[PATCH 0/2] net/af_packet: Fix kernel BUG in __skb_gso_segment

[Tudor Ambarus]

The series is intended for stable@xxxxxxxxxxxxxxx # 5.4+

Syzkaller reported the following bug on linux-5.{4, 10, 15}.y:
https://syzkaller.appspot.com/bug?id=ce5575575f074c33ff80d104f5baee26f22e95f5

The upstream commit that introduces this bug is:
1ed1d5921139 ("net: skip virtio_net_hdr_set_proto if protocol already set")

Upstream fixes the bug with the following commits, one of which introduces
new support:
e9d3f80935b6 ("net/af_packet: make sure to pull mac header")
dfed913e8b55 ("net/af_packet: add VLAN support for AF_PACKET SOCK_RAW GSO") 

The additional logic and risk backported seems manageable.

The blammed commit introduces a kernel BUG in __skb_gso_segment for
AF_PACKET SOCK_RAW GSO VLAN tagged packets. What happens is that
virtio_net_hdr_set_proto() exists early as skb->protocol is already set to
ETH_P_ALL. Then in packet_parse_headers() skb->protocol is set to
ETH_P_8021AD, but neither the network header position is adjusted, nor the
mac header is pulled. Thus when we get to validate the xmit skb and enter
skb_mac_gso_segment(), skb->mac_len has value 14, but vlan_depth gets
updated to 18 after skb_network_protocol() is called. This causes the
BUG_ON from __skb_pull(skb, vlan_depth) to be hit, as the mac header has
not been pulled yet.

The fixes from upstream backported cleanly without conflicts. I updated
the commit message of the first patch to describe the problem encountered,
and added Cc, Fixes, Reported-by and Tested-by tags. For the second patch
I just added Cc to stable indicating the versions to be fixed, and added
my Tested and Signed-off-by tags.

I tested the patches on linux-5.{4, 10, 15}.y.

Eric Dumazet (1):
  net/af_packet: make sure to pull mac header

Hangbin Liu (1):
  net/af_packet: add VLAN support for AF_PACKET SOCK_RAW GSO

 net/packet/af_packet.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

-- 
2.34.1