Re: [PATCH net-next v3 1/1] net: fec: Enable imx6 enet checksum acceleration.

[Ben Hutchings <bhutchings@solarflare.com>]

On Thu, 2013-04-18 at 22:27 +0100, Jim Baxter wrote:
> On 18/04/13 18:21, Jim Baxter wrote:
> > On 18/04/13 18:12, Ben Hutchings wrote:
> >> On Thu, 2013-04-18 at 18:08 +0100, Jim Baxter wrote:
> >>> On 18/04/13 17:16, Ben Hutchings wrote:
> >>>> On Thu, 2013-04-18 at 13:49 +0100, Jim Baxter wrote:
> >>>>> On 17/04/13 22:37, Eric Dumazet wrote:
> >>>>>> On Wed, 2013-04-17 at 21:07 +0100, Jim Baxter wrote:
> >>>>>>
> >>>>>>> +			skb_set_transport_header(skb,
> >>>>>>> +					ETH_HLEN + ip_hdrlen(skb));
> >>>>>>> +			udp_hdr(skb)->check = 0;
> >>>>>>> +			break;
> >>>>>>> +		case IPPROTO_TCP:
> >>>>>>> +			hdr_len = (ETH_HLEN +
> >>>>>>> +					(ip_hdr(skb)->ihl << 2) +
> >>>>>>> +					sizeof(struct tcphdr));
> >>>>>>> +			if (skb->len < hdr_len)
> >>>>>>> +				return;
> >>>>>>> +			skb_cow_head(skb, hdr_len);
> >>>>>>
> >>>>>> same here
> >>>>> Do I need to call skb_cow_head here, I am not changing the size of the
> >>>>> header?
> >>>> [...]
> >>>>
> >>>> The length passed to skb_cow_head() may be significant when the skb has
> >>>> paged fragments.  Since you aren't (yet) implementing scatter-gather,
> >>>> that won't happen.  And the headers shouldn't be in paged fragments
> >>>> anyway.  I think you can safely use skb_cow_head(skb, 0).
> >>>>
> >>>> But you don't actually need to check protocol numbers at all, as the
> >>>> kernel already specifies where the checksum should be.
> >>>>
> >>>> So I think this function should look like:
> >>>> {
> >>>> 	if (skb->ip_summed != CHECKSUM_PARTIAL)
> >>>> 		return 0;
> >>>>
> >>>> 	if (unlikely(skb_cow_head(skb, 0)))
> >>>> 		return -1;
> >>>>
> >>>> 	*(__sum16 *)(skb->head + skb->csum_start + skb->csum_offset) = 0;
> >>>> 	return 0;
> >>>> }
> >>>>
> >>>> The caller needs to check for the failure, and free the skb
> >>>> (kfree_skb()) rather than transmitting it.
> >>>>
> >>>> Ben.
> >>>>
> >>>
> >>> Which checksum does skb->csum (skb->csum_start + skb->csum_offset)
> >>> relate to?
> >>
> >> TCP or UDP.
> >>
> >>> The network card can generate IP header and protocol (UDP/TCP/ICMP)
> >>> checksums as long as the checksums are zeroed?
> >>
> >> You don't need to offload the IPv4 header checksum.  TX checksum offload
> >> will not be requested for ICMP.
> > 
> > So the kernel only offloads TCP and UDP checksum's. All other transport
> > protocols (ICMP, ICMPIPV6 etc.) will be done by the kernel.
> > 
> >>
> >> Ben.
> >>
> > 
> > Thanks, though I saw you replied to my question before I could send it.
> > 
> > It sounds like there is no reason wasting CPU cycles clearing the IP
> > header checksum, so I can keep the code efficient.
> > 
> > Jim
> > 
> 
> Does anybody know if the kernel uses checksum offloading for IPV6, if I
> enable the feature NETIF_F_HW_CSUM, I only receive CHECKSUM_NONE when
> the protocol is 0x86DD?

NETIF_F_HW_CSUM means you can generate TCP-style checksums from an
arbitrary starting offset.  You would then use skb->csum_start and
skb->csum_offset to construct the TX descriptor(s).

If the hardware generates checksums for specific protocols based on
parsing the packet contents, you should set NETIF_F_IP_CSUM (TCP and UDP
on IPv4) and/or NETIF_F_IPV6_CSUM (TCP and UDP on IPv6).  It appears
from your original patch that the hardware behaves like this.

Ben.

> I realise there is no IP header checksum, but I thought the protocols
> still use a checksum.
> 
> I am using iperf with the -V option in UDP mode, which offloads for IPV4.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.