Re: [PATCH net-next 0/2] net: two updates related to UDP GSO

[tanhuazhong]

On 2020/9/7 17:22, Willem de Bruijn wrote:
> On Sun, Sep 6, 2020 at 8:42 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> > On Sat, 5 Sep 2020 14:11:11 +0800 Huazhong Tan wrote:
> > > There are two updates relates to UDP GSO.
> > > #1 adds a new GSO type for UDPv6
> > > #2 adds check for UDP GSO when csum is disable in netdev_fix_features().
> > >
> > > Changes since RFC V2:
> > > - modifies the timing of setting UDP GSO type when doing UDP GRO in #1.
> > >
> > > Changes since RFC V1:
> > > - updates NETIF_F_GSO_LAST suggested by Willem de Bruijn.
> > >    and add NETIF_F_GSO_UDPV6_L4 feature for each driver who support UDP GSO in #1.
> > >    - add #2 who needs #1.
> >
> > Please CC people who gave you feedback (Willem).
> >
> > I don't feel good about this series. IPv6 is not optional any more.
> > AFAIU you have some issues with csum support in your device? Can you
> > use .ndo_features_check() to handle this?
> >
> > The change in semantics of NETIF_F_GSO_UDP_L4 from "v4 and v6" to
> > "just v4" can trip people over; this is not a new feature people
> > may be depending on the current semantics.
> >
> > Willem, what are your thoughts on this?
>
> If that is the only reason, +1 on fixing it up in the driver's
> ndo_features_check.

Hi, Willem & Jakub.

This series mainly fixes the feature dependency between hardware 
checksum and UDP GSO.
When turn off hardware checksum offload, run 'ethtool -k [devname]'
we can see TSO is off as well, but udp gso still is on.

[root@localhost ~]# ethtool -K eth0 tx off
Actual changes:
tx-checksumming: off
	tx-checksum-ipv4: off
	tx-checksum-ipv6: off
	tx-checksum-sctp: off
tcp-segmentation-offload: off
	tx-tcp-segmentation: off [requested on]
	tx-tcp-ecn-segmentation: off [requested on]
	tx-tcp6-segmentation: off [requested on]
[root@localhost ~]# ethtool -k eth0
Features for eth0:
rx-checksumming: on
tx-checksumming: off
	tx-checksum-ipv4: off
	tx-checksum-ip-generic: off [fixed]
	tx-checksum-ipv6: off
	tx-checksum-fcoe-crc: off [fixed]
	tx-checksum-sctp: off
...
tcp-segmentation-offload: off
	tx-tcp-segmentation: off [requested on]
	tx-tcp-ecn-segmentation: off [requested on]
	tx-tcp-mangleid-segmentation: off
	tx-tcp6-segmentation: off [requested on]
udp-fragmentation-offload: off
generic-segmentation-offload: on
generic-receive-offload: on
...
tx-udp-segmentation: on
...

.ndo_feature_check seems unnecessary.
Because the stack has already do this check.
in __ip_append_data() below if branch will not run if hardware checksum 
offload is off.
        if (transhdrlen &&
            length + fragheaderlen <= mtu &&
            rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
            (!(flags & MSG_MORE) || cork->gso_size) &&
            (!exthdrlen || (rt->dst.dev->features & 
NETIF_F_HW_ESP_TX_CSUM)))
                csummode = CHECKSUM_PARTIAL;
...
so skb->ip_summed set as CHECKSUM_NONE,
then in udp_send_skb()
        if (cork->gso_size) {
		...
                if (skb->ip_summed != CHECKSUM_PARTIAL || is_udplite ||
                    dst_xfrm(skb_dst(skb))) {
                        kfree_skb(skb); 


                        return -EIO;
                }
the packet who needs udp gso will return ERROR.

For this kind of problem how could we fix it?

Thanks.
Huazhong.

> .