Re: KMSAN: uninit-value in can_receive

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

On 11/19/19 12:24 PM, Oliver Hartkopp wrote:
> Hi Eric,
> 
> On 19/11/2019 17.53, Eric Dumazet wrote:
>>
>>
>> On 11/18/19 11:35 PM, Oliver Hartkopp wrote:
>>>
>>
>>>
>>> See ioctl$ifreq https://syzkaller.appspot.com/x/log.txt?x=14563416e00000
>>>
>>> 23:11:34 executing program 2:
>>> r0 = socket(0x200000000000011, 0x3, 0x0)
>>> ioctl$ifreq_SIOCGIFINDEX_vcan(r0, 0x8933, &(0x7f0000000040)={'vxcan1\x00', <r1=>0x0})
>>> bind$packet(r0, &(0x7f0000000300)={0x11, 0xc, r1}, 0x14)
>>> sendmmsg(r0, &(0x7f0000000d00), 0x400004e, 0x0)
>>>
>>> We only can receive skbs from (v(x))can devices.
>>> No matter if someone wrote to them via PF_CAN or PF_PACKET.
>>> We check for ETH_P_CAN(FD) type and ARPHRD_CAN dev type at rx time.
>>
>> And what entity sets the can_skb_prv(skb)->skbcnt to zero exactly ?
>>
>>>
>>>>> We additionally might think about introducing a check whether we have a
>>>>> can_skb_reserve() created skbuff.
>>>>>
>>>>> But even if someone forged a skbuff without this reserved space the
>>>>> access to can_skb_prv(skb)->skbcnt would point into some CAN frame
>>>>> content - which is still no access to uninitialized content, right?
>>>
>>> So this question remains still valid whether we have a false positive from KMSAN here.
>>
>> I do not believe it is a false positive.
>>
>> It seems CAN relies on some properties of low level drivers using alloc_can_skb() or similar function.
>>
>> Why not simply fix this like that ?
>>
>> diff --git a/net/can/af_can.c b/net/can/af_can.c
>> index 128d37a4c2e0ba5d8db69fcceec8cbd6a79380df..3e71a78d82af84caaacd0ef512b5e894efbf4852 100644
>> --- a/net/can/af_can.c
>> +++ b/net/can/af_can.c
>> @@ -647,8 +647,9 @@ static void can_receive(struct sk_buff *skb, struct net_device *dev)
>>          pkg_stats->rx_frames_delta++;
>>            /* create non-zero unique skb identifier together with *skb */
>> -       while (!(can_skb_prv(skb)->skbcnt))
>> +       do {
>>                  can_skb_prv(skb)->skbcnt = atomic_inc_return(&skbcounter);
>> +       } while (!(can_skb_prv(skb)->skbcnt));
>>            rcu_read_lock();
>>   
> 
> Please check commit d3b58c47d330d ("can: replace timestamp as unique skb attribute").

Oh well... This notion of 'unique skb attribute' is interesting...

> 
> can_skb_prv(skb)->skbcnt is set to 0 at skb creation time when sending CAN frames from local host or receiving CAN frames from a real CAN interface.

We can not enforce this to happen with a virtual interface.

> 
> When a CAN skb is received by the net layer the *first* time it gets a unique value which we need for a per-cpu filter mechanism in raw_rcv().
> 
> So where's the problem to check for (!(can_skb_prv(skb)->skbcnt)) in a while statement? I can't see a chance for an uninitialized value there.

You have to make sure the packet has been properly cooked by a 'real CAN interface' then, and reject them if not.