Re: [PATCH v2] netfilter: xt_owner: Add supplementary groups option

[Eric Dumazet]

On 5/9/19 6:47 AM, Lukasz Pawelczyk wrote:
> On Wed, 2019-05-08 at 09:53 -0700, Eric Dumazet wrote:
>>
>> On 5/8/19 11:56 AM, Lukasz Pawelczyk wrote:
>>> On Wed, 2019-05-08 at 08:41 -0700, Eric Dumazet wrote:
>>>> On 5/8/19 11:25 AM, Lukasz Pawelczyk wrote:
>>>>> On Wed, 2019-05-08 at 07:58 -0700, Eric Dumazet wrote:
>>>>>> On 5/8/19 10:12 AM, Lukasz Pawelczyk wrote:
>>>>>>> The XT_SUPPL_GROUPS flag causes GIDs specified with
>>>>>>> XT_OWNER_GID to
>>>>>>> be also checked in the supplementary groups of a process.
>>>>>>>
>>>>>>> Signed-off-by: Lukasz Pawelczyk <l.pawelczyk@xxxxxxxxxxx>
>>>>>>> ---
>>>>>>>  include/uapi/linux/netfilter/xt_owner.h |  1 +
>>>>>>>  net/netfilter/xt_owner.c                | 23
>>>>>>> ++++++++++++++++++++-
>>>>>>> --
>>>>>>>  2 files changed, 21 insertions(+), 3 deletions(-)
>>>>>>>
>>>>>>> diff --git a/include/uapi/linux/netfilter/xt_owner.h
>>>>>>> b/include/uapi/linux/netfilter/xt_owner.h
>>>>>>> index fa3ad84957d5..d646f0dc3466 100644
>>>>>>> --- a/include/uapi/linux/netfilter/xt_owner.h
>>>>>>> +++ b/include/uapi/linux/netfilter/xt_owner.h
>>>>>>> @@ -8,6 +8,7 @@ enum {
>>>>>>>  	XT_OWNER_UID    = 1 << 0,
>>>>>>>  	XT_OWNER_GID    = 1 << 1,
>>>>>>>  	XT_OWNER_SOCKET = 1 << 2,
>>>>>>> +	XT_SUPPL_GROUPS = 1 << 3,
>>>>>>>  };
>>>>>>>  
>>>>>>>  struct xt_owner_match_info {
>>>>>>> diff --git a/net/netfilter/xt_owner.c
>>>>>>> b/net/netfilter/xt_owner.c
>>>>>>> index 46686fb73784..283a1fb5cc52 100644
>>>>>>> --- a/net/netfilter/xt_owner.c
>>>>>>> +++ b/net/netfilter/xt_owner.c
>>>>>>> @@ -91,11 +91,28 @@ owner_mt(const struct sk_buff *skb,
>>>>>>> struct
>>>>>>> xt_action_param *par)
>>>>>>>  	}
>>>>>>>  
>>>>>>>  	if (info->match & XT_OWNER_GID) {
>>>>>>> +		unsigned int i, match = false;
>>>>>>>  		kgid_t gid_min = make_kgid(net->user_ns, info-
>>>>>>>> gid_min);
>>>>>>>  		kgid_t gid_max = make_kgid(net->user_ns, info-
>>>>>>>> gid_max);
>>>>>>> -		if ((gid_gte(filp->f_cred->fsgid, gid_min) &&
>>>>>>> -		     gid_lte(filp->f_cred->fsgid, gid_max)) ^
>>>>>>> -		    !(info->invert & XT_OWNER_GID))
>>>>>>> +		struct group_info *gi = filp->f_cred-
>>>>>>>> group_info;
>>>>>>> +
>>>>>>> +		if (gid_gte(filp->f_cred->fsgid, gid_min) &&
>>>>>>> +		    gid_lte(filp->f_cred->fsgid, gid_max))
>>>>>>> +			match = true;
>>>>>>> +
>>>>>>> +		if (!match && (info->match & XT_SUPPL_GROUPS)
>>>>>>> && gi) {
>>>>>>> +			for (i = 0; i < gi->ngroups; ++i) {
>>>>>>> +				kgid_t group = gi->gid[i];
>>>>>>> +
>>>>>>> +				if (gid_gte(group, gid_min) &&
>>>>>>> +				    gid_lte(group, gid_max)) {
>>>>>>> +					match = true;
>>>>>>> +					break;
>>>>>>> +				}
>>>>>>> +			}
>>>>>>> +		}
>>>>>>> +
>>>>>>> +		if (match ^ !(info->invert & XT_OWNER_GID))
>>>>>>>  			return false;
>>>>>>>  	}
>>>>>>>  
>>>>>>>
>>>>>>
>>>>>> How can this be safe on SMP ?
>>>>>>
>>>>>
>>>>> From what I see after the group_info rework some time ago this
>>>>> struct
>>>>> is never modified. It's replaced. Would
>>>>> get_group_info/put_group_info
>>>>> around the code be enough?
>>>>
>>>> What prevents the data to be freed right after you fetch filp-
>>>>> f_cred->group_info ?
>>>
>>> I think the get_group_info() I mentioned above would. group_info
>>> seems
>>> to always be freed by put_group_info().
>>
>> The data can be freed _before_ get_group_info() is attempted.
>>
>> get_group_info() would do a use-after-free
>>
>> You would need something like RCU protection over this stuff,
>> this is not really only a netfilter change.
>>
> 
> sk_socket keeps reference to f_cred. f_cred keeps reference to
> group_info. As long as f_cred is alive and it doesn't seem to be the
> issue in the owner_mt() function, group_info should be alive as well as
> far as I can see. Its refcount will go down only when f_cred is freed
> (put_cred_rcu()).
> 
> If there is something I'm missing please correct me.

The problem is that you canÂt clearly explain why the code is safe :/

Why would get_group_info() be needed then ?

You need to explain this in the changelog, so that future bug hunters do not have
to guess.

Note to netfilter maintainers : 

owner_mt() reads sk->sk_socket multiple times, this looks racy to me.

(sock_orphan() could be done in the middle from another cpu)


diff --git a/net/netfilter/xt_owner.c b/net/netfilter/xt_owner.c
index 46686fb73784bf71c79282e87e3f01f2c0411f5c..6adfb992bfe1765c57430b4bb98212786086d379 100644
--- a/net/netfilter/xt_owner.c
+++ b/net/netfilter/xt_owner.c
@@ -66,8 +66,10 @@ owner_mt(const struct sk_buff *skb, struct xt_action_param *par)
        const struct file *filp;
        struct sock *sk = skb_to_full_sk(skb);
        struct net *net = xt_net(par);
+       struct socket *sock;
 
-       if (!sk || !sk->sk_socket || !net_eq(net, sock_net(sk)))
+       sock = sk ? READ_ONCE(sk->sk_socket) : NULL;
+       if (!sock || !net_eq(net, sock_net(sk)))
                return (info->match ^ info->invert) == 0;
        else if (info->match & info->invert & XT_OWNER_SOCKET)
                /*
@@ -76,7 +78,7 @@ owner_mt(const struct sk_buff *skb, struct xt_action_param *par)
                 */
                return false;
 
-       filp = sk->sk_socket->file;
+       filp = sock->file;
        if (filp == NULL)
                return ((info->match ^ info->invert) &
                       (XT_OWNER_UID | XT_OWNER_GID)) == 0;