On 2019/4/24 17:11, Jason Wang wrote:
>
> On 2019/4/24 上午12:41, Cong Wang wrote:
>> On Mon, Apr 22, 2019 at 11:42 PM Jason Wang <jasow...@redhat.com> wrote:
>>>
>>> On 2019/4/23 下午2:00, Cong Wang wrote:
>>>> On Mon, Apr 22, 2019 at 2:41 AM Jason Wang <jasow...@redhat.com> wrote:
>>>>> On 2019/4/22 上午11:57, YueHaibing wrote:
>>>>>> We get a KASAN report as below, but don't have any reproducer.
>>>>>>
>>>>>> Any comments are appreciated.
>>>>>>
>>>>>> ==================================================================
>>>>>> BUG: KASAN: use-after-free in tun_net_xmit+0x1670/0x1750
>>>>>> drivers/net/tun.c:1104
>>>>>> Read of size 8 at addr ffff88836cc26a70 by task swapper/3/0
>>>>> Which kernel version did you use? The calltrace points out the a use
>>>>> after free for tun_file structure which should be synchronized through
>>>>> RCU + RTNL lock.
>>>> The tfile socket has to be marked with SOCK_RCU_FREE in order
>>>> to fully respect the RCU grace period.
>>>>
>>>> diff --git a/drivers/net/tun.c b/drivers/net/tun.c
>>>> index e9ca1c088d0b..31c3210288cb 100644
>>>> --- a/drivers/net/tun.c
>>>> +++ b/drivers/net/tun.c
>>>> @@ -3431,6 +3431,7 @@ static int tun_chr_open(struct inode *inode,
>>>> struct file * file)
>>>> file->private_data = tfile;
>>>> INIT_LIST_HEAD(&tfile->next);
>>>>
>>>> + sock_set_flag(&tfile->sk, SOCK_RCU_FREE);
>>>> sock_set_flag(&tfile->sk, SOCK_ZEROCOPY);
>>>>
>>>> return 0;
>>>
>>> We did a synchronize_net() when socket is detached from netdevice in
>>> __tun_detach() so it looks to me this is unnecessary.
>> I knew, but it is only called conditionally, that is:
>>
>> 695 if (tun && !tfile->detached) {
>> ...
>> 710
>> 711 synchronize_net();
>>
>> And it looks like syzbot just skipped this condition,
>
>
> If tfile is detached, it should have gone for the path of synchronize_net()
> before. If tfile is never attached, tun_net_xmit() doesn't have the chance to
> access that. I wonder whether or not we should use WRITE_ONCE() for
> tun->numqueues-- in this fucntion. If the value was not committed to memory
> before synchronize_net(), we may race with tun_net_xmit() which check txq
> against tun->numqueues.
>
>
>> this is why I believe
>> you still need to respect RCU grace period _unconditionally_ for tfile.
>
>
> This is true if I miss subtle race in the code.
>
>
> Haibing: could you please try the following test?
>
> 1) start VM with multiple queue
>
> 2) using pktgen to inject packets to all queues through tap
>
> 3) using ethtool to change the combined channels in guest in a loop
>
> 4) kill the guest
>
>
Ok, I will try this.
> Thanks
>
>
>> Thanks.
>
> .
>
