net: mv643xx: interface does not transmit after some time

[Adam Baker]

On 11/02/16 14:38, Ezequiel Garcia wrote:
> (let's expand the Cc a bit)
>
> On 10 February 2016 at 19:57, Andrew Lunn <andrew at lunn.ch> wrote:
>> On Wed, Feb 10, 2016 at 07:40:54PM +0100, Thomas Schlöter wrote:
>>>
>>>> Am 08.02.2016 um 19:49 schrieb Thomas Schlöter <thomas at schloeter.net>:
>>>>
>>>>
>>>>> Am 07.02.2016 um 22:07 schrieb Thomas Schlöter <thomas at schloeter.net>:
>>>>>
>>>>> Am 07.02.2016 um 21:35 schrieb Andrew Lunn <andrew at lunn.ch>:
>>>>>>
>>>>>>>> FWIW, we had a similar bug report in Debian recently:
>>>>>>>> https://lists.debian.org/debian-arm/2016/01/msg00098.html
>>>>>>
>>>>>> Hi Thomas
>>>>>>
>>>>>> I this thread, Ian Campbell mentions a patch. Please could you try
>>>>>> that patch and see if it fixes your problem.
>>>>>>
>>>>>> Thanks
>>>>>>    Andrew
>>>>>
>>>>> Hi Andrew,
>>>>>
>>>>> I just applied the patch and the NAS is now running it. I???ll try to crash it tonight and keep you informed whether it worked.
>>>>>
>>>>> Thanks
>>>>>     Thomas
>>>>
>>>> Hi Andrew,
>>>>
>>>> the patch did not fix the problem. After 1.2 GiB RX and 950 MiB TX, the interface crashed again.
>>>>
>>>> Now I switched off RX/TX offload just to make sure we are talking about the same problem. If we are, the interface should be stable without offload, right?
>>>>
>>>>      Thomas
>>>
>>> Okay, so I have installed ethtool and switched off all offload features available. Now the NAS is running rock solid for two days. I backed up my Mac using Time Machine / netatalk (450 GiB transferred) and some Linux machines via NFS (100 GiB total) without a problem.
>>>
>>> How much code is used for mv643xx offload functionality?
>>> Is it possible to debug things in the driver and figure out what happens during the crash?
>>> Is the hardware offload interface proprietary or reverse engineered or is it a well known API that can be analyzed?
>>
>> Hi Thomas
>>
>> Ezequiel Garcia probably knows this part of the driver and hardware
>> the best...
>>
>
> The TCP segmentation offload (TSO) implemented in this driver is
> mostly a software thing.
>
> I'm CCing Karl and Philipp, who have fixed subtle issues in the TSO
> path, and may be able to help figure this one out.
>

Hi,

Had this issue occur again today. In my case it seems to be triggered by 
large NFSv4 transfers.

I'm running 4.4 plus Nicolas Schichan's patch at
https://patchwork.ozlabs.org/patch/573334/

There is a thread a http://forum.doozan.com/read.php?2,17404 suggesting 
that this has been broken since at least 3.16.

I first spotted the issue when upgrading from 3.11 to 4.4.

Looking at 
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/log/drivers/net/ethernet/marvell/mv643xx_eth.c 
I see 2014-05-22 as the date TSO support was first added which is 
shortly before the merge window opened for 3.16. I'm therefore guessing 
that TSO has been problematic since it's introduction.

Regards

Adam