Re: [PATCH] sched/deadline: fix switching to -deadline

[Juri Lelli]

On 24/04/17 12:36, Luca Abeni wrote:
> On Mon, 24 Apr 2017 11:16:24 +0100
> Juri Lelli <juri.lelli@xxxxxxx> wrote:
> 
> > On 21/04/17 21:08, Luca Abeni wrote:

[...]

> > > 
> > > Well, double thinking about it, this is an interesting problem...
> > > What do we want to do with do_set_cpus_allowed()? (I mean: what is
> > > the expected behaviour?)
> > > 
> > > With this patch, if a task is moved to a different runqueue when its
> > > deadline is in the past (because we are doing gEDF, or because of
> > > timer granularity issues) its scheduling deadline is reinitialized
> > > to current time + relative deadline... I think this makes perfect
> > > sense, doesn't it?
> > >   
> > 
> > Mmm, I don't think we will (with this patch) actually reinitialize the
> > deadline when a "normal" gEDF migration happen (push/pull), as
> > (de)activate_task() have no flag set. Which brings the question,
> > should we actually take care of this corner case (as what you say
> > makes sense to me too)?
> 
> I might be misunderstanding the problem, here... Are you talking about
> do_set_cpus_allowed()? Or about push/pull migrations happening because
> of the gEDF algorithm?
> 

My concern was about do_set_cpus_allowed(), but then you mentioned
"because we are doing gEDF" and that made me think of what happens when
we do push/pull. :)

> If you are referring to do_set_cpus_allowed, this is my understanding:
> 1) If do_set_cpus_allowed() is called on a queued task, then
>    dequeue_task() with DEQUEUE_SAVE is called, followed by
>    enqueue_task() with ENQUEUE_RESTORE... So, if the deadline is in the
>    past it is correctly reinitialized
> 2) If do_set_cpus_allowed() is called on a non-queued task, this means
>    the task is blocked, no? So, when it will wake up enqueue_dl_entity()
>    will invoke update_dl_entity() that will check if the deadline is in
>    the past.
> 

OK. I think it makes sense, and your patch should cure the problem.
Maybe add a comment to note this down.

> If you are referring to push/pull migrations due to gEDF, then
> enqueue_dl_entity() will be invoked with "flags" = 0, so the deadline
> will not be changed (and this is correct: we do not want to
> initialize / change tasks' deadlines during gEDF migrations).
> 

Ok, but I was wondering about the (admittedly) corner case in which we
migrate (via push/pull) a task on a rq, the rq_clock of which is after
the task's deadline (because clocks on src_rq and dst_rq are not in
sync). Anyway, maybe it's so corner case that we don't really want to
deal with it right now? I guess bigger things to fix first. :)

> In my previous email, with "a task is moved to a different runqueue" I
> wanted to say that the taks is forced to moved to a different runqueue
> because its affinity is changed; I did not want to talk about "regular
> migrations" due to the push/pull (gEDF) mechanism.
> 

Thanks for claryfing. As said, I just got distracted by what you
mentioned as examples between parenthesis.