Re: [PATCH net v2] net: vlan: fix a UAF in vlan_dev_real_dev()

[Petr Machata <petrm@nvidia.com>]

Petr Machata <petrm@nvidia.com> writes:

> Ziyang Xuan (William) <william.xuanziyang@huawei.com> writes:
>
>>> 
>>> Ziyang Xuan (William) <william.xuanziyang@huawei.com> writes:
>>> 
>>>> I need some time to test my some ideas. And anyone has good ideas, please
>>>> do not be stingy.
>>> 
>>> Jakub Kicinski <kuba@kernel.org> writes:
>>> 
>>>> I think we should move the dev_hold() to ndo_init(), otherwise it's
>>>> hard to reason if destructor was invoked or not if
>>>> register_netdevice() errors out.
>>> 
>>> That makes sense to me. We always put real_dev in the destructor, so we
>>> should always hold it in the constructor...
>>
>> Inject error before dev_hold(real_dev) in register_vlan_dev(), and execute
>> the following testcase:
>>
>> ip link add dev dummy1 type dummy
>> ip link add name dummy1.100 link dummy1 type vlan id 100 // failed for error injection
>> ip link del dev dummy1
>>
>> Make the problem repro. The problem is solved using the following fix
>> according to the Jakub's suggestion:
>>
>> diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
>> index a3a0a5e994f5..abaa5d96ded2 100644
>> --- a/net/8021q/vlan.c
>> +++ b/net/8021q/vlan.c
>> @@ -184,9 +184,6 @@ int register_vlan_dev(struct net_device *dev, struct netlink_ext_ack *extack)
>>         if (err)
>>                 goto out_unregister_netdev;
>>
>> -       /* Account for reference in struct vlan_dev_priv */
>> -       dev_hold(real_dev);
>> -
>>         vlan_stacked_transfer_operstate(real_dev, dev, vlan);
>>         linkwatch_fire_event(dev); /* _MUST_ call rfc2863_policy() */
>>
>> diff --git a/net/8021q/vlan_dev.c b/net/8021q/vlan_dev.c
>> index ab6dee28536d..a54535cbcf4c 100644
>> --- a/net/8021q/vlan_dev.c
>> +++ b/net/8021q/vlan_dev.c
>> @@ -615,6 +615,9 @@ static int vlan_dev_init(struct net_device *dev)
>>         if (!vlan->vlan_pcpu_stats)
>>                 return -ENOMEM;
>>
>> +       /* Get vlan's reference to real_dev */
>> +       dev_hold(real_dev);
>>
>>
>> If there is not any other idea and objection, I will send the fix patch later.
>>
>> Thank you!
>
> This fixes the issue in my repro, and does not seems to break anything.
> We'll take it to full regression overnight, I'll report back tomorrow
> about the result.

Sorry, was AFK yesterday. It does look good.