From: Luca Abeni <luca.abeni@unitn.it>
To: Juri Lelli <juri.lelli@arm.com>
Cc: linux-kernel@vger.kernel.org,
Peter Zijlstra <peterz@infradead.org>,
Ingo Molnar <mingo@redhat.com>,
Claudio Scordino <claudio@evidence.eu.com>,
Steven Rostedt <rostedt@goodmis.org>
Subject: Re: [RFC v3 1/6] Track the active utilisation
Date: Tue, 8 Nov 2016 20:09:56 +0100 [thread overview]
Message-ID: <20161108200956.35b3e7c7@utopia> (raw)
In-Reply-To: <20161108185309.GG16920@e106622-lin>
Hi again,
On Tue, 8 Nov 2016 18:53:09 +0000
Juri Lelli <juri.lelli@arm.com> wrote:
[...]
> > > Also, AFAIU, do_exit() works on current and the TASK_DEAD case is
> > > handled in finish_task_switch(), so I don't think we are taking
> > > care of the "task is dying" condition.
> > Ok, so I am missing something... The state is set to TASK_DEAD, and
> > then schedule() is called... So, __schedule() sees the dying task as
> > "prev" and invokes deactivate_task() with the DEQUEUE_SLEEP flag...
> > After that, finish_task_switch() calls task_dead_dl(). Is this
> > wrong? If not, why aren't we taking care of the "task is dying"
> > condition?
> >
>
> No, I think you are right. But, semantically this cleanup goes in
> task_dead_dl(), IMHO.
Just to be sure I understand correctly: you suggest to add a check for
"state == TASK_DEAD" (skipping the cleanup if the condition is true) in
dequeue_task_dl(), and to add a sub_running_bw() in task_dead_dl()...
Is this understanding correct?
> It's most probably moot if it complicates
> things, but it might be helpful to differentiate the case between a
> task that is actually going to sleep (and for which we want to
> activate the timer) and a task that is dying (and for which we want
> to release bw immediately).
I suspect the two cases should be handled in the same way :)
> So, it actually matters for next patch,
> not here. But, maybe we want to do things clean from start?
You mean, because patch 2/6 adds
+ if (hrtimer_active(&p->dl.inactive_timer)) {
+ raw_spin_lock_irq(&task_rq(p)->lock);
+ sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
+ raw_spin_unlock_irq(&task_rq(p)->lock);
+ }
in task_dead_dl()? I suspect this hunk is actually unneeded (worse, it
is wrong :). I am trying to remember why it is there, but I cannot find
any reason... In the next days, I'll run some tests to check if that
hunk is actually needed. If yes, then I'll modify patch 1/6 as you
suggest; if it is not needed, I'll remove it from patch 2/6 and I'll
not do this change to patch 1/6... Is this ok?
Thanks,
Luca
>
> >
> > > Peter, does what I'm saying make any sense? :)
> > >
> > > I still have to set up things here to test these patches (sorry,
> > > I was travelling), but could you try to create some tasks and
> > > that kill them from another shell to see if the accounting
> > > deviates or not? Or did you already do this test?
> > I think this is one of the tests I tried...
> > I have to check if I changed this code after the test (but I do not
> > think I did). Anyway, tomorrow I'll write a script for automating
> > this test, and I'll leave it running for some hours.
> >
>
> OK, thanks. As said I think that you actually handle the case already,
> but I'll try to setup testing as well soon.
>
> Thanks,
>
> - Juri
next prev parent reply other threads:[~2016-11-08 19:10 UTC|newest]
Thread overview: 45+ messages / expand[flat|nested] mbox.gz Atom feed top
2016-10-24 14:06 [RFC v3 0/6] CPU reclaiming for SCHED_DEADLINE Luca Abeni
2016-10-24 14:06 ` [RFC v3 1/6] Track the active utilisation Luca Abeni
2016-10-25 9:09 ` Daniel Bristot de Oliveira
2016-10-25 9:29 ` luca abeni
2016-10-25 13:58 ` Steven Rostedt
2016-10-25 18:04 ` Luca Abeni
2016-11-18 14:23 ` Peter Zijlstra
2016-11-18 15:10 ` luca abeni
2016-11-18 15:28 ` Peter Zijlstra
2016-11-18 16:42 ` Steven Rostedt
2016-12-05 22:30 ` luca abeni
2016-12-06 8:35 ` Peter Zijlstra
2016-12-06 8:57 ` luca abeni
2016-12-06 13:47 ` luca abeni
2016-11-01 16:45 ` Juri Lelli
2016-11-01 21:10 ` luca abeni
2016-11-08 17:56 ` Juri Lelli
2016-11-08 18:17 ` Luca Abeni
2016-11-08 18:53 ` Juri Lelli
2016-11-08 19:09 ` Luca Abeni [this message]
2016-11-08 20:02 ` Juri Lelli
2016-11-09 15:25 ` luca abeni
2016-11-09 16:29 ` luca abeni
2016-11-18 14:55 ` Peter Zijlstra
2016-11-18 13:55 ` Peter Zijlstra
2016-11-18 15:06 ` luca abeni
2016-10-24 14:06 ` [RFC v3 2/6] Improve the tracking of " Luca Abeni
2016-11-01 16:46 ` Juri Lelli
2016-11-01 21:46 ` luca abeni
2016-11-02 2:35 ` luca abeni
2016-11-10 10:04 ` Juri Lelli
2016-11-10 11:56 ` Juri Lelli
2016-11-10 12:15 ` luca abeni
2016-11-10 12:34 ` Juri Lelli
2016-11-10 12:45 ` luca abeni
2016-11-02 2:41 ` luca abeni
2016-11-18 15:36 ` Peter Zijlstra
2016-11-18 15:56 ` luca abeni
2016-11-18 15:47 ` Peter Zijlstra
2016-11-18 16:06 ` luca abeni
2016-11-18 18:49 ` Peter Zijlstra
2016-10-24 14:06 ` [RFC v3 3/6] Fix the update of the total -deadline utilization Luca Abeni
2016-10-24 14:06 ` [RFC v3 4/6] GRUB accounting Luca Abeni
2016-10-24 14:06 ` [RFC v3 5/6] Do not reclaim the whole CPU bandwidth Luca Abeni
2016-10-24 14:06 ` [RFC v3 6/6] Make GRUB a task's flag Luca Abeni
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20161108200956.35b3e7c7@utopia \
--to=luca.abeni@unitn.it \
--cc=claudio@evidence.eu.com \
--cc=juri.lelli@arm.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@redhat.com \
--cc=peterz@infradead.org \
--cc=rostedt@goodmis.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).