Re: [PATCH] sched/deadline: Remove redundant code replenishing runtime

[Juri Lelli]

On 14/02/17 11:06, Byungchul Park wrote:
> On Tue, Feb 14, 2017 at 08:42:43AM +0900, Byungchul Park wrote:
> > On Mon, Feb 13, 2017 at 03:24:55PM +0000, Juri Lelli wrote:
> > > > > > I think we actually want to replenish and set the next deadline at this
> > > > > > point of time, not the one that we get when the task will eventually wake up.
> > > > > 
> > > > > Hello juri,
> > > > > 
> > > > > But I wonder if it's meaningful to set a next deadline for a 'sleeping
> > > > > task', which, rather, could be worse because its bandwidth might be
> > > > > distorted at the time it's woken up.
> > > > > 
> > > 
> > > What you mean by 'distorted'. AFAIU, we just want to replenish when
> > > needed. The instant of time when the task will eventually wake up it is
> > > something we cannot rely upon, and could introduce errors.
> > > 
> > > IIUC, your situation looks like the below
> > 
> > Exactly.
> > 
> > > 
> > >    oooo|-------------------vxxx^ooo
> > >        |                   |   |
> > >        |                   |   |
> > >     sleep/throttle         |   |
> > >                      r. timer  |
> > >    		           wakeup
> > 
> > Sorry for bothering you..
> > 
> > > The task gets throttled while going to sleep, when the replenishment
> > > timer fires you are proposing we do nothing and we actually replenishing
> > > using the wakeup rq_clock() as reference. My worry is that, by doing so,
> > > we make the task potentially loose some of its bandwidth, as we will
> > > have lost some time (the 3 x-es in the diagram above) when calculating
> > > its next dynamic deadline.
> > 
> > I meant, when we decide whether it's overflowed in dl_entiry_overflow(),
> > 'right' might be smaller than 'left' because 't' is the time the 3 x-es
> > already passed.
> > 
> > Of course, here I assumed that runtime ~= 0 and deadline ~= rq_clock
> > when it was throttled, if scheduler works nicely.
> > 
> > > > > IMHO, it's neat to set its deadline and runtime when being woken up, in
> > > > > the case already passed its deadline. Am I wrong?
> > > > 
> > > > And I found that dl_entity_overflow() returns true and replenishes the
> > > > task unconditionally in update_dl_entity() again when the task is woken
> > > > up, because 'runtime / (deadline - t) > dl_runtime / dl_period' is true.
> > > > 
> > > 
> > > Why 'unconditionally'? It will postpone and replenish if the task is
> > 
> > Not exactly 'unconditially' if my assumption is broken. Sorry for
> > choosing a word that is not careful.
> > 
> > > going to overflow, if not, it will keep its runtime and deadline we set
> > 
> > I meant the task will be almost always considered 'overflow', as I
> > explained above. So it will be overwritten again when waking up the task
> > than keep what we set in timer callback.
> 
> I don't want to argue strongly, keeping current code unchanged is ok.
> I just wanted to say it will be replenished twice in most cases if:
> 
>    1. The task was throttled and passed its deadline while sleeping.
> 
> Of course, it also depends on how much negative runtime it had when
> throttled. Sorry for bothering you and thanks for explaining it.

No bothering at all! Thanks for raising a potential problem, but I guess
we need to be correct 100% of the times, without trying to optimize for
the most common case.

Best,

- Juri