[Samba] file differences when copying files to linux, using samba

[Douglas VanLeuven]

jeffunit wrote:
> 
>> >>         Have you tried copying the file over with "cp" from
>> >> windows to your server? (cp from 'cygwin')?
>> >
>> > No, but I will try that today.
>> >
>> >>         Have you tried comparing some of the differing files and
>> >> looking for a pattern?
>> >
>> > Yes. I wrote a modified version of cmp that tries to list all byte
>> > differences.
>> > I was looking at an iso image of some linux distribution.
>> > There were three bytes that differed, and if I recall correctly,
>> > they were all one bit differences.
>> >
>> Hi,
>> I would think one bit differences should be picked up by the TCP
>> transport layer.
>>
>> You probably have Rx checksum offload on the receiving box.  You could
>> try turning that off and recopying to see if the error persists.  Could
>> be a bad card.
> 
> How do I turn rx checksums off under linux?  This is an intel pro1000 ct.
> I am happy to try it.
> The gigabit ethernet is on-board, but I have several spare gigabit nics
> available.

If eth0 is the name:
   ethtool -K eth0 rx off

I got good service out of these settings in modprobe.conf (one line)
options e1000 RxDescriptors=1024 TxDescriptors=1024
    InterruptThrottleRate=1

I have lots of memory, so I upped the buffers, probably overkill.
Default is 256
InterruptThrottleRate defaults to 3 (dynamic conservative) and I changed
it to 1 (dynamic).

see /usr/src/<current kernel>/Documentaion/networking/e1000.txt

Along with larger buffers in smb.conf, I regularly hit my hard disk
limits on gigabyte. And bit error free.  I never did it your way, but I
have run tripwire with checksums and after disaster recovery, haven't
had any issues or noticed any discrepancies with entire drives going
across the wires.

> 
>> Seems unlikely the Tx sender could send an incorrect checksum unless the
>> buffer memory flipped a bit before checksum computation which seems
>> ruled out by the ECC.  Still, I'm a believer in memtest.
> 
> I ran memtest-86+ through about 5 iterations, and there were no problems.
> 
> thanks,
> jeff

Good luck (whatever it is), Doug