Re: [PATCH 6/8] sched: task_sched_runtime introduce micro optimization

[KOSAKI Motohiro <kosaki.motohiro@gmail.com>]

(6/18/13 10:18 AM), Frederic Weisbecker wrote:
> On Tue, Jun 18, 2013 at 11:17:41AM -0400, KOSAKI Motohiro wrote:
>>>> +#ifdef CONFIG_64BIT
>>>> +     /*
>>>> +      * 64-bit doesn't need locks to atomically read a 64bit value. So we
>>>> +      * have two optimization chances, 1) when caller doesn't need
>>>> +      * delta_exec and 2) when the task's delta_exec is 0. The former is
>>>> +      * obvious. The latter is complicated. reading ->on_cpu is racy, but
>>>> +      * this is ok. If we race with it leaving cpu, we'll take a lock. So
>>>> +      * we're correct. If we race with it entering cpu, unaccounted time
>>>> +      * is 0. This is indistinguishable from the read occurring a few
>>>> +      * cycles earlier.
>>>> +      */
>>>> +     if (!add_delta || !p->on_cpu)
>>>> +             return p->se.sum_exec_runtime;
>>>
>>> I'm not sure this is correct from an smp ordering POV. p->on_cpu may appear
>>> to be 0 whereas the task is actually running for a while and p->se.sum_exec_runtime
>>> can then be past the actual value on the remote CPU.
>>
>> Quate form Paul's last e-mail
>>
>>> Stronger:
>>>
>>> +#ifdef CONFIG_64BIT
>>> +       if (!p->on_cpu)
>>> +               return p->se.sum_exec_runtime;
>>> +#endif
>>>
>>> [ Or !p->on_cpu || !add_delta ].
>>>
>>> We can take the racy read versus p->on_cpu since:
>>>   If we race with it leaving cpu: we take lock, we're correct
>>>   If we race with it entering cpu: unaccounted time ---> 0, this is
>>> indistinguishable from the read occurring a few cycles earlier.
>
> Yeah, my worry was more about both p->on_cpu and p->se.sum_exec_runtime being
> stale for too long. How much time can happen in the worst case before CPU X sees
> the updates done by a CPU Y under rq(Y)->lock considering that CPU X doesn't take rq(Y)
> to read that update? I guess it depends on the hardware, locking and ordering
> that happened before.

Right. The worst case depend on memory access cost on remote node.
If memory access cost is m, The worst scenario is:

t
-------------
0: CPU X start to fetch p->on_cpu. (now it's 0)
0: CPU Y changes p->on_cpu to 1.
m: CPU X sees p->on_cpu is 0.

Then CPU X uses p->se.sum_exec_runtime even though p has delta m. And, in worst case,
all cpus make the same scenario. So, inaccuracy should be m x nr_online_cpus.
In this case, we can ignore cpu cache because we don't hit anyway in worst case.

However, I think it's ok. m is us order and our run_posx_cpu_timer() only has 1/HZ accuracy.
Moreover,taking lock needs more time than m because it need lock prefixed op. Then, It's free lunch.

I have no seen any ordering issue in this code because each CPU have independent time and no dependency. Please tell me if I'm missing something.

> Bah it probably doesn't matter in practice.

I'm ok to drop this "!p->on_cpu" optimization if you really dislike it.
It is just a micro optimization and the benefit is not so much.