Re: [Bug 209423] WARN_ON_ONCE() at rtl8169_tso_csum_v2()

[Eric Dumazet <eric.dumazet@gmail.com>]

On 10/8/20 10:54 PM, Heiner Kallweit wrote:
> On 08.10.2020 21:07, Eric Dumazet wrote:
>>
>>
>> On 10/8/20 8:50 PM, Eric Dumazet wrote:
>>>
>>>
>>> OK, it would be nice to know what is the input interface
>>>
>>> if4 -> look at "ip link | grep 4:"
>>>
>>> Then identifying the driver that built such a strange packet (32000
>>> bytes allocated in skb->head)
>>>
>>> ethtool -i ifname
>>>
>>
>> According to https://bugzilla.kernel.org/show_bug.cgi?id=209423
>>
>> iif4 is the tun200 interface used by openvpn.
>>
>> So this might be a tun bug, or lack of proper SKB_GSO_DODGY validation
>> in our stack for buggy/malicious packets.
>>
>>
> 
> Following old commit sounds like it might be related:
> 622e0ca1cd4d ("gro: Fix bogus gso_size on the first fraglist entry")
> 
> This code however was removed later in 58025e46ea2d ("net: gro: remove
> obsolete code from skb_gro_receive()")
> 

GRO wont keep in its queues a GSO packet
dev_gro_receive()
...
NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
...

Also note that tun no longer can inject a packet with a length of 134 bytes pretending
to have gso_size == 538

Look at virtio_net_hdr_to_skb() and commits
6dd912f82680 ("net: check untrusted gso_size at kernel entry")
7c6d2ecbda83 ("net: be more gentle about silly gso requests coming from user")

Really looking at the skb layout I suspect some usbnet bug and a use-after-free.

ASAN build might help.