Linux network gurus wanted - TX errors on eth0?

Inspector Gadget

Hi folks:

I've been asked to look at a web server box that's misbehaving quite badly - it's on a college network and network transfers are quire sluggish - this i s particularly obvious when retrieving web pages from the machine.

The machine in question is a Celeron 1.2GHz with 256MB of RAM, running RH Linux 7.3; load averages rarely exceed 0.03 on any figure, so the machine certainly isn't overloaded. I can't find anything physically wrong, except for this when I look at ifconfig:

eth0      Link encap:Ethernet  HWaddr []
          inet addr:[]  Bcast:[]  Mask:255.255.0.0
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:17272716 errors:0 dropped:0 overruns:0 frame:0
          TX packets:21993 errors:2150 dropped:0 overruns:0 carrier:4290
          collisions:3339 txqueuelen:100
          RX bytes:1419324212 (1353.5 Mb)  TX bytes:8529015 (8.1 Mb)
          Interrupt:5 Base address:0xdc00

(some info removed for obvious reasons)

What alarms me out of this lot is "TX errors: 2150", or about 100 TX errors an hour (the machine is up for 21h).

I assume that a certain amount of lost/damaged packets are inevitable (the network protocols are designed to correct for this sort of thing), but if I compare this to the same from the RH 9 box six feet away from me, the statistics become more worrying:

eth0      Link encap:Ethernet  HWaddr []
          inet addr:[] Bcast:[] Mask:255.255.0.0
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:7428490 errors:103 dropped:0 overruns:0 frame:0
          TX packets:9132578 errors:1 dropped:0 overruns:11 carrier:2
          collisions:127267 txqueuelen:100
          RX bytes:2716825979 (2590.9 Mb)  TX bytes:708595111 (675.7 Mb)
          Interrupt:10 Base address:0x5000

...which looks a lot healthier (uptime for that box is ~13 days).

Both of these machines are using cheapo cards based on the Realtek RTL-8139 controller - the problem box originally had a Macronics MX98715-based no name NIC in it, the Realtek (U.S. Robotics badged Compustore muck - specifically, an RTL8139D) was installed yesterday, and is performing better than the MX card, but it's still not what I'd consider "right". Or is it?

Another worrying thing is what the problem box reports when I run mii-tool:

eth0: 10 Mbit, half duplex, no link
  product info: vendor 00:00:00, model 0 rev 0
  basic mode:   10 Mbit, half duplex
  basic status: no link
  capabilities:
  advertising:

...even though I'm running mii-tool remotely ("no link"?), which means the network card must be at least sort-of working!

Any suggestions greatly appreciated...
Gadget

fatherdougalmag

Tried a different cable?

Tried a different slot?

Also try swapping in a different (known good) card.

Inspector Gadget

Cable: yes.
Slot: yes - all five (assuming you mean PCI slots).
Known-good card: not yet. Need to "liberate" one from a machine I'm not using. Seems unlikely that I'd buy a brand new card with exactly the same fault though, doesn't it?

Thanks,
Gadget

sirlinux

Could you try another socket on your switch or a different wall socket it looks like cable damage or noise.

spongebob

could be bent pins in the card if a new cable and switch port change together do not fix it.

/me would never put a Realtek in a server, thats what Intel Pro or 3Com905(b or c) cards are for TBH

Much better hardware and MUCH better drivers esp in Linux

M

Inspector Gadget

Hmm...

While there are two ports in the office (both in use) that I could theoretically swap to test this, I have a problem; I can't change the port switch without considerable bureaucracy (switch not controlled by me, plus each port on the switch is mated to a specific MAC address), all of which leads to forms, signatures and permissions.

Personally, I'd tend to agree with you about the terrible choice of network cards; I use 3C905's myself at home, but I'm damned if I'm swapping that for a Realtek... also, the guy I'm dealing with asks another guy intermittently about various things like this, following which I'm instantly overruled, so (for what it's worth) I'm not responsible for recommending the new Realtek.

Thanks,
Gadget

sirlinux

beg/borrow/steal a small hub, switch or router to test it out seperate from the exisiting network if your other ports are mac locked.

hazbot

Card set to 10mbit half, port set to 100mbit?

Apologies if this was already checked, didn't see any mention of it yet

Inspector Gadget

Thanks for the replies, people.

sirlinux: The server is "live" at the moment, so I have to keep downtime to an absolute minimum - and as I'd need to leave the system hooked up to the hub for a few hours to get a feel, that's not really practicable.

(Yes, this means that I was called in to fix it a little late in the day...) :rolleyes:

hazbot: The port is 10mbit full duplex, and the card is currently (it seems) set to 10mbit half-duplex. Need to change that lively, but mii-tool won't change it.

Thanks
Gadget

sirlinux

you could try a newer version of the 8139 module

http://www.scyld.com/network/rtl8139.html

or at least compile the diagnostic tools, there was some versions for this card around with MII disabled.

Inspector Gadget

Just read the page you linked to, sirlinux - and as the card that's in this machine is based on an RTL8139_D_ (as opposed to a _B_ which the driver was designed for) it may not solve the problem.

What might, on the other hand, is to swap the card completely for a 3Com. Just found a 3c905-tx in an old P200 thrown in the corner, so that should help things.

I'll post my progress...
Thanks,
Gadget