[9fans] Intel 8254[17] and switches

[Brantley Coile <brantley@coraid.com>]

Final update on the etherigbe.c effort.

The 82541 and 82547 have phy's that don't like inexpensive switches!
After a couple of weeks banging my head on the wall, I discover that
the PXE firmware on the things can't even drive 1GbE.  I couldn't find
any mention of this fact in anything from Intel, unless one of the
erata's on a sheet I got states this fact in an obscure way.

At first I noticed that at 100Mb I had to insert a delay after each
frame.  Otherwise it the frames that followed would disappear.  Counts
showed that they were transmitted, but snoopy didn't see anything.  I
shared my work in progress with several people, and they commented on
the resends.  This was because the stack would send 4 frames, drop
three, get acked for the first, wait and send the first 3 over and
another one.  This would finally work, but was very slow.  IL was less
patient.  I added a delay of 150µs and they all get out.

At 1 Gb things got really screwy.  The phy would state that it had to
use MDI-X, that is, it had to switch the cable to get a link.  This
obviously wasn't needed.  The cables were good, and they were plugged
into a switch.  Phy showed that it was linked at Gb, light on switch
concurred.  The MAC registered also agreed, but no data.  Nice
activity lights but nothing that could be snoopyed or forward by the
switch.  Buggers!

So, I would obviously discourage anyone from using these parts.

Does anyone have a driver for the Broadcom BCM5721?

  Brantley