.28(3)@picton.eecg.toronto.edu>:
   > > > Any last words before I test, then send-pr the following?
   > > 
   > > I think the problem is the Hydra driver. So you should fix it and
   > > not fiddle arround with "param.h".
   > 
   > The Hydra driver is an exact (more or less) copy of the ISA wd8003 driver.
   > It expects to be able to get an entire packet in one mbuf.  "Fixing"
   > this is non-trivial.
   > 
  Other ports use the 2K default for performance reasons, i.e.
  ethernet packets fit in one mbuf and you avoid the performance
  loss of mbuf chains. It might not be much of a difference but I
  don't have enough time free to take measurements.

Well, its ok for me to avoid chains if possible. But the tuning
decision about how big MCLBYTES is (which is kernel global) has
nothing to do with the sanity check for the chip ring buffer pointers.
In this case, the driver is in error, independent of the size of
MCLBYTES. I submitted a patch for this the day before yesterday.

  I did up it to 2K once and rebuild EVERYTHING. There was a little
  improvement in NFS but not enough for me to retain the change when
  param.h was altered later.

The problem is, that for medium sized packets (len > MINCLSIZE && len
< old MCLBYTES) you use up 100% more buffer space ;-) The decision
about this should not be based on buggy drivers.


Regards,
	Ignatios Souvatzis