Re: [PATCH net] net: iptunnel: fix stale transport header during tunnel decapsulation

On Fri, Aug 21, 2026 at 4:21 AM dongchenchen (A)
<[email protected]> wrote:
>
>
>
> 在 2026/8/21 0:32, Eric Dumazet 写道:
> > On Thu, Aug 20, 2026 at 12:04 PM dongchenchen (A)
> > <[email protected]> wrote:
> >>
> >>
> >>
> >> 在 2026/8/20 17:16, Eric Dumazet 写道:
> >>> On Thu, Aug 20, 2026 at 10:58 AM dongchenchen (A)
> >>> <[email protected]> wrote:
> >>>>
> >>>>
> >>>>> On Thu, Aug 13, 2026 at 5:30 AM Dong Chenchen <[email protected]> wrote:
> >>>>>>
> >>>>>> Syzbot reported a crash in qdisc_pkt_len_segs_init() caused by a stale
> >>>>>> transport_header offset after tunnel decapsulation.
> >>>>>>
> >>>>>> BUG: unable to handle page fault for address: ffffed102091a42e
> >>>>>> Oops: Oops: 0000 [#1] SMP KASAN NOPTI
> >>>>>> CPU: 0 UID: 0 PID: 340 Comm: qdisc_uaf_repro Not tainted 7.2.0-rc4-00061-g248951ddc14d #256 PREEMPT(full)
> >>>>>> Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> >>>>>> RIP: 0010:__asan_load2
> >>>>>> <IRQ>
> >>>>>> qdisc_pkt_len_segs_init (net/core/dev.c:4145)
> >>>>>> __dev_queue_xmit (net/core/dev.c:4787)
> >>>>>> br_dev_queue_push_xmit (net/bridge/br_forward.c:53)
> >>>>>> br_handle_frame_finish (net/bridge/br_input.c:229)
> >>>>>> br_handle_frame (net/bridge/br_input.c:315)
> >>>>>> __netif_receive_skb_core.constprop.0 (net/core/dev.c:6099)
> >>>>>> __netif_receive_skb_list_core (net/core/dev.c:6287)
> >>>>>> netif_receive_skb_list_internal (net/core/dev.c:6445)
> >>>>>> napi_complete_done (net/core/dev.c:6813)
> >>>>>> gro_cell_poll (net/core/gro_cells.c:74)
> >>>>>> __napi_poll (net/core/dev.c:7735)
> >>>>>> net_rx_action (net/core/dev.c:7798 net/core/dev.c:7955)
> >>>>>> handle_softirqs (kernel/softirq.c:622)
> >>>>>> do_softirq (kernel/softirq.c:523  kernel/softirq.c:510 )
> >>>>>> __local_bh_enable_ip (kernel/softirq.c:450)
> >>>>>> tun_get_user (drivers/net/tun.c:1986 (discriminator 1))
> >>>>>> tun_chr_write_iter (drivers/net/tun.c:2032)
> >>>>>>
> >>>>>> The crash requires four conditions to line up:
> >>>>>>
> >>>>>> 1. The incoming packet is encapsulated and carries GSO metadata.  The outer
> >>>>>>       transport header offset is stored in skb->transport_header while the
> >>>>>>       packet is still in the outer tunnel context.
> >>>>>> 2. The tunnel receiver strips the outer headers.  skb->data is advanced to
> >>>>>>       the inner frame, but skb->transport_header is left pointing to the
> >>>>>>       now-removed outer L4 header, so it becomes a negative offset relative to
> >>>>>>       the new data.
> >>>>>> 3. The inner frame is not delivered to the local IP stack.  Instead, it
> >>>>>>       is forwarded at L2 by a bridge or HSR, so ip_rcv_core() never runs and
> >>>>>>       the transport header is not reset to the inner L4 offset.
> >>>>>> 4. The forwarding path calls __dev_queue_xmit(), which enters
> >>>>>>       qdisc_pkt_len_segs_init().  That function computes the GSO header length
> >>>>>>       from skb_transport_offset(skb).  Because the offset is negative, the
> >>>>>>       unsigned cast overflows and pskb_may_pull(skb, hdr_len +
> >>>>>>       sizeof(struct tcphdr)) reads past the end of the skb, triggering a
> >>>>>>       KASAN fault or page fault.
> >>>>>>
> >>>>>> Fix this by clearing skb->transport_header to the ~0U sentinel at the
> >>>>>> tunnel decapsulation boundary, after each tunnel receive function has
> >>>>>> finished all processing that needs the outer L4 header and before the skb
> >>>>>> is handed to GRO or the stack. The IP/GRO receive paths then set the
> >>>>>> transport header correctly when they parse the inner packet.
> >>>>>>
> >>>>>> Fixes: 7fb4c1967011 ("net: pull headers in qdisc_pkt_len_segs_init()")
> >>>>>
> >>>>> This Fixes: tag seems pretty random to me. Issue was present long
> >>>>> before that commit.
> >>>>> Next time, please be precise with the `Fixes:` tag; it's not decoration.
> >>>>>
> >>>> Thanks for your review. Sorry for the late reply.
> >>>> The stale transport header has been present since 1da177e4c3f4
> >>>> ("Linux-2.6.12-rc2"), which is completely latent until
> >>>> 7fb4c1967011 ("net: pull headers in qdisc_pkt_len_segs_init()"
> >>>> was merged.
> >>>>
> >>>>    From the perspective of the root cause introduced by the issue,
> >>>> we can use 1da177e4c3f4 ("Linux-2.6.12-rc2") as fixtag.
> >>>>
> >>>>> Also, calling skb_unset_transport_header(skb) directly from
> >>>>> gro_cells_receive() would make more sense and
> >>>>> your patch would be less intrusive.
> >>>>>
> >>>>
> >>>> gtp，sit use netif_rx to receive pkt. so modifying only gro_cells_receive
> >>>> cannot solve all the problems. netif_rx is widely
> >>>> used and is not only invoked by the tunnel driver, and serval path
> >>>> (such as genve encap) will reset transport header before netif_rx.
> >>>> Therefore, should we retain unset in the GTP/SIT tunnel?
> >>>
> >>> Hi Dong,
> >>>
> >>> 1) Regarding SIT:
> >>>
> >>> SIT decapsulates IPv6 over IPv4 (IPPROTO_IPV6). There is no outer L4 transport
> >>> header (no UDP/TCP), so skb->transport_header is not pointing to an
> >>> outer L4 header.
> >>>
> >>> 2) Regarding GTP:
> >>>
> >>> GTP is an L3-only point-to-point device (ARPHRD_NONE). It cannot be
> >>> enslaved by a bridge
> >>> or forwarded at L2; its packets always go through ip_rcv()/ip6_rcv()
> >>> where the headers
> >>> are properly parsed/reset,
> >>>
> >> Hi, Eric. Thanks for the review!
> >>
> >> I missed the rcv path in the commit message.
> >> With an ingress qdisc attached to gtp, the stale transport header
> >> will been consumed in sch_handle_ingress() and its not covered by
> >> gro_cells_receive() either. So maybe gtp is not safe by
> >> construction.
> >>
> >> gtp_encap_rcv
> >>     gtp_rx //stale transport header
> >>       __netif_rx
> >>         __netif_receive_skb_core
> >>           sch_handle_ingress
> >>             qdisc_pkt_len_segs_init
> >>
> >
> > Let's step back a bit :)
> >
> > For non-GSO packets, qdisc_pkt_len_segs_init() returns immediately:
> >
> > qdisc_skb_cb(skb)->pkt_len = skb->len;
> > if (!shinfo->gso_size) {
> >      qdisc_skb_cb(skb)->pkt_segs = 1;
> >      return SKB_NOT_DROPPED_YET;
> > }
> >
> > It does not inspect skb->transport_header at all. It will be set
> > eventually later if needed,
> >
> > GTP does not implement GRO on RX and does not use gro_cells, so incoming
> > packets in gtp_rx() have gso_size == 0.
> >
>
> Hi, Eric. Thanks for your patient explanation.
>
> Normal GTP traffic indeed cannot trigger this issue and pkt without
> SKB_GSO_UDP_TUNNEL will be caught by udp_unexpected_gso.
> The only way to reproduce this for gtp is by constructing and injecting
> packets, such as TAP device injects a GSO UDP-tunnel packet (with
> GSO_TCPV4 and GSO_UDP_TUNNEL flag, Ethernet/IPv4/UDP/GTP-U/IPV4/TCP).
> Should we add interception for such illegal pkts?
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index af260ff5462a..8620428b950f 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -4138,6 +4138,10 @@ static enum skb_drop_reason
> qdisc_pkt_len_segs_init(struct sk_buff *skb)
>          if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
>                  const struct tcphdr *th;
>
> +               /* Defensive: stale offsets from buggy tunnel decap can
> underflow. */
> +               if (unlikely(hdr_len > INT_MAX))
> +                       return 0;
> +
>                  if (!pskb_may_pull(skb, hdr_len + sizeof(struct tcphdr)))
>                          return SKB_DROP_REASON_SKB_BAD_GSO;

Certainly not. Do not add defensive code in the fast path for GTP that
nobody uses.

Add everything you think of in GTP itself, I do not care.

 I suggest you look at skb_gso_reset(), this might be missing in GTP.

commit e3c674db356c4303804b2415e7c2b11776cdd8c3
Author: Antoine Tenart <[email protected]>
Date:   Thu Sep 4 14:53:50 2025 +0200

    tunnels: reset the GSO metadata before reusing the skb