[PATCH] sched: fair scheduler should not resched rt tasks

[PATCH] sched: fair scheduler should not resched rt tasks

[Steven Rostedt]

Using ftrace, I noticed latencies in real-time tasks where they were
needlessly calling schedule due to sched_fair sending out time slices.

This patch prevents a call to resched_task by the sched fair class if
the task it wants to reschedule is an rt task.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
---
 kernel/sched_fair.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Index: linux-compile.git/kernel/sched_fair.c
===================================================================
--- linux-compile.git.orig/kernel/sched_fair.c	2008-10-11 02:54:01.000000000 -0400
+++ linux-compile.git/kernel/sched_fair.c	2008-10-11 02:55:52.000000000 -0400
@@ -889,7 +889,7 @@ static void hrtick_start_fair(struct rq 
 		s64 delta = slice - ran;
 
 		if (delta < 0) {
-			if (rq->curr == p)
+			if (rq->curr == p && !rt_task(p))
 				resched_task(p);
 			return;
 		}

Re: [PATCH] sched: fair scheduler should not resched rt tasks

[Peter Zijlstra]

On Sat, 2008-10-11 at 03:01 -0400, Steven Rostedt wrote:
> Using ftrace, I noticed latencies in real-time tasks where they were
> needlessly calling schedule due to sched_fair sending out time slices.
> 
> This patch prevents a call to resched_task by the sched fair class if
> the task it wants to reschedule is an rt task.

Right, thats not a good thing, however this patch looks wrong, we should
never call hrtick_start_fair() on a rt task to begin with.

The way I can see that happening is through enqueue/dequeue_task_fair()
where we want to re-programm the hrtick because nr_running changes (and
thus the current tasks desired runtime).

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index 5d39d92..6bcceec 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -73,6 +73,8 @@ unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
 
 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
 
+static const struct sched_class fair_sched_class;
+
 /**************************************************************
  * CFS operations on generic schedulable entities:
  */
@@ -849,11 +851,31 @@ static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
 		hrtick_start(rq, delta);
 	}
 }
+
+/*
+ * called from enqueue/dequeue and updates the hrtick when the
+ * current task is from our class and nr_running is low enough
+ * to matter.
+ */
+static void hrtick_update(struct rq *rq)
+{
+	struct task_struct *curr = rq->curr;
+
+	if (curr->sched_class != &fair_sched_class)
+		return;
+
+	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
+		hrtick_start_fair(rq, curr);
+}
 #else /* !CONFIG_SCHED_HRTICK */
 static inline void
 hrtick_start_fair(struct rq *rq, struct task_struct *p)
 {
 }
+
+static inline void hrtick_update(struct rq *rq)
+{
+}
 #endif
 
 /*
@@ -874,7 +896,7 @@ static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
 		wakeup = 1;
 	}
 
-	hrtick_start_fair(rq, rq->curr);
+	hrtick_update(rq);
 }
 
 /*
@@ -896,7 +918,7 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
 		sleep = 1;
 	}
 
-	hrtick_start_fair(rq, rq->curr);
+	hrtick_update(rq);
 }
 
 /*
@@ -1002,8 +1024,6 @@ static inline int wake_idle(int cpu, struct task_struct *p)
 
 #ifdef CONFIG_SMP
 
-static const struct sched_class fair_sched_class;
-
 #ifdef CONFIG_FAIR_GROUP_SCHED
 /*
  * effective_load() calculates the load change as seen from the root_task_group

Re: [PATCH] sched: fair scheduler should not resched rt tasks

[Steven Rostedt]

On Sat, 11 Oct 2008, Peter Zijlstra wrote:

> On Sat, 2008-10-11 at 03:01 -0400, Steven Rostedt wrote:
> > Using ftrace, I noticed latencies in real-time tasks where they were
> > needlessly calling schedule due to sched_fair sending out time slices.
> > 
> > This patch prevents a call to resched_task by the sched fair class if
> > the task it wants to reschedule is an rt task.
> 
> Right, thats not a good thing, however this patch looks wrong, we should
> never call hrtick_start_fair() on a rt task to begin with.

Great! I don't care which way we fix this. This was making noticeable
latencies against rt tasks.

> 
> The way I can see that happening is through enqueue/dequeue_task_fair()
> where we want to re-programm the hrtick because nr_running changes (and
> thus the current tasks desired runtime).

My original patch was more like this one, but I wasn't sure if I was 
skipping any calculations that were assumed to be made. Thus, I wrote the 
bare minimum and sent it out CC'ing the experts. Looks like my plan worked 
;-)

-- Steve


> 
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>

Acked-by: Steven Rostedt <srostedt@redhat.com>