freebsd stable - Jan 2010 - General problems with checksums (txcsum/rxcsum) on FreeBSD 8.0?

I just read a different thread about problems with checksums on vge (and nfe in
the replies).

I'll just chime in here with some more information - I have a couple other
message threads going about some weird high packet volumes on my new FreeBSD
8.0-Release NFS server.  I thought it might be an issue with the igb driver so I
put in a new card using em instead and got the exact same behavior.  I'm
currently sifting through a tcpdump in wireshark and there are all sorts of
messages in there about checksums being incorrect - both TCP and UDP.  This is
for communications between this client machine (FreeBSD 7.0-Release) and any of
the 8.0 machines I have.  The packets going to non-8.0 machines (at least so
far) appear to be fine.

I'll defer to those who know more than I about the networking code, but is
there perhaps an issue in general with the checksuming and not specific to one
card or driver - is that even possible?  That's now 4 different drivers all
with various checksum problem reports.

I'm going to be working on this all day today (and likely over the weekend)
so if I can help by supplying information please let me know what you need.

Hi--

On Jan 15, 2010, at 11:12 AM, alan bryan wrote:
[ ... ]
> I'm currently sifting through a tcpdump in wireshark and there are all
sorts of messages in there about checksums being incorrect - both TCP and UDP.
If you run tcpdump on a machine, it normally will receive the traffic being sent
from that machine before the checksums are computed, especially if HW
checksumming is being used.  For reliable detection of these problems, you need
to look at the traffic either on a hub or via the monitoring or span port of a
smart switch, although simply glancing at the checksum stats from "netstat
-s" on both sides should indicate whether significant error rates are
happening.

Regards,
-- 
-Chuck

On Fri, Jan 15, 2010 at 11:12:43AM -0800, alan bryan
wrote:
> I just read a different thread about problems with checksums on vge (and
nfe in the replies).
> 
> I'll just chime in here with some more information - I have a couple
other message threads going about some weird high packet volumes on my new
FreeBSD 8.0-Release NFS server.  I thought it might be an issue with the igb
driver so I put in a new card using em instead and got the exact same behavior. 
I'm currently sifting through a tcpdump in wireshark and there are all sorts
of messages in there about checksums being incorrect - both TCP and UDP.  This
is for communications between this client machine (FreeBSD 7.0-Release) and any
of the 8.0 machines I have.  The packets going to non-8.0 machines (at least so
far) appear to be fine.
> 
> I'll defer to those who know more than I about the networking code, but
is there perhaps an issue in general with the checksuming and not specific to
one card or driver - is that even possible?  That's now 4 different drivers
all with various checksum problem reports.
> 
> I'm going to be working on this all day today (and likely over the
weekend) so if I can help by supplying information please let me know what you
need.
> 
If you are seeing bad checksum reported by tcpdump/wireshark for TX
frames on checksum capable controller, it's normal. bpf(4) just
sees TX frames before inserting checksum computed by hardware so
tcpdump/wireshark reports invalid checksum. You can safely ignore
that. If you want to verify whether sending host generated correct
checksum, you should capture the frame on receive side. If tcpdump/
wireshark reports bad checksummed frame on received frames it's
real bad checksummed frame.