Re: [PATCH] ip6_tunnel: Fix uninit-value in ip6_tnl_xmit

[Eric Dumazet]

On Tue, Dec 17, 2024 at 4:09 AM Mazin Al Haddad <mazin@xxxxxxxxxxxx> wrote:
>
> When taking the branch with skb_realloc_headroom, pskb_expand_head is
> called, as such, pointers referencing content within the new skb's header
> are invalid. Currently, the assignment of hop_limit accesses the now
> invalid pointer in the network header of this "new" skb. Fix this by
> moving the logic to assign hop_limit earlier so that the assignment
> references the original un-resized skb instead.

Unfortunately this is not fixing anything.

If the IPv6 header was in the skb head before skb_realloc_headroom()
and/or pskb_expand_head(),
it would be copied in the new skb head.

Note how the repro is sending a packet with vlan tag (88A8 : ETH_P_8021AD)

endto$packet(r0, &(0x7f0000000180)="a6bea8a120e5f8320c30ce5088a8",
0x12, 0x0, &(0x7f0000000140)={0x11, 0x0, r3, 0x1, 0x0, 0x6, @local},
0x14)

Current code, using pskb_inet_may_pull() is not ready yet.

My patch has been tested by syzbot and I was about to submit it.

diff --git a/net/ipv6/ip6_gre.c b/net/ipv6/ip6_gre.c
index 235808cfec705032b545d6f396f8e58f4693e8d8..c4f0383a136cf5f5e6846293078ec8b826c754c9
100644
--- a/net/ipv6/ip6_gre.c
+++ b/net/ipv6/ip6_gre.c
@@ -910,7 +910,7 @@ static netdev_tx_t ip6gre_tunnel_xmit(struct sk_buff *skb,
        __be16 payload_protocol;
        int ret;

-       if (!pskb_inet_may_pull(skb))
+       if (!skb_vlan_inet_prepare(skb, false))
                goto tx_err;

        if (!ip6_tnl_xmit_ctl(t, &t->parms.laddr, &t->parms.raddr))
@@ -958,7 +958,7 @@ static netdev_tx_t ip6erspan_tunnel_xmit(struct
sk_buff *skb,
        __u32 mtu;
        int nhoff;

-       if (!pskb_inet_may_pull(skb))
+       if (!skb_vlan_inet_prepare(skb, false))
                goto tx_err;

        if (!ip6_tnl_xmit_ctl(t, &t->parms.laddr, &t->parms.raddr))