[PATCH] audit: Only use the syscall slowpath when syscall audit rules exist

[Andy Lutomirski]

On Mon, Feb 3, 2014 at 10:11 AM, Oleg Nesterov <oleg at redhat.com> wrote:
> On 02/03, Andy Lutomirski wrote:
>>
>> @@ -911,6 +918,47 @@ static inline struct audit_context *audit_alloc_context(enum audit_state state)
>>       return context;
>>  }
>>
>> +void audit_inc_n_rules()
>> +{
>> +     struct task_struct *p, *g;
>> +
>> +     write_lock(&n_rules_lock);
>> +
>> +     if (audit_n_rules++ != 0)
>> +             goto out;  /* The overall state isn't changing. */
>> +
>> +     read_lock(&tasklist_lock);
>> +     do_each_thread(g, p) {
>> +             if (p->audit_context)
>> +                     set_tsk_thread_flag(p, TIF_SYSCALL_AUDIT);
>> +     } while_each_thread(g, p);
>> +     read_unlock(&tasklist_lock);
>
> Cosmetic, but I'd suggest to use for_each_process_thread() instead
> of do_each_thread/while_each_thread.

I didn't notice that option :)

>
> And I am not sure why n_rules_lock is rwlock_t... OK, to make
> audit_alloc() more scalable, I guess. Please see below.
>

Yes.  I should probably also use the irqsave variant.

>> @@ -942,8 +995,14 @@ int audit_alloc(struct task_struct *tsk)
>>       }
>>       context->filterkey = key;
>>
>> +     read_lock(&n_rules_lock);
>>       tsk->audit_context  = context;
>> -     set_tsk_thread_flag(tsk, TIF_SYSCALL_AUDIT);
>> +     if (audit_n_rules)
>> +             set_tsk_thread_flag(tsk, TIF_SYSCALL_AUDIT);
>> +     else
>> +             clear_tsk_thread_flag(tsk, TIF_SYSCALL_AUDIT);
>> +     read_unlock(&n_rules_lock);
>
> Perhaps this is fine, but n_rules_lock can't prevent the race with
> audit_inc/dec_n_rules(). The problem is, this is called before the
> new task is visible to for_each_process_thread().

Hmm.  I missed that.

>
> If we want to fix this race, we need something like audit_sync_flags()
> called after copy_process() drops tasklist, or from tasklist_lock
> protected section (in this case it doesn't need n_rules_lock).
>
> Or perhaps audit_alloc() should not try to clear TIF_SYSCALL_AUDIT at all.
> In both cases n_rules_lock can be spinlock_t.

Here are two options:

1. Sync the flags inside tasklist_lock, as I think you're suggesting.
This seems simple.

2. Make this whole thing lazy -- always set TIF_SYSCALL_AUDIT for new
tasks, but clear it on the first syscall.

I'm leaning toward number 1.

Thanks for the instant review!

--Andy