Re: [RFC] sched/deadline: Prevent rt_time growth to infinity

[Kirill Tkhai <tkhai@yandex.ru>]

21.02.2014, 16:44, "Juri Lelli" <juri.lelli@gmail.com>:
> On Fri, 21 Feb 2014 16:09:25 +0400
> Kirill Tkhai <tkhai@yandex.ru> wrote:
>
>>  21.02.2014, 15:39, "Kirill Tkhai" <tkhai@yandex.ru>:
>>>  21.02.2014, 14:37, "Peter Zijlstra" <peterz@infradead.org>:
>>>>   On Thu, Feb 20, 2014 at 02:16:00AM +0400, Kirill Tkhai wrote:
>>>>>    Since deadline tasks share rt bandwidth, we must care about
>>>>>    bandwidth timer set. Otherwise rt_time may grow up to infinity
>>>>>    in update_curr_dl(), if there are no other available RT tasks
>>>>>    on top level bandwidth.
>>>>>
>>>>>    I'm going to decide the problem the way below. Almost untested
>>>>>    because of I skipped almost all of recent patches which haveto be applied from lkml.
>>>>>
>>>>>    Please say, if I skipped anything in idea. Maybe better put
>>>>>    start_top_rt_bandwidth() into set_curr_task_dl()?
>>>>   How about we only increment rt_time when there's an RT bandwidth timer
>>>>   active?
>>>  This case RT and DL may eat all the time:
>>>
>>>  --------------  time ------------------>
>>>  |RT's working |DL's working|
>>>  ----------------------------
>>>  |rt_runtime   |            |
>>>  ----------------------------
>>>  |        rt_period         |
>>>
>>>  Or at least more, than it's allowed.
>>>
>>>  It looks like, if we want to limit time of high priority classes
>>>  execution, we have to set the timer anyway.
>>  Oh, above is confusing. Sorry.
>>
>>  I mean one RT task and
>>
>>  -----------------  time -------------------------->
>>  |DL's working    |RT's working| |DL's working    |
>>  --------------------------------------------------
>>  |                |rt_runtime  | |                |
>>  --------------------------------------------------
>>  |                |   rt_period  |                |
>>
>>  In this case FAIR receives less ratio, than (rt_period-rt_runtime)/rt_period.
>
> DL tasks won't be allowed to run in this situation, as their bw exceedes
> rt_runtime/rt_period.

Maybe, I don't uderstand. Where does DL control summary (DL+RT) runtime?

RT does not do this too. But it looks like, RT has to do this.

Kirill