Network timing (?) problems since kernel 3.12

[mk@drkrebs.de (Markus Krebs)]

Am 12.11.2014 um 20:48 schrieb Simon Baatz:
> Hi Markus,
>
> On Sun, Nov 09, 2014 at 05:56:37PM +0100, Markus Krebs wrote:
>> Am 09.11.2014 um 16:55 schrieb Andrew Lunn:
>>
>>>> I'm using a Sheevaplug (Marvell Kirkwood) with an USB-Stick for
>>>> storage as a server. Since kernel 3.12 I'm experiencing various
>>>> erratic problems when transferring files, especially large or many
>>>> files at the same time, over the network (ssh complaining about
>>>> "Corrupted MAC on input" or Windows [over Samba] about corrupt
>>>> files).
>>>> Turning "offload check summing" off with ethtool didn't really help.
>>>
>>> Try turning off gso.
>>
>> Thanks, Andrew; I've tried now but that didn't do the trick either.
>
> I am not able to provoke this behaviour on my Sheevaplug using a
> stock 3.12.32 kernel (but that does not mean much.  Just did some
> quick tests by copying large files from an USB stick using ssh or
> netcat).
>
> Could you elaborate on the actual kernel and config you are using?
> Is there a simple testcase that reliably reproduces the behaviour on
> your box?
>
> - Simon
>

Hi Simon,

the actual working kernel I'm using is "3.11-2-kirkwood #1 Debian 
3.11.10-1 (2013-12-04) armv5tel". It never fails on anything.

When I upgrade to a newer kernel (3.15, 3.16, 3.17 ..), my testcase 
is/was to do a unison synchronisation run: I synchronize a large 
directory on the server with a backup directory on a Windows machine. 
Unison is invoked from Windows, while the Linux unison server is 
waiting, over ssh. Immediately after invoking unison from Windows, it 
crashes, when the server runs a newer kernel. This is reproduceable.
Being curious what could be the problem, I have synchronized two Linux 
servers with unison, again over ssh. It then gave the error "Corrupted 
MAC on input".
This "unison over ssh"-scenario is my testcase, as when copying files, 
the error is not consistently reproduceable.

- Markus