On Mon, Jun 27, 2016 at 2:47 PM, Tom Herbert [off-list ref] wrote:

On Mon, Jun 27, 2016 at 2:44 PM, Cong Wang [off-list ref] wrote:

quoted

On Mon, Jun 27, 2016 at 2:08 PM, Or Gerlitz [off-list ref] wrote:

quoted

On Mon, Jun 27, 2016 at 9:22 PM, Cong Wang [off-list ref] wrote:

quoted

The stack doesn't trust the complete csum by hardware
even when it is correct.

Can you explain that a little further?

Sure, here is the code in __skb_checksum_complete():

        /* skb->csum holds pseudo checksum */
        sum = csum_fold(csum_add(skb->csum, csum));
        if (likely(!sum)) {
                if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
                    !skb->csum_complete_sw)
                        netdev_rx_csum_fault(skb->dev);
        }

So when sum == 0, it means the checksum is correct. And
we already set ->ip_summed to CHECKSUM_COMPLETE
after check_csum(), and ->csum_complete_sw is initialized
to 0 when we allocate the skb. This is why we trigger
netdev_rx_csum_fault().

Yes, but this also means that the driver gave the stack a checksum
complete value that was incorrect. That's an error.

That is the whole purpose of commit f8c6455bb04b944edb69e,
isn't it?