Re: [PATCHv2 net-next] ip_gre: remove CRC flag from dev features in gre_gso_segment

[Xin Long <lucien.xin@gmail.com>]

On Tue, Jan 12, 2021 at 12:48 AM Alexander Duyck
<alexander.duyck@gmail.com> wrote:
>
> On Mon, Jan 11, 2021 at 5:22 AM Xin Long <lucien.xin@gmail.com> wrote:
> >
> > This patch is to let it always do CRC checksum in sctp_gso_segment()
> > by removing CRC flag from the dev features in gre_gso_segment() for
> > SCTP over GRE, just as it does in Commit 527beb8ef9c0 ("udp: support
> > sctp over udp in skb_udp_tunnel_segment") for SCTP over UDP.
> >
> > It could set csum/csum_start in GSO CB properly in sctp_gso_segment()
> > after that commit, so it would do checksum with gso_make_checksum()
> > in gre_gso_segment(), and Commit 622e32b7d4a6 ("net: gre: recompute
> > gre csum for sctp over gre tunnels") can be reverted now.
> >
> > Note that the current HWs like igb NIC can only handle the SCTP CRC
> > when it's in the outer packet, not in the inner packet like in this
> > case, so here it removes CRC flag from the dev features even when
> > need_csum is false.
>
> So the limitation in igb is not the hardware but the driver
> configuration. When I had coded things up I put in a limitation on the
> igb_tx_csum code that it would have to validate that the protocol we
> are requesting an SCTP CRC offload since it is a different calculation
> than a 1's complement checksum. Since igb doesn't support tunnels we
> limited that check to the outer headers.
Ah.. I see, thanks.
>
> We could probably enable this for tunnels as long as the tunnel isn't
> requesting an outer checksum offload from the driver.
I think in igb_tx_csum(), by checking skb->csum_not_inet would be enough
to validate that is a SCTP request:
-               if (((first->protocol == htons(ETH_P_IP)) &&
-                    (ip_hdr(skb)->protocol == IPPROTO_SCTP)) ||
-                   ((first->protocol == htons(ETH_P_IPV6)) &&
-                    igb_ipv6_csum_is_sctp(skb))) {
+               if (skb->csum_not_inet) {
                        type_tucmd = E1000_ADVTXD_TUCMD_L4T_SCTP;
                        break;
                }

Otherwise, we will need to parse the packet a little bit, as it does in
hns3_get_l4_protocol().

>
> > v1->v2:
> >   - improve the changelog.
> >   - fix "rev xmas tree" in varibles declaration.
> >
> > Signed-off-by: Xin Long <lucien.xin@gmail.com>
> > ---
> >  net/ipv4/gre_offload.c | 15 ++++-----------
> >  1 file changed, 4 insertions(+), 11 deletions(-)
> >
> > diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
> > index e0a2465..a681306 100644
> > --- a/net/ipv4/gre_offload.c
> > +++ b/net/ipv4/gre_offload.c
> > @@ -15,10 +15,10 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
> >                                        netdev_features_t features)
> >  {
> >         int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
> > -       bool need_csum, need_recompute_csum, gso_partial;
> >         struct sk_buff *segs = ERR_PTR(-EINVAL);
> >         u16 mac_offset = skb->mac_header;
> >         __be16 protocol = skb->protocol;
> > +       bool need_csum, gso_partial;
> >         u16 mac_len = skb->mac_len;
> >         int gre_offset, outer_hlen;
> >
> > @@ -41,10 +41,11 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
> >         skb->protocol = skb->inner_protocol;
> >
> >         need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_GRE_CSUM);
> > -       need_recompute_csum = skb->csum_not_inet;
> >         skb->encap_hdr_csum = need_csum;
> >
> >         features &= skb->dev->hw_enc_features;
> > +       /* CRC checksum can't be handled by HW when SCTP is the inner proto. */
> > +       features &= ~NETIF_F_SCTP_CRC;
> >
> >         /* segment inner packet. */
> >         segs = skb_mac_gso_segment(skb, features);
>
> Do we have NICs that are advertising NETIF_S_SCTP_CRC as part of their
> hw_enc_features and then not supporting it? Based on your comment
Yes, igb/igbvf/igc/ixgbe/ixgbevf, they have a similar code of SCTP
proto validation.

> above it seems like you are masking this out because hardware is
> advertising features it doesn't actually support. I'm just wondering
> if that is the case or if this is something where this should be
> cleared if need_csum is set since we only support one level of
> checksum offload.
Since only these drivers only do SCTP proto validation, and "only
one level checksum offload" issue only exists when inner packet
is SCTP packet, clearing NETIF_F_SCTP_CRC should be enough.

But seems to fix the drivers will be better, as hw_enc_features should
tell the correct features for inner proto. wdyt?

(Just note udp tunneling SCTP doesn't have this issue, as the outer
 udp checksum is always required by RFC)

>
> > @@ -99,15 +100,7 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
> >                 }
> >
> >                 *(pcsum + 1) = 0;
> > -               if (need_recompute_csum && !skb_is_gso(skb)) {
> > -                       __wsum csum;
> > -
> > -                       csum = skb_checksum(skb, gre_offset,
> > -                                           skb->len - gre_offset, 0);
> > -                       *pcsum = csum_fold(csum);
> > -               } else {
> > -                       *pcsum = gso_make_checksum(skb, 0);
> > -               }
> > +               *pcsum = gso_make_checksum(skb, 0);
> >         } while ((skb = skb->next));
> >  out:
> >         return segs;
> > --
> > 2.1.0
> >