Re: [PATCH 8/8] sched/fair: Relax task_hot() for misfit tasks

[Valentin Schneider]

On 08/02/21 17:21, Vincent Guittot wrote:
> On Thu, 28 Jan 2021 at 19:32, Valentin Schneider
> <valentin.schneider@xxxxxxx> wrote:
>>
>> Misfit tasks can and will be preempted by the stopper to migrate them over
>> to a higher-capacity CPU. However, when runnable but not current misfit
>> tasks are scanned by the load balancer (i.e. detach_tasks()), the
>> task_hot() ratelimiting logic may prevent us from enqueuing said task onto
>> a higher-capacity CPU.
>>
>> Align detach_tasks() with the active-balance logic and let it pick a
>> cache-hot misfit task when the destination CPU can provide a capacity
>> uplift.
>>
>> Signed-off-by: Valentin Schneider <valentin.schneider@xxxxxxx>
>> ---
>>  kernel/sched/fair.c | 11 +++++++++++
>>  1 file changed, 11 insertions(+)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index cba9f97d9beb..c2351b87824f 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -7484,6 +7484,17 @@ static int task_hot(struct task_struct *p, struct lb_env *env)
>>         if (env->sd->flags & SD_SHARE_CPUCAPACITY)
>>                 return 0;
>>
>> +       /*
>> +        * On a (sane) asymmetric CPU capacity system, the increase in compute
>> +        * capacity should offset any potential performance hit caused by a
>> +        * migration.
>> +        */
>> +       if (sd_has_asym_cpucapacity(env->sd) &&
>> +           env->idle != CPU_NOT_IDLE &&
>> +           !task_fits_capacity(p, capacity_of(env->src_cpu)) &&
>> +           cpu_capacity_greater(env->dst_cpu, env->src_cpu))
>
> Why not using env->migration_type to directly detect that it's a
> misfit task active migration ?
>

This is admittedly a kludge. Consider the scenario described in patch 7/8,
i.e.:
- there's a misfit task running on a LITTLE CPU
- a big CPU completes its work and is about to go through newidle_balance()

Now, consider by the time that big CPU gets into load_balance(), the misfit
task on the LITTLE CPU got preempted by a percpu kworker. As of right now,
it's quite likely the imbalance won't be classified as group_misfit_task,
but as group_overloaded (depends on the topology / workload, but that's a
symptom I've been seeing).

Unfortunately, even if we e.g. change the misfit load-balance logic to also
track preempted misfit tasks (rather than just the rq's current), this
could still happen AFAICT.

Long story short, we already trigger an active-balance to upmigrate running
misfit tasks, this changes task_hot() to allow any preempted task that
doesn't fit on its CPU to be upmigrated (regardless of the imbalance
classification).

>> +               return 0;
>> +
>>         /*
>>          * Buddy candidates are cache hot:
>>          */
>> --
>> 2.27.0
>>