On Tue, 2017-08-08 at 18:07 +0300, Saeed Mahameed wrote:
> > {
> > __u16 length_for_csum = 0;
> > __wsum csum_pseudo_header = 0;
> > + __u8 ipproto = iph->protocol;
> > +
> > + if (unlikely(ipproto == IPPROTO_SCTP))
> > + return -1;
> >
>
> Hi Davide
>
hi Saeed,
thank you for looking at this!
> If you got to here then it means this is a non UDP/TCP ipv4 packet and
> the HW failed to validate it's checksum but
> you get from the connectx3 HW a 1's complement 16-bit sum of IP
> Payload + IP pseudo-header.
> so if you return -1 here the driver will report checksum none for this
> packet (and you will abandon any checsum offload/help from HW).
>
> I am not an SCTP expert but it seems that you decided here that
> connectX3 hw checksum (described above) can't be used to calculate
> SCTP packet checksum
> is that correct?
>
Yes, that's correct. SCTP uses CRC32c, not 1's complement (and does not use
pseudo-headers): therefore, the checksum computed by the NIC hardware can't
be used.
The issue I observed is skb->ip_summed set to CHECKSUM_COMPLETE, that made
CRC32c validation fail in my setup (that was a netfilter REJECT rule, matching
SCTP packets). AFAIK, CHECKSUM_COMPLETE is valid only for the Internet Checksum;
setting CHECKSUM_NONE on rx packets carrying IPPROTO_SCTP fixed my test
scenario.
> If so, then i am ok with this patch.
I planned to post this some weeks ago, after agreeing v2 with Tariq
(https://www.spinics.net/lists/netdev/msg441231.html), but took some time to
find a ConnectX-3 (from what I saw the issue is not present on ConnectX-3 Pro,
since it has MLX4_RX_CSUM_MODE_VAL_NON_TCP_UDP bit set to 0).
regards,
--
davide
