2022-06-01 15:27:33

by Bing Huang

[permalink] [raw]
Subject: [PATCH v2] sched/fair: static cpumasks for load balance

The both cpu mask load_balance_mask and select_idle_mask just only used
in fair.c, but allocation in core.c in CONFIG_CPUMASK_OFFSTACK=y case,
and global via declare per cpu variations. More or less, it looks wired.

Signed-off-by: Bing Huang <[email protected]>
---

v2: move load_balance_mask and select_idle_mask allocation from
sched_init() to init_sched_fair_class()

kernel/sched/core.c | 11 -----------
kernel/sched/fair.c | 14 ++++++++++++--
2 files changed, 12 insertions(+), 13 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 696c6490bd5b..707df2aeebf8 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -9503,9 +9503,6 @@ LIST_HEAD(task_groups);
static struct kmem_cache *task_group_cache __read_mostly;
#endif

-DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
-DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
-
void __init sched_init(void)
{
unsigned long ptr = 0;
@@ -9549,14 +9546,6 @@ void __init sched_init(void)

#endif /* CONFIG_RT_GROUP_SCHED */
}
-#ifdef CONFIG_CPUMASK_OFFSTACK
- for_each_possible_cpu(i) {
- per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
- cpumask_size(), GFP_KERNEL, cpu_to_node(i));
- per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
- cpumask_size(), GFP_KERNEL, cpu_to_node(i));
- }
-#endif

init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 8c5b74f66bd3..377d908866ab 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5843,8 +5843,8 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
#ifdef CONFIG_SMP

/* Working cpumask for: load_balance, load_balance_newidle. */
-DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
-DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
+static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
+static DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);

#ifdef CONFIG_NO_HZ_COMMON

@@ -11841,6 +11841,16 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
__init void init_sched_fair_class(void)
{
#ifdef CONFIG_SMP
+
+#ifdef CONFIG_CPUMASK_OFFSTACK
+ for_each_possible_cpu(i) {
+ per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
+ cpumask_size(), GFP_KERNEL, cpu_to_node(i));
+ per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
+ cpumask_size(), GFP_KERNEL, cpu_to_node(i));
+ }
+#endif
+
open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);

#ifdef CONFIG_NO_HZ_COMMON
--
2.25.1


No virus found
Checked by Hillstone Network AntiVirus


2022-06-01 19:14:23

by Dietmar Eggemann

[permalink] [raw]
Subject: Re: [PATCH v2] sched/fair: static cpumasks for load balance

On 31/05/2022 05:12, Bing Huang wrote:
> The both cpu mask load_balance_mask and select_idle_mask just only used
> in fair.c, but allocation in core.c in CONFIG_CPUMASK_OFFSTACK=y case,
> and global via declare per cpu variations. More or less, it looks wired.
>
> Signed-off-by: Bing Huang <[email protected]>
> ---
>
> v2: move load_balance_mask and select_idle_mask allocation from
> sched_init() to init_sched_fair_class()

This would align CFS with RT (local_cpu_mask) and DL
(local_cpu_mask_dl).

[...]

> @@ -11841,6 +11841,16 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
> __init void init_sched_fair_class(void)
> {
> #ifdef CONFIG_SMP

`int i` missing for DEBUG_PER_CPU_MAPS/CONFIG_CPUMASK_OFFSTACK case.

> +
> +#ifdef CONFIG_CPUMASK_OFFSTACK
> + for_each_possible_cpu(i) {
> + per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
> + cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> + per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
> + cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> + }
> +#endif
> +

What about:

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 98319b654788..9ef5133c72d6 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -11811,15 +11811,14 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
__init void init_sched_fair_class(void)
{
#ifdef CONFIG_SMP
+ int i;

-#ifdef CONFIG_CPUMASK_OFFSTACK
for_each_possible_cpu(i) {
- per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
- cpumask_size(), GFP_KERNEL, cpu_to_node(i));
- per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
- cpumask_size(), GFP_KERNEL, cpu_to_node(i));
+ zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i),
+ GFP_KERNEL, cpu_to_node(i));
+ zalloc_cpumask_var_node(&per_cpu(select_idle_mask, i),
+ GFP_KERNEL, cpu_to_node(i));
}
-#endif

to get rid of the #ifdef ? We do the same for RT (local_cpu_mask) and DL
(local_cpu_mask_dl).

[...]

2022-06-02 08:27:12

by Bing Huang

[permalink] [raw]
Subject: Re: [PATCH v2] sched/fair: static cpumasks for load balance


On 6/1/22 16:45, Dietmar Eggemann wrote:
> On 31/05/2022 05:12, Bing Huang wrote:
>> The both cpu mask load_balance_mask and select_idle_mask just only used
>> in fair.c, but allocation in core.c in CONFIG_CPUMASK_OFFSTACK=y case,
>> and global via declare per cpu variations. More or less, it looks wired.
>>
>> Signed-off-by: Bing Huang <[email protected]>
>> ---
>>
>> v2: move load_balance_mask and select_idle_mask allocation from
>> sched_init() to init_sched_fair_class()
> This would align CFS with RT (local_cpu_mask) and DL
> (local_cpu_mask_dl).
>
> [...]
>
>> @@ -11841,6 +11841,16 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
>> __init void init_sched_fair_class(void)
>> {
>> #ifdef CONFIG_SMP
> `int i` missing for DEBUG_PER_CPU_MAPS/CONFIG_CPUMASK_OFFSTACK case.
sorry,my fault.
>
>> +
>> +#ifdef CONFIG_CPUMASK_OFFSTACK
>> + for_each_possible_cpu(i) {
>> + per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
>> + cpumask_size(), GFP_KERNEL, cpu_to_node(i));
>> + per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
>> + cpumask_size(), GFP_KERNEL, cpu_to_node(i));
>> + }
>> +#endif
>> +
> What about:
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 98319b654788..9ef5133c72d6 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -11811,15 +11811,14 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
> __init void init_sched_fair_class(void)
> {
> #ifdef CONFIG_SMP
> + int i;
>
> -#ifdef CONFIG_CPUMASK_OFFSTACK
> for_each_possible_cpu(i) {
> - per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
> - cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> - per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
> - cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> + zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i),
> + GFP_KERNEL, cpu_to_node(i));
> + zalloc_cpumask_var_node(&per_cpu(select_idle_mask, i),
> + GFP_KERNEL, cpu_to_node(i));
> }
> -#endif
>
> to get rid of the #ifdef ? We do the same for RT (local_cpu_mask) and DL
> (local_cpu_mask_dl).
good suggestion, thanks very much. happy to do that.
>
> [...]