Re: [PATCH v2] ftgmac100: Disable HW checksum generation on AST2500

From: Benjamin Herrenschmidt
Date: Thu Oct 17 2019 - 19:35:36 EST


On Fri, 2019-10-18 at 10:14 +1100, Benjamin Herrenschmidt wrote:
>
> I don't understand what you are saying. You reported a problem with
> IPV6 checksums generation. The HW doesn't support it. What's "not a
> matter of unsupported csum" ?
>
> Your patch uses a *deprecated* bit to tell the network stack to only do
> HW checksum generation on IPV4.
>
> This bit is deprecated for a reason, again, see skbuff.h. The right
> approach, *which the driver already does*, is to tell the stack that we
> support HW checksuming using NETIF_F_HW_CSUM, and then, in the transmit
> handler, to call skb_checksum_help() to have the SW calculate the
> checksum if it's not a supported type.
>
> This is exactly what ftgmac100_prep_tx_csum() does. It only enables HW
> checksum generation on supported types and uses skb_checksum_help()
> otherwise, supported types being protocol ETH_P_IP and IP protocol
> being raw IP, TCP and UDP.
>
> So this *should* have fallen back to SW for IPV6. So either something
> in my code there is making an incorrect assumption, or something is
> broken in skb_checksum_help() for IPV6 (which I somewhat doubt) or
> something else I can't think of, but setting a *deprecated* flag is
> definitely not the right answer, neither is completely disabling HW
> checksumming.
>
> So can you investigate what's going on a bit more closely please ? I
> can try myself, though I have very little experience with IPV6 and
> probably won't have time before next week.

I did get that piece of information from Aspeed: The HW checksum
generation is supported if:

- The length of UDP header is always 20 bytes.
- The length of TCP and IP header have 4 * N bytes (N is 5 to 15).

Now these afaik are also the protocol limits, so it *should* work.

Or am I missing something or some funky encaspulation/header format
that can be used under some circumstances ?

Cheers,
Ben.