* [PATCH net] ipv4: don't let the transport header offset wrap
@ 2026-11-18 0:48 Chris Mason
2026-11-18 11:16 ` Ido Schimmel
2026-11-18 15:25 ` netdev-bot+sashiko
0 siblings, 2 replies; 3+ messages in thread
From: Chris Mason @ 2026-11-18 0:48 UTC (permalink / raw)
To: mason, dsahern, idosch, davem, edumazet, kuba, pabeni, horms,
willemb, kuniyu, netdev, josef
From: Josef Bacik <josef@toxicpanda.com>
skb->network_header and skb->transport_header are 16-bit offsets from
skb->head. ip_rcv_core() computes transport_header as
network_header + ihl * 4.
These offsets are normally small. But a tunnel pulls its outer headers
and hands the same skb back to the stack. skb->head stays where it is,
so skb->data moves further away from it with each decapsulation. After
nested decapsulations the end of the inner IP header can lie beyond what
16 bits can hold.
The fix used here is to set the offset with
skb_set_transport_header_careful(), and to drop the packet if the offset
does not fit.
The helper computes the offset from skb->data, which points to the IP
header in ip_rcv_core(). It does not use network_header. So it also
catches a packet whose network_header was already truncated, where
network_header + ihl * 4 would look valid.
Found by LLM code inspection, and confirmed with LLM generated
reproducers on an unmodified kernel. Fixed by LLM, and tested with
the reproducers and existing kernel selftests.
Fixes: 21d1196a35f5 ("ipv4: set transport header earlier")
Assisted-by: LLM
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Chris Mason <mason@kernel.org>
---
net/ipv4/ip_input.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c
index 9860178752b8..397043893376 100644
--- a/net/ipv4/ip_input.c
+++ b/net/ipv4/ip_input.c
@@ -572,7 +572,11 @@ static struct sk_buff *ip_rcv_core(struct sk_buff *skb, struct net *net)
}
iph = ip_hdr(skb);
- skb->transport_header = skb->network_header + iph->ihl*4;
+ if (!skb_set_transport_header_careful(skb, iph->ihl * 4)) {
+ drop_reason = SKB_DROP_REASON_PKT_TOO_BIG;
+ __IP_INC_STATS(net, IPSTATS_MIB_INDISCARDS);
+ goto drop;
+ }
/* Remove any debris in the socket control block */
memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
--
2.55.0
^ permalink raw reply related [flat|nested] 3+ messages in thread
* Re: [PATCH net] ipv4: don't let the transport header offset wrap
2026-11-18 0:48 [PATCH net] ipv4: don't let the transport header offset wrap Chris Mason
@ 2026-11-18 11:16 ` Ido Schimmel
2026-11-18 15:25 ` netdev-bot+sashiko
1 sibling, 0 replies; 3+ messages in thread
From: Ido Schimmel @ 2026-11-18 11:16 UTC (permalink / raw)
To: Chris Mason
Cc: dsahern, davem, edumazet, kuba, pabeni, horms, willemb, kuniyu,
netdev, josef, carges
On Tue, Nov 17, 2026 at 07:48:50PM -0500, Chris Mason wrote:
> From: Josef Bacik <josef@toxicpanda.com>
>
> skb->network_header and skb->transport_header are 16-bit offsets from
> skb->head. ip_rcv_core() computes transport_header as
> network_header + ihl * 4.
>
> These offsets are normally small. But a tunnel pulls its outer headers
> and hands the same skb back to the stack. skb->head stays where it is,
> so skb->data moves further away from it with each decapsulation. After
> nested decapsulations the end of the inner IP header can lie beyond what
> 16 bits can hold.
>
> The fix used here is to set the offset with
> skb_set_transport_header_careful(), and to drop the packet if the offset
> does not fit.
>
> The helper computes the offset from skb->data, which points to the IP
> header in ip_rcv_core(). It does not use network_header. So it also
> catches a packet whose network_header was already truncated, where
> network_header + ihl * 4 would look valid.
>
> Found by LLM code inspection, and confirmed with LLM generated
> reproducers on an unmodified kernel. Fixed by LLM, and tested with
> the reproducers and existing kernel selftests.
>
> Fixes: 21d1196a35f5 ("ipv4: set transport header earlier")
> Assisted-by: LLM
> Signed-off-by: Josef Bacik <josef@toxicpanda.com>
> Signed-off-by: Chris Mason <mason@kernel.org>
Chris Arges is going to post a more comprehensive patch soon.
> ---
> net/ipv4/ip_input.c | 6 +++++-
> 1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c
> index 9860178752b8..397043893376 100644
> --- a/net/ipv4/ip_input.c
> +++ b/net/ipv4/ip_input.c
> @@ -572,7 +572,11 @@ static struct sk_buff *ip_rcv_core(struct sk_buff *skb, struct net *net)
> }
>
> iph = ip_hdr(skb);
> - skb->transport_header = skb->network_header + iph->ihl*4;
> + if (!skb_set_transport_header_careful(skb, iph->ihl * 4)) {
> + drop_reason = SKB_DROP_REASON_PKT_TOO_BIG;
> + __IP_INC_STATS(net, IPSTATS_MIB_INDISCARDS);
> + goto drop;
> + }
>
> /* Remove any debris in the socket control block */
> memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
> --
> 2.55.0
>
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH net] ipv4: don't let the transport header offset wrap
2026-11-18 0:48 [PATCH net] ipv4: don't let the transport header offset wrap Chris Mason
2026-11-18 11:16 ` Ido Schimmel
@ 2026-11-18 15:25 ` netdev-bot+sashiko
1 sibling, 0 replies; 3+ messages in thread
From: netdev-bot+sashiko @ 2026-11-18 15:25 UTC (permalink / raw)
To: mason
Cc: dsahern, idosch, davem, edumazet, kuba, pabeni, horms, willemb,
kuniyu, netdev, josef
Thank you for your contribution! Sashiko AI review found 3 potential
issue(s) to consider.
Critical: 0 · High: 1 · Medium: 1 · Low: 1
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 IPv4 xfrm transport-mode (ESP/AH) receive path still lets
the transport header offset wrap, and the wrapped value still reaches…
- [Medium] Sibling receive paths still compute the transport header with
the unchecked 16-bit addition this patch removes from ip_rcv_core().
- [Low] The new drop path in ip_rcv_core() (net/ipv4/ip_input.c:576)
reports SKB_DROP_REASON_PKT_TOO_BIG.
--
ipv4: don't let the transport header offset wrap
This patch replaces the unchecked 16-bit transport_header computation in
ip_rcv_core() with skb_set_transport_header_careful(). If the offset from
skb->head doesn't fit, the packet is dropped. The helper computes the
offset from skb->data, so it also catches a network_header that was
already truncated.
> diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c
> index 9860178752b8c..3970438933764 100644
> --- a/net/ipv4/ip_input.c
> +++ b/net/ipv4/ip_input.c
> @@ -572,7 +572,11 @@ static struct sk_buff *ip_rcv_core(struct sk_buff *skb, struct net *net)
> }
>
> iph = ip_hdr(skb);
> - skb->transport_header = skb->network_header + iph->ihl*4;
> + if (!skb_set_transport_header_careful(skb, iph->ihl * 4)) {
[Severity: High]
After this check passes, can the transport header offset still wrap on the
IPv4 xfrm transport-mode (ESP/AH) receive path?
ip_rcv_core() only validates the outer header. After that,
esp_input_done2() pulls the ESP header and IV, and then sets the transport
header relative to the new skb->data:
net/ipv4/esp4.c:esp_input_done2() {
...
skb_pull_rcsum(skb, hlen);
if (x->props.mode == XFRM_MODE_TUNNEL ||
x->props.mode == XFRM_MODE_IPTFS)
skb_reset_transport_header(skb);
else
skb_set_transport_header(skb, -ihl);
...
}
Next, xfrm4_transport_input() moves the IP header and resets the transport
header without a check:
net/xfrm/xfrm_input.c:xfrm4_transport_input() {
...
skb->network_header = skb->transport_header;
}
ip_hdr(skb)->tot_len = htons(skb->len + ihl);
skb_reset_transport_header(skb);
...
}
This stores (u16)(N + ihl + hlen), where N is the outer IP header offset
from skb->head. Suppose N + ihl <= 65534, which passes the new check, but
N + ihl + hlen >= 65536. For ihl = 20 and hlen = 16 that is N in
[65500, 65514]. Then transport_header wraps to a small value while
network_header still fits.
The skb is then re-injected without going back through ip_rcv_core():
xfrm4_transport_finish()
NF_HOOK(NFPROTO_IPV4, NF_INET_PRE_ROUTING, ...)
xfrm4_rcv_encap_finish()
dst_input()
ip_local_deliver()
ip_local_deliver_finish()
__skb_pull(skb, skb_network_header_len(skb));
With the wrapped transport_header, skb_network_header_len() returns
0xFFFF0000 + ihl as a u32. __skb_pull() then raises skb->len by
65536 - ihl, so the skb->len < skb->data_len BUG() never fires. It also
moves skb->data about 4 GiB forward on 64-bit, or wraps it back toward
skb->head on 32-bit. The protocol handler then dereferences skb->data.
This needs a transport-mode SA and an ESP/AH packet that authenticates
against it. It also needs a head offset near 64 KiB, which the commit
message says nested decapsulation produces. It hasn't been reproduced at
runtime.
The subject says this wrap is fixed for IPv4. Should the series also
cover xfrm4_transport_input() and the esp4/ah4 transport-mode paths?
[Severity: Medium]
The unchecked 16-bit addition that this patch removes from ip_rcv_core()
is still present in several sibling receive paths:
net/bridge/br_netfilter_hooks.c:br_nf_pre_routing() {
...
skb->transport_header = skb->network_header + ip_hdr(skb)->ihl * 4;
...
}
net/bridge/br_netfilter_ipv6.c:br_nf_pre_routing_ipv6() {
...
skb->transport_header = skb->network_header + sizeof(struct ipv6hdr);
...
}
net/ipv6/ip6_input.c:ip6_rcv_core() {
...
skb->transport_header = skb->network_header + sizeof(*hdr);
...
}
The commit message says:
So it also catches a packet whose network_header was already
truncated, where network_header + ihl * 4 would look valid.
Does that hold only for native IPv4 input?
After VXLAN, GENEVE or gretap decapsulation, bridged traffic reaches
NF_HOOK(NF_INET_PRE_ROUTING) from br_nf_pre_routing() and never passes
through ip_rcv_core(). The outer header can pass the new check just under
the limit, and the tunnel then pulls about 50 more bytes.
br_validate_ipv4() is a copy of the ip_rcv_core() checks, but it still
validates ip_hdr(skb), which is based on network_header. If network_header
was truncated, it validates stale bytes, and the netfilter hooks see a
wrapped transport_header.
For IPv6, transport_header can wrap to 0..39, below network_header. That
looks like it would cause:
- ipv6_parse_hopopts() in net/ipv6/exthdrs.c reads
skb_transport_header(skb)[1] from stale bytes at the start of the head
buffer.
- ip6_parse_tlv() starts from a negative skb_network_header_len() offset,
so the jumbo, RA and IOAM handlers act on stale bytes.
- Local delivery drops the packet, because
pskb_pull(skb, skb_transport_offset(skb)) in ip6_protocol_deliver_rcu()
gets a huge length.
If network_header is truncated, IPv6 validation, route lookup and
PRE_ROUTING/FORWARD filtering all run on stale bytes. Forwarding still
sends the real bytes at skb->data.
Could filtering and routing decisions be made on one header while
different bytes are forwarded? Should br_nf_pre_routing(),
br_nf_pre_routing_ipv6() and ip6_rcv_core() get the same treatment in
this series?
> + drop_reason = SKB_DROP_REASON_PKT_TOO_BIG;
[Severity: Low]
This isn't a bug, but is SKB_DROP_REASON_PKT_TOO_BIG the right reason
here? include/net/dropreason-core.h documents it as:
* @SKB_DROP_REASON_PKT_TOO_BIG: packet size is too big (maybe exceed the
* MTU)
By this point the packet has been trimmed to tot_len and has passed every
length check, so it can be a minimal 20-byte datagram. The failing
condition is that skb->data - skb->head + ihl * 4 doesn't fit in the
16-bit transport_header. That is about buffer layout, not packet size.
The other in-tree users of this reason are size or MTU checks, for example
in ip6_output.c, openvswitch actions.c and xfrm_iptfs.c. As a result,
dropwatch and the kfree_skb tracepoint would report these drops as
size/MTU drops.
> + __IP_INC_STATS(net, IPSTATS_MIB_INDISCARDS);
> + goto drop;
> + }
[ ... ]
--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261118004850.52808-1-mason%40kernel.org
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2026-11-18 15:25 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 0:48 [PATCH net] ipv4: don't let the transport header offset wrap Chris Mason
2026-11-18 11:16 ` Ido Schimmel
2026-11-18 15:25 ` netdev-bot+sashiko
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox