lkml.org 
[lkml]   [2019]   [May]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2] netfilter: xt_owner: Add supplementary groups option
From
Date


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@samsung.com>
>>>>> ---
>>>>> 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.


\
 
 \ /
  Last update: 2019-05-08 18:55    [W:0.076 / U:2.268 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site