Re: [PATCH net v2 1/1] ip: orphan prefetched skbs before multicast forwarding

[Ido Schimmel <idosch@nvidia.com>]

On Thu, Aug 06, 2026 at 11:18:56PM +0800, Zhiling Zou wrote:
> IPv4 and IPv6 input preserve an skb->sk association installed by
> bpf_sk_assign() so that local delivery can use the selected socket under
> RCU. IPv6 can also get such an association from udp_v6_early_demux(),
> which runs in ip6_rcv_finish_core(), after ip6_rcv_core().

Last sentence is not accurate since it doesn't cover IPv4 which is also
fixed here. I suggest something like:

"Both address families can also prefetch a socket in UDP early demux. In
both paths (BPF and UDP early demux) a reference is not guaranteed to be
held on the socket."

> 
> The reproduced UDPv6 packet has a multicast IP destination but a unicast
> destination MAC address. It is therefore classified as PACKET_HOST and
> passes the UDP early-demux check.

This is again specific to IPv6 although IPv4 suffers from the same
problem. I think we can drop this paragraph.

> 
> When the multicast packet is not locally deliverable, IPv6 hands the

s/When the/When a/

> original skb to ip6_mr_input(). IPv4's ip_mr_input() similarly keeps the
> original skb when local delivery is not needed. Either path can put the
> skb on an unresolved multicast route queue or forward it after the
> receive-side RCU section ends.
> 
> After the prefetched socket is destroyed, a later skb free invokes
> sock_pfree() and dereferences the stale skb->sk. Orphan the skb before
> each non-local multicast forwarding path. Local delivery retains the
> original skb; the existing skb_clone() calls provide multicast forwarding
> with a socket-free clone.
> 
> Fixes: cf7fbe660f2d ("bpf: Add socket assign support")
> Fixes: 08842c43d016 ("udp: no longer touch sk->sk_refcnt in early demux")
> Cc: stable@vger.kernel.org
> Reported-by: Vega <vega@nebusec.ai>
> Signed-off-by: Zhiling Zou <zhilinz@nebusec.ai>

The rest looks fine to me.

Both sashikos mention other possible instances of the bug, but I believe
they require bpf_sk_assign(), unlike UDP early demux which is used here.
Assuming these issues are real and reproducible, they should be fixed in
a separate patchset.