Re: [PATCH net, v3] net: nsh: Use correct mac_offset to unwind gso skb in nsh_gso_segment()

[Eric Dumazet]

On Tue, May 9, 2023 at 3:55 PM Dong Chenchen <dongchenchen2@xxxxxxxxxx> wrote:
>
> As the call trace shows, skb_panic was caused by wrong skb->mac_header
> in nsh_gso_segment():
>
> invalid opcode: 0000 [#1] PREEMPT SMP KASAN PTI
> CPU: 3 PID: 2737 Comm: syz Not tainted 6.3.0-next-20230505 #1
> RIP: 0010:skb_panic+0xda/0xe0
> call Trace:
>  skb_push+0x91/0xa0
>  nsh_gso_segment+0x4f3/0x570
>  skb_mac_gso_segment+0x19e/0x270
>  __skb_gso_segment+0x1e8/0x3c0
>  validate_xmit_skb+0x452/0x890
>  validate_xmit_skb_list+0x99/0xd0
>  sch_direct_xmit+0x294/0x7c0
>  __dev_queue_xmit+0x16f0/0x1d70
>  packet_xmit+0x185/0x210
>  packet_snd+0xc15/0x1170
>  packet_sendmsg+0x7b/0xa0
>  sock_sendmsg+0x14f/0x160
>
> The root cause is:
> nsh_gso_segment() use skb->network_header - nhoff to reset mac_header
> in skb_gso_error_unwind() if inner-layer protocol gso fails.
> However, skb->network_header may be reset by inner-layer protocol
> gso function e.g. mpls_gso_segment. skb->mac_header reset by the
> inaccurate network_header will be larger than skb headroom.
>
> nsh_gso_segment
>     nhoff = skb->network_header - skb->mac_header;
>     __skb_pull(skb,nsh_len)
>     skb_mac_gso_segment
>         mpls_gso_segment
>             skb_reset_network_header(skb);//skb->network_header+=nsh_len
>             return -EINVAL;
>     skb_gso_error_unwind
>         skb_push(skb, nsh_len);
>         skb->mac_header = skb->network_header - nhoff;
>         // skb->mac_header > skb->headroom, cause skb_push panic
>
> Use correct mac_offset to restore mac_header to fix it.
>
> Fixes: c411ed854584 ("nsh: add GSO support")
> Signed-off-by: Dong Chenchen <dongchenchen2@xxxxxxxxxx>
>
> ---
> v2:
>   - Use skb->mac_header not skb->network_header-nhoff for mac_offset.
>
> v3:
>   - 'net' is noted in the subject.
>   - arrange local variable following reverse xmas tree order
> ---
>  net/nsh/nsh.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/net/nsh/nsh.c b/net/nsh/nsh.c
> index e9ca007718b7..793e0bd94558 100644
> --- a/net/nsh/nsh.c
> +++ b/net/nsh/nsh.c
> @@ -77,6 +77,7 @@ static struct sk_buff *nsh_gso_segment(struct sk_buff *skb,
>                                        netdev_features_t features)
>  {
>         struct sk_buff *segs = ERR_PTR(-EINVAL);
> +       u16 mac_offset = skb->mac_header;
>         unsigned int nsh_len, mac_len;
>         __be16 proto;
>         int nhoff;
> @@ -108,8 +109,7 @@ static struct sk_buff *nsh_gso_segment(struct sk_buff *skb,
>         segs = skb_mac_gso_segment(skb, features);
>         if (IS_ERR_OR_NULL(segs)) {
>                 skb_gso_error_unwind(skb, htons(ETH_P_NSH), nsh_len,
> -                                    skb->network_header - nhoff,
> -                                    mac_len);
> +                                    mac_offset, mac_len);
>                 goto out;
>         }
>

I do not think this patch is enough ?

This is still not nice, because mac_header == 0xFFFF

nhoff = skb->network_header - skb->mac_header;
...
skb_set_mac_header(skb, -nhoff);

I would simply restore mac_header with "skb->mac_header =  mac_offset"
and get rid of nhoff.

(Accept the fact that GSO layer should not rely on skb mac_header being set)

In the future, we might be able to rewrite GSO without any assumptions
on skb->mac_header.