svn commit: r242739 - stable/9/sys/dev/ti

[Andre Oppermann]

On 09.11.2012 17:51, Adrian Chadd wrote:
> On 9 November 2012 00:08, Andre Oppermann <andre at freebsd.org> wrote:
>
>> Firewalling doesn't change the packet and no checksum is needed.
>> NAT does change the packet and the pesky pseudo-header in the TCP/
>> UDP checksum.  However here only the pseudo-header checksum is
>> recalculated and reintegrated into the one-complement payload checksum.
>> The payload itself is not being looked at, except for protocols that
>> do contain IP addresses in their internal commands or such.  There
>> the payload is modified.  The same reintegration trick can be used.
>> In the majority of cases these packets are very small though and
>> the entire checksum is simply recalculated.  As the packets are very
>> small no fragmentation is occuring.
>> The IPv4 header checksum is never ever a problem and always works.
>>
>> Can we please put this to rest now.
>
> Andre,
>
> I'm assuming that Pyun fixed a real bug that he really did see. My day
> job is dealing with these kinds of corner cases with embedded hardware
> that does offload-y type stuff.
> So yes, I'd really like to know what's going on under the hood and fix it.

There isn't anything to fix other than removing a poorly supported
and implemented feature.  Nothings breaks if it goes away.

> You may be able to make it all go away with a bunch of logical
> reasoning but if it weren't a problem, Pyun would've never stumbled
> across it and had to disable checksumming.

I found the problem and wanted to remove this particular mis-feature
from the stack.  To do that he removed it from two drivers, one of them
very old with no known users.

-- 
Andre

> I'll chase this up privately with him and see what was really going on.
>
> Thanks,
>
>
>
> Adrian
>
>