Jakub Kicinski wrote:
> A not-so-careful NAT46 BPF program can crash the kernel
> if it indiscriminately flips ingress packets from v4 to v6:
>
> BUG: kernel NULL pointer dereference, address: 0000000000000000
> ip6_rcv_core (net/ipv6/ip6_input.c:190:20)
> ipv6_rcv (net/ipv6/ip6_input.c:306:8)
> process_backlog (net/core/dev.c:6186:4)
> napi_poll (net/core/dev.c:6906:9)
> net_rx_action (net/core/dev.c:7028:13)
> do_softirq (kernel/softirq.c:462:3)
> netif_rx (net/core/dev.c:5326:3)
> dev_loopback_xmit (net/core/dev.c:4015:2)
> ip_mc_finish_output (net/ipv4/ip_output.c:363:8)
> NF_HOOK (./include/linux/netfilter.h:314:9)
> ip_mc_output (net/ipv4/ip_output.c:400:5)
> dst_output (./include/net/dst.h:459:9)
> ip_local_out (net/ipv4/ip_output.c:130:9)
> ip_send_skb (net/ipv4/ip_output.c:1496:8)
> udp_send_skb (net/ipv4/udp.c:1040:8)
> udp_sendmsg (net/ipv4/udp.c:1328:10)
>
> The output interface has a 4->6 program attached at ingress.
> We try to loop the multicast skb back to the sending socket.
> Ingress BPF runs as part of netif_rx(), pushes a valid v6 hdr
> and changes skb->protocol to v6. We enter ip6_rcv_core which
> tries to use skb_dst(). But the dst is still an IPv4 one left
> after IPv4 mcast output.
>
> Clear the dst in all BPF helpers which change the protocol.
> Try to preserve metadata dsts, those may carry non-routing
> metadata.
>
> Cc: stable@xxxxxxxxxxxxxxx
> Reviewed-by: Maciej Żenczykowski <maze@xxxxxxxxxx>
> Acked-by: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
> Fixes: d219df60a70e ("bpf: Add ipip6 and ip6ip decap support for bpf_skb_adjust_room()")
> Fixes: 1b00e0dfe7d0 ("bpf: update skb->protocol in bpf_skb_net_grow")
> Fixes: 6578171a7ff0 ("bpf: add bpf_skb_change_proto helper")
> Signed-off-by: Jakub Kicinski <kuba@xxxxxxxxxx>
Reviewed-by: Willem de Bruijn <willemb@xxxxxxxxxx>
