Re: [PATCH 01/16] freezer: fix current->state restoration race in refrigerator()

[Tejun Heo]

Hello, Oleg.

On Fri, Aug 19, 2011 at 5:52 PM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> Indeed, we can miss kthread->should_stop, and the patch fixes this
> case.
>
> But please look at, say, kauditd_thread(), it does
>
>                DECLARE_WAITQUEUE(wait, current);
>                set_current_state(TASK_INTERRUPTIBLE);
>                add_wait_queue(&kauditd_wait, &wait);
>
>                if (!skb_queue_len(&audit_skb_queue)) {
>                        try_to_freeze();
>                        schedule();
>                }
>
> Now suppose that wake_up_interruptible(&kauditd_wait) happens, and
> after that refrigerator() restores TASK_INTERRUPTIBLE.
>
> Any reason refrigerator() should try to restore? Shouldn't we simply
> change the rules? Yes, probably we will have to fix some users.
>
> But it seems to me it is simply not possible to make this ->state
> restoration correct.

Yes, it's broken, but it's not the only thing broken. There are race
conditions everywhere. Even without freezer, a lot of kthread users
use kthread_should_stop() incorrectly - they test them without setting
sleep state and freezable kthreads are worse as the two conditions can
race each other. I think kthread is just too difficult to use. We can
try to salvage the situation and make kthread more friendly but I
think the better solution would be just converting everyone to use
workqueue or the kthread_wq wrapper. So, yeah, the whole thing is
vastly broken. Let's fix things in baby steps.

Thanks.

-- 
tejun
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/