Re: [PATCH v2] ksm: React on changing "sleep_millisecs" parameter faster

[Kirill Tkhai <ktkhai@virtuozzo.com>]

On 11.12.2018 15:34, Cyrill Gorcunov wrote:
> On Tue, Dec 11, 2018 at 03:22:42PM +0300, Kirill Tkhai wrote:
>> On 11.12.2018 14:13, Cyrill Gorcunov wrote:
>>> On Tue, Dec 11, 2018 at 01:26:59PM +0300, Kirill Tkhai wrote:
>>>> ksm thread unconditionally sleeps in ksm_scan_thread()
>>>> after each iteration:
>>>>
>>>> 	schedule_timeout_interruptible(
>>>> 		msecs_to_jiffies(ksm_thread_sleep_millisecs))
>>>>
>>>> The timeout is configured in /sys/kernel/mm/ksm/sleep_millisecs.
>>>>
>>>> In case of user writes a big value by a mistake, and the thread
>>>> enters into schedule_timeout_interruptible(), it's not possible
>>>> to cancel the sleep by writing a new smaler value; the thread
>>>> is just sleeping till timeout expires.
>>>>
>>>> The patch fixes the problem by waking the thread each time
>>>> after the value is updated.
>>>>
>>>> This also may be useful for debug purposes; and also for userspace
>>>> daemons, which change sleep_millisecs value in dependence of
>>>> system load.
>>>>
>>>> Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
>>>>
>>>> v2: Use wait_event_interruptible_timeout() instead of unconditional
>>>>     schedule_timeout().
>>> ...
>>>> @@ -2844,7 +2849,10 @@ static ssize_t sleep_millisecs_store(struct kobject *kobj,
>>>>  	if (err || msecs > UINT_MAX)
>>>>  		return -EINVAL;
>>>>  
>>>> +	mutex_lock(&ksm_thread_mutex);
>>>>  	ksm_thread_sleep_millisecs = msecs;
>>>> +	mutex_unlock(&ksm_thread_mutex);
>>>> +	wake_up_interruptible(&ksm_iter_wait);
>>>
>>> Btw, just thought -- if we start using this mutex here don't we
>>> open a window for force attack on the thread self execution,
>>> iow if there gonna be a million of writers do we have a guarantee
>>> thread ksm_scan_thread will grab the mutex earlier than writers
>>> (or somewhere inbetween)?
>>
>> This file is permitted for global root only. I don't think there is
>> a problem.
>>
>> If someone wants to make ksm helpless, a person may just write a big
>> "sleep_millisecs" value. KSM thread won't be executed almost all the time
>> in this case.
> 
> True. Still I think if we can leave without taking a lock it a rule of thumb.
> Something like
> 
> 	if (msecs != ksm_thread_sleep_millisecs)
> 		wake_up_interruptable(&ksm_iter_wait);
> 
> Thoughts?

Ok, good idea.