CVE-2024-26804

In the Linux kernel, the following vulnerability has been resolved:

net: ip_tunnel: prevent perpetual headroom growth

syzkaller triggered following kasan splat:
BUG: KASAN: use-after-free in __skb_flow_dissect+0x19d1/0x7a50 net/core/flow_dissector.c:1170
Read of size 1 at addr ffff88812fb4000e by task syz-executor183/5191
[..]
kasan_report+0xda/0x110 mm/kasan/report.c:588
__skb_flow_dissect+0x19d1/0x7a50 net/core/flow_dissector.c:1170
skb_flow_dissect_flow_keys include/linux/skbuff.h:1514 [inline]
___skb_get_hash net/core/flow_dissector.c:1791 [inline]
__skb_get_hash+0xc7/0x540 net/core/flow_dissector.c:1856
skb_get_hash include/linux/skbuff.h:1556 [inline]
ip_tunnel_xmit+0x1855/0x33c0 net/ipv4/ip_tunnel.c:748
ipip_tunnel_xmit+0x3cc/0x4e0 net/ipv4/ipip.c:308
__netdev_start_xmit include/linux/netdevice.h:4940 [inline]
netdev_start_xmit include/linux/netdevice.h:4954 [inline]
xmit_one net/core/dev.c:3548 [inline]
dev_hard_start_xmit+0x13d/0x6d0 net/core/dev.c:3564
__dev_queue_xmit+0x7c1/0x3d60 net/core/dev.c:4349
dev_queue_xmit include/linux/netdevice.h:3134 [inline]
neigh_connected_output+0x42c/0x5d0 net/core/neighbour.c:1592
...
ip_finish_output2+0x833/0x2550 net/ipv4/ip_output.c:235
ip_finish_output+0x31/0x310 net/ipv4/ip_output.c:323
..
iptunnel_xmit+0x5b4/0x9b0 net/ipv4/ip_tunnel_core.c:82
ip_tunnel_xmit+0x1dbc/0x33c0 net/ipv4/ip_tunnel.c:831
ipgre_xmit+0x4a1/0x980 net/ipv4/ip_gre.c:665
__netdev_start_xmit include/linux/netdevice.h:4940 [inline]
netdev_start_xmit include/linux/netdevice.h:4954 [inline]
xmit_one net/core/dev.c:3548 [inline]
dev_hard_start_xmit+0x13d/0x6d0 net/core/dev.c:3564
...

The splat occurs because skb->data points past skb->head allocated area.
This is because neigh layer does:
__skb_pull(skb, skb_network_offset(skb));

... but skb_network_offset() returns a negative offset and __skb_pull()
arg is unsigned. IOW, we skb->data gets "adjusted" by a huge value.

The negative value is returned because skb->head and skb->data distance is
more than 64k and skb->network_header (u16) has wrapped around.

The bug is in the ip_tunnel infrastructure, which can cause
dev->needed_headroom to increment ad infinitum.

The syzkaller reproducer consists of packets getting routed via a gre
tunnel, and route of gre encapsulated packets pointing at another (ipip)
tunnel. The ipip encapsulation finds gre0 as next output device.

This results in the following pattern:

1). First packet is to be sent out via gre0.
Route lookup found an output device, ipip0.

2).
ip_tunnel_xmit for gre0 bumps gre0->needed_headroom based on the future
output device, rt.dev->needed_headroom (ipip0).

3).
ip output / start_xmit moves skb on to ipip0. which runs the same
code path again (xmit recursion).

4).
Routing step for the post-gre0-encap packet finds gre0 as output device
to use for ipip0 encapsulated packet.

tunl0->needed_headroom is then incremented based on the (already bumped)
gre0 device headroom.

This repeats for every future packet:

gre0->needed_headroom gets inflated because previous packets' ipip0 step
incremented rt->dev (gre0) headroom, and ipip0 incremented because gre0
needed_headroom was increased.

For each subsequent packet, gre/ipip0->needed_headroom grows until
post-expand-head reallocations result in a skb->head/data distance of
more than 64k.

Once that happens, skb->network_header (u16) wraps around when
pskb_expand_head tries to make sure that skb_network_offset() is unchanged
after the headroom expansion/reallocation.

After this skb_network_offset(skb) returns a different (and negative)
result post headroom expansion.

The next trip to neigh layer (or anything else that would __skb_pull the
network header) makes skb->data point to a memory location outside
skb->head area.

v2: Cap the needed_headroom update to an arbitarily chosen upperlimit to
prevent perpetual increase instead of dropping the headroom increment
completely.