Re: [PATCH v2] sched: cpuset: Don't rebuild sched domains on suspend-resume

[Qais Yousef]

On 01/29/23 21:49, Waiman Long wrote:
> On 1/25/23 11:35, Qais Yousef wrote:
> > On 01/20/23 17:16, Waiman Long wrote:
> > > On 1/20/23 14:48, Qais Yousef wrote:
> > > > Commit f9a25f776d78 ("cpusets: Rebuild root domain deadline accounting information")
> > > > enabled rebuilding sched domain on cpuset and hotplug operations to
> > > > correct deadline accounting.
> > > > 
> > > > Rebuilding sched domain is a slow operation and we see 10+ ms delay on
> > > > suspend-resume because of that.
> > > > 
> > > > Since nothing is expected to change on suspend-resume operation; skip
> > > > rebuilding the sched domains to regain the time lost.
> > > > 
> > > > Debugged-by: Rick Yiu <rickyiu@xxxxxxxxxx>
> > > > Signed-off-by: Qais Yousef (Google) <qyousef@xxxxxxxxxxx>
> > > > ---
> > > > 
> > > >       Changes in v2:
> > > >       	* Remove redundant check in update_tasks_root_domain() (Thanks Waiman)
> > > >       v1 link:
> > > >       	https://lore.kernel.org/lkml/20221216233501.gh6m75e7s66dmjgo@airbuntu/
> > > > 
> > > >    kernel/cgroup/cpuset.c  | 3 +++
> > > >    kernel/sched/deadline.c | 3 +++
> > > >    2 files changed, 6 insertions(+)
> > > > 
> > > > diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> > > > index a29c0b13706b..9a45f083459c 100644
> > > > --- a/kernel/cgroup/cpuset.c
> > > > +++ b/kernel/cgroup/cpuset.c
> > > > @@ -1088,6 +1088,9 @@ static void rebuild_root_domains(void)
> > > >    	lockdep_assert_cpus_held();
> > > >    	lockdep_assert_held(&sched_domains_mutex);
> > > > +	if (cpuhp_tasks_frozen)
> > > > +		return;
> > > > +
> > > >    	rcu_read_lock();
> > > >    	/*
> > > > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> > > > index 0d97d54276cc..42c1143a3956 100644
> > > > --- a/kernel/sched/deadline.c
> > > > +++ b/kernel/sched/deadline.c
> > > > @@ -2575,6 +2575,9 @@ void dl_clear_root_domain(struct root_domain *rd)
> > > >    {
> > > >    	unsigned long flags;
> > > > +	if (cpuhp_tasks_frozen)
> > > > +		return;
> > > > +
> > > >    	raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
> > > >    	rd->dl_bw.total_bw = 0;
> > > >    	raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
> > > cpuhp_tasks_frozen is set when thaw_secondary_cpus() or
> > > freeze_secondary_cpus() is called. I don't know the exact suspend/resume
> > > calling sequences, will cpuhp_tasks_frozen be cleared at the end of resume
> > > sequence? Maybe we should make sure that rebuild_root_domain() is called at
> > > least once at the end of resume operation.
> > Very good questions. It made me look at the logic again and I realize now that
> > the way force_build behaves is causing this issue.
> > 
> > I *think* we should just make the call rebuild_root_domains() only if
> > cpus_updated in cpuset_hotplug_workfn().
> > 
> > cpuset_cpu_active() seems to be the source of force_rebuild in my case; which
> > seems to be called only after the last cpu is back online (what you suggest).
> > In this case we can end up with cpus_updated = false, but force_rebuild = true.
> > 
> > Now you added a couple of new users to force_rebuild in 4b842da276a8a; I'm
> > trying to figure out what the conditions would be there. It seems we can have
> > corner cases for cpus_update might not trigger correctly?
> > 
> > Could the below be a good cure?
> > 
> > AFAICT we must rebuild the root domains if something has changed in cpuset.
> > Which should be captured by either having:
> > 
> > 	* cpus_updated = true
> > 	* force_rebuild && !cpuhp_tasks_frozen
> > 
> > /me goes to test the patch
> > 
> > --->8---
> > 
> > 	diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> > 	index a29c0b13706b..363e4459559f 100644
> > 	--- a/kernel/cgroup/cpuset.c
> > 	+++ b/kernel/cgroup/cpuset.c
> > 	@@ -1079,6 +1079,8 @@ static void update_tasks_root_domain(struct cpuset *cs)
> > 		css_task_iter_end(&it);
> > 	 }
> > 
> > 	+static bool need_rebuild_rd = true;
> > 	+
> > 	 static void rebuild_root_domains(void)
> > 	 {
> > 		struct cpuset *cs = NULL;
> > 	@@ -1088,6 +1090,9 @@ static void rebuild_root_domains(void)
> > 		lockdep_assert_cpus_held();
> > 		lockdep_assert_held(&sched_domains_mutex);
> > 
> > 	+       if (!need_rebuild_rd)
> > 	+               return;
> > 	+
> > 		rcu_read_lock();
> > 
> > 		/*
> > 	@@ -3627,7 +3632,9 @@ static void cpuset_hotplug_workfn(struct work_struct *work)
> > 		/* rebuild sched domains if cpus_allowed has changed */
> > 		if (cpus_updated || force_rebuild) {
> > 			force_rebuild = false;
> > 	+               need_rebuild_rd = cpus_updated || (force_rebuild && !cpuhp_tasks_frozen);
> > 			rebuild_sched_domains();
> > 	+               need_rebuild_rd = true;
> 
> You do the force_check check after it is set to false in the previous
> statement which is definitely not correct. So it will be false whenever
> cpus_updated is false.
> 
> If you just want to skip rebuild_sched_domains() call for hotplug, why don't

We just need to skip rebuild_root_domains(). I think rebuild_sched_domains()
should still happen.

The issue, AFAIU, is that we assume this hotplug operation results in changes
in cpuset and the DEADLINE accounting is now wrong and must be re-calculated.
But s2ram will cause hotplug operation without actually changing the cpuset
configuration - the re-calculation is not required. But it'd be good to get
a confirmation from Juri.

> 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(). force_rebuild will only be set after the last cpu
is brought online though - so this should happen once at the end.

(will update the comment too)

It seems I still need more time to study the code. What appeared simple, looks
is actually not..


Cheers

--
Qais Yousef