[PATCH] sched/deadline: Fix sched_dl_global_validate()

[PATCH] sched/deadline: Fix sched_dl_global_validate()

[Peng Liu]

When user changes sched_rt_{runtime, period}_us, then

  sched_rt_handler()
    -->	sched_dl_bandwidth_validate()
	{
		new_bw = global_rt_runtime()/global_rt_period();

		for_each_possible_cpu(cpu) {
			dl_b = dl_bw_of(cpu);
			if (new_bw < dl_b->total_bw)
				ret = -EBUSY;
		}
	}

Under CONFIG_SMP, dl_bw is per root domain , but not per CPU,
dl_b->total_bw is the allocated bandwidth of the whole root domain.
we should compare dl_b->total_bw against cpus*new_bw, where 'cpus'
is the number of CPUs of the root domain.

Also, below annotation(in kernel/sched/sched.h) implied implementation
only appeared in SCHED_DEADLINE v2[1], then deadline scheduler kept
evolving till got merged(v9), but the annotation remains unchanged,
meaningless and misleading, correct it.

* With respect to SMP, the bandwidth is given on a per-CPU basis,
* meaning that:
*  - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
*  - dl_total_bw array contains, in the i-eth element, the currently
*    allocated bandwidth on the i-eth CPU.

[1] https://lkml.org/lkml/2010/2/28/119

Signed-off-by: Peng Liu <iwtbavbm@gmail.com>
---
In fact, I'm not 100% sure that's a bug, since it's too 'obvious' and
not newly introduced code.

Also, the introduced #ifdef...#endif pairs look ugly, I have no idea
how to eliminate them. Ideas and comments are welcome. Thanks.

 kernel/sched/deadline.c | 48 ++++++++++++++++++++++++++++-------------
 kernel/sched/sched.h    | 17 +++++----------
 2 files changed, 38 insertions(+), 27 deletions(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index 3862a28cd05d..6524cb31148e 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -2511,33 +2511,43 @@ const struct sched_class dl_sched_class
 	.update_curr		= update_curr_dl,
 };
 
+#ifdef CONFIG_SMP
+static struct cpumask dl_local_possible_mask;
+#endif /* CONFIG_SMP */
+
 int sched_dl_global_validate(void)
 {
 	u64 runtime = global_rt_runtime();
 	u64 period = global_rt_period();
 	u64 new_bw = to_ratio(period, runtime);
 	struct dl_bw *dl_b;
-	int cpu, ret = 0;
+	int cpu, cpus = 1, ret = 0;
 	unsigned long flags;
-
+	cpumask_t *possible_mask = NULL;
+#ifdef CONFIG_SMP
+	cpumask_t *span;
+#endif /* CONFIG_SMP */
 	/*
 	 * Here we want to check the bandwidth not being set to some
 	 * value smaller than the currently allocated bandwidth in
 	 * any of the root_domains.
-	 *
-	 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
-	 * cycling on root_domains... Discussion on different/better
-	 * solutions is welcome!
 	 */
-	for_each_possible_cpu(cpu) {
+#ifdef CONFIG_SMP
+	possible_mask = &dl_local_possible_mask;
+	cpumask_copy(possible_mask, cpu_possible_mask);
+#endif /* CONFIG_SMP */
+	for_each_cpu(cpu, possible_mask) {
 		rcu_read_lock_sched();
 		dl_b = dl_bw_of(cpu);
-
+#ifdef CONFIG_SMP
+		span = cpu_rq(cpu)->rd->span;
+		cpus = cpumask_weight(span);
+		cpumask_andnot(possible_mask, possible_mask, span);
+#endif /* CONFIG_SMP */
 		raw_spin_lock_irqsave(&dl_b->lock, flags);
-		if (new_bw < dl_b->total_bw)
+		if (new_bw * cpus < dl_b->total_bw)
 			ret = -EBUSY;
 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
-
 		rcu_read_unlock_sched();
 
 		if (ret)
@@ -2566,6 +2576,10 @@ void sched_dl_do_global(void)
 	struct dl_bw *dl_b;
 	int cpu;
 	unsigned long flags;
+	cpumask_t *possible_mask = NULL;
+#ifdef CONFIG_SMP
+	cpumask_t *span;
+#endif /* CONFIG_SMP */
 
 	def_dl_bandwidth.dl_period = global_rt_period();
 	def_dl_bandwidth.dl_runtime = global_rt_runtime();
@@ -2573,17 +2587,21 @@ void sched_dl_do_global(void)
 	if (global_rt_runtime() != RUNTIME_INF)
 		new_bw = to_ratio(global_rt_period(), global_rt_runtime());
 
-	/*
-	 * FIXME: As above...
-	 */
-	for_each_possible_cpu(cpu) {
+#ifdef CONFIG_SMP
+	possible_mask = &dl_local_possible_mask;
+	cpumask_copy(possible_mask, cpu_possible_mask);
+#endif /* CONFIG_SMP */
+	for_each_cpu(cpu, possible_mask) {
 		rcu_read_lock_sched();
 		dl_b = dl_bw_of(cpu);
 
 		raw_spin_lock_irqsave(&dl_b->lock, flags);
 		dl_b->bw = new_bw;
 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
-
+#ifdef CONFIG_SMP
+		span = cpu_rq(cpu)->rd->span;
+		cpumask_andnot(possible_mask, possible_mask, span);
+#endif /* CONFIG_SMP */
 		rcu_read_unlock_sched();
 		init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
 	}
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 28709f6b0975..2602544e06ff 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -258,9 +258,9 @@ struct rt_bandwidth {
 void __dl_clear_params(struct task_struct *p);
 
 /*
- * To keep the bandwidth of -deadline tasks and groups under control
+ * To keep the bandwidth of -deadline tasks under control
  * we need some place where:
- *  - store the maximum -deadline bandwidth of the system (the group);
+ *  - store the maximum -deadline bandwidth of each root domain;
  *  - cache the fraction of that bandwidth that is currently allocated.
  *
  * This is all done in the data structure below. It is similar to the
@@ -269,17 +269,10 @@ void __dl_clear_params(struct task_struct *p);
  * do not decrease any runtime while the group "executes", neither we
  * need a timer to replenish it.
  *
- * With respect to SMP, the bandwidth is given on a per-CPU basis,
+ * With respect to SMP, the bandwidth is given on a per root domain basis,
  * meaning that:
- *  - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
- *  - dl_total_bw array contains, in the i-eth element, the currently
- *    allocated bandwidth on the i-eth CPU.
- * Moreover, groups consume bandwidth on each CPU, while tasks only
- * consume bandwidth on the CPU they're running on.
- * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw
- * that will be shown the next time the proc or cgroup controls will
- * be red. It on its turn can be changed by writing on its own
- * control.
+ *  - bw (< 100%) is the bandwidth of the system on each CPU;
+ *  - total_bw is the currently allocated bandwidth on each root domain.
  */
 struct dl_bandwidth {
 	raw_spinlock_t		dl_runtime_lock;
-- 
2.20.1

Re: [PATCH] sched/deadline: Fix sched_dl_global_validate()

[Juri Lelli]

Hi,

On 15/09/20 23:20, Peng Liu wrote:
> When user changes sched_rt_{runtime, period}_us, then
> 
>   sched_rt_handler()
>     -->	sched_dl_bandwidth_validate()
> 	{
> 		new_bw = global_rt_runtime()/global_rt_period();
> 
> 		for_each_possible_cpu(cpu) {
> 			dl_b = dl_bw_of(cpu);
> 			if (new_bw < dl_b->total_bw)
> 				ret = -EBUSY;
> 		}
> 	}
> 
> Under CONFIG_SMP, dl_bw is per root domain , but not per CPU,
> dl_b->total_bw is the allocated bandwidth of the whole root domain.
> we should compare dl_b->total_bw against cpus*new_bw, where 'cpus'
> is the number of CPUs of the root domain.

Hummm, I think you are right. Guess nobody played a lot with changing
global values. :-/

> Also, below annotation(in kernel/sched/sched.h) implied implementation
> only appeared in SCHED_DEADLINE v2[1], then deadline scheduler kept
> evolving till got merged(v9), but the annotation remains unchanged,
> meaningless and misleading, correct it.
> 
> * With respect to SMP, the bandwidth is given on a per-CPU basis,
> * meaning that:
> *  - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
> *  - dl_total_bw array contains, in the i-eth element, the currently
> *    allocated bandwidth on the i-eth CPU.
> 
> [1] https://lkml.org/lkml/2010/2/28/119
> 
> Signed-off-by: Peng Liu <iwtbavbm@gmail.com>
> ---
> In fact, I'm not 100% sure that's a bug, since it's too 'obvious' and
> not newly introduced code.
> 
> Also, the introduced #ifdef...#endif pairs look ugly, I have no idea
> how to eliminate them. Ideas and comments are welcome. Thanks.

Can't we just use dl_bw_cpus() ?

Thanks,
Juri

Re: [PATCH] sched/deadline: Fix sched_dl_global_validate()

[Peng Liu]

On Wed, Sep 16, 2020 at 11:42:03AM +0200, Juri Lelli wrote:
> Hi,
> 
> On 15/09/20 23:20, Peng Liu wrote:
> > When user changes sched_rt_{runtime, period}_us, then
> > 
> >   sched_rt_handler()
> >     -->	sched_dl_bandwidth_validate()
> > 	{
> > 		new_bw = global_rt_runtime()/global_rt_period();
> > 
> > 		for_each_possible_cpu(cpu) {
> > 			dl_b = dl_bw_of(cpu);
> > 			if (new_bw < dl_b->total_bw)
> > 				ret = -EBUSY;
> > 		}
> > 	}
> > 
> > Under CONFIG_SMP, dl_bw is per root domain , but not per CPU,
> > dl_b->total_bw is the allocated bandwidth of the whole root domain.
> > we should compare dl_b->total_bw against cpus*new_bw, where 'cpus'
> > is the number of CPUs of the root domain.
> 
> Hummm, I think you are right. Guess nobody played a lot with changing
> global values. :-/
> 
> > Also, below annotation(in kernel/sched/sched.h) implied implementation
> > only appeared in SCHED_DEADLINE v2[1], then deadline scheduler kept
> > evolving till got merged(v9), but the annotation remains unchanged,
> > meaningless and misleading, correct it.
> > 
> > * With respect to SMP, the bandwidth is given on a per-CPU basis,
> > * meaning that:
> > *  - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU;
> > *  - dl_total_bw array contains, in the i-eth element, the currently
> > *    allocated bandwidth on the i-eth CPU.
> > 
> > [1] https://lkml.org/lkml/2010/2/28/119
> > 
> > Signed-off-by: Peng Liu <iwtbavbm@gmail.com>
> > ---
> > In fact, I'm not 100% sure that's a bug, since it's too 'obvious' and
> > not newly introduced code.
> > 
> > Also, the introduced #ifdef...#endif pairs look ugly, I have no idea
> > how to eliminate them. Ideas and comments are welcome. Thanks.
> 
> Can't we just use dl_bw_cpus() ?
> 

Oh, good idea, I will have a try. Thanks for your time!

> Thanks,
> Juri
>