Re: Data corruption with checksum offloading enabled

Hello,

Dmitry Marakasov <amdmi3_at_amdmi3.ru> writes:

> * Dmitry Marakasov (amdmi3_at_amdmi3.ru) wrote:
>
>> 3. 512MB random bytes with NFS: 2/5 correct
>
> Just for the record: this no longer seems to be a problem. Recent
> 8-STABLE, ale works fine with rxcsum/txcsum, at least I could not
> reproduce the problem with similar tests as before.

bon, I upgraded the box with this problem to *7*-STABLE this
weekend, and no luck, the problem still persists.

To refresh memory, nfs-client has data corrupt in a particular way :

  - just one byte (per file) in my case, a single 128 byte block for Dimitry
  - independent of network driver and NFS-[options|client versions]
  - seems nfs-limited (netcat file transfers work OK (zero-only nfs as well))
  - disabling checksum offloading at least makes it much harder to provoke
     (cann't remembre whether everyone confirmed this)
  - disabling cpufreq makes it impossible to provoke
     (at least for me)

So far the bad news, the better news is that I cannot reproduce
this problem on a similar setup running *8*-STABLE indeed.

The 7-STABLE box is in production, but involved in some
maintenaince/upgrade-shuffle I am responsable for, I will try
to let it at least boot a 8-kernel and 7-world sometime next
week and see if I can still reproduce the problem.

Best, Arno