mpls_gso_segment() assumes skb_inner_network_header() returns
a valid result:

  mpls_hlen = skb_inner_network_header(skb) - skb_network_header(skb);
  if (unlikely(!mpls_hlen || mpls_hlen % MPLS_HLEN))
        goto out;
  if (unlikely(!pskb_may_pull(skb, mpls_hlen)))

With syzbot reproducer, skb_inner_network_header() yields 0,
skb_network_header() returns 108, so this will
"pskb_may_pull(skb, -108)))" which triggers a newly added
DEBUG_NET_WARN_ON_ONCE() check:

------------[ cut here ]------------
WARNING: CPU: 0 PID: 5068 at include/linux/skbuff.h:2723 pskb_may_pull_reason include/linux/skbuff.h:2723 [inline]
WARNING: CPU: 0 PID: 5068 at include/linux/skbuff.h:2723 pskb_may_pull include/linux/skbuff.h:2739 [inline]
WARNING: CPU: 0 PID: 5068 at include/linux/skbuff.h:2723 mpls_gso_segment+0x773/0xaa0 net/mpls/mpls_gso.c:34
[..]
 skb_mac_gso_segment+0x383/0x740 net/core/gso.c:53
 nsh_gso_segment+0x40a/0xad0 net/nsh/nsh.c:108
 skb_mac_gso_segment+0x383/0x740 net/core/gso.c:53
 __skb_gso_segment+0x324/0x4c0 net/core/gso.c:124
 skb_gso_segment include/net/gso.h:83 [inline]
 [..]
 sch_direct_xmit+0x11a/0x5f0 net/sched/sch_generic.c:327
 [..]
 packet_sendmsg+0x46a9/0x6130 net/packet/af_packet.c:3113
 [..]

First iteration of this patch made mpls_hlen signed and changed
test to error out to "mpls_hlen <= 0 || ..".

Eric Dumazet said:
 > I was thinking about adding a debug check in skb_inner_network_header()
 > if inner_network_header is zero (that would mean it is not 'set' yet),
 > but this would trigger even after your patch.

So add new skb_inner_network_header_was_set() helper and use that.

The syzbot reproducer injects data via packet socket. The skb that gets
allocated and passed down the stack has ->protocol set to NSH (0x894f)
and gso_type set to SKB_GSO_UDP | SKB_GSO_DODGY.

This gets passed to skb_mac_gso_segment(), which sees NSH as ptype to
find a callback for.  nsh_gso_segment() retrieves next type:

        proto = tun_p_to_eth_p(nsh_hdr(skb)->np);

... which is MPLS (TUN_P_MPLS_UC). It updates skb->protocol and then
calls mpls_gso_segment().  Inner offsets are all 0, so mpls_gso_segment()
ends up with a negative header size.

In case more callers rely on silent handling of such large may_pull values
we could also 'legalize' this behaviour, either replacing the debug check
with (len > INT_MAX) test or removing it and instead adding a comment
before existing

 if (unlikely(len > skb->len))
    return SKB_DROP_REASON_PKT_TOO_SMALL;

test in pskb_may_pull_reason(), saying that this check also implicitly
takes care of callers that miscompute header sizes.

Cc: Simon Horman <horms@kernel.org>
Fixes: 219eee9c ("net: skbuff: add overflow debug check to pull/push helpers")
Reported-by: syzbot+99d15fcdb0132a1e1a82@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/netdev/00000000000043b1310611e388aa@google.com/rawSigned-off-by: Florian Westphal <fw@strlen.de>
Link: https://lore.kernel.org/r/20240222140321.14080-1-fw@strlen.deSigned-off-by: Jakub Kicinski <kuba@kernel.org>