]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
sched: Fix sched rt group scheduling when hierachy is enabled
authorBalbir Singh <balbir@linux.vnet.ibm.com>
Thu, 3 Mar 2011 11:34:35 +0000 (17:04 +0530)
committerGreg Kroah-Hartman <gregkh@suse.de>
Mon, 14 Mar 2011 21:17:39 +0000 (14:17 -0700)
commit 0c3b9168017cbad2c4af3dd65ec93fe646eeaa62 upstream.

The current sched rt code is broken when it comes to hierarchical
scheduling, this patch fixes two problems

1. It adds redundant enqueuing (harmless) when it finds a queue
   has tasks enqueued, but it has no run time and it is not
   throttled.

2. The most important change is in sched_rt_rq_enqueue/dequeue.
   The code just picks the rt_rq belonging to the current cpu
   on which the period timer runs, the patch fixes it, so that
   the correct rt_se is enqueued/dequeued.

Tested with a simple hierarchy

/c/d, c and d assigned similar runtimes of 50,000 and a while
1 loop runs within "d". Both c and d get throttled, without
the patch, the task just stops running and never runs (depends
on where the sched_rt b/w timer runs). With the patch, the
task is throttled and runs as expected.

[ bharata, suggestions on how to pick the rt_se belong to the
  rt_rq and correct cpu ]

Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20110303113435.GA2868@balbir.in.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
kernel/sched_rt.c

index c5f1d2370336ebb19aca13842e5f64234ab8b2eb..acc5b8b42fd13ebc21d5cb6e477cc8565e419924 100644 (file)
@@ -199,11 +199,12 @@ static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 
 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 {
-       int this_cpu = smp_processor_id();
        struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
        struct sched_rt_entity *rt_se;
 
-       rt_se = rt_rq->tg->rt_se[this_cpu];
+       int cpu = cpu_of(rq_of_rt_rq(rt_rq));
+
+       rt_se = rt_rq->tg->rt_se[cpu];
 
        if (rt_rq->rt_nr_running) {
                if (rt_se && !on_rt_rq(rt_se))
@@ -215,10 +216,10 @@ static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 
 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 {
-       int this_cpu = smp_processor_id();
        struct sched_rt_entity *rt_se;
+       int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 
-       rt_se = rt_rq->tg->rt_se[this_cpu];
+       rt_se = rt_rq->tg->rt_se[cpu];
 
        if (rt_se && on_rt_rq(rt_se))
                dequeue_rt_entity(rt_se);
@@ -546,8 +547,11 @@ static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
                        if (rt_rq->rt_time || rt_rq->rt_nr_running)
                                idle = 0;
                        raw_spin_unlock(&rt_rq->rt_runtime_lock);
-               } else if (rt_rq->rt_nr_running)
+               } else if (rt_rq->rt_nr_running) {
                        idle = 0;
+                       if (!rt_rq_throttled(rt_rq))
+                               enqueue = 1;
+               }
 
                if (enqueue)
                        sched_rt_rq_enqueue(rt_rq);