Re: [PATCH net-next] net: skip offload for NETIF_F_IPV6_CSUM if ipv6 header contains extension
From: Willem de Bruijn
Date: Mon Oct 21 2024 - 18:05:05 EST
Benoît Monin wrote:
> 10/10/2024 Benoît Monin wrote:
> > 07/10/2024 Willem de Bruijn wrote:
> > > Benoît Monin wrote:
> > > > 07/10/2024 Willem de Bruijn wrote :
> > > > > Benoît Monin wrote:
> [...]
> > > > > > Signed-off-by: Benoît Monin <benoit.monin@xxxxxx>
> > > > > > ---
> > > > > > net/core/dev.c | 4 ++++
> > > > > > 1 file changed, 4 insertions(+)
> > > > > >
> > > > > > diff --git a/net/core/dev.c b/net/core/dev.c
> > > > > > index ea5fbcd133ae..199831d86ec1 100644
> > > > > > --- a/net/core/dev.c
> > > > > > +++ b/net/core/dev.c
> > > > > > @@ -3639,6 +3639,9 @@ int skb_csum_hwoffload_help(struct sk_buff *skb,
> > > > > > return 0;
> > > > > >
> > > > > > if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
> > > > > > + if (ip_hdr(skb)->version == 6 &&
> > > > > > + skb_network_header_len(skb) != sizeof(struct ipv6hdr))
> > > > > > + goto sw_checksum;
> > >
> > > This check depends on skb->transport_header and skb->network_header
> > > being set. This is likely true for all CHECKSUM_PARTIAL packets that
> > > originate in the local stack. As well as for the injected packets and
> > > forwarded packets, as far as I see, so Ack.
> > >
> > > Access to the network header at this point likely requires
> > > skb_header_pointer, however. As also used in qdisc_pkt_len_init called
> > > from the same __dev_queue_xmit_nit.
> > >
> > > Perhaps this test should be in can_checksum_protocol, which already
> > > checks that the packet is IPv6 when testing NETIF_F_IPV6_CSUM.
> > >
> > You're right, moving this to can_checksum_protocol() makes more sense. I will
> > do that, retest and post a new version of the patch.
> >
> Looking more into it, can_checksum_protocol() is called from multiple places
> where network header length cannot easily extracted, in particular from
> vxlan_features_check().
>
> How about keeping the length check in skb_csum_hwoffload_help() but using
> vlan_get_protocol() to check for IPv6 instead of ip_hdr(skb)->version?
Yes, both sound good to me.