Re: [PATCH] net: prevent receive header offset overflow during decapsulation

From: Chris Arges

Date: Wed Nov 18 2026 - 20:40:19 EST


On 2026-11-18 17:39:27, netdev-bot+sashiko@xxxxxxxxxx wrote:
> Thank you for your contribution! Sashiko AI review found 1 potential
> issue(s) to consider.
>
> Critical: 0 · High: 1 · Medium: 0 · Low: 0
>
> If you believe none of the issues are legitimate, please reply to each
> issue of Medium severity or higher and explain why the AI is wrong.
> If any of them are legitimate and you plan to respin, please let patchwork
> know by including "pw-bot: cr" as a separate line at the end of your reply
> (one such reply per series is enough).
>
> - [High] The new SKB_MAX_RX_HEADROOM (32 KiB) checks run in only three
> places: when __netif_receive_skb_core() is entered…
>
> --
>
> net: prevent receive header offset overflow during decapsulation
>
> This patch limits receive headroom to 32 KiB (SKB_MAX_RX_HEADROOM). The
> goal is to keep the 16-bit skb header offsets from being truncated after
> repeated decapsulation. The check runs on entry to
> __netif_receive_skb_core() and at the resubmit points of
> ip_protocol_deliver_rcu() and ip6_protocol_deliver_rcu(), and drops
> packets with the new SKB_RX_HEADROOM drop reason.
>
> [ ... ]
>
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index 3c58db732761a..6140fd9d11cd0 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -6223,6 +6223,10 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
> > trace_netif_receive_skb(skb);
> >
> > orig_dev = skb->dev;
> > + if (unlikely(skb_headroom(skb) > SKB_MAX_RX_HEADROOM)) {
> > + drop_reason = SKB_DROP_REASON_SKB_RX_HEADROOM;
> > + goto drop;
> > + }
>
> [Severity: High]
>
> This check runs once, when __netif_receive_skb_core() is entered. That is
> before the another_round label, sch_handle_ingress(), vlan_do_receive()
> and the rx_handler call.
>
> Can headers stripped later in the same receive pass still push the 16-bit
> header offsets past U16_MAX?
>
> For example, tc act_mpls with TCA_MPLS_ACT_POP calls skb_mpls_pop() once
> per action. Each call stores the current headroom into the u16 header
> fields:
>
> net/core/skbuff.c:skb_mpls_pop() {
> ...
> __skb_pull(skb, MPLS_HLEN);
> skb_reset_mac_header(skb);
> skb_set_network_header(skb, mac_len);
> ...
> }
>
> Outside CONFIG_DEBUG_NET, the only guard in those helpers is
> DEBUG_NET_WARN_ON_ONCE(), so the value gets truncated silently.
>
> The number of pops per pass doesn't look bounded:
>
> - a filter can hold up to 32 actions
> - a filter whose action control is continue returns a negative result,
> and __tcf_classify() moves on to the next tcf_proto without counting
> it toward max_reclassify_loop (that limit only covers the
> TC_ACT_RECLASSIFY / TC_ACT_GOTO_CHAIN reset path)
> - about 512 matchall or flower filters with 32 pops each comes to
> roughly 16K pops in one pass
>
> veth_setup() sets dev->max_mtu = ETH_MAX_MTU. A frame of about 65.5 KB
> carrying about 16K MPLS labels, sent with AF_PACKET on one veth end, would
> reach __netif_receive_skb_core() on the peer with about 30 bytes of
> headroom and pass this check.
>
> tc ingress could then pop the labels while skb_ensure_writable() keeps
> pulling data from the frags into the linear head. At some point
> mac_header and network_header would wrap past U16_MAX.
>
> After the wrap, the remaining skb_mpls_pop() calls would run memmove(),
> skb_postpull_rcsum() and skb_mod_eth_type() at the truncated offsets.
> ip_rcv_core() would then read ip_hdr() from stale bytes and set
> transport_header from that position. L4 code could then misparse the
> packet, for example in tcp_data_queue():
>
> __skb_pull(skb, tcp_hdr(skb)->doff * 4);
>
> Here doff comes from the wrong bytes, so could this wrap skb->len or hit
> the BUG() in __skb_pull()?
>
> This setup needs only CAP_NET_ADMIN and CAP_NET_RAW in a netns owned by a
> user namespace.
>
> Similar in-pass header strips look possible elsewhere:
>
> - openvswitch pop_mpls with recirculation, where
> ovs_dp_process_packet() runs again without re-entering
> __netif_receive_skb_core()
> - bpf_skb_adjust_room() shrinking, which adds len directly to
> mac_header and network_header
>

These additional paths can be addressed in a follow-up patchset. I'd
prefer to fix the currently known cases first.

> The commit message says:
>
> Together, these checks
> prevent 16-bit header-offset truncation in cases of repeated
> decapsulation.
>
> Is that still accurate given these paths?
>
> Would it be more robust to bound the offset where it is stored, either in
> the skb_set_*_header() / skb_reset_*_header() helpers or in the pop/shrink
> helpers, instead of only at the chosen re-entry points?
>
> [ ... ]
>
> --
> Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261118-prevent_rx_header_offset_overflow-v1-1-92a5be55f994%40cloudflare.com

If the checks were moved to where the offset is stored (using something like
skb_set_transport_header_careful()), then we could introduce potential
performance issues.

--chris