tcp/udp checksum on loopback interface

tcp/udp checksum on loopback interface

[Pierre Louis Aublin]

Hello everybody

I am interested in the reliability of TCP and UDP using the loopback 
interface.
I found that there is no checksum verification on the body of packets 
transmitted through the loopback interface : 
http://kerneltrap.org/mailarchive/linux-netdev/2007/10/2/327241
I ran a simple test using iperf and wireshark and I indeed found the 
validation of the checksum is disabled at packets reception.

At the reception, I guess the piece of code responsible for this 
behaviour is (line 1196 of net/ipv4/udp.c, in udp_recvmsg() for UDP) :
if (skb_csum_unnecessary(skb))
     err = skb_copy_datagram_iovec(skb, sizeof(struct udphdr), 
msg->msg_iov, len);
Is this the responsible piece of code? If yes, where does the 
CHECKSUM_UNNECESSARY flag is set?

However, I am not sure whether the checksum is computed on the whole 
packet at emission.
In __ip_append_data (net/ipv4/ip_output.c), I found the flag csummode is 
at the beginning at CHECKSUM_NONE.
Then, it is augmented using the dev->features flag of the network device 
(e.g. the loopback interface).
As the loopback interface has the NETIF_F_NO_CSUM flag, csummode stays 
at CHECKSUM_NONE and the checksum is not computed on the body of the 
message.
Can you confirm the checksum is computed only on the header of the 
packet when sending it through the loopback interface?

Finally, why this behaviour? Is it because you assume message can not 
get corrupted while staying on the same machine?

Thank you in advance
Pierre Louis Aublin

Re: tcp/udp checksum on loopback interface

[Chris Friesen]

On 07/22/2011 02:01 AM, Pierre Louis Aublin wrote:
> Hello everybody
>
> I am interested in the reliability of TCP and UDP using the loopback
> interface.
> I found that there is no checksum verification on the body of packets
> transmitted through the loopback interface :

> Finally, why this behaviour? Is it because you assume message can not
> get corrupted while staying on the same machine?

That's correct.  We can save cpu time by not doing the checksum because 
we assume that our own hardware won't introduce errors (or if it does 
and we care about them we'll be monitoring the hardware for ECC errors 
anyways).

Chris

-- 
Chris Friesen
Software Developer
GENBAND
chris.friesen@genband.com
www.genband.com