Skip to contentSkip to conversation

e1000 TX unit hang (redux)

DISCUSSION [email protected] 3 participants 3 messages Jul 11, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hello All,

I have an e1000 card periodically misbehaving with the message 'Detected Tx
unit hang'. I've noticed this problem come up on netdev a couple of times
and found the link to the bug tracking page--
http://sourceforge.net/tracker/index.php?func=detail&aid=1463045&group_id=42302&atid=447449

I've also seen the patch that I believe was placed in 2.6.16 and subsequently brought down to 2.4.2? that seems to address this problem by creating a tx_timeout_factor relative to the speed of the NIC. However, there is no mention of this workaround/fix on the bug at the link above and I haven't
found any discussion of it here on netdev. Auke recommends turning off tso
to see if that resolves the problem and this also seems to work, though I have as yet not been able to confirm this and would prefer a more performance friendly fix..if available ;)

Would one of you pplease give an update on the status of the bug? If a cause was ever found and if the tx_timeout_factor was intended as a fix or
temporary workaround? I feel like I must have missed something, because I
never saw the tx_timeout_factor patch go through netdev at all..

Thanks again for your help,
Shaw

#1 lore raw
Quoted 4 lines · [email protected] wrote:

I have an e1000 card periodically misbehaving with the message 'Detected Tx unit hang'. I've noticed this problem come up on netdev a couple of times and found the link to the bug tracking page--
http://sourceforge.net/tracker/index.php?func=detail&aid=1463045&group_id=42302&atid=447449

I've also seen the patch that I believe was placed in 2.6.16 and subsequently brought down to 2.4.2?

that's not only impossible but also unlikely - we don't push changes to 2.4 kernels anymore a lot, I think the last change is likely older than 2.4.28 or so.

Quoted 2 lines

that seems to address this problem by creating a
tx_timeout_factor relative to the speed of the NIC. However, there is no mention of this workaround/fix on the bug at the link above and I haven't found any discussion of it here on netdev.

I wouldn't even know what patch you are talking about (?!)

Quoted 3 lines

Auke recommends turning off tso to see if that resolves the problem and this also seems to work, though I have as yet not been able to confirm this and would prefer a more performance friendly fix..if available ;)

Would one of you pplease give an update on the status of the bug? If a cause was ever found and if the tx_timeout_factor was intended as a fix or temporary workaround? I feel like I must have missed something, because I never saw the tx_timeout_factor patch go through netdev at all..

One possible problem is a bad EEPROM bit, where the hardware might have been misconfigured. This only affects _some_ older e1000's. Any bugreport therefore should include the output of ethtool -e ethX (as well as the lspci -vv output of course. If you haven't already done so, please submit this to the bugtracker or to us by e-mail

Cheers,

Auke

#2 lore raw

Hi Auke,

Quoted 8 lines · Auke Kok
Quoted 6 lines

that seems to address this problem by creating a

tx_timeout_factor relative to the speed of the NIC. However, there is no mention of this workaround/fix on the bug at the link above and I haven't found any discussion of it here on netdev.

I wouldn't even know what patch you are talking about (?!)

Ok, well, the patch is in 2.6.17.4 and looks to have been announced in the 2.6.16-c2 changelog -- http://lwn.net/Articles/170529/ -- and written by Jeff Kirsher. I haven't been able to find a link to the original patch submission anywhere. The code looks something like this now:

        /* Detect a transmit hang in hardware, this serializes the
         * check with the clearing of time_stamp and movement of i */
        adapter->detect_tx_hung = FALSE;
        if (tx_ring->buffer_info[eop].dma &&
            time_after(jiffies, tx_ring->buffer_info[eop].time_stamp +
                       (adapter->tx_timeout_factor * HZ))
            && !(E1000_READ_REG(&adapter->hw, STATUS) &
                 E1000_STATUS_TXOFF)) {

..where the tx_timeout_factor has been added and is set in the watchdog code based on the link speed.

Quoted 2 lines

that's not only impossible but also unlikely - we don't push changes to 2.4 kernels anymore a lot, I think the last change is likely older than 2.4.28.

I'm sure you're right. Jumped to conclusions on a patch I saw posted at redhat.. I'll be more careful next time :)

I'll also try to get some better debugging info from my side.

Thanks.
Shaw