linux-kernel - Re: [PATCH v2] sched/deadline: remove useless param from setup_new_dl

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20160706104411.0632d89f@utopia>
Date:	Wed, 6 Jul 2016 10:44:11 +0200
From:	luca abeni <luca.abeni@...tn.it>
To:	Juri Lelli <juri.lelli@....com>
Cc:	Steven Rostedt <rostedt@...dmis.org>, peterz@...radead.org,
	linux-kernel@...r.kernel.org, mingo@...hat.com
Subject: Re: [PATCH v2] sched/deadline: remove useless param from
 setup_new_dl_entity

On Tue, 5 Jul 2016 17:58:30 +0100
Juri Lelli <juri.lelli@....com> wrote:

> On 05/07/16 12:47, Steven Rostedt wrote:
> > On Tue, 5 Jul 2016 15:39:33 +0100
> > Juri Lelli <juri.lelli@....com> wrote:
> > 
> > 		return;  
> > > > >  
> > > > >  	/*
> > > > > +	 * Use the scheduling parameters of the top
> > > > > pi-waiter task,
> > > > > +	 * if we have one from which we can inherit a
> > > > > deadline.
> > > > > +	 */
> > > > > +	if (pi_task && dl_se->dl_boosted &&
> > > > > dl_prio(pi_task->normal_prio))
> > > > > +		pi_se = &pi_task->dl;
> > > > > +    
> > > > 
> > > > OK, I'm micro-optimizing now, but hey, isn't this a fast path?
> > > > 
> > > > What about changing the above to:
> > > > 
> > > > 	struct task_struct *pi_task;
> > > > 	[...]
> > > > 
> > > > 	if (dl_se->dl_boosted && dl_prio(pi_task->normal_prio
> > > > &&    
> > >                                     ^
> > > OK, we need to reorder these two
> > >                                     V  
> > > > 	    (pi_task = rt_mutex_get_top_task(dl_task_of(dl_se)))
> > > > 		pe_se = &pi_task->dl;  
> > 
> > Opps, you're right.
> >   
> > > > 
> > > > This way we don't need to do any work of looking at
> > > > rt_mutex_get_top_task() for the normal case.
> > > >     
> > > 
> > > But, yes. Looks good to me. I'll shoot a v3 ASAP.  
> > 
> > I have to ask, should there be any check if the dl_se has a shorter
> > deadline than the pi one?
> >   
> 
> Yeah. I wondered the same actually. I convinced myself that, since the
> task is boosted, we assume that the donor will have a shorter
> deadline.

Do you mean relative deadline (dl_se->dl_deadline) or absolute
(scheduling) dealine (dl_se->deadline)?

If I understand well, here we are in setup_new_dl_entity(), right?
This should be called only from switched_to_dl(); so, dl_se is from a
task that is switching to -deadline. If it is dl_boosted, it means that
it is switching from SCHED_OTHER (or RT) to -deadline because of
inheritance... So, it is very likely that dl_se->dl_deadline is not
meaningful.

Moreover, setup_new_dl_entity() is only called if the current
scheduling deadline of the task is not usable (that is, if
"dl_time_before(p->dl.deadline, rq_clock(rq)"). So, dl_se->deadline
will be surely smaller than pi_se->deadline... But the inheritance has
to happen anyway.

> We seem to be doing the same elsewhere, but Luca was saying
> some time ago that the DI thing my have some problems and needs to be
> revised.

My doubts regarding the inheritance code currently used for -deadline
tasks are due to the fact that it is not clear which kind of
inheritance algorithm is used...
I think it should use deadline inheritance, that, AFAIK, says that when
task T1 block waiting for task T2, T2 can inherit T1's _absolute_
deadline - if it is earlier than T2's one.
But the current code seems to be using relative deadlines (dl_deadline)
to decide the inheritance...

Having a better look at this is in my TODO list... But I still need to
find some time :)

				Luca

> Is is fair enough fixing this bit in accordance with the
> current (maybe broken) behaviour and then spend time reviewing the
> whole thing, or do we want to do both at the same time (which will of
> course require more time)?
> 
> Best,
> 
> - Juri