Re: TAHI CN-6-4-1 failed on Linux 2.6.32 kernel

On Fri, Aug 20, 2010 at 12:35 PM, Jesse Gross [off-list ref] wrote:

On Fri, Aug 20, 2010 at 1:16 PM, Steve Chen [off-list ref] wrote:

quoted

On Thu, Aug 19, 2010 at 7:06 PM, David Miller [off-list ref] wrote:

quoted

From: Steve Chen <redacted>
Date: Thu, 19 Aug 2010 13:35:14 -0500

quoted

I trace through the code.  It appears that the network driver (e1000e
for my setup) always set ip_summed to CHECKSUM_UNNECESSARY.  I have
been unsuccessful to get the driver to take the other branch where
ip_summed is set to CHECKSUM_COMPLETE.  Even when I hard code
ip_summed to CHECKSUM_COMPLETE, __skb_checksum_complete_head set
ip_summed to CHECKSUM_UNNECESSARY after recomputing the checksum.

So far the only way I'm able to get ICMP to recompute checksum is
through the attached hack.  Even though I can get all the tests to
pass, but it just seem wrong.

If turning off hardware RX checksumming with ethtool has no effect,
and the problem is seen with multiple ethernet cards, the problem
is elsewhere.

First of all, if you turn RX checksumming off, the checksum field
of the SKB should always be skb->ip_summed = 0.  If this is not
happening, find out why.

Ahhh, thats my problem.  I incorrectly thought the ip_summed should be
2.  The ip_summed is set to 1 in
__skb_checksum_complete_head.  Looking at the code, shouldn't

if (likely(!sum))

be

if (likely(sum))

Since sum == 0 would indicate an error?

sum == 0 indicates that the checksum is correct.
If you compute the checksum of a packet containing the correct checksum the
result is 0.  It's like a slightly more complicated varient of a parity bit.

It appears the issue lies somewhere within CONNTRACK.  A co-worker
tested with CONNTRACK disable, and the test passed.  Looks like I have
a bit of homework to do.  Thank you for all the helpful hints that got
me this far.

Steve