Re: [PATCH v4 10/16] ima: Implement hierarchical processing of file accesses

[Stefan Berger <stefanb@linux.ibm.com>]

On 12/8/21 11:50, Stefan Berger wrote:
>
> On 12/8/21 07:23, Christian Brauner wrote:
>> On Wed, Dec 08, 2021 at 01:09:54PM +0100, Christian Brauner wrote:
>>> On Tue, Dec 07, 2021 at 03:21:21PM -0500, Stefan Berger wrote:
>>>> Implement hierarchical processing of file accesses in IMA 
>>>> namespaces by
>>>> walking the list of IMA namespaces towards the init_ima_ns. This way
>>>> file accesses can be audited in an IMA namespace and also be evaluated
>>>> against the IMA policies of parent IMA namespaces.
>>>>
>>>> Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
>>>> ---
>>>>   security/integrity/ima/ima_main.c | 29 +++++++++++++++++++++++++----
>>>>   1 file changed, 25 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/security/integrity/ima/ima_main.c 
>>>> b/security/integrity/ima/ima_main.c
>>>> index 2121a831f38a..e9fa46eedd27 100644
>>>> --- a/security/integrity/ima/ima_main.c
>>>> +++ b/security/integrity/ima/ima_main.c
>>>> @@ -200,10 +200,10 @@ void ima_file_free(struct file *file)
>>>>       ima_check_last_writer(iint, inode, file);
>>>>   }
>>>>   -static int process_measurement(struct ima_namespace *ns,
>>>> -                   struct file *file, const struct cred *cred,
>>>> -                   u32 secid, char *buf, loff_t size, int mask,
>>>> -                   enum ima_hooks func)
>>>> +static int _process_measurement(struct ima_namespace *ns,
>>> Hm, it's much more common to use double underscores then single
>>> underscores to
>>>
>>> __process_measurement()
>>>
>>> reads a lot more natural to people perusing kernel code quite often.
>>>
>>>> +                struct file *file, const struct cred *cred,
>>>> +                u32 secid, char *buf, loff_t size, int mask,
>>>> +                enum ima_hooks func)
>>>>   {
>>>>       struct inode *inode = file_inode(file);
>>>>       struct integrity_iint_cache *iint = NULL;
>>>> @@ -405,6 +405,27 @@ static int process_measurement(struct 
>>>> ima_namespace *ns,
>>>>       return 0;
>>>>   }
>>>>   +static int process_measurement(struct ima_namespace *ns,
>>>> +                   struct file *file, const struct cred *cred,
>>>> +                   u32 secid, char *buf, loff_t size, int mask,
>>>> +                   enum ima_hooks func)
>>>> +{
>>>> +    int ret = 0;
>>>> +    struct user_namespace *user_ns;
>>>> +
>>>> +    do {
>>>> +        ret = _process_measurement(ns, file, cred, secid, buf, 
>>>> size, mask, func);
>>>> +        if (ret)
>>>> +            break;
>>>> +        user_ns = ns->user_ns->parent;
>>>> +        if (!user_ns)
>>>> +            break;
>>>> +        ns = user_ns->ima_ns;
>>>> +    } while (1);
>>> I'd rather write this as:
>>>
>>>     struct user_namespace *user_ns = ns->user_ns;
>>>
>>>     while (user_ns) {
>>>         ns = user_ns->ima_ns;
>>>
>>>             ret = __process_measurement(ns, file, cred, secid, buf, 
>>> size, mask, func);
>>>             if (ret)
>>>                 break;
>>>         user_ns = user_ns->parent;
>>>
>>>     }
>>>
>>> because the hierarchy is only an implicit property inherited by ima
>>> namespaces from the implementation of user namespaces. In other words,
>>> we're only indirectly walking a hierarchy of ima namespaces because
>>> we're walking a hierarchy of user namespaces. So the ima ns actually
>>> just gives us the entrypoint into the userns hierarchy which the double
>>> deref writing it with a while() makes obvious.
>> Which brings me to another point.
>>
>> Technically nothing seems to prevent an ima_ns to survive the
>> destruction of its associated userns in ima_ns->user_ns?
>>
>> One thread does get_ima_ns() and mucks around with it while another one
>> does put_user_ns().
>>
>> Assume it's the last reference to the userns which is now -
>> asynchronously - cleaned up from ->work. So at some point you're ending
>> with a dangling pointer in ima_ns->user_ns eventually causing a UAF.
>>
>> If I'm thinking correct than you need to fix this. I can think of two
>> ways right now where one of them I'm not sure how well that would work:
>> 1. ima_ns takes a reference count to userns at creation. Here you need
>>     to make very sure that you're not ending up with reference counting
>>     cycles where the two structs keep each other alive.
>
> Right. I am not sure what the trigger would be for ima_ns to release 
> that one reference.
>
>
>> 2. rcu trickery. That's the one I'm not sure how well that would work
>>     where you'd need rcu_read_lock()/rcu_read_unlock() with a
>>     get_user_ns() in the middle whenever you're trying to get a ref to
>>     the userns from an ima_ns and handle the case where the userns is
>>     gone.
>>
>> Or maybe I'me missing something in the patch series that makes this all
>> a non-issue.
>
> I suppose one can always call current_user_ns() to get a pointer to 
> the current user namespace that the process is accessing the file in 
> that IMA now reacts to. With the hierarchical processing we are 
> walking backwards towards init_user_ns. The problem should only exist 
> if something else frees the current user namespace (or its parents) so 
> that the hierarchy collapses. Assuming we are always in a process 
> context then 'current' should protect us, no ?
>
All existing callers to process_measurements call it at least once with 
current_cred().

The only problem that I see where we are accessing the IMA namespace 
outside a process context is in 4/16 'ima: Move delayed work queue and 
variables into ima_namespace' where a delayed work queue is used. I 
fixed this now by getting an additional reference to the user namesapce  
before scheduling the delayed work and release it when it ran or when it 
is canceled (cancel_delayed_work_sync()) but it didn't run.