skb_gso_transport_seglen() derives the TCP header length with tcp_hdrlen() or inner_tcp_hdrlen(). Both helpers dereference transport header metadata without validating it first. A TUN user can supply a TCP GSO packet without NEEDS_CSUM and with an invalid IP header. The skb remains GSO while transport_header keeps the unset sentinel. TBF and police can then reach the length validator and read tcp->doff outside the skb head. On the RX path, CONFIG_DEBUG_NET currently lets the unset marker survive to ingress while non-debug builds still apply a temporary compatibility reset. Validate the consumer instead of relying on that reset. Validate TCP header ordering, linear bounds and fixed header presence before either public GSO length check. For encapsulated TCP, validate the inner offsets while allowing the outer and inner transport offsets to be equal, as required by IPIP. Also validate the MAC header for the MAC length variant. Leave non-TCP GSO behavior unchanged. Those paths do not dereference a TCP header, and valid FCoE skbs can have no transport header. KASAN reports: BUG: KASAN: slab-out-of-bounds in skb_gso_transport_seglen Read of size 2 by task poc/133 skb_gso_transport_seglen (net/core/gso.c:155) skb_gso_validate_mac_len (net/core/gso.c:270) tbf_enqueue (net/sched/sch_tbf.c:260) dev_qdisc_enqueue (net/core/dev.c:4227) __dev_queue_xmit (net/core/dev.c:4884) Cc: stable@vger.kernel.org Fixes: 4d0820cf6a55 ("sch_tbf: handle too small burst") Reported-by: Assisted-by: LLM Signed-off-by: Weiming Shi --- net/core/gso.c | 33 ++++++++++++++++++++++++++++++++- 1 file changed, 32 insertions(+), 1 deletion(-) diff --git a/net/core/gso.c b/net/core/gso.c index bcd156372f4df..7c76f721fe24e 100644 --- a/net/core/gso.c +++ b/net/core/gso.c @@ -240,6 +240,29 @@ static inline bool skb_gso_size_check(const struct sk_buff *skb, return true; } +/* TCP segment length reads doff, so validate its header offsets first. */ +static bool skb_gso_tcp_header_valid(const struct sk_buff *skb) +{ + unsigned int transport = skb->transport_header; + unsigned int tail = skb_tail_pointer(skb) - skb->head; + + if (!skb_is_gso_tcp(skb)) + return true; + + if (!skb_transport_header_was_set(skb) || + transport <= skb->network_header || transport > tail) + return false; + + if (skb->encapsulation) { + transport = skb->inner_transport_header; + if (transport <= skb->inner_network_header || + transport < skb->transport_header || transport > tail) + return false; + } + + return sizeof(struct tcphdr) <= tail - transport; +} + /** * skb_gso_validate_network_len - Will a split GSO skb fit into a given MTU? * @@ -252,6 +275,9 @@ static inline bool skb_gso_size_check(const struct sk_buff *skb, */ bool skb_gso_validate_network_len(const struct sk_buff *skb, unsigned int mtu) { + if (unlikely(!skb_gso_tcp_header_valid(skb))) + return false; + return skb_gso_size_check(skb, skb_gso_network_seglen(skb), mtu); } EXPORT_SYMBOL_GPL(skb_gso_validate_network_len); @@ -267,7 +293,12 @@ EXPORT_SYMBOL_GPL(skb_gso_validate_network_len); */ bool skb_gso_validate_mac_len(const struct sk_buff *skb, unsigned int len) { + if (unlikely(!skb_gso_tcp_header_valid(skb) || + (skb_is_gso_tcp(skb) && + (!skb_mac_header_was_set(skb) || + skb->transport_header <= skb->mac_header)))) + return false; + return skb_gso_size_check(skb, skb_gso_mac_seglen(skb), len); } EXPORT_SYMBOL_GPL(skb_gso_validate_mac_len); - -- 2.55.0