Trying to page this back in...
On 07/01/22 11:46, Dietmar Eggemann wrote:
> On 22/12/2021 20:48, Valentin Schneider wrote:
>> On 22/12/21 18:46, Dietmar Eggemann wrote:
>>> On 21.12.21 17:45, John Keeping wrote:
>>>> On Tue, 21 Dec 2021 16:11:34 +0000
>>>> Valentin Schneider <valentin.schneider@xxxxxxx> wrote:
>>>>
>>>>> On 20/12/21 18:35, Dietmar Eggemann wrote:
>
> [...]
>
>>> switched_from_rt() -> rt_queue_pull_task(, pull_rt_task)
>>> pull_rt_task()->tell_cpu_to_push()->irq_work_queue_on(&rq->rd->rto_push_work,)
>>> rto_push_irq_work_func() -> push_rt_task(rq, true)
>>>
>>> seems to be the only way with pull=true.
>>>
>>> In my tests, rq->rt.rt_nr_running seems to be 0 when it happens.
>>>
>>> [ 22.288537] CPU3 switched_to_rt: p=[ksoftirqd/3 35]
>>> [ 22.288554] rt_mutex_setprio: CPU3 p=[ksoftirqd/3 35] pi_task=[rcu_preempt 11] queued=1 running=0 prio=98 oldprio=120
>>> [ 22.288636] CPU3 switched_from_rt: p=[ksoftirqd/3 35] rq->rt.rt_nr_running=0
>>> ^^^^^^^^^^^^^^^^^^^^^^
>>> [ 22.288649] rt_mutex_setprio: CPU3 p=[ksoftirqd/3 35] queued=1 running=1 prio=120 oldprio=98
>>> [ 22.288681] CPU3 push_rt_task: next_task=[rcu_preempt 11] migr_dis=1 rq->curr=[ksoftirqd/3 35] pull=1
>>> ^^^^^^^^^^ ^^^^^^
>>
>> mark_wakeup_next_waiter() first deboosts the previous owner and then
>> wakeups the next top waiter. Seems like you somehow have the wakeup happen
>> before the push_rt_task IRQ work is run. Also, tell_cpu_to_push() should
>> only pick a CPU that is in rq->rd->rto_mask, which requires having at least
>> 2 RT tasks there...
>
> True, this_cpu has rt_nr_running = 0 and *cpu* has rt_nr_running >= 2:
>
> mark_wakeup_next_waiter()
>
> (1) /* deboost */
> rt_mutex_adjust_prio()
>
> rt_mutex_setprio(current, ...)
>
> rq = __task_rq_lock(current, );
> check_class_changed(rq, p, prev_class, oldprio)
>
> switched_from_rt()
>
> if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
> return;
>
> rt_queue_pull_task(rq)
>
> queue_balance_callback(rq, ..., pull_rt_task);
>
> pull_rt_task()
>
> tell_cpu_to_push()
>
> *cpu* = rto_next_cpu(rq->rd)
> irq_work_queue_on(&rq->rd->rto_push_work, *cpu*)
>
> rto_push_irq_work_func()
> push_rt_task(rq, true) <-- !!!
>
> (2) /* waking the top waiter */
> rt_mutex_wake_q_add(wqh, waiter);
>
>> Now, that wakeup from the rtmutex unlock would give us a resched_curr() via
>> check_preempt_curr() if required, which is good, though I think we are
>> still missing some for sched_setscheduler() (there are no wakeups
>> there). So if we just have to live with an IRQ work popping in before we
>> get to preempt_schedule_irq() (or somesuch), then perhaps the below would
>> be sufficient.
>
> I think that's the case here but we are on the RT overloaded CPU (*cpu*).
>
So one thing I wasn't entirely clear on (and holidays didn't fix that) is
the rt_queue_pull_task() from switched_from_rt() only happens if that rq
has no other runnable RT tasks, so I don't quite see how the
irq_work_queue_on() can end up as a self-IPI...
>> diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
>> index ef8228d19382..8f3e3a1367b6 100644
>> --- a/kernel/sched/rt.c
>> +++ b/kernel/sched/rt.c
>> @@ -1890,6 +1890,16 @@ static int push_rt_task(struct rq *rq, bool pull)
>> if (!next_task)
>> return 0;
>>
>> + /*
>> + * It's possible that the next_task slipped in of higher priority than
>> + * current, or current has *just* changed priority. If that's the case
>> + * just reschedule current.
>> + */
>> + if (unlikely(next_task->prio < rq->curr->prio)) {
>> + resched_curr(rq);
>> + return 0;
>> + }
>
> IMHO, that's the bit which prevents the BUG.
>
> But this would also prevent the case in which rq->curr is an RT task
> with lower prio than next_task.
>
I think that's what we want - if current isn't the (logical) highest
priority task on the runqueue, we should forgo push/pull and reschedule
ASAP.
> Also `rq->curr = migration/X` goes still though which is somehow fine
> since find_lowest_rq() bails out for if (task->nr_cpus_allowed == 1).
>
> And DL tasks (like sugov:X go through and they can have
> task->nr_cpus_allowed > 1 (arm64 slow-switching boards with shared
> freuency domains with schedutil). cpupri_find_fitness()->convert_prio()
> can handle task_pri, p->prio = -1 (CPUPRI_INVALID) although its somehow
> by coincidence.
>
Right. This reminds me of:
https://lore.kernel.org/lkml/jhjblbx7glh.mognet@xxxxxxx/
> So maybe something like this:
>
Ah, so you explicitely prevent rt.c::find_lowest_rq() invocations with a
non-RT task... But what if current is an RT task that just got deboosted,
so that next_task->prio < rq->curr->prio? IMO we should reschedule ASAP (as
I already blabbered about above). If next_task is migration_disabled but
higher (logical) prio than current, we don't need to do any of the
migration_disabled specific crud, we just reschedule.
> @ -1898,6 +1898,11 @@ static int push_rt_task(struct rq *rq, bool pull)
> if (!pull || rq->push_busy)
> return 0;
>
> + if (rq->curr->sched_class != &rt_sched_class) {
> + resched_curr(rq);
> + return 0;
> + }
> +
> cpu = find_lowest_rq(rq->curr);
>
> [...]
