On Thu, Jan 28, 2021 at 4:42 PM Alexander Duyck
<alexander.du...@gmail.com> wrote:
>
> On Thu, Jan 28, 2021 at 12:00 PM Willem de Bruijn
> <willemdebruijn.ker...@gmail.com> wrote:
> >
> > On Thu, Jan 28, 2021 at 2:46 PM Alexander Duyck
> > <alexander.du...@gmail.com> wrote:
> > >
> > > On Thu, Jan 28, 2021 at 6:07 AM Willem de Bruijn
> > > <willemdebruijn.ker...@gmail.com> wrote:
> > > >
> > > > On Thu, Jan 28, 2021 at 4:29 AM Xin Long <lucien....@gmail.com> wrote:
> > > > >
> > > > > NETIF_F_IP|IPV6_CSUM feature flag indicates UDP and TCP csum offload
> > > > > while NETIF_F_HW_CSUM feature flag indicates ip generic csum offload
> > > > > for HW, which includes not only for TCP/UDP csum, but also for other
> > > > > protocols' csum like GRE's.
> > > > >
> > > > > However, in skb_csum_hwoffload_help() it only checks features against
> > > > > NETIF_F_CSUM_MASK(NETIF_F_HW|IP|IPV6_CSUM). So if it's a non TCP/UDP
> > > > > packet and the features doesn't support NETIF_F_HW_CSUM, but supports
> > > > > NETIF_F_IP|IPV6_CSUM only, it would still return 0 and leave the HW
> > > > > to do csum.
> > > > >
> > > > > This patch is to support ip generic csum processing by checking
> > > > > NETIF_F_HW_CSUM for all protocols, and check (NETIF_F_IP_CSUM |
> > > > > NETIF_F_IPV6_CSUM) only for TCP and UDP.
> > > > >
> > > > > Note that we're using skb->csum_offset to check if it's a TCP/UDP
> > > > > proctol, this might be fragile. However, as Alex said, for now we
> > > > > only have a few L4 protocols that are requesting Tx csum offload,
> > > > > we'd better fix this until a new protocol comes with a same csum
> > > > > offset.
> > > > >
> > > > > v1->v2:
> > > > > - not extend skb->csum_not_inet, but use skb->csum_offset to tell
> > > > > if it's an UDP/TCP csum packet.
> > > > > v2->v3:
> > > > > - add a note in the changelog, as Willem suggested.
> > > > >
> > > > > Suggested-by: Alexander Duyck <alexander.du...@gmail.com>
> > > > > Signed-off-by: Xin Long <lucien....@gmail.com>
> > > > > ---
> > > > > net/core/dev.c | 13 ++++++++++++-
> > > > > 1 file changed, 12 insertions(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/net/core/dev.c b/net/core/dev.c
> > > > > index 6df3f1b..aae116d 100644
> > > > > --- a/net/core/dev.c
> > > > > +++ b/net/core/dev.c
> > > > > @@ -3621,7 +3621,18 @@ int skb_csum_hwoffload_help(struct sk_buff
> > > > > *skb,
> > > > > return !!(features & NETIF_F_SCTP_CRC) ? 0 :
> > > > > skb_crc32c_csum_help(skb);
> > > > >
> > > > > - return !!(features & NETIF_F_CSUM_MASK) ? 0 :
> > > > > skb_checksum_help(skb);
> > > > > + if (features & NETIF_F_HW_CSUM)
> > > > > + return 0;
> > > > > +
> > > > > + if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
> > > >
> > > > Should this check the specific feature flag against skb->protocol? I
> > > > don't know if there are actually instances that only support one of
> > > > the two flags.
> > >
> > > The issue is at a certain point we start excluding devices that were
> > > previously working.
> > >
> > > All this patch is really doing is using the checksum offset to
> > > identify the cases that were previously UDP or TCP offloads and
> > > letting those through with the legacy path, while any offsets that are
> > > not those two, such as the GRE checksum will now have to be explicitly
> > > caught by the NETIF_F_HW_CSUM case and not accepted by the other
> > > cases.
> >
> > I understand. But letting through an IPv6 packet to a nic that
> > advertises NETIF_F_IP_CSUM, but not NETIF_F_IPV6_CSUM, is still
> > incorrect, right?
>
> That all depends. The problem is if we are going to look at protocol
> we essentially have to work our way through a number of fields and
> sort out if there are tunnels or not and if so what the protocol for
> the inner headers are and if that is supported. It might make more
> sense in that case to look at incorporating a v4/v6 specific check
> into netif_skb_features so we could mask off the bit there.
>
> The question i would have is how has this code been working up until
> now without that check? If we are broken outright and need to add it
> then maybe this should be deemed more of a fix and pushed for net with
> the added protocol bit masking added.
Fair enough. I agree that this patch does not make anything worse, as
it actually tightens the rules.
