Re: IB pkey policy problem found via the selinux-testsuite

On 9/4/19 6:32 PM, Paul Moore wrote:
> On Tue, Sep 3, 2019 at 10:45 AM Stephen Smalley <[email protected]> wrote:
>> On 9/3/19 10:30 AM, Daniel Jurgens wrote:
>>>
>>> On 8/27/2019 12:24 PM, Paul Moore wrote:
>>>> On Thu, Feb 28, 2019 at 4:58 PM Paul Moore <[email protected]> wrote:
>>>>> On Wed, Feb 13, 2019 at 4:35 PM Paul Moore <[email protected]> wrote:
>>>>>> Hello all,
>>>>>>
>>>>>> On a fully up-to-date Rawhide system you need the following line added
>>>>>> to the policy/test_ibpkey.te file to get a clean run of the
>>>>>> selinux-testsuite:
>>>>>>
>>>>>>     allow test_ibpkey_access_t self:capability { ipc_lock };
>>>>>>
>>>>>> The breakage doesn't appear to be due to a kernel change (previously
>>>>>> working kernels now fail), or a Fedora Rawhide policy change (nothing
>>>>>> relevant changed since the last clean run), but I did notice that my
>>>>>> libibverbs package was updated just prior to the breakage.  I haven't
>>>>>> had the time to dig into the library code, but I expect that to be the
>>>>>> source of the problem.
>>>>> Just to be clear, I don't believe this breakage is limited to the test
>>>>> suite, I expect any users of the SELinux IB hooks will run into this
>>>>> problem.  I believe we need to update the upstream and distro
>>>>> policies.
>>>> A ping to bring this issue back to the top of the mailing list.
>>>
>>> Hi Paul, I looked in the libraries and don't see explicit use of mlock. Maybe there was a change to use that access control for get_user_pages? That doesn't really jive with previously working kernels no longer working though.
>>
>> It would be useful to see the audit messages for that ipc_lock denial,
>> including the SYSCALL record.
>>
>> There are a number of kernel operations that can trigger ipc_lock checks,
>> https://elixir.bootlin.com/linux/latest/ident/CAP_IPC_LOCK
>>
>> Several of those are infiniband-specific.
> 
> Hi all,
> 
> Sorry for the delay in responding.  Here is what I see when running
> the infiniband_pkey test on a current Fedora Rawhide system (run
> individually to capture the output):
> 
> 1..4
> # Running under perl version 5.030000 for linux
> # Current time local: Wed Sep  4 18:24:39 2019
> # Current time GMT:   Wed Sep  4 22:24:39 2019
> # Using Test.pm version 1.31
> ok 1
> Unable to create cq.
> not ok 2
> # Test 2 got: "256" (./test at line 50)
> #   Expected: "0"
> #  ./test line 50 is:     ok( $result, 0 );
> Unable to create cq.
> not ok 3
> # Test 3 got: "1" (./test at line 71)
> #   Expected: "13"
> #  ./test line 71 is: if (@unlabeled_pkeys) {
> ok 4
> 
> ... and I see the following AVCs:
> 
> type=AVC msg=audit(1567635879.312:15018): avc:  denied  { ipc_lock } for
>    pid=15726 comm="create_modify_q" capability=14
>    scontext=unconfined_u:unconfined_r:test_ibpkey_access_t:s0-s0:c0.c1023
>    tcontext=unconfined_u:unconfined_r:test_ibpkey_access_t:s0-s0:c0.c1023
>    tclass=capability permissive=0
> type=AVC msg=audit(1567635884.022:15020): avc:  denied  { ipc_lock } for
>    pid=15732 comm="create_modify_q" capability=14
>    scontext=unconfined_u:unconfined_r:test_ibpkey_access_t:s0-s0:c0.c1023
>    tcontext=unconfined_u:unconfined_r:test_ibpkey_access_t:s0-s0:c0.c1023
>    tclass=capability permissive=0
> 
> If I apply the workaround patch I mentioned earlier, the test succeeds
> and the only AVC is a infiniband_pkey:access denial (which is expected
> given the test).

I don't suppose you have SYSCALL (and possibly other auxiliary) audit 
records from those denials?
Trying to figure out what system call triggered the CAP_IPC_LOCK check.