Re: [PATCH v4 6/7] sched/deadline: Deferrable dl server

[Daniel Bristot de Oliveira]

On 9/6/23 22:08, Peter Zijlstra wrote:
> On Wed, Sep 06, 2023 at 10:04:06PM +0200, Peter Zijlstra wrote:
>> On Wed, Sep 06, 2023 at 04:58:11PM +0200, Daniel Bristot de Oliveira wrote:
>>
>>>> So one thing we could do is have update_curr_fair() decrement
>>>> fair_server's runtime and yield the period then it hits 0 (and capping
>>>> it at 0, not allowing it to go negative or so).
>>>>
>>>> That way you only force the situation when FAIR hasn't had it's allotted
>>>> time this perio, and only for as much as to make up for the time it
>>>> lacks.
>>>
>>> We can also decrease the runtime to a negative number while in
>>> defer/throttle state, and let the while in replenish_dl_entity() to
>>> replenish with the += runtime;
>>
>> Yes, but my point was that fair_server gives a lower bound of runtime
>> per period, more -- if available -- is fine.
>>
>> If we allow negative runtime, you'll affect future periods, and that is
>> not desired in this case.
>>
>> Or am I still confused?
> 
> That is, let update_curr_fair() decrement fair_server runtime
> *unconditionally* -- even if the task was not selected through the
> server.

Ah, I see! but then we would have to also consider the period, and control a
period... without SCHED_DEADLINE watching us...

I was considering only the "waiting for the 0-lag time to start running (after
being armed)"

If there is no need for the server to be armed... do nothing :-) If it is armed,
reduce the amount of time the fair server could get.

> Specifically, if the fair task is selected normally due to lack of
> deadline tasks, that runtime *still* counts towards the fair-server and
> have the server yield the period when zero.
> 
> This means that fair_server is only effective IFF 'normal' execution
> doesn't match the fair_server.runtime execution.

I see! I *think* it will cause more overhead than doing nothing unless there
is something that can cause starvation. But I need to think more.

-- Daniel