Date:   Fri, 20 Nov 2020 08:10:27 -0800
From:   Alexander Duyck <alexander.duyck@...il.com>
To:     Xin Long <lucien.xin@...il.com>
Cc:     network dev <netdev@...r.kernel.org>,
        "linux-sctp @ vger . kernel . org" <linux-sctp@...r.kernel.org>,
        Marcelo Ricardo Leitner <marcelo.leitner@...il.com>,
        David Miller <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        Guillaume Nault <gnault@...hat.com>,
        Paolo Abeni <pabeni@...hat.com>,
        Lorenzo Bianconi <lorenzo@...nel.org>
Subject: Re: [PATCH net-next] ip_gre: remove CRC flag from dev features in gre_gso_segment

On Fri, Nov 20, 2020 at 2:23 AM Xin Long <lucien.xin@...il.com> wrote:
>
> On Fri, Nov 20, 2020 at 1:24 AM Alexander Duyck
> <alexander.duyck@...il.com> wrote:
> >
> > On Wed, Nov 18, 2020 at 9:53 PM Xin Long <lucien.xin@...il.com> wrote:
> > >
> > > On Thu, Nov 19, 2020 at 4:35 AM Alexander Duyck
> > > <alexander.duyck@...il.com> wrote:
> > > >
> > > > On Mon, Nov 16, 2020 at 1:17 AM Xin Long <lucien.xin@...il.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.
> > > > >
> > > > > Signed-off-by: Xin Long <lucien.xin@...il.com>
> > > > > ---
> > > > >  net/ipv4/gre_offload.c | 14 +++-----------
> > > > >  1 file changed, 3 insertions(+), 11 deletions(-)
> > > > >
> > > > > diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
> > > > > index e0a2465..a5935d4 100644
> > > > > --- a/net/ipv4/gre_offload.c
> > > > > +++ b/net/ipv4/gre_offload.c
> > > > > @@ -15,12 +15,12 @@ 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;
> > > > >         u16 mac_len = skb->mac_len;
> > > > >         int gre_offset, outer_hlen;
> > > > > +       bool need_csum, gso_partial;
> > > > >
> > > > >         if (!skb->encapsulation)
> > > > >                 goto out;
> > > > > @@ -41,10 +41,10 @@ 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;
> > > > > +       features &= ~NETIF_F_SCTP_CRC;
> > > > >
> > > > >         /* segment inner packet. */
> > > > >         segs = skb_mac_gso_segment(skb, features);
> > > >
> > > > Why just blindly strip NETIF_F_SCTP_CRC? It seems like it would make
> > > > more sense if there was an explanation as to why you are stripping the
> > > > offload. I know there are many NICs that could very easily perform
> > > > SCTP CRC offload on the inner data as long as they didn't have to
> > > > offload the outer data. For example the Intel NICs should be able to
> > > > do it, although when I wrote the code up enabling their offloads I
> > > > think it is only looking at the outer headers so that might require
> > > > updating to get it to not use the software fallback.
> > > >
> > > > It really seems like we should only be clearing NETIF_F_SCTP_CRC if
> > > > need_csum is true since we must compute the CRC before we can compute
> > > > the GRE checksum.
> > > Right, it's also what Jakub commented, thanks.
> > >
> > > >
> > > > > @@ -99,15 +99,7 @@ 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;
> > > >
> > > > This change doesn't make much sense to me. How are we expecting
> > > > gso_make_checksum to be able to generate a valid checksum when we are
> > > > dealing with a SCTP frame? From what I can tell it looks like it is
> > > > just setting the checksum to ~0 and checksum start to the transport
> > > > header which isn't true because SCTP is using a CRC, not a 1's
> > > > complement checksum, or am I missing something? As such in order to
> > > > get the gre checksum we would need to compute it over the entire
> > > > payload data wouldn't we? Has this been tested with an actual GRE
> > > > tunnel that had checksums enabled? If so was it verified that the GSO
> > > > frames were actually being segmented at the NIC level and not at the
> > > > GRE tunnel level?
> > > Hi Alex,
> > >
> > > I think you're looking at net.git? As on net-next.git, sctp_gso_make_checksum()
> > > has been fixed to set csum/csum_start properly by Commit 527beb8ef9c0 ("udp:
> > > support sctp over udp in skb_udp_tunnel_segment"), so that we compute it over
> > > the entire payload data, as you said above:
> >
> > No. I believe the code is still wrong. That is what I was pointing
> > out. The GSO_CB->csum is supposed to be the checksum of the header
> > from csum_start up to the end of the payload. In the case of the 1's
> > complement checksum that is normally the inverse of the pseudo-header
> > checksum. We don't normally compute it and instead assume it since it
> > is offloaded. For a CRC based checksum you would need to compute the
> > checksum over the entire packet since CRC and checksum are very
> > different computations. That is why we were calling skb_checksum in
> > the original code.
> Hi, Alex, sorry for having confused you, see below.
>
> >
> > > @@ -27,7 +27,11 @@ static __le32 sctp_gso_make_checksum(struct sk_buff *skb)
> > >  {
> > >         skb->ip_summed = CHECKSUM_NONE;
> > >         skb->csum_not_inet = 0;
> > > -       gso_reset_checksum(skb, ~0);
> > > +       /* csum and csum_start in GSO CB may be needed to do the UDP
> > > +        * checksum when it's a UDP tunneling packet.
> > > +        */
> > > +       SKB_GSO_CB(skb)->csum = (__force __wsum)~0;
> > > +       SKB_GSO_CB(skb)->csum_start = skb_headroom(skb) + skb->len;
> > >         return sctp_compute_cksum(skb, skb_transport_offset(skb));
> > >  }
> > >
> > > And yes, this patch has been tested with GRE tunnel checksums enabled.
> > > (... icsum ocsum).
> > > And yes, it was segmented in the lower NIC level, and we can make it by:
> > >
> > > # ethtool -K gre1 tx-sctp-segmentation on
> > > # ethtool -K veth0 tx-sctp-segmentation off
> > >
> > > (Note: "tx-checksum-sctp" and "gso" are on for both devices)
> > >
> > > Thanks.
> >
> > I would also turn off Tx and Rx checksum offload on your veth device
> > in order to make certain you aren't falsely sending data across
> > indicating that the checksums are valid when they are not. It might be
> > better if you were to run this over an actual NIC as that could then
> > provide independent verification as it would be a fixed checksum test.
> >
> > I'm still not convinced this is working correctly. Basically a crc32c
> > is not the same thing as a 1's complement checksum so you should need
> > to compute both if you have an SCTP packet tunneled inside a UDP or
> > GRE packet with a checksum. I don't see how computing a crc32c should
> > automatically give you a 1's complement checksum of ~0.
>
> On the tx Path [1] below, the sctp crc checksum is calculated in
> sctp_gso_make_checksum() [a], where it calls *sctp_compute_cksum()*
> to do that, and as for the code in it:
>
>     SKB_GSO_CB(skb)->csum = (__force __wsum)~0;
>     SKB_GSO_CB(skb)->csum_start = skb_headroom(skb) + skb->len;

Okay, so I think I know how this is working, but the number of things
relied on is ugly. Normally assuming skb_headroom(skb) + skb->len
being valid for this would be a non-starter. I was assuming you
weren't doing the 1's compliment checksum because you weren't using
__skb_checksum to generate it.

If I am not mistaken SCTP GSO uses the GSO_BY_FRAGS and apparently
none of the frags are using page fragments within the skb. Am I
understanding correctly? One thing that would help to address some of
my concerns would be to clear instead of set NETIF_F_SG in
sctp_gso_segment since your checksum depends on linear skbs.

> is prepared for doing 1's complement checksum (for outer UDP/GRE), and used
> in gre_gso_segment() [b], where it calls gso_make_checksum() to do that
> when need_csum is set. Note that, here it played a TRICK:
>
> I set SKB_GSO_CB->csum_start to the end of this packet and
> SKB_GSO_CB->csum = ~0 manually, so that in gso_make_checksum() it will do
> the 1's complement checksum for outer UDP/GRE by summing all packet bits up.
> See gso_make_checksum() (called by gre_gso_segment()):
>
>  unsigned char *csum_start = skb_transport_header(skb);
>  int plen = (skb->head + SKB_GSO_CB(skb)->csum_start) - csum_start;
>  /* now plen is from udp header to the END of packet. */
>  __wsum partial = SKB_GSO_CB(skb)->csum;
>
>  return csum_fold(csum_partial(csum_start, plen, partial));
>
> So yes, here it does compute both if I have an SCTP packet tunnelled inside
> a UDP or GRE packet with a checksum.

Assuming you have the payload data in the skb->data section. Normally
payload is in page frags. That is why I was concerned. You have to
have guarantees in place that there will not be page fragments
attached to the skb.

> But you're right that "the GSO_CB->csum is supposed to be the checksum
> of the header from csum_start up to the end of the payload". In this
> TRICK, csum_start is set to the end of packet,  and csum should be
> set to 0, and I will fix it in sctp_gso_make_checksum(), thanks!
>
> -       SKB_GSO_CB(skb)->csum = (__force __wsum)~0;
> +       SKB_GSO_CB(skb)->csum = (__force __wsum)0;
>
> Does it make sense to you now?

For a 1's compliment checksum ~0 and 0 are the same thing. So that
value doesn't matter. The issue as I have pointed out is the fact that
you are assuming a linear skb, and I am not certain that is what you
will actually get out of the call to skb_segment that you make in
sctp_gso_segment.

> Path [1]:
>  sctp_gso_segment.cold.6+0x3c/0x87 [sctp] <----- [a]
>  inet_gso_segment+0x152/0x3c0
>  skb_mac_gso_segment+0xa2/0x110
>  gre_gso_segment+0x138/0x380 <---- [b]
>  inet_gso_segment+0x152/0x3c0
>  skb_mac_gso_segment+0xa2/0x110
>  __skb_gso_segment+0xba/0x160
>  validate_xmit_skb+0x147/0x300
>  __dev_queue_xmit+0x569/0x920
>  ip_finish_output2+0x264/0x570
>  ip_output+0x6d/0x100
>  iptunnel_xmit+0x16e/0x200
>  ip_tunnel_xmit+0x437/0x870 [ip_tunnel]
>  ipgre_xmit+0x17b/0x210 [ip_gre]
>  dev_hard_start_xmit+0xd4/0x200
>  __dev_queue_xmit+0x78c/0x920
>  ip_finish_output2+0x194/0x570
>  ip_output+0x6d/0x100
>  __ip_queue_xmit+0x15d/0x430
>  sctp_packet_transmit+0x706/0x9b0 [sctp]
>  sctp_outq_flush+0xd7/0x8d0 [sctp]
>  sctp_cmd_interpreter.isra.21+0x721/0x1a20 [sctp]
>  sctp_do_sm+0xc3/0x2a0 [sctp]
>  sctp_primitive_SEND+0x2f/0x40 [sctp]
>  sctp_sendmsg_to_asoc+0x5fa/0x870 [sctp]
>  sctp_sendmsg+0x692/0x9d0 [sctp]
>  sock_sendmsg+0x54/0x60

Powered by blists - more mailing lists