Re: Add rq->nr_uninterruptible count to dest cpu's rq while CPU goes down.

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Aug 20, 2012 at 09:26:57AM -0700, Paul E. McKenney wrote:
> On Mon, Aug 20, 2012 at 11:26:57AM +0200, Peter Zijlstra wrote:

[ . . . ]

> > OK, so how about something like the below, it would also solve Paul's
> > issue with that code.
> > 
> > 
> > Please do double check the logic, I've had all of 4 hours sleep and its
> > far too warm for a brain to operate in any case.
> > 
> > ---
> > Subject: sched: Fix load avg vs cpu-hotplug
> > 
> > Rabik and Paul reported two different issues related to the same few
> > lines of code.
> > 
> > Rabik's issue is that the nr_uninterruptible migration code is wrong in
> > that he sees artifacts due to this (Rabik please do expand in more
> > detail).
> > 
> > Paul's issue is that this code as it stands relies on us using
> > stop_machine() for unplug, we all would like to remove this assumption
> > so that eventually we can remove this stop_machine() usage altogether.
> > 
> > The only reason we'd have to migrate nr_uninterruptible is so that we
> > could use for_each_online_cpu() loops in favour of
> > for_each_possible_cpu() loops, however since nr_uninterruptible() is the
> > only such loop and its using possible lets not bother at all.
> > 
> > The problem Rabik sees is (probably) caused by the fact that by
> > migrating nr_uninterruptible we screw rq->calc_load_active for both rqs
> > involved.
> > 
> > So don't bother with fancy migration schemes (meaning we now have to
> > keep using for_each_possible_cpu()) and instead fold any nr_active delta
> > after we migrate all tasks away to make sure we don't have any skewed
> > nr_active accounting.
> > 
> > 
> > Reported-by: Rakib Mullick <rakib.mullick@gmail.com>
> > Reported-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > ---
> >  kernel/sched/core.c | 31 ++++++++++---------------------
> >  1 file changed, 10 insertions(+), 21 deletions(-)
> > 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 4376c9f..06d23c6 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -5338,27 +5338,17 @@ void idle_task_exit(void)
> >  }
> >  
> >  /*
> > - * While a dead CPU has no uninterruptible tasks queued at this point,
> > - * it might still have a nonzero ->nr_uninterruptible counter, because
> > - * for performance reasons the counter is not stricly tracking tasks to
> > - * their home CPUs. So we just add the counter to another CPU's counter,
> > - * to keep the global sum constant after CPU-down:
> > - */
> > -static void migrate_nr_uninterruptible(struct rq *rq_src)
> > -{
> > -	struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
> > -
> > -	rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
> > -	rq_src->nr_uninterruptible = 0;
> > -}
> > -
> > -/*
> > - * remove the tasks which were accounted by rq from calc_load_tasks.
> > + * Since this CPU is going 'away' for a while, fold any nr_active delta
> > + * we might have. Assumes we're called after migrate_tasks() so that the
> > + * nr_active count is stable.
> > + *
> > + * Also see the comment "Global load-average calculations".
> >   */
> > -static void calc_global_load_remove(struct rq *rq)
> > +static void calc_load_migrate(struct rq *rq)
> >  {
> > -	atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
> > -	rq->calc_load_active = 0;
> > +	long delta = calc_load_fold_active(rq);
> > +	if (delta)
> > +		atomic_long_add(delta, &calc_load_tasks);
> >  }
> >  
> >  /*
> > @@ -5652,8 +5642,7 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
> >  		BUG_ON(rq->nr_running != 1); /* the migration thread */
> >  		raw_spin_unlock_irqrestore(&rq->lock, flags);
> >  
> > -		migrate_nr_uninterruptible(rq);
> > -		calc_global_load_remove(rq);
> > +		calc_load_migrate(rq);
> 
> Not sure that it matters, but...
> 
> This is called from the CPU_DYING notifier, which runs with irqs
> disabled, but in process context.  As I understand it, this means that
> ->nr_running==1.  If my understanding is correct (ha!), this means that
> this change sets ->calc_load_active to one (rather than zero as in the
> original) and that it subtracts one fewer from calc_load_tasks than did
> the original.  Of course, I have no idea whether this matters.
> 
> If I am correct and if it does matter, one straightforward fix
> is to add a "CPU_DEAD" branch to the switch statement and move the
> "calc_load_migrate(rq)" to that new branch.  Given that "rq" references
> the outgoing CPU, my guess is that locking is not needed, but you would
> know better than I.

How about the following updated patch?

							Thanx, Paul

------------------------------------------------------------------------

sched: Fix load avg vs cpu-hotplug

Rabik and Paul reported two different issues related to the same few
lines of code.

Rabik's issue is that the nr_uninterruptible migration code is wrong in
that he sees artifacts due to this (Rabik please do expand in more
detail).

Paul's issue is that this code as it stands relies on us using
stop_machine() for unplug, we all would like to remove this assumption
so that eventually we can remove this stop_machine() usage altogether.

The only reason we'd have to migrate nr_uninterruptible is so that we
could use for_each_online_cpu() loops in favour of
for_each_possible_cpu() loops, however since nr_uninterruptible() is the
only such loop and its using possible lets not bother at all.

The problem Rabik sees is (probably) caused by the fact that by
migrating nr_uninterruptible we screw rq->calc_load_active for both rqs
involved.

So don't bother with fancy migration schemes (meaning we now have to
keep using for_each_possible_cpu()) and instead fold any nr_active delta
after we migrate all tasks away to make sure we don't have any skewed
nr_active accounting.

[ paulmck: Move call to calc_load_migration to CPU_DEAD to avoid
miscounting noted by Rakib. ]

Reported-by: Rakib Mullick <rakib.mullick@gmail.com>
Reported-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index e841dfc..a8807f2 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -5309,27 +5309,17 @@ void idle_task_exit(void)
 }
 
 /*
- * While a dead CPU has no uninterruptible tasks queued at this point,
- * it might still have a nonzero ->nr_uninterruptible counter, because
- * for performance reasons the counter is not stricly tracking tasks to
- * their home CPUs. So we just add the counter to another CPU's counter,
- * to keep the global sum constant after CPU-down:
- */
-static void migrate_nr_uninterruptible(struct rq *rq_src)
-{
-	struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
-
-	rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
-	rq_src->nr_uninterruptible = 0;
-}
-
-/*
- * remove the tasks which were accounted by rq from calc_load_tasks.
+ * Since this CPU is going 'away' for a while, fold any nr_active delta
+ * we might have. Assumes we're called after migrate_tasks() so that the
+ * nr_active count is stable.
+ *
+ * Also see the comment "Global load-average calculations".
  */
-static void calc_global_load_remove(struct rq *rq)
+static void calc_load_migrate(struct rq *rq)
 {
-	atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
-	rq->calc_load_active = 0;
+	long delta = calc_load_fold_active(rq);
+	if (delta)
+		atomic_long_add(delta, &calc_load_tasks);
 }
 
 /*
@@ -5622,9 +5612,18 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
 		migrate_tasks(cpu);
 		BUG_ON(rq->nr_running != 1); /* the migration thread */
 		raw_spin_unlock_irqrestore(&rq->lock, flags);
+		break;
 
-		migrate_nr_uninterruptible(rq);
-		calc_global_load_remove(rq);
+	case CPU_DEAD:
+		{
+			struct rq *dest_rq;
+
+			local_irq_save(flags);
+			dest_rq = cpu_rq(smp_processor_id());
+			raw_spin_lock(&dest_rq->lock);
+			calc_load_migrate(rq);
+			raw_spin_unlock_irqrestore(&dest_rq->lock, flags);
+		}
 		break;
 #endif
 	}