[vortex] Problem with 3c982 Cyclone Network Card, Dual AMD.

[Heflin, Roger A.]

> Synopsis of Problem:
> 
> Currently, we have a series of Dual AMD systems, and we are
> experiencing network problems
> with the onboard 3c982 Cyclone 3Com NIC's.  Following is a synopsis of
> the problems, hardware
> involved, and different drivers and methods we have used to determine
> the problems.
> 
> Hardware:
> 
> 3c982 3Com Integrated NIC's (Dual Server Version)
> Tyan Thunder K7 (S2462) with AMD 760MP chipset
> Integrated ATI video
> Standard EIDE
> 
> We are currently using a modified RedHat 6.x installation with the
> 2.2.19 kernel.  We also have
> several patches applied to the kernel (NFS, autofs, serial console
> with serial request keys,
> Dual AMD IDE boot patch).  
> 
> Here is what we are seeing:
> 
> We started with 3Com's 3c90x-102 driver, which is stable, but produces
> hideously slow network
> performance, especially across NFS (700KB - 1.5 MB writes).  Once this
> problem was noticed,
> we then changed to Don Becker's latest 3c59x driver (Ver 0.99Uc).
> This driver provides much better
> performance (at least 5x's faster), but is unstable.  If a large
> write, or several writes, occur to an
> NFS share from a node with this driver, the 3c59x driver will hang,
> producing the following error output:
It has been verified that non-nfs traffic does produce slower results,
but are not as badly affected as
with nfs.

> eth0: transmit timed out, tx_status 00 status e000.
> 
> I have also tried the latest 3c59x non-test driver (Ver 0.99U) and
> have the same problem.
> 
> Eventually, we settle on 3Com's 3c90x-1.0.0i driver.  This still
> provides horrible performance, but at
> least is stable compared to the 3c59x driver.
> 
> Other things we have tried:
> 
> Supplied the "noapic" keyword on lilo boot prompt.
> Checked ECC and ECC Scrub versus no ECC settings for memory.
> Tried mii-diag to determine if auto-negotiation is working (appears to
> be).
> 
> Is there something that we are missing, some later level patch needed
> for this card or for our
> setup?  We are also using the pci-scan that comes with the 0.99U
> driver.
> 
We are using Cicso 2948G switches on the other end set to auto on duplex
and speed.
The swiches appear to have selected 100-FD, and output from mii-diag is:

Using the default interface 'eth0'.
Basic registers of MII PHY #24:  3000 782d 0041 6800 0541 41e1 0007
2801.
 Basic mode control register 0x3000: Auto-negotiation enabled.
 You have link beat, and everything is working OK.
 Your link partner can do 41e1: 100baseTx-FD 100baseTx 10baseT-FD
10baseT

> Thank you for your effort in this matter.