[PATCH v18 05/23] net: Prepare UDS for security module stacking

[Casey Schaufler]

On 7/9/2020 9:28 AM, John Johansen wrote:
> On 7/9/20 9:11 AM, Stephen Smalley wrote:
>> On Wed, Jul 8, 2020 at 8:23 PM Casey Schaufler <casey at schaufler-ca.com> wrote:
>>> Change the data used in UDS SO_PEERSEC processing from a
>>> secid to a more general struct lsmblob. Update the
>>> security_socket_getpeersec_dgram() interface to use the
>>> lsmblob. There is a small amount of scaffolding code
>>> that will come out when the security_secid_to_secctx()
>>> code is brought in line with the lsmblob.
>>>
>>> The secid field of the unix_skb_parms structure has been
>>> replaced with a pointer to an lsmblob structure, and the
>>> lsmblob is allocated as needed. This is similar to how the
>>> list of passed files is managed. While an lsmblob structure
>>> will fit in the available space today, there is no guarantee
>>> that the addition of other data to the unix_skb_parms or
>>> support for additional security modules wouldn't exceed what
>>> is available.
>>>
>>> Reviewed-by: Kees Cook <keescook at chromium.org>
>>> Signed-off-by: Casey Schaufler <casey at schaufler-ca.com>
>>> Cc: netdev at vger.kernel.org
>>> ---
>>> diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
>>> index 3385a7a0b231..d246aefcf4da 100644
>>> --- a/net/unix/af_unix.c
>>> +++ b/net/unix/af_unix.c
>>> @@ -138,17 +138,23 @@ static struct hlist_head *unix_sockets_unbound(void *addr)
>>>  #ifdef CONFIG_SECURITY_NETWORK
>>>  static void unix_get_secdata(struct scm_cookie *scm, struct sk_buff *skb)
>>>  {
>>> -       UNIXCB(skb).secid = scm->secid;
>>> +       UNIXCB(skb).lsmdata = kmemdup(&scm->lsmblob, sizeof(scm->lsmblob),
>>> +                                     GFP_KERNEL);
>>>  }
>>>
>>>  static inline void unix_set_secdata(struct scm_cookie *scm, struct sk_buff *skb)
>>>  {
>>> -       scm->secid = UNIXCB(skb).secid;
>>> +       if (likely(UNIXCB(skb).lsmdata))
>>> +               scm->lsmblob = *(UNIXCB(skb).lsmdata);
>>> +       else
>>> +               lsmblob_init(&scm->lsmblob, 0);
>>>  }
>>>
>>>  static inline bool unix_secdata_eq(struct scm_cookie *scm, struct sk_buff *skb)
>>>  {
>>> -       return (scm->secid == UNIXCB(skb).secid);
>>> +       if (likely(UNIXCB(skb).lsmdata))
>>> +               return lsmblob_equal(&scm->lsmblob, UNIXCB(skb).lsmdata);
>>> +       return false;
>>>  }
>> I don't think that this provides sensible behavior to userspace.  On a
>> transient memory allocation failure, instead of returning an error to
>> the sender and letting them handle it, this will just proceed with
>> sending the message without its associated security information, and
>> potentially split messages on arbitrary boundaries because it cannot
>> tell whether the sender had the same security information.  I think
>> you instead need to change unix_get_secdata() to return an error on
>> allocation failure and propagate that up to the sender.

Can't say that I think that would go over especially well.
You're right about that being a better, or at least more correct,
change.

>>   Not a fan of
>> this change in general both due to extra overhead on this code path
>> and potential for breakage on allocation failures.  I know it was
>> motivated by paul's observation that we won't be able to fit many more
>> secids into the cb but not sure we have to go there prematurely,

Paul wasn't completely against the original approach. His objection
was that using a struct lsmblob, which was already close to the maximum
size it can be and that can grow over time, might be a hard sell.

>> especially absent its usage by upstream AA (no unix_stream_connect
>> hook implementation upstream).  Also not sure how the whole bpf local
> I'm not sure how premature it is, I am running late for 5.9 but would
> like to land apparmor unix mediation in 5.10

Which means that scaffolding around the UNIXCB.secid wouldn't
suffice for very long.

>
>> storage approach to supporting security modules (or at least bpf lsm)
>> might reduce need for expanding these structures?

I think the allocation failure case would still be an issue,
and it could be much more complicated to deal with using the
local storage model. The fundamental problem comes back to fitting
more that 32 bits of information into 32 bits without having to
perform an operation that might fail.

At this point I'm inclined to revert to the original implementation
and see if it doesn't turn out to be acceptable after all. I remain
open to better ideas.