Problem with checksum offloading on RPi3 (PF + Jails involved)

[Kristof Provost]

On 19 Nov 2020, at 3:22, YongHyeon PYUN wrote:
> On Wed, Nov 18, 2020 at 10:47:08AM +0100, Carsten Bäcker wrote:
>
> [...]
>
>> Sorry, that's my fault - i overlooked your request for a test with pf
>> enabled.
>> The example-ruleset is attached again. I added a line to allow the
>> incoming ping to 5555.
>>
>> I don't see a difference until i enable the redirection to the jail
>> which makes the packet with extended data-length fail.
>> Once i disable RXCSUM it works.
>>
>
> Thanks for testing.  Unfortunately I couldn't find a driver issue
> on patched smsc(4).  Quick looking pf(4) makes me wonder how it
> verifies UDP datagrams.
>
> Kristof, does pf(4) take advantage of H/W checksummed result for
> UDP datagrams?
Yes. Pf updates checksums if they’re present, rather than fully 
recalculating them.

> It seems pf_check_proto_cksum() always pass
> IPPROTO_TCP such that UDP pseudo checksum is not computed in the
> function.
Note that pf_check_proto_cksum() is only ever called from pf_return(), 
and then only if pf->proto == IPPROTO_TCP.
Checksum updates are mostly done through pf_proto_cksum_fixup(), in 
different parts of the code.

Best egards,
Kristof