Re: sleep in selinux_audit_rule_init

[Stephen Smalley <sds@tycho.nsa.gov>]

On 5/30/19 9:27 AM, Janne Karhunen wrote:
> On Thu, May 30, 2019 at 3:08 PM Stephen Smalley <sds@tycho.nsa.gov> wrote:
> 
>>> @@ -269,11 +269,23 @@ static void ima_lsm_update_rules(void)
>>>                                                              Audit_equal,
>>>                                                              entry->lsm[i].args_p,
>>>                                                              &entry->lsm[i].rule);
>>> -                       BUG_ON(!entry->lsm[i].rule);
>>> +                       if (result == -EINVAL)
>>> +                               pr_warn("ima: rule for LSM \'%d\' is invalid\n",
>>> +                                       entry->lsm[i].type);
>>
>> I could be wrong, but I think there is still a problem here in that you
>> are modifying entry->lsm[i].rule in-place, but it is protected under RCU
>> and therefore needs to be duplicated and then modified?  Also you are
>> leaking the old rule?
> 
> Right. Bit too fast tapping the keyboard without thinking, will fix
> and post in the proper form. But I guess the original point was to
> verify if that 'notifier_block' is indeed the right way to get the
> update notification?

Yes.

> 
>>   Both of those issues also exist prior to your
>> patch but you aren't fixing them here. And lastly, it looks like lsm
>> notifiers are atomic notifiers (not clear to me why) so you can't block
>> in the callback, thereby requiring scheduling the work as is done in
>> infiniband.
> 
> Great catch, thank you. That's an easy fix if no-one objects pushing
> these through the system-wq for example.

I think you can switch the lsm notifier over to using blocking notifiers 
instead; there seems to be no valid reason for making it atomic.