Date:   Thu, 2 Feb 2023 19:20:34 +0000
From:   Qais Yousef <qyousef@...alina.io>
To:     Waiman Long <longman@...hat.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...nel.org>,
        Juri Lelli <juri.lelli@...hat.com>,
        Steven Rostedt <rostedt@...dmis.org>, tj@...nel.org,
        linux-kernel@...r.kernel.org, luca.abeni@...tannapisa.it,
        claudio@...dence.eu.com, tommaso.cucinotta@...tannapisa.it,
        bristot@...hat.com, mathieu.poirier@...aro.org,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        cgroups@...r.kernel.org,
        Vincent Guittot <vincent.guittot@...aro.org>,
        Wei Wang <wvw@...gle.com>, Rick Yiu <rickyiu@...gle.com>,
        Quentin Perret <qperret@...gle.com>
Subject: Re: [PATCH v2] sched: cpuset: Don't rebuild sched domains on
 suspend-resume

On 01/31/23 14:33, Waiman Long wrote:
> 
> On 1/31/23 14:22, Qais Yousef wrote:
> > On 01/30/23 14:57, Waiman Long wrote:
> > > On 1/30/23 14:48, Qais Yousef wrote:
> > > > On 01/30/23 11:29, Waiman Long wrote:
> > > > > On 1/30/23 08:00, Qais Yousef wrote:
> > > > > 
> > > > >           just skip the call here if the condition is right? Like
> > > > > 
> > > > >                   /* rebuild sched domains if cpus_allowed has changed */
> > > > >                   if (cpus_updated || (force_rebuild && !cpuhp_tasks_frozen)) {
> > > > >                           force_rebuild = false;
> > > > >                           rebuild_sched_domains();
> > > > >                   }
> > > > > 
> > > > >           Still, we will need to confirm that cpuhp_tasks_frozen will be cleared
> > > > >           outside of the suspend/resume cycle.
> > > > > 
> > > > >       I think it's fine to use this variable from the cpuhp callback context only.
> > > > >       Which I think this cpuset workfn is considered an extension of.
> > > > > 
> > > > >       But you're right, I can't use cpuhp_tasks_frozen directly in
> > > > >       rebuild_root_domains() as I did in v1 because it doesn't get cleared after
> > > > >       calling the last _cpu_up().
> > > > > 
> > > > > That is what I suspect. So we can't use that cpuhp_tasks_frozen variable here
> > > > > in cpuset.
> > > > > 
> > > > >        force_rebuild will only be set after the last cpu
> > > > >       is brought online though - so this should happen once at the end.
> > > > > 
> > > > > Perhaps you can add another tracking variable for detecting if suspend/resume
> > > > > is in progress.
> > > > I think cpuhp_tasks_frozen is meant for that. All users who cared so far
> > > > belonged to the cpuhp callback. I think reading it from cpuset_hotplug_workfn()
> > > > is fine too as this function will only run as a consequence of the cpuhp
> > > > callback AFAICS. cpuset_cpu_active() takes care of not forcing a rebuild of
> > > > sched_domains until the last cpu becomes active - so the part of it being done
> > > > once at the end at resume is handled too.
> > > Well we will have to add code to clear cpuhp_tasks_frozen at the end of
> > > resume then. We don't want to affect other callers unless we are sure that
> > > it won't affect them.
> > Actually I think since the cpuset_hotplug_workfn() is called later, there's
> > a chance to race with another cpuhp operation just after resume.
> > 
> > Anyway. I think we don't have to use this flag. But we'd have to better distill
> > the reasons of why we force_rebuild.
> > 
> > Your 2 new users are tripping me so far - do they handle errors where the shape
> > of cpuset changes? If yes, then we must take dl accounting update into
> > consideration for these errors.
> The 2 new users is for the cpuset cpu partition which is used to create a
> secondary scheduling domain and hence have to call rebuilds_sched_domains()
> to set it up. Those should not be used that frequently.

Okay, thanks. So honouring these looks important, unlike the force_rebuild on
suspend/resume.

> 
> > 
> > Juri, I'd still would appreciate a confirmation from you that I'm not
> > understanding things completely wrong.
> > 
> > > > It's just rebuild_sched_domains() will always assume it needs to clear and
> > > > rebuild deadline accounting - which is not true for suspend/resume case. But
> > > > now looking at other users of rebuild_sched_domains(), others might be getting
> > > > the hit too. For example rebuild_sched_domains_locked() is called on
> > > > update_relax_domain_level() which AFAIU should not impact dl accounting.
> > > > 
> > > > FWIW, I did capture a worst case scenario of 21ms because of
> > > > rebuild_root_domains().
> > > > 
> > > > /me thinks rebuild_root_domains() is a misleading name too as it just fixes
> > > > dl accounting but not rebuild the rd itself.
> > > > 
> > > > What makes sense to me now is to pass whether dl accounting requires updating
> > > > to rebuild_sched_domains() as an arg so that the caller can decide whether the
> > > > reason can affect dl accounting.
> > > > 
> > > > Or maybe pull rebuild_root_domains() out of the chain and let the caller call
> > > > it directly. And probably rename it to update_do_rd_accounting() or something.
> > > > 
> > > > I'll continue to dig more..
> > > Looking forward to see that.
> > Another thought I had is maybe worth trying to optimize the rebuild root domain
> > process. Interestingly in my system there are no dl tasks but
> > 
> > 	rebuilds_sched_domains()
> > 	  cpuset_for_each_descendant_pre()
> > 	    update_tasks_root_domain()
> > 	      css_task_iter_next()
> > 	        dl_add_task_root_domain()
> > 
> > seems to be going through every task in the hierarchy anyway which would
> > explain the slow down. We can have special variants to iterate through
> > hierarchies that ever seen a dl task attached to them and a special variant to
> > iterate through dl tasks only in a css - but I'm not sure if I'm brave enough
> > to go down this rabbit hole :D
> 
> Yes, it seems like we have to check every tasks in the system to see if they
> are dl tasks. It can be expensive if there are a large number of tasks.
> Maybe we should track the # of dl tasks in each cgroup and skip this
> operation if there is none.

Yep, would be nice to have, hehe.


Cheers

--
Qais Yousef

Powered by blists - more mailing lists