Re: TCP checksum problem

[Arvid Brodin <arvid.brodin@alten.se>]

On 2014-04-12 02:49, Eric Dumazet wrote:
> On Sat, 2014-04-12 at 02:25 +0200, Arvid Brodin wrote:
>> Hi,
>>
>> I'm trying to insert a 6-byte HSR header into a newly copied skb,
>> between the ethernet header and the IP header. The skb comes from my
>> virtual device's ndo_start_xmit() function and is to be sent on to a
>> physical interface after modification. Essentially, this is what I'm
>> doing:
>>
>> 	skb = skb_copy_expand(skb, skb_headroom(skb) + HSR_HLEN, 0,
>> GFP_ATOMIC);
>>
>> 	/* Move ethernet header to make room for the HSR header */
>> 	src = skb_mac_header(skb);
>> 	WARN_ON_ONCE(src != skb->data);
>> 	dst = skb_push(skb, HSR_HLEN);
>> 	memmove(dst, src, ETH_HLEN);
>> 	skb_reset_mac_header(skb);
>>
>> 	/* Fill in the 6-byte header */
>> 	hsr_fill_tag(skb, frame, port);
>>
>> 	/* Point the skb to the physical interface */
>> 	skb->dev = port->dev; 
>>
>> 	dev_queue_xmit(skb);
>>
>> But something goes wrong with the TCP checksum on the way out. Arp,
>> ping etc works fine, but anything using TCP is discarded by the
>> receiver due to incorrect TCP checksums.
>>
>> Using wireshark and some printks, I've managed to decide that the
>> updated TCP checksum is written offset by -6 bytes from where it
>> should be (into what wireshark calls the TCP header "acknowledgement
>> number" field).
>>
>> Any idea why this happens? I've spent a few days on this now and I'm
>> running out of ideas...
>>
>> This is run on the cadence/macb ethernet device. Before the call to
>> skb_copy_expand(), skb->ip_summed is CHECKSUM_PARTIAL, skb->csum_start
>> is 120, and skb->csum_offset is 16. The total frame length is 80 bytes
>> (on the wire as reported by wireshark - I think this is excluding the
>> FCS).
>>
>>
> 
> You need to set skb->csum_start to the new offset.
> 
> ( skb_transport_header(skb) - skb->head )
> 
> Check how __tcp_v4_send_check() sets csum_start & csum_offset
> 
> (csum_offset you can leave it as is)
> 

I recompiled my code today and ran it again, and it worked. I have no 
idea what change I made, althought I know I first used 

skb = __pskb_copy(skb, skb_headroom(skb) + HSR_HLEN, GFP_ATOMIC);

instead of 

skb = skb_copy_expand(skb, skb_headroom(skb) + HSR_HLEN, 0, GFP_ATOMIC);

for the allocation of the new skb. The latter adjusts csum_start, the 
former doesn't. I may have changed to the latter and simultaneously 
added a manual adjustment to csum_start, because I know I was surprised 
that the checksum moved 2*HSR_HLEN bytes when I did the adjustment. :)

Anyway, it seems to work now. Thanks for the help!


-- 
Arvid Brodin | Consultant (Linux)
ALTEN | Knarrarnäsgatan 7 | SE-164 40 Kista | Sweden
arvid.brodin@alten.se | www.alten.se/en/