Possible sge(4)/atphy(4) regression on RELENG_9?

[Nikolay Denev]

On Nov 28, 2011, at 2:43 AM, YongHyeon PYUN wrote:

> On Sat, Nov 26, 2011 at 04:39:03PM +0200, Nikolay Denev wrote:
>> 
>> On Oct 24, 2011, at 9:18 PM, Nikolay Denev wrote:
>> 
>>> 
>>> On Oct 24, 2011, at 8:52 PM, YongHyeon PYUN wrote:
>>> 
>>>> On Mon, Oct 24, 2011 at 04:43:57PM +0300, Nikolay Denev wrote:
>>>>> Hello,
>>>>> 
>>>>> I've recently upgraded a box running RELENG_8 to RELENG_9 and immediately I noticed much slower network connection.
>>>>> Running iperf shows about 20-30Mbits which was almost full GigE (~900Mbits) speed before.
>>>>> 
>>>>> I'm noticing interface errors :
>>>>> 
>>>>> [16:37]ndenev at nas:~% netstat -I sge0
>>>>> Name    Mtu Network       Address              Ipkts Ierrs Idrop    Opkts Oerrs  Coll
>>>>> sge0   1500 <Link#4>      00:0a:e4:86:62:fa 76114295 42197     0 103559806 10324     0
>>>>> sge0   1500 10.0.0.0      nas               76109575     -     - 119109557     -     -
>>>>> 
>>>>> Both the switch and the card show 1000 full-duplex.
>>>>> I've tried playing with rxcsum,txcsum,vlanhwtag,tso but disabling even all of them do not change anything.
>>>>> I've tried different switch port and changed the cable.
>>>>> 
>>>>> Here is devinfo for my hardware :
>>>>> 
>>>>> sge0 pnpinfo vendor=0x1039 device=0x0191 subvendor=0x103c subdevice=0x2a70 class=0x020000
>>>>> atphy0 pnpinfo oui=0xc82e model=0x1 rev=0x6 at phyno=0
>>>>> 
>>>>> Of course all of this can mean hardware problem, I just want to ask if somebody is seeing something similar, since
>>>>> there are quite a lot minibus related changes as far as I can see.
>>>>> 
>>>>> I'll boot RELENG_8 again tomorrow and do a quick test again to verify that this is not a hardware issue.
>>>>> 
>>>> 
>>>> I don't have sge(4) controller so it would be better to let us know
>>>> which revision introduced the regression.  Just looking over the
>>>> code change didn't reveal the possible cause.
>>>> BTW, I thought sge(4) shall use rgephy(4). Can you also verify
>>>> whether sge(4) in stable/8 also use atphy(4)?
>>> 
>>> I've just checked my logs and I can confirm that it was atphy(4) even in stable/8.
>>> 
>>> Sep 26 15:55:19 nas kernel: atphy0: <Atheros F1 10/100/1000 PHY> PHY 0 on miibus0
>>> Sep 26 15:55:19 nas kernel: atphy0:  none, 10baseT, 10baseT-FDX, 100baseTX, 100baseTX-FDX, 1000baseT-FDX, 1000baseT-FDX-master, auto
>>> 
>>> I'll post more info when I try again stable/8 on this hardware.
>>> 
>>> Thanks!
>>> 
>> 
>> Just for the sake of completeness I'm reporting that the problem turned out to be not hardware related.
>> The thread "TCP Reassembly Issues" in freebsd-stable list describes the issue.
>> 
> 
> Thanks for letting us know that.
> 
> BTW, it seems sge(4) controllers support flow control as well as
> jumbo frame by scattering incoming jumbo frame into multiple RX
> buffers. Jumbo frame support code may require more time in my side
> but flow control support code could be implemented within a week.
> Can you test the patch if I manage to write support code?

Yes, I can test it, my switch supports both flow control and jumbo frames.

Thanks,
Nikolay