[PATCH] Fix tasks being forgotten for a long time on SMP

[Yuriy Romanenko]

Observed occasional very high latency on an embedded SMP system between
a task becoming ready to run and actually running with low system load,
impacting interactive usage.

A sched_wake() from CPUx on CPUy puts the task into the run queue and
marks it runnable, but does not trigger an IPI to have the scheduler
re-run on CPUy and see if the current task needs to get pre-empted and
does not wake up CPUy if it is asleep.

This is especially evident when a CPU is in SWFI and simply does not
wake up even though it now has a runnable task.

This is probably not the most elegant fix and definitely generates some
unnecessary scheduler runs, but it's better for overall latency.

Signed-off-by: Yuriy Romanenko <yromanenko@xxxxxxxxxxxxxx>
---
 kernel/sched/core.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 860070f..7c334b7 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1686,6 +1686,14 @@ static void ttwu_do_wakeup(struct rq *rq,
struct task_struct *p, int wake_flags,
  trace_sched_wakeup(p);

 #ifdef CONFIG_SMP
+ /*
+ * If the task is not on the current cpu, there is a chance
+ * the other cpu might be asleep and will not get to our task
+ * for a really long time. Send an IPI to avoid that
+ */
+ if (task_cpu(p) != smp_processor_id())
+ smp_send_reschedule(task_cpu(p));
+
  if (p->sched_class->task_woken) {
  /*
  * Our task @p is fully woken up and running; so its safe to
-- 
2.7.4