I've been trying to track down some unexpected realtime latencies and
believe one source is a bug in the wakeup code. Specifically, this is
within the try_to_wake_up() routine. Within this routine there is the
following code segment:
/*
* If a newly woken up RT task cannot preempt the
* current (RT) task (on a target runqueue) then try
* to find another CPU it can preempt:
*/
if (rt_task(p) && !TASK_PREEMPTS_CURR(p, rq)) {
struct rq *this_rq = cpu_rq(this_cpu);
/*
* Special-case: the task on this CPU can be
* preempted. In that case there's no need to
* trigger reschedules on other CPUs, we can
* mark the current task for reschedule.
*
* (Note that it's safe to access this_rq without
* extra locking in this particular case, because
* we are on the current CPU.)
*/
if (TASK_PREEMPTS_CURR(p, this_rq))
set_tsk_need_resched(this_rq->curr);
else
/*
* Neither the intended target runqueue
* nor the current CPU can take this task.
* Trigger a reschedule on all other CPUs
* nevertheless, maybe one of them can take
* this task:
*/
smp_send_reschedule_allbutself_cpumask(p->cpus_allowed);
schedstat_inc(this_rq, rto_wakeup);
}
This logic seems appropriate. But, the task 'p' is most likely not on
the runqueue when sending the IPI. It gets added to the runqueue a
little later in the routine. As a result, the 'rt_overload' global may
not be set (based on the count of RT tasks on the runqueue) and other
CPUs may 'pass over' the runqueue when doing RT load balancing.
My observations/debugging/conclusions are based on an earlier version
of the code. It appears the same code/issue still exists in the most
version. But, I have not not done any work with the latest version.
--
Mike
-
