Re: [PATCH 1/2] lkdtm: use atomic_t to replace count_lock

[Cong Wang <xiyou.wangcong@gmail.com>]

On 02/01/2012 11:27 PM, Arnd Bergmann wrote:
> On Wednesday 01 February 2012, Cong Wang wrote:
>>   static void lkdtm_handler(void)
>>   {
>> -       unsigned long flags;
>> -
>> -       spin_lock_irqsave(&count_lock, flags);
>> -       count--;
>>          printk(KERN_INFO "lkdtm: Crash point %s of type %s hit, trigger in %d rounds\n",
>> -                       cp_name_to_str(cpoint), cp_type_to_str(cptype), count);
>> +                       cp_name_to_str(cpoint), cp_type_to_str(cptype), atomic_dec_return(&count));
>>
>> -       if (count == 0) {
>> +       if (!atomic_cmpxchg(&count, 0, cpoint_count))
>>                  lkdtm_do_action(cptype);
>> -               count = cpoint_count;
>> -       }
>> -       spin_unlock_irqrestore(&count_lock, flags);
>>   }
>
> This use is not atomic, you could have two threads doing atomic_dec_return
> at the same time, and after that the value will be -1 so the atomic_cmpxchg
> does not trigger.


Yeah, simply combining two atomic operations is not atomic. :-/

>
> In order to have an atomic here, you have to use a loop around
> atomic_cmpxchg, like
>
>
> 	int old, new;
> 	old = atomic_read(&count);
> 	do {
> 		new = old ? old - 1 : cpoint_count;
> 		old = cmpxchg(&count, old, new);
> 	} while (old != new);
>
> I suppose you could also just keep the spinlock and move lkdtm_do_action()
> outside of it?

If we still need spinlock, I think we don't need to bother atomic_t at all.