Re: [PATCH] bnx2x: fix checksum validation

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2012-06-13 at 16:42 +0300, Eilon Greenstein wrote:

> So what you are saying is that the indication that the checksum is valid
> is interpreted as the encapsulated checksum and not just the IP
> header... This was not the intention of this code, it was meant to
> indicate that the IP header is valid.
> 

IP header is always checked by our stack.

Its so fast that there is no point spending a bit in skb, and testing
this bit.

Yes, CHECKSUM_UNNECESSARY really means everything was checked up to L4

> I must admit that the code looks much better with this change. The only
> down side is that there is no longer IP checksum offload for pure IP
> packets, but that's negligible. The only thing that bothers me is that
> there is no way to indicate anything about the encapsulated packet
> separately from the outer header. Is that the way we want to keep it?

Some NIC allow for csum L3 offloading, providing the 16bit csum.

We then can :

skb->csum = csum;
skb->ip_summed = CHECKSUM_COMPLETE;

Check drivers/net/ethernet/ibm/ehea/ehea_main.c,
drivers/net/ethernet/marvell/skge.c ,
drivers/net/ethernet/marvell/sky2.c , ... for examples.


When a layer is removed(pulled), we can use skb_postpull_rcsum() and
friends (check net/ipv4/ip_gre.c for example) to keep skb->csum up2date.


If bnx2x provides this csum (for non TCP/UCP packets or fragments), you
could fallback to this CHECKSUM_COMPLETE stuff.