e1000 problems in 2.6.x

e1000 problems in 2.6.x

[Steve Simitzis]

hello. since upgrading to 2.6 from 2.4.22, my e1000 devices have
been angry. i've tried running both 2.6.2 and 2.6.3-rc2, but the
results have been the same. when i return to 2.4.22, all of the
problems go away.

i'm getting these over and over again:

Feb 11 19:57:55 sg1 kernel: NETDEV WATCHDOG: eth1: transmit timed out
Feb 11 19:57:57 sg1 kernel: e1000: eth1 NIC Link is Up 100 Mbps Full Duplex
Feb 11 19:58:17 sg1 kernel: NETDEV WATCHDOG: eth1: transmit timed out
Feb 11 19:58:18 sg1 kernel: e1000: eth1 NIC Link is Up 100 Mbps Full Duplex

i'm also seeing about 10-20% of my RX packets as errors when running 2.6.x,
and 0 errors when running 2.4.22.

another oddity is that even after forcing my interfaces to 100 Mbps
full duplex, my switch is reporting half duplex. again, this only happens
in 2.6.x. when running 2.4.22, full duplex is properly negotiated between
the e1000 and my switch.

i tried turning off TSO, as suggested elsewhere by Scott Feldman, but that
had no effect.

i would love to run 2.6, so if there's anything else i can try (or if
this is a known problem and about to be fixed), please let me know.

thanks in advance.

:)

-- 

steve simitzis : /sim' - i - jees/
          pala : saturn5 productions
 www.steve.org : 415.282.9979
  hath the daemon spawn no fire?

RE: e1000 problems in 2.6.x

[Feldman, Scott]

> another oddity is that even after forcing my interfaces to 
> 100 Mbps full duplex, my switch is reporting half duplex. 
> again, this only happens in 2.6.x. when running 2.4.22, full 
> duplex is properly negotiated between the e1000 and my switch.

Are you forcing both the e1000 interfaces and the switch ports to the
same forced settings?  A duplex mismatch would cause problems, but I'm
not sure why this is happening for 2.6 only.  What happens if you don't
force settings, and just rely on autoneg?  (Again, on both ends of the
wire).

-scott

Re: e1000 problems in 2.6.x

[Steve Simitzis]

i should have mentioned in my email that i tried every combination of
settings: auto-neg on the box and forced on the switch, both forced
(to the same settings, of course), forced on the box with auto-neg on
the switch, and auto-neg on both sides. in all cases, the result was
the same: RX packet errors and the same watchdog messages. what i thought
was particularly strange was that the switch refused to auto-negotiate
full duplex.

(someone on another list suggested to try booting with the noapic
option, which i tried for good measure, even though i don't see
how that could have helped. again, no improvement.)

i was initially tempted to assume they were hardware problems, until
i noticed that the problems only appeared when running a 2.6 kernel.
also, the machines are only a few months old, so i'm assuming that
all the hardware is modern enough to do the right thing.

if there's anything you'd recommend trying out, please let me know.
i considered trying to run the 2.4.22 driver in a 2.6 installation,
but i didn't know if any of the driver code depended on anything else
elsewhere in the kernel source.

On 02/14/04, "Feldman, Scott" <scott.feldman@intel.com> wrote: 

> > another oddity is that even after forcing my interfaces to 
> > 100 Mbps full duplex, my switch is reporting half duplex. 
> > again, this only happens in 2.6.x. when running 2.4.22, full 
> > duplex is properly negotiated between the e1000 and my switch.
> 
> Are you forcing both the e1000 interfaces and the switch ports to the
> same forced settings?  A duplex mismatch would cause problems, but I'm
> not sure why this is happening for 2.6 only.  What happens if you don't
> force settings, and just rely on autoneg?  (Again, on both ends of the
> wire).
> 
> -scott

-- 

steve simitzis : /sim' - i - jees/
          pala : saturn5 productions
 www.steve.org : 415.282.9979
  hath the daemon spawn no fire?

Re: e1000 problems in 2.6.x

[Bill Davidsen]

Steve Simitzis wrote:
> i should have mentioned in my email that i tried every combination of
> settings: auto-neg on the box and forced on the switch, both forced
> (to the same settings, of course), forced on the box with auto-neg on
> the switch, and auto-neg on both sides. in all cases, the result was
> the same: RX packet errors and the same watchdog messages. what i thought
> was particularly strange was that the switch refused to auto-negotiate
> full duplex.
> 
> (someone on another list suggested to try booting with the noapic
> option, which i tried for good measure, even though i don't see
> how that could have helped. again, no improvement.)
> 
> i was initially tempted to assume they were hardware problems, until
> i noticed that the problems only appeared when running a 2.6 kernel.
> also, the machines are only a few months old, so i'm assuming that
> all the hardware is modern enough to do the right thing.
> 
> if there's anything you'd recommend trying out, please let me know.
> i considered trying to run the 2.4.22 driver in a 2.6 installation,
> but i didn't know if any of the driver code depended on anything else
> elsewhere in the kernel source.
> 
> On 02/14/04, "Feldman, Scott" <scott.feldman@intel.com> wrote: 
> 
> 
>>>another oddity is that even after forcing my interfaces to 
>>>100 Mbps full duplex, my switch is reporting half duplex. 
>>>again, this only happens in 2.6.x. when running 2.4.22, full 
>>>duplex is properly negotiated between the e1000 and my switch.
>>
>>Are you forcing both the e1000 interfaces and the switch ports to the
>>same forced settings?  A duplex mismatch would cause problems, but I'm
>>not sure why this is happening for 2.6 only.  What happens if you don't
>>force settings, and just rely on autoneg?  (Again, on both ends of the
>>wire).

1 - check your cables in case 2.6 is checking (or not) something
2 - set your NIC half to match the switch and see if there's a different 
problem.

Switches and NICs, especially nice new ones, don't always or even 
usually auto-auto, at least with any of the stuff I have. I have one 
system which works well full when the switch negotiates full and the NIC 
gets half. Any other force usually works poorly. If the NIC comes up 
helf just kick it into full in rc.local (ethtool, mii-tool, whatever) as 
long as the switch convinces itself that it's full.

I see this with 2.4 as well, so I'm familiar if not expert.

-- 
bill davidsen <davidsen@tmr.com>
   CTO TMR Associates, Inc
   Doing interesting things with small computers since 1979

Re: e1000 problems in 2.6.x

[Steve Simitzis]

On 02/14/04, Bill Davidsen <davidsen@tmr.com> wrote: 

> 1 - check your cables in case 2.6 is checking (or not) something
> 2 - set your NIC half to match the switch and see if there's a different 
> problem.

rebooting to 2.4.22 results in a perfectly working network connection,
even with auto-negotiate on both the card and the switch. so i am
hestitant to blame the cables or the hardware. unless, of course, my
cables have the ability to detect which OS i'm running. :)

the result is 100% predictable: boot to 2.6.x, network problems. change
the settings on the device or the switch - network problems. boot
back to 2.4.22 with auto-negotiate everywhere, perfect connection. alas.

-- 

steve simitzis : /sim' - i - jees/
          pala : saturn5 productions
 www.steve.org : 415.282.9979
  hath the daemon spawn no fire?

Re: e1000 problems in 2.6.x

[Steve Simitzis]

that doesn't explain it for me. i have NAPI enabled in the kernel, and
i still have problems in 2.6.x, and not in 2.4.22.

CONFIG_E1000=m
CONFIG_E1000_NAPI=y

:(

On 02/15/04, Elikster <elik@webspires.com> wrote: 

> Hello Klaus,
> 
>    Hmmm..that might explains it, since I don't use NAPI enabled on the E1000 and it works fine without in the 2.4.x series kernels, but under 2.6, it barfs half the time.
> 
> Sunday, February 15, 2004, 8:20:57 AM, you wrote:
> 
> KD> I have a Tyan-S2665 mobo which has an intergrated
> KD> e1000 and I never saw such errors with 2.6 kernels. 
> 
> KD> I use it with 100-MBit Full-Duplex in a switched 
> KD> private network.
> 
> KD> CONFIG_IP_MULTICAST=y
> KD> CONFIG_E1000=y
> KD> CONFIG_E1000_NAPI=y
> 
> 
> 
> 
> -- 
> Best regards,
>  Elikster                            mailto:elik@webspires.com
> 
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

-- 

steve simitzis : /sim' - i - jees/
          pala : saturn5 productions
 www.steve.org : 415.282.9979
  hath the daemon spawn no fire?

Re: e1000 problems in 2.6.x

[Paul Blazejowski]

[-- Attachment #1: Type: text/plain, Size: 420 bytes --]

Hi Steve,

I am experiencing same problem with e1000 on 2.6.3-rc2-mm1 and all
previous 2.6 kernels with getting the transmit timed out messages.I do
not get the transmission errors though.The same setup works fine on 2.4
kernel just like yours.

This happens on Slackware 9.1 on NFORCE2 based Gigabyte 7NNXP board with
OEM Intel 1000MT builtin chip.

Have you found any solutions yet? 

Regards,

Paul B.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: e1000 problems in 2.6.x

[Anton Blanchard]

 
Hi,

> I am experiencing same problem with e1000 on 2.6.3-rc2-mm1 and all
> previous 2.6 kernels with getting the transmit timed out messages.I do
> not get the transmission errors though.The same setup works fine on 2.4
> kernel just like yours.

We beat a recent 2.6 up with specweb on ppc64 and got the transmit
timeout errors within about 30 seconds. Disabling TSO made the problem
go away. (I actually hacked the driver forgetting we can do it via
ethtool these days).

Scott: it smells like the TSO early write back issue, perhaps the new
version of the fix isnt working properly. I need to verify that turning
TSO on/off via ethtool has the same effect.

Anton