Re: [RESEND] In cgroup v2, setting a smaller value for sched_rt_runtime_us fails.

[chenjinghuang <chenjinghuang2@huawei.com>]

On 3/27/2026 11:53 AM, chenjinghuang wrote:
> On 3/20/2026 9:34 AM, Chen Jinghuang wrote:
>> For example:
>>
>> 	# cat /proc/sys/kernel/sched_rt_runtime_us
>> 	950000
>> 	# echo 940000 > /proc/sys/kernel/sched_rt_runtime_us
>> 	-bash: echo: write error: Invalid argument
>>
>> This occurs because when the global RT quota is reduced while the root
>> cgroup’s ratio remains unchanged, the root cgroup’s proportion ends up
>> exceeding the global RT ratio. This is unreasonable: the root cgroup’s
>> ratio should be updated when the global RT ratio changes.
>>
>> This patch ensures the root cgroup’s RT quota is adjusted whenever the
>> global RT quota is modified.
>>
>> Signed-off-by: Chen Jinghuang <chenjinghuang2@huawei.com>
>> Tested-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com>
>> Reviewed-by: Chen Ridong <chenridong@huawei.com>
>> ---
>>  kernel/sched/rt.c | 12 +++++++++++-
>>  1 file changed, 11 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
>> index f1867fe8e5c5..30770ba7be4e 100644
>> --- a/kernel/sched/rt.c
>> +++ b/kernel/sched/rt.c
>> @@ -2814,9 +2814,11 @@ long sched_group_rt_period(struct task_group *tg)
>>  static int sched_rt_global_constraints(void)
>>  {
>>  	int ret = 0;
>> +	u64 period = global_rt_period();
>> +	u64 runtime = global_rt_runtime();
>>  
>>  	mutex_lock(&rt_constraints_mutex);
>> -	ret = __rt_schedulable(NULL, 0, 0);
>> +	ret = __rt_schedulable(&root_task_group, period, runtime);
>>  	mutex_unlock(&rt_constraints_mutex);
>>  
>>  	return ret;
>> @@ -2856,6 +2858,14 @@ static int sched_rt_global_validate(void)
>>  
>>  static void sched_rt_do_global(void)
>>  {
>> +#ifdef CONFIG_RT_GROUP_SCHED
>> +	if (!rt_group_sched_enabled())
>> +		return;
>> +
>> +	WARN_ON_ONCE(tg_set_rt_bandwidth(&root_task_group,
>> +					global_rt_period(),
>> +					global_rt_runtime()));
>> +#endif
>>  }
>>  
>>  static int sched_rt_handler(const struct ctl_table *table, int write, void *buffer,
> ping
Hi,
Gentle ping on this patch.

I'd appreciate any feedback whenever you get time, or let me know if I
should resend/rework anything.

Regards,
Chen Jinghuang