lkml.org 
[lkml]   [2021]   [Dec]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v4 10/16] ima: Implement hierarchical processing of file accesses
From

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 ?


\
 
 \ /
  Last update: 2021-12-08 17:51    [W:0.073 / U:0.520 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site