[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

link-up link-down ethernet switch tale of woe



Thought I would post this to debian as someone on the list or out there
in the webiverse might find it useful.

My story begins with an Intel MB with Realtek R8169 ethernet device, an
MPC55 nvidia ethernet device on a different PC and, finally, what turns
out to be a piece of sh*t linksys ethernet switch.  Everything is
supposed to be operating 1000.

I had a heck of a time with the Intel MB because Debian stable's kernel
was too old to support it properly but the unstable install did the
trick.  The other PC goes smoothly, once I give 'noapic' as a boot
param. *sigh*.

Now everything is "working" except both computers are giving me
link-up and link-down messages in /var/syslog.  NFS is going away for a
minute at time and iperf says I'm getting about 70MB/s. Well that ain't
1000 MB/s.

After looking around a bit I discover ethtool and use it to attempt to
turn auto-negotation off.  Well it seems to work on the realtek but the
forcedeth driver gives me messages that are totally inconsistent with
what it's telling me should be happening. I upgrade the install and it
becomes MORE broken. Now it's giving me errors that it didn't give me
before. So much for ethtool, it's a great idea, if it would only
work.  The distressing part is that the forcedeth is on a 3 year old
pc.  So much for the linux and old hardware meme.  Although in
fairness, I think that forcedeth is a relatively recent driver which
allows the driver to be genuine free software instead of nvidia
binary-only. A virtual beer to the device driver writer if that's the
case.

Finally I give-up and plug in both ethernet connections into my
wrt54g.  I figure reliable 100 is better than sh*t 1000.

Well that fixes the problem completely.  I'm getting reliable 100MB/s
connection (iperf says 90MB/s) without ethtool or link-up/link-down
messages.

So I order up a new switch (HP procurve 1400-8G). It arrives, I cross
my fingers and plug in the ethernet cables.  1000 lights come on and
iperf tells me I'm getting 700MB/s, solidly.  Yeah for HP, and
supposedly I've got a lifetime warranty (sorry for the commercial
message, but hey, it worked !).

I take the linksys apart, and here's where things get kind of
interesting.  Those of you who have any hardware experience with
ethernet phy's are probably aware that they typically use a 25MHz clock.
This crazy piece of dung has a 25.0006 MHz crystal.  That is not a
typo.

I've heard rumors of manufacturers overseas grabbing anything in the
way of parts so they can ship it out the door, and I _know_ they don't
test.  Too expensive, it's cheaper for you to test it. Probably happens
in the US too - I mean it would if we actually made anything anymore.

I'm really interested if anybody with some hardware experience can shed
light on the crystal I found.    I even measured it in the lab and it's
25.0006MHz alright.  I'm 99% sure that it's the wrong frequency for
proper ethernet operation, but the level of incompetence that it implies
is amazing.

So morals of the story:

iperf is your friend !
Good hardware is STILL hard to find.
Give the linux drivers the benefit of the doubt (unless it's video).

HTH some poor soul who finds themself in a similar situation.

Brian


Reply to: