Re: [PATCH] kernel/sched: Update schedstats when migrating threads

[Carlos Bilbao]

On 2/23/2022 4:19 AM, Peter Zijlstra wrote:
> On Wed, Jan 26, 2022 at 09:22:23AM -0600, Carlos Bilbao wrote:
>> The kernel manages per-task scheduler statistics or schedstats. Such
>> counters should be reinitialized when the thread is migrated to a
>> different core rq, except for the values recording number of migrations.
> 
> I'm confused, why should we reset schedstats on migrate? I'm thinking
> this breaks per-task, since tasks tend to bounce around quite a lot.
> 

Thanks for your comments, Peter. 

Looking at the documentation of schedstats I see that most values are 
actually linked to the particular CPU: time spent on the cpu, timeslices 
run on this cpu, number of times _something_ was called when the cpu was 
idle, and so forth. Those values lose their meaning after migration and we 
should reinitialize their counters. However, reviewing sched_statistics I 
identify two fields that we should definitely keep increasing even after 
migration (nr_migrations_cold, nr_forced_migrations).

So this patch will have to be upgraded if there's some other value(s) in
schedstats that we do not want to reinitialize either.

>> Signed-off-by: Carlos Bilbao <carlos.bilbao@xxxxxxx>
>> ---
>>  kernel/sched/core.c | 10 +++++++++-
>>  1 file changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index fe53e510e711..d64c2a290176 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -8757,6 +8757,7 @@ bool sched_smp_initialized __read_mostly;
>>  int migrate_task_to(struct task_struct *p, int target_cpu)
>>  {
>>  	struct migration_arg arg = { p, target_cpu };
>> +	uint64_t forced_migrations, migrations_cold;
>>  	int curr_cpu = task_cpu(p);
>>  
>>  	if (curr_cpu == target_cpu)
>> @@ -8765,7 +8766,14 @@ int migrate_task_to(struct task_struct *p, int target_cpu)
>>  	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
>>  		return -EINVAL;
>>  
>> -	/* TODO: This is not properly updating schedstats */
>> +	if (schedstat_enabled()) {
>> +		forced_migrations = schedstat_val(p->stats.nr_forced_migrations);
>> +		migrations_cold = schedstat_val(p->stats.nr_migrations_cold);
>> +		memset(&p->stats, 0, sizeof(p->stats));
>> +		schedstat_set(p->stats.nr_forced_migrations, forced_migrations);
>> +		schedstat_set(p->stats.nr_migrations_cold, migrations_cold);
>> +		schedstat_inc(p->stats.nr_migrations_cold);
>> +	}
>>  
>>  	trace_sched_move_numa(p, curr_cpu, target_cpu);
>>  	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
>> -- 
>> 2.27.0
>>

Thanks,
Carlos