Skip to contentSkip to conversation

sched/deadline: Prevent rt_time growth to infinity

RFC In review Kirill Tkhai 5 participants 14 messages Feb 19, 2014 – Feb 27, 2014 Scheduler

lore
Share & tools
lore
#0 lore raw

RFC 1/1 sched/deadline: Prevent rt_time growth to infinity

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

---
 kernel/sched/deadline.c |  1 +
 kernel/sched/rt.c       | 16 +++++++++++++++-
 kernel/sched/sched.h    |  1 +
 3 files changed, 17 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +1 -0
index ed31ef6..f1d2304 100644
@@ -720,6 +720,7 @@ void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
 
 	inc_dl_deadline(dl_rq, deadline);
 	inc_dl_migration(dl_se, dl_rq);
+	start_top_rt_bandwidth(rq_of_dl_rq(dl_rq));
 }
 
 static inline
kernel/sched/rt.c +15 -1
index 72f9ec7..a187eb8 100644
@@ -726,7 +726,7 @@ static inline int balance_runtime(struct rt_rq *rt_rq)
 
 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 {
-	int i, idle = 1, throttled = 0;
+	int i, idle = 1, throttled = 0, has_dl_tasks = 0;
 	const struct cpumask *span;
 
 	span = sched_rt_period_mask();
@@ -781,9 +781,15 @@ static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 
 		if (enqueue)
 			sched_rt_rq_enqueue(rt_rq);
+
+		if (rt_rq == &rq->rt && rq->dl.dl_nr_running)
+			has_dl_tasks = 1;
 		raw_spin_unlock(&rq->lock);
 	}
 
+	if (has_dl_tasks)
+		return 0;
+
 	if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
 		return 1;
 
@@ -1005,6 +1011,10 @@ dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 	WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 }
 
+void start_top_rt_bandwidth(struct rq *rq)
+{
+	start_rt_bandwidth(&rq->rt.tg->rt_bandwidth);
+}
 #else /* CONFIG_RT_GROUP_SCHED */
 
 static void
@@ -1016,6 +1026,10 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 static inline
 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 
+void start_top_rt_bandwidth(struct rq *rq)
+{
+	start_rt_bandwidth(&def_rt_bandwidth);
+}
 #endif /* CONFIG_RT_GROUP_SCHED */
 
 static inline
kernel/sched/sched.h +1 -0
index 1bf34c2..a20f8f7 100644
@@ -1294,6 +1294,7 @@ static inline void sched_avg_update(struct rq *rq) { }
 #endif
 
 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
+extern void start_top_rt_bandwidth(struct rq *rq);
 
 #ifdef CONFIG_SMP
 #ifdef CONFIG_PREEMPT
#1 lore raw

Peter, Juri,

Have you seen this?

-- Steve

Quoted 91 lines · On Thu, 20 Feb 2014 02:16:00 +0400 Kirill Tkhai <[email protected]> wrote:

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

---
 kernel/sched/deadline.c |  1 +
 kernel/sched/rt.c       | 16 +++++++++++++++-
 kernel/sched/sched.h    |  1 +
 3 files changed, 17 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +1 -0
index ed31ef6..f1d2304 100644
@@ -720,6 +720,7 @@ void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
 
 	inc_dl_deadline(dl_rq, deadline);
 	inc_dl_migration(dl_se, dl_rq);
+	start_top_rt_bandwidth(rq_of_dl_rq(dl_rq));
 }
 
 static inline
kernel/sched/rt.c +15 -1
index 72f9ec7..a187eb8 100644
@@ -726,7 +726,7 @@ static inline int balance_runtime(struct rt_rq *rt_rq)
 
 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 {
-	int i, idle = 1, throttled = 0;
+	int i, idle = 1, throttled = 0, has_dl_tasks = 0;
 	const struct cpumask *span;
 
 	span = sched_rt_period_mask();
@@ -781,9 +781,15 @@ static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 
 		if (enqueue)
 			sched_rt_rq_enqueue(rt_rq);
+
+		if (rt_rq == &rq->rt && rq->dl.dl_nr_running)
+			has_dl_tasks = 1;
 		raw_spin_unlock(&rq->lock);
 	}
 
+	if (has_dl_tasks)
+		return 0;
+
 	if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
 		return 1;
 
@@ -1005,6 +1011,10 @@ dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 	WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 }
 
+void start_top_rt_bandwidth(struct rq *rq)
+{
+	start_rt_bandwidth(&rq->rt.tg->rt_bandwidth);
+}
 #else /* CONFIG_RT_GROUP_SCHED */
 
 static void
@@ -1016,6 +1026,10 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 static inline
 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 
+void start_top_rt_bandwidth(struct rq *rq)
+{
+	start_rt_bandwidth(&def_rt_bandwidth);
+}
 #endif /* CONFIG_RT_GROUP_SCHED */
 
 static inline
kernel/sched/sched.h +1 -0
index 1bf34c2..a20f8f7 100644
@@ -1294,6 +1294,7 @@ static inline void sched_avg_update(struct rq *rq) { }
 #endif
 
 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period);
+extern void start_top_rt_bandwidth(struct rq *rq);
 
 #ifdef CONFIG_SMP
 #ifdef CONFIG_PREEMPT
maintainer
#2 lore raw
In reply to Kirill Tkhai
Quoted 10 lines · Kirill Tkhai

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

How about we only increment rt_time when there's an RT bandwidth timer active?

---

kernel/sched/rt.c +6 -0
@@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+	return hrtimer_active(&rt_b->rt_period_timer);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.
kernel/sched/deadline.c +6 -2
@@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_active(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }
#3 lore raw

21.02.2014, 14:37, "Peter Zijlstra" <[email protected]>:

Quoted 7 lines · Kirill Tkhai
Quoted 6 lines

 Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

Quoted 8 lines
Quoted 5 lines

 I'm going to decide the problem the way below. Almost untested
 because of I skipped almost all of recent patches which haveto be applied from lkml.

 Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

How about we only increment rt_time when there's an RT bandwidth timer active?

This case RT and DL may eat all the time:

-------------- time ------------------>
|RT's working |DL's working|
----------------------------
|rt_runtime | |
----------------------------
| rt_period |

Or at least more, than it's allowed.

It looks like, if we want to limit time of high priority classes execution, we have to set the timer anyway.

kernel/sched/rt.c static inline struct rt_bandwidth *sched

Quoted 4 lines

---

kernel/sched/rt.c +0 -0
@@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched

[snipped by dumb mail client]

#4 lore raw

21.02.2014, 15:39, "Kirill Tkhai" <[email protected]>:

Quoted 25 lines

21.02.2014, 14:37, "Peter Zijlstra" <[email protected]>:

Quoted 6 lines · Kirill Tkhai
Quoted 5 lines

  Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

Quoted 7 lines
Quoted 5 lines

  I'm going to decide the problem the way below. Almost untested
  because of I skipped almost all of recent patches which haveto be applied from lkml.

  Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

 How about we only increment rt_time when there's an RT bandwidth timer active?

This case RT and DL may eat all the time:

--------------  time ------------------>
|RT's working |DL's working|
----------------------------
|rt_runtime   |            |
----------------------------
|        rt_period         |

Quoted 4 lines

Or at least more, than it's allowed.

It looks like, if we want to limit time of high priority classes execution, we have to set the timer anyway.

Oh, above is confusing. Sorry.

I mean one RT task and

----------------- time -------------------------->
|DL's working |RT's working| |DL's working |
--------------------------------------------------
| |rt_runtime | | |
--------------------------------------------------
| | rt_period | |

In this case FAIR receives less ratio, than (rt_period-rt_runtime)/rt_period.

Quoted 11 lines
Quoted 4 lines

 ---
 --- a/kernel/sched/rt.c
 +++ b/kernel/sched/rt.c
 @@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched

[snipped by dumb mail client]
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

#5 lore raw
Quoted 37 lines · On Fri, 21 Feb 2014 16:09:25 +0400 Kirill Tkhai <[email protected]> wrote:

21.02.2014, 15:39, "Kirill Tkhai" <[email protected]>:

Quoted 25 lines

21.02.2014, 14:37, "Peter Zijlstra" <[email protected]>:

Quoted 6 lines · Kirill Tkhai
Quoted 5 lines

  Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

Quoted 7 lines
Quoted 5 lines

  I'm going to decide the problem the way below. Almost untested
  because of I skipped almost all of recent patches which haveto be applied from lkml.

  Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

 How about we only increment rt_time when there's an RT bandwidth timer active?

This case RT and DL may eat all the time:

--------------  time ------------------>
|RT's working |DL's working|
----------------------------
|rt_runtime   |            |
----------------------------
|        rt_period         |

Quoted 4 lines

Or at least more, than it's allowed.

It looks like, if we want to limit time of high priority classes execution, we have to set the timer anyway.

Oh, above is confusing. Sorry.

I mean one RT task and

Quoted 9 lines

----------------- time -------------------------->
|DL's working |RT's working| |DL's working |
--------------------------------------------------
| |rt_runtime | | |
--------------------------------------------------
| | rt_period | |

In this case FAIR receives less ratio, than (rt_period-rt_runtime)/rt_period.

DL tasks won't be allowed to run in this situation, as their bw exceedes rt_runtime/rt_period.

Thanks,

- Juri

#6 lore raw

21.02.2014, 16:44, "Juri Lelli" <[email protected]>:

Quoted 42 lines
Quoted 39 lines · On Fri, 21 Feb 2014 16:09:25 +0400 Kirill Tkhai <[email protected]> wrote:

 21.02.2014, 15:39, "Kirill Tkhai" <[email protected]>:

Quoted 23 lines

 21.02.2014, 14:37, "Peter Zijlstra" <[email protected]>:

Quoted 6 lines · Kirill Tkhai
Quoted 5 lines

   Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

Quoted 7 lines
Quoted 5 lines

   I'm going to decide the problem the way below. Almost untested
   because of I skipped almost all of recent patches which haveto be applied from lkml.

   Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

  How about we only increment rt_time when there's an RT bandwidth timer active?

 This case RT and DL may eat all the time:

 --------------  time ------------------>
 |RT's working |DL's working|
 ----------------------------
 |rt_runtime   |            |
 ----------------------------
 |        rt_period         |

Quoted 4 lines

 Or at least more, than it's allowed.

 It looks like, if we want to limit time of high priority classes execution, we have to set the timer anyway.

 Oh, above is confusing. Sorry.

 I mean one RT task and

 -----------------  time --------------------------> |DL's working    |RT's working| |DL's working    |
 --------------------------------------------------
 |                |rt_runtime  | |                |
 --------------------------------------------------
 |                |   rt_period  |                |

Quoted 4 lines
Quoted 1 line

 In this case FAIR receives less ratio, than (rt_period-rt_runtime)/rt_period.

DL tasks won't be allowed to run in this situation, as their bw exceedes rt_runtime/rt_period.

Maybe, I don't uderstand. Where does DL control summary (DL+RT) runtime?

RT does not do this too. But it looks like, RT has to do this.

Kirill

#7 lore raw
In reply to Peter Zijlstra
Quoted 49 lines · On Fri, 21 Feb 2014 11:37:15 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 10 lines · Kirill Tkhai

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

How about we only increment rt_time when there's an RT bandwidth timer active?

---

kernel/sched/rt.c +6 -0
@@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+	return hrtimer_active(&rt_b->rt_period_timer);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.
kernel/sched/deadline.c +2 -1
@@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*

kernel/sched/deadline.c static void update_curr_dl(struct rq *rq · line 664

Quoted 10 lines
kernel/sched/deadline.c +4 -1
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_active(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }

So, I ran some tests with the above and I'd like to share with you what I've found. You can find here a trace-cmd trace that should be feeded to kernelshark to be able to understand what follows (or feel free to reproduce same scenario :)):
http://retis.sssup.it/~jlelli/traces/trace_rt_time.dat

Here you have a DL task (4/10) and a while(1) RT task, both running inside a rt_bw of 0.5. RT tasks is activated 500ms after DL. As I filtered in sched_rt_period_timer(), you can search for time instants when the rt_bw is replenished. It is evident that the first time after rt timer is activated back (search for start_bandwidth_timer), we can eat some bw to FAIR tasks (if any). This is due to the fact that we reset rt_bw budget at this time, start decrementing rt_time for both DL and RT tasks, throttle RT tasks when rt_time > runtime, but, since DL tasks acually executes inside their own server, they don't care about rt_bw. Good news is that steady state is ok: keeping track of overruns we are able to stop eating bw to other guys.

My thougths:

- Peter's patch is an easy fix to Kirill's problem (RT tasks were throttled too early);
- something to add to this solution could be to pre-calculate bw of ready DL tasks and subtract it to rt_bw at replenishment time, but it sounds quite awkward, pessimistic, and I'm not sure it is gonna work;
- we are stealing bw to best-effort tasks, and just at the beginning of the transistion, is it really a problem?
- I mean, if you want guarantees make your tasks DL! :);
- in the long run we are gonna have RT tasks scheduled inside CBS servers, and all this will be properly fixed up.

Comments?

BTW, rt timer activation/deactivation should probably be fixed for !RT_GROUP_SCHED with something like this:

---
 kernel/sched/rt.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
kernel/sched/rt.c +7 -3
index 6161de8..274f992 100644
@@ -86,12 +86,12 @@ void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
 	raw_spin_lock_init(&rt_rq->rt_runtime_lock);
 }
 
-#ifdef CONFIG_RT_GROUP_SCHED
 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 {
 	hrtimer_cancel(&rt_b->rt_period_timer);
 }
 
+#ifdef CONFIG_RT_GROUP_SCHED
 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
 
 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
@@ -1017,8 +1017,12 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 	start_rt_bandwidth(&def_rt_bandwidth);
 }
 
-static inline
-void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
+static void
+dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
+{
+	if (!rt_rq->rt_nr_running)
+		destroy_rt_bandwidth(&def_rt_bandwidth);
+}
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
Signature · 1 line
1.7.9.5
#8 lore raw
Quoted 69 lines · On Fri, 21 Feb 2014 17:36:41 +0100 Juri Lelli <[email protected]> wrote:
Quoted 49 lines · On Fri, 21 Feb 2014 11:37:15 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 10 lines · Kirill Tkhai

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

How about we only increment rt_time when there's an RT bandwidth timer active?

---

kernel/sched/rt.c +6 -0
@@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+	return hrtimer_active(&rt_b->rt_period_timer);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.
kernel/sched/deadline.c +2 -1
@@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*

kernel/sched/deadline.c static void update_curr_dl(struct rq *rq · line 664

Quoted 10 lines
kernel/sched/deadline.c +4 -1
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_active(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }

So, I ran some tests with the above and I'd like to share with you what I've found. You can find here a trace-cmd trace that should be feeded to kernelshark to be able to understand what follows (or feel free to reproduce same scenario :)):
http://retis.sssup.it/~jlelli/traces/trace_rt_time.dat

Quoted 7 lines

Here you have a DL task (4/10) and a while(1) RT task, both running inside a rt_bw of 0.5. RT tasks is activated 500ms after DL. As I filtered in sched_rt_period_timer(), you can search for time instants when the rt_bw is replenished. It is evident that the first time after rt timer is activated back (search for start_bandwidth_timer), we can eat some bw to FAIR tasks (if any). This is due to the fact that we reset rt_bw budget at this time, start decrementing rt_time for both DL

The reset happens when rt_bw replenishment timer fires, after a bit:

sched_rt_period_timer <-- __run_hrtimer

Apologies,

- Juri

Quoted 61 lines

and RT tasks, throttle RT tasks when rt_time > runtime, but, since DL tasks acually executes inside their own server, they don't care about rt_bw. Good news is that steady state is ok: keeping track of overruns we are able to stop eating bw to other guys.

My thougths:

- Peter's patch is an easy fix to Kirill's problem (RT tasks were throttled too early);
- something to add to this solution could be to pre-calculate bw of ready DL tasks and subtract it to rt_bw at replenishment time, but it sounds quite awkward, pessimistic, and I'm not sure it is gonna work;
- we are stealing bw to best-effort tasks, and just at the beginning of the transistion, is it really a problem?
- I mean, if you want guarantees make your tasks DL! :);
- in the long run we are gonna have RT tasks scheduled inside CBS servers, and all this will be properly fixed up.

Comments?

BTW, rt timer activation/deactivation should probably be fixed for !RT_GROUP_SCHED with something like this:

---
 kernel/sched/rt.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
kernel/sched/rt.c +7 -3
index 6161de8..274f992 100644
@@ -86,12 +86,12 @@ void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
 	raw_spin_lock_init(&rt_rq->rt_runtime_lock);
 }
 
-#ifdef CONFIG_RT_GROUP_SCHED
 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 {
 	hrtimer_cancel(&rt_b->rt_period_timer);
 }
 
+#ifdef CONFIG_RT_GROUP_SCHED
 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
 
 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
@@ -1017,8 +1017,12 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 	start_rt_bandwidth(&def_rt_bandwidth);
 }
 
-static inline
-void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
+static void
+dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
+{
+	if (!rt_rq->rt_nr_running)
+		destroy_rt_bandwidth(&def_rt_bandwidth);
+}
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
#9 lore raw

21.02.2014, 20:52, "Juri Lelli" <[email protected]>:

Quoted 76 lines
Quoted 65 lines · On Fri, 21 Feb 2014 17:36:41 +0100 Juri Lelli <[email protected]> wrote:
Quoted 44 lines · On Fri, 21 Feb 2014 11:37:15 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 10 lines · Kirill Tkhai

 Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

 I'm going to decide the problem the way below. Almost untested
 because of I skipped almost all of recent patches which haveto be applied from lkml.

 Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

 How about we only increment rt_time when there's an RT bandwidth timer active?

 ---
 --- a/kernel/sched/rt.c
 +++ b/kernel/sched/rt.c
 @@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched

  #endif /* CONFIG_RT_GROUP_SCHED */

 +bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
 +{
 + struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 + return hrtimer_active(&rt_b->rt_period_timer);
 +}
 +
  #ifdef CONFIG_SMP
  /*
   * We ran out of runtime, see if we can borrow some from our neighbours. --- a/kernel/sched/deadline.c
 +++ b/kernel/sched/deadline.c
 @@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s return 1;
  }

 +extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq); +
  /*
   * Update the current task's runtime statistics (provided it is still
   * a -deadline task and has not been removed from the dl_rq). @@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq struct rt_rq *rt_rq = &rq->rt;

Quoted 13 lines

                  raw_spin_lock(&rt_rq->rt_runtime_lock);
 - rt_rq->rt_time += delta_exec;
                  /*
                   * We'll let actual RT tasks worry about the overflow here, we
 - * have our own CBS to keep us inline -- see above.
 + * have our own CBS to keep us inline; only account when RT
 + * bandwidth is relevant.
                   */
 + if (sched_rt_bandwidth_active(rt_rq))
 + rt_rq->rt_time += delta_exec; raw_spin_unlock(&rt_rq->rt_runtime_lock);
          }
  }

 So, I ran some tests with the above and I'd like to share with you what I've found. You can find here a trace-cmd trace that should be feeded to kernelshark to be able to understand what follows (or feel free to reproduce same scenario :)):
 http://retis.sssup.it/~jlelli/traces/trace_rt_time.dat

Quoted 7 lines

 Here you have a DL task (4/10) and a while(1) RT task, both running inside a rt_bw of 0.5. RT tasks is activated 500ms after DL. As I filtered in sched_rt_period_timer(), you can search for time instants when the rt_bw is replenished. It is evident that the first time after rt timer is activated back (search for start_bandwidth_timer), we can eat some bw to FAIR tasks (if any). This is due to the fact that we reset rt_bw budget at this time, start decrementing rt_time for both DL

Quoted 3 lines

The reset happens when rt_bw replenishment timer fires, after a bit:

 sched_rt_period_timer <-- __run_hrtimer

Juri, sorry, I forgot to wrote I mean the situation when only one task is on_rq at every moment.

DL, RT, DL, RT, ...

rt_runtime = n;
rt_period = 2n;

| DL's working, RT's sleeping  | RT's working, DL's sleeping  |   all sleep               |
------------------------------------------------------------------------------------------|
| (1)     duration = n         | (2)     duration = n         | (3)     duration = n      |  (repeat)
|------------------------------|------------------------------|---------------------------|
| (rt_bw timer is not running) |                   (rt_bw timer is running)               |

According to the patch, rt_bw timer is working only if we have queued RT task.

In the case above part (1) has no queued RT tasks, so timer is not working. rt_time is not being increased too.

We have ratio 2/3.

Thanks,
Kirill

Quoted 65 lines

Apologies,

- Juri

Quoted 60 lines

 and RT tasks, throttle RT tasks when rt_time > runtime, but, since DL tasks acually executes inside their own server, they don't care about rt_bw. Good news is that steady state is ok: keeping track of overruns we are able to stop eating bw to other guys.

 My thougths:

  - Peter's patch is an easy fix to Kirill's problem (RT tasks were throttled too early);
  - something to add to this solution could be to pre-calculate bw of ready DL tasks and subtract it to rt_bw at replenishment time, but it sounds quite awkward, pessimistic, and I'm not sure it is gonna work;
  - we are stealing bw to best-effort tasks, and just at the beginning of the transistion, is it really a problem?
  - I mean, if you want guarantees make your tasks DL! :);
  - in the long run we are gonna have RT tasks scheduled inside CBS servers, and all this will be properly fixed up.

 Comments?

 BTW, rt timer activation/deactivation should probably be fixed for !RT_GROUP_SCHED with something like this:

 ---
  kernel/sched/rt.c |   10 +++++++---

  1 file changed, 7 insertions(+), 3 deletions(-)

 diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
 index 6161de8..274f992 100644
 --- a/kernel/sched/rt.c
 +++ b/kernel/sched/rt.c
 @@ -86,12 +86,12 @@ void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) raw_spin_lock_init(&rt_rq->rt_runtime_lock);
  }

 -#ifdef CONFIG_RT_GROUP_SCHED
  static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) {
          hrtimer_cancel(&rt_b->rt_period_timer);
  }

 +#ifdef CONFIG_RT_GROUP_SCHED
  #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)

  static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) @@ -1017,8 +1017,12 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) start_rt_bandwidth(&def_rt_bandwidth);
  }

 -static inline
 -void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} +static void
 +dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) +{
 + if (!rt_rq->rt_nr_running)
 + destroy_rt_bandwidth(&def_rt_bandwidth);
 +}

  #endif /* CONFIG_RT_GROUP_SCHED */

#10 lore raw
In reply to Juri Lelli
Quoted 127 lines · Juri Lelli
Quoted 49 lines · On Fri, 21 Feb 2014 11:37:15 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 10 lines · Kirill Tkhai

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

How about we only increment rt_time when there's an RT bandwidth timer active?

---

kernel/sched/rt.c +6 -0
@@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+	return hrtimer_active(&rt_b->rt_period_timer);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.
kernel/sched/deadline.c +2 -1
@@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*

kernel/sched/deadline.c static void update_curr_dl(struct rq *rq · line 664

Quoted 10 lines
kernel/sched/deadline.c +4 -1
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_active(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }

So, I ran some tests with the above and I'd like to share with you what I've found. You can find here a trace-cmd trace that should be feeded to kernelshark to be able to understand what follows (or feel free to reproduce same scenario :)):
http://retis.sssup.it/~jlelli/traces/trace_rt_time.dat

Here you have a DL task (4/10) and a while(1) RT task, both running inside a rt_bw of 0.5. RT tasks is activated 500ms after DL. As I filtered in sched_rt_period_timer(), you can search for time instants when the rt_bw is replenished. It is evident that the first time after rt timer is activated back (search for start_bandwidth_timer), we can eat some bw to FAIR tasks (if any). This is due to the fact that we reset rt_bw budget at this time, start decrementing rt_time for both DL and RT tasks, throttle RT tasks when rt_time > runtime, but, since DL tasks acually executes inside their own server, they don't care about rt_bw. Good news is that steady state is ok: keeping track of overruns we are able to stop eating bw to other guys.

My thougths:

- Peter's patch is an easy fix to Kirill's problem (RT tasks were throttled too early);
- something to add to this solution could be to pre-calculate bw of ready DL tasks and subtract it to rt_bw at replenishment time, but it sounds quite awkward, pessimistic, and I'm not sure it is gonna work;
- we are stealing bw to best-effort tasks, and just at the beginning of the transistion, is it really a problem?
- I mean, if you want guarantees make your tasks DL! :);
- in the long run we are gonna have RT tasks scheduled inside CBS servers, and all this will be properly fixed up.

Comments?

BTW, rt timer activation/deactivation should probably be fixed for !RT_GROUP_SCHED with something like this:

---
 kernel/sched/rt.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
kernel/sched/rt.c +1 -2
index 6161de8..274f992 100644
@@ -86,12 +86,12 @@ void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
 	raw_spin_lock_init(&rt_rq->rt_runtime_lock);
 }
 
-#ifdef CONFIG_RT_GROUP_SCHED
 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 {
 	hrtimer_cancel(&rt_b->rt_period_timer);
 }
 
+#ifdef CONFIG_RT_GROUP_SCHED
 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
 
 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
@@ -1017,8 +1017,12 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 	start_rt_bandwidth(&def_rt_bandwidth);
 }
 
-static inline

kernel/sched/rt.c inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) · line 1028

Quoted 10 lines
kernel/sched/rt.c +6 -1
-void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
+static void
+dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
+{
+	if (!rt_rq->rt_nr_running)
+		destroy_rt_bandwidth(&def_rt_bandwidth);
+}
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 

It looks with both patches applied, we may get into a situation, when all CPU time is shared between RT and DL tasks:

rt_runtime = n
rt_period = 2n

| RT working, DL sleeping | DL working, RT sleeping |
-----------------------------------------------------------

| (1)     duration = n     | (2)     duration = n         | (repeat)
|--------------------------|------------------------------|
| (rt_bw timer is running) | (rt_bw timer is not running) |

No time for fair tasks at all.

Kirill

#11 lore raw
Quoted 145 lines · On Sat, 22 Feb 2014 04:56:59 +0400 Kirill Tkhai <[email protected]> wrote:
Quoted 127 lines · Juri Lelli
Quoted 49 lines · On Fri, 21 Feb 2014 11:37:15 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 10 lines · Kirill Tkhai

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

I'm going to decide the problem the way below. Almost untested
because of I skipped almost all of recent patches which haveto be applied from lkml.

Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

How about we only increment rt_time when there's an RT bandwidth timer active?

---

kernel/sched/rt.c +6 -0
@@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+	return hrtimer_active(&rt_b->rt_period_timer);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.
kernel/sched/deadline.c +2 -1
@@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*

kernel/sched/deadline.c static void update_curr_dl(struct rq *rq · line 664

Quoted 10 lines
kernel/sched/deadline.c +4 -1
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_active(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }

So, I ran some tests with the above and I'd like to share with you what I've found. You can find here a trace-cmd trace that should be feeded to kernelshark to be able to understand what follows (or feel free to reproduce same scenario :)):
http://retis.sssup.it/~jlelli/traces/trace_rt_time.dat

Here you have a DL task (4/10) and a while(1) RT task, both running inside a rt_bw of 0.5. RT tasks is activated 500ms after DL. As I filtered in sched_rt_period_timer(), you can search for time instants when the rt_bw is replenished. It is evident that the first time after rt timer is activated back (search for start_bandwidth_timer), we can eat some bw to FAIR tasks (if any). This is due to the fact that we reset rt_bw budget at this time, start decrementing rt_time for both DL and RT tasks, throttle RT tasks when rt_time > runtime, but, since DL tasks acually executes inside their own server, they don't care about rt_bw. Good news is that steady state is ok: keeping track of overruns we are able to stop eating bw to other guys.

My thougths:

- Peter's patch is an easy fix to Kirill's problem (RT tasks were throttled too early);
- something to add to this solution could be to pre-calculate bw of ready DL tasks and subtract it to rt_bw at replenishment time, but it sounds quite awkward, pessimistic, and I'm not sure it is gonna work;
- we are stealing bw to best-effort tasks, and just at the beginning of the transistion, is it really a problem?
- I mean, if you want guarantees make your tasks DL! :);
- in the long run we are gonna have RT tasks scheduled inside CBS servers, and all this will be properly fixed up.

Comments?

BTW, rt timer activation/deactivation should probably be fixed for !RT_GROUP_SCHED with something like this:

---
 kernel/sched/rt.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
kernel/sched/rt.c +1 -2
index 6161de8..274f992 100644
@@ -86,12 +86,12 @@ void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
 	raw_spin_lock_init(&rt_rq->rt_runtime_lock);
 }
 
-#ifdef CONFIG_RT_GROUP_SCHED
 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 {
 	hrtimer_cancel(&rt_b->rt_period_timer);
 }
 
+#ifdef CONFIG_RT_GROUP_SCHED
 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
 
 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
@@ -1017,8 +1017,12 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 	start_rt_bandwidth(&def_rt_bandwidth);
 }
 
-static inline

kernel/sched/rt.c inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) · line 1028

Quoted 10 lines
kernel/sched/rt.c +6 -1
-void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
+static void
+dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
+{
+	if (!rt_rq->rt_nr_running)
+		destroy_rt_bandwidth(&def_rt_bandwidth);
+}
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 

It looks with both patches applied, we may get into a situation, when all CPU time is shared between RT and DL tasks:

rt_runtime = n
rt_period = 2n

Quoted 7 lines

| RT working, DL sleeping | DL working, RT sleeping |
-----------------------------------------------------------

| (1)     duration = n     | (2)     duration = n         | (repeat)
|--------------------------|------------------------------|
| (rt_bw timer is running) | (rt_bw timer is not running) |

No time for fair tasks at all.

Ok, this situation is pathological. DL bandwidth is guaranteed at admission control, while RT isn't. In this case RT tasks are doomed by construction. Still you'd like to let FAIR tasks execute :).

I argumented on a slightly different solution in what follows, what you think?

Thanks,

- Juri

Quoted 1 line

From e44fe2eef34433a7799cfc153f467f7c62813596 Mon Sep 17 00:00:00 2001

From: Juri Lelli <[email protected]>
Date: Fri, 21 Feb 2014 11:37:15 +0100
Subject: [PATCH] sched/deadline: Prevent rt_time growth to infinity

Kirill Tkhai noted:
Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

RT task were in fact throttled right after they got enqueued, and never executed again (rt_time never again went below rt_runtime).

Peter than proposed to accrue DL execution on rt_time only when rt timer is active, and proposed a patch (this patch is a slight modification of that) to implement that behavior. While this solves Kirill problem, it has a drawback.

Indeed, Kirill noted again:
It looks we may get into a situation, when all CPU time is shared between RT and DL tasks:

rt_runtime = n
rt_period = 2n

| RT working, DL sleeping | DL working, RT sleeping |
-----------------------------------------------------------

| (1)     duration = n     | (2)     duration = n         | (repeat)
|--------------------------|------------------------------|
| (rt_bw timer is running) | (rt_bw timer is not running) |

No time for fair tasks at all.

While this can happen during the first period, if rq is always backlogged, RT tasks won't have the opportunity to execute anymore: rt_time reached rt_runtime during (1), suppose after (2) RT is enqueued back, it gets throttled since rt timer didn't fire, replenishment is from now on eaten up by DL tasks that accrue their execution on rt_time (while rt timer is active - we have an RT task waiting for replenishment). FAIR tasks are not touched after this first period. Ok, this is not ideal, and the situation is even worse!

What above (the nice case), practically never happens in reality, where your rt timer is not aligned to tasks periods, tasks are in general not periodic, etc.. Long story short, you always risk to overload your system.

This patch is based on Peter's idea, but exploits an additional fact:
if you don't have RT tasks enqueued, it makes little sense to continue incrementing rt_time once you reached the upper limit (DL tasks have their own mechanism for throttling).

This cures both problems:

- no matter how many DL instances in the past, you'll have an rt_time slightly above rt_runtime when an RT task is enqueued, and from that point on (after the first replenishment), the task will normally execute;

- you can still eat up all bandwidth during the first period, but not anymore after that, remember that DL execution will increment rt_time till the upper limit is reached.

The situation is still not perfect! But, we have a simple solution for now, that limits how much you can jeopardize your system, as we keep working towards the right answer: RT groups scheduled using deadline servers.

---
 kernel/sched/deadline.c |    8 ++++++--
 kernel/sched/rt.c       |    8 ++++++++
 2 files changed, 14 insertions(+), 2 deletions(-)
kernel/sched/deadline.c +6 -2
index 15cbc17..f59d774 100644
@@ -564,6 +564,8 @@ int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -627,11 +629,13 @@ static void update_curr_dl(struct rq *rq)
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_account(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }
kernel/sched/rt.c +8 -0
index 7dba25a..7f372e1 100644
@@ -538,6 +538,14 @@ static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+
+	return (hrtimer_active(&rt_b->rt_period_timer) ||
+		rt_rq->rt_time < rt_b->rt_runtime);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.
Signature · 1 line
1.7.9.5
#12 lore raw

25.02.2014, 18:14, "Juri Lelli" <[email protected]>:

Quoted 219 lines
Quoted 139 lines · On Sat, 22 Feb 2014 04:56:59 +0400 Kirill Tkhai <[email protected]> wrote:
Quoted 131 lines · Juri Lelli
Quoted 44 lines · On Fri, 21 Feb 2014 11:37:15 +0100 Peter Zijlstra <[email protected]> wrote:
Quoted 10 lines · Kirill Tkhai

 Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

 I'm going to decide the problem the way below. Almost untested
 because of I skipped almost all of recent patches which haveto be applied from lkml.

 Please say, if I skipped anything in idea. Maybe better put start_top_rt_bandwidth() into set_curr_task_dl()?

 How about we only increment rt_time when there's an RT bandwidth timer active?

 ---
 --- a/kernel/sched/rt.c
 +++ b/kernel/sched/rt.c
 @@ -568,6 +568,12 @@ static inline struct rt_bandwidth *sched

  #endif /* CONFIG_RT_GROUP_SCHED */

 +bool sched_rt_bandwidth_active(struct rt_rq *rt_rq)
 +{
 + struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 + return hrtimer_active(&rt_b->rt_period_timer);
 +}
 +
  #ifdef CONFIG_SMP
  /*
   * We ran out of runtime, see if we can borrow some from our neighbours. --- a/kernel/sched/deadline.c
 +++ b/kernel/sched/deadline.c
 @@ -587,6 +587,8 @@ int dl_runtime_exceeded(struct rq *rq, s return 1;
  }

 +extern bool sched_rt_bandwidth_active(struct rt_rq *rt_rq); +
  /*
   * Update the current task's runtime statistics (provided it is still
   * a -deadline task and has not been removed from the dl_rq). @@ -650,11 +652,13 @@ static void update_curr_dl(struct rq *rq struct rt_rq *rt_rq = &rq->rt;

Quoted 13 lines

                  raw_spin_lock(&rt_rq->rt_runtime_lock);
 - rt_rq->rt_time += delta_exec;
                  /*
                   * We'll let actual RT tasks worry about the overflow here, we
 - * have our own CBS to keep us inline -- see above.
 + * have our own CBS to keep us inline; only account when RT
 + * bandwidth is relevant.
                   */
 + if (sched_rt_bandwidth_active(rt_rq))
 + rt_rq->rt_time += delta_exec; raw_spin_unlock(&rt_rq->rt_runtime_lock);
          }
  }

 So, I ran some tests with the above and I'd like to share with you what I've found. You can find here a trace-cmd trace that should be feeded to kernelshark to be able to understand what follows (or feel free to reproduce same scenario :)):
 http://retis.sssup.it/~jlelli/traces/trace_rt_time.dat

 Here you have a DL task (4/10) and a while(1) RT task, both running inside a rt_bw of 0.5. RT tasks is activated 500ms after DL. As I filtered in sched_rt_period_timer(), you can search for time instants when the rt_bw is replenished. It is evident that the first time after rt timer is activated back (search for start_bandwidth_timer), we can eat some bw to FAIR tasks (if any). This is due to the fact that we reset rt_bw budget at this time, start decrementing rt_time for both DL and RT tasks, throttle RT tasks when rt_time > runtime, but, since DL tasks acually executes inside their own server, they don't care about rt_bw. Good news is that steady state is ok: keeping track of overruns we are able to stop eating bw to other guys.

 My thougths:

  - Peter's patch is an easy fix to Kirill's problem (RT tasks were throttled too early);
  - something to add to this solution could be to pre-calculate bw of ready DL tasks and subtract it to rt_bw at replenishment time, but it sounds quite awkward, pessimistic, and I'm not sure it is gonna work;
  - we are stealing bw to best-effort tasks, and just at the beginning of the transistion, is it really a problem?
  - I mean, if you want guarantees make your tasks DL! :);
  - in the long run we are gonna have RT tasks scheduled inside CBS servers, and all this will be properly fixed up.

 Comments?

 BTW, rt timer activation/deactivation should probably be fixed for !RT_GROUP_SCHED with something like this:

 ---
  kernel/sched/rt.c |   10 +++++++---

  1 file changed, 7 insertions(+), 3 deletions(-)

 diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
 index 6161de8..274f992 100644
 --- a/kernel/sched/rt.c
 +++ b/kernel/sched/rt.c
 @@ -86,12 +86,12 @@ void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) raw_spin_lock_init(&rt_rq->rt_runtime_lock);
  }

 -#ifdef CONFIG_RT_GROUP_SCHED
  static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) {
          hrtimer_cancel(&rt_b->rt_period_timer);
  }

 +#ifdef CONFIG_RT_GROUP_SCHED
  #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)

  static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) @@ -1017,8 +1017,12 @@ inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) start_rt_bandwidth(&def_rt_bandwidth);
  }

 -static inline
 -void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} +static void
 +dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) +{
 + if (!rt_rq->rt_nr_running)
 + destroy_rt_bandwidth(&def_rt_bandwidth);
 +}

Quoted 1 line

  #endif /* CONFIG_RT_GROUP_SCHED */

 It looks with both patches applied, we may get into a situation, when all CPU time is shared between RT and DL tasks:

 rt_runtime = n
 rt_period  = 2n

Quoted 7 lines

 | RT working, DL sleeping  | DL working, RT sleeping      |
 -----------------------------------------------------------

 | (1)     duration = n     | (2)     duration = n         | (repeat)
 |--------------------------|------------------------------|
 | (rt_bw timer is running) | (rt_bw timer is not running) |

 No time for fair tasks at all.

Ok, this situation is pathological. DL bandwidth is guaranteed at admission control, while RT isn't. In this case RT tasks are doomed by construction. Still you'd like to let FAIR tasks execute :).

I argumented on a slightly different solution in what follows, what you think?

Thanks,

- Juri

From e44fe2eef34433a7799cfc153f467f7c62813596 Mon Sep 17 00:00:00 2001
From: Juri Lelli <[email protected]>
Date: Fri, 21 Feb 2014 11:37:15 +0100
Subject: [PATCH] sched/deadline: Prevent rt_time growth to infinity

Kirill Tkhai noted:
Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

RT task were in fact throttled right after they got enqueued, and never executed again (rt_time never again went below rt_runtime).

Peter than proposed to accrue DL execution on rt_time only when rt timer is active, and proposed a patch (this patch is a slight modification of that) to implement that behavior. While this solves Kirill problem, it has a drawback.

Indeed, Kirill noted again:
It looks we may get into a situation, when all CPU time is shared between RT and DL tasks:

rt_runtime = n
rt_period  = 2n

| RT working, DL sleeping  | DL working, RT sleeping      |
-----------------------------------------------------------

| (1)     duration = n     | (2)     duration = n         | (repeat)
|--------------------------|------------------------------|
| (rt_bw timer is running) | (rt_bw timer is not running) |

No time for fair tasks at all.

While this can happen during the first period, if rq is always backlogged, RT tasks won't have the opportunity to execute anymore: rt_time reached rt_runtime during (1), suppose after (2) RT is enqueued back, it gets throttled since rt timer didn't fire, replenishment is from now on eaten up by DL tasks that accrue their execution on rt_time (while rt timer is active - we have an RT task waiting for replenishment). FAIR tasks are not touched after this first period. Ok, this is not ideal, and the situation is even worse!

What above (the nice case), practically never happens in reality, where your rt timer is not aligned to tasks periods, tasks are in general not periodic, etc.. Long story short, you always risk to overload your system.

This patch is based on Peter's idea, but exploits an additional fact:
if you don't have RT tasks enqueued, it makes little sense to continue incrementing rt_time once you reached the upper limit (DL tasks have their own mechanism for throttling).

This cures both problems:

 - no matter how many DL instances in the past, you'll have an rt_time slightly above rt_runtime when an RT task is enqueued, and from that point on (after the first replenishment), the task will normally execute;

Quoted 7 lines

 - you can still eat up all bandwidth during the first period, but not anymore after that, remember that DL execution will increment rt_time till the upper limit is reached.

The situation is still not perfect! But, we have a simple solution for now, that limits how much you can jeopardize your system, as we keep working towards the right answer: RT groups scheduled using deadline servers.

Excellent, Juri! This is almost perfect.

Thanks,
Kirill

Quoted 56 lines

---
 kernel/sched/deadline.c |    8 ++++++--
 kernel/sched/rt.c       |    8 ++++++++

 2 files changed, 14 insertions(+), 2 deletions(-)
kernel/sched/deadline.c +6 -2
index 15cbc17..f59d774 100644
@@ -564,6 +564,8 @@ int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
          return 1;
  }
 
+extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
+
  /*
   * Update the current task's runtime statistics (provided it is still
   * a -deadline task and has not been removed from the dl_rq).
@@ -627,11 +629,13 @@ static void update_curr_dl(struct rq *rq)
                  struct rt_rq *rt_rq = &rq->rt;
 
                  raw_spin_lock(&rt_rq->rt_runtime_lock);
- rt_rq->rt_time += delta_exec;
                  /*
                   * We'll let actual RT tasks worry about the overflow here, we
- * have our own CBS to keep us inline -- see above.
+ * have our own CBS to keep us inline; only account when RT
+ * bandwidth is relevant.
                   */
+ if (sched_rt_bandwidth_account(rt_rq))
+ rt_rq->rt_time += delta_exec;
                  raw_spin_unlock(&rt_rq->rt_runtime_lock);
          }
  }
kernel/sched/rt.c +8 -1
index 7dba25a..7f372e1 100644
@@ -538,6 +538,14 @@ static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 
  #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
+{
+ struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+
+ return (hrtimer_active(&rt_b->rt_period_timer) ||
+ rt_rq->rt_time < rt_b->rt_runtime);
+}
+
  #ifdef CONFIG_SMP
  /*
   * We ran out of runtime, see if we can borrow some from our neighbours.
--

1.7.9.5

#13 lore raw
In reply to Juri Lelli
sched/deadline: Prevent rt_time growth to infinity
Commit-ID:  faa5993736d9b44b508cab4f1f3a77d66641c6f4
Gitweb:     http://git.kernel.org/tip/faa5993736d9b44b508cab4f1f3a77d66641c6f4
Author:     Juri Lelli <[email protected]>
AuthorDate: Fri, 21 Feb 2014 11:37:15 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 27 Feb 2014 12:29:41 +0100

sched/deadline: Prevent rt_time growth to infinity

Kirill Tkhai noted:

Since deadline tasks share rt bandwidth, we must care about bandwidth timer set. Otherwise rt_time may grow up to infinity in update_curr_dl(), if there are no other available RT tasks on top level bandwidth.

RT task were in fact throttled right after they got enqueued, and never executed again (rt_time never again went below rt_runtime).

Peter then proposed to accrue DL execution on rt_time only when rt timer is active, and proposed a patch (this patch is a slight modification of that) to implement that behavior. While this solves Kirill problem, it has a drawback.

Indeed, Kirill noted again:

It looks we may get into a situation, when all CPU time is shared between RT and DL tasks:

rt_runtime = n
rt_period = 2n

| RT working, DL sleeping | DL working, RT sleeping |
-----------------------------------------------------------

  | (1)     duration = n     | (2)     duration = n         | (repeat)
  |--------------------------|------------------------------|
  | (rt_bw timer is running) | (rt_bw timer is not running) |

No time for fair tasks at all.

While this can happen during the first period, if rq is always backlogged, RT tasks won't have the opportunity to execute anymore: rt_time reached rt_runtime during (1), suppose after (2) RT is enqueued back, it gets throttled since rt timer didn't fire, replenishment is from now on eaten up by DL tasks that accrue their execution on rt_time (while rt timer is active - we have an RT task waiting for replenishment). FAIR tasks are not touched after this first period. Ok, this is not ideal, and the situation is even worse!

What above (the nice case), practically never happens in reality, where your rt timer is not aligned to tasks periods, tasks are in general not periodic, etc.. Long story short, you always risk to overload your system.

This patch is based on Peter's idea, but exploits an additional fact:
if you don't have RT tasks enqueued, it makes little sense to continue incrementing rt_time once you reached the upper limit (DL tasks have their own mechanism for throttling).

This cures both problems:

- no matter how many DL instances in the past, you'll have an rt_time slightly above rt_runtime when an RT task is enqueued, and from that point on (after the first replenishment), the task will normally execute;

- you can still eat up all bandwidth during the first period, but not anymore after that, remember that DL execution will increment rt_time till the upper limit is reached.

The situation is still not perfect! But, we have a simple solution for now, that limits how much you can jeopardize your system, as we keep working towards the right answer: RT groups scheduled using deadline servers.

---
 kernel/sched/deadline.c | 8 ++++++--
 kernel/sched/rt.c       | 8 ++++++++
 2 files changed, 14 insertions(+), 2 deletions(-)
kernel/sched/deadline.c +6 -2
index aecf930..6e79b3f 100644
@@ -562,6 +562,8 @@ int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
 	return 1;
 }
 
+extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
+
 /*
  * Update the current task's runtime statistics (provided it is still
  * a -deadline task and has not been removed from the dl_rq).
@@ -625,11 +627,13 @@ static void update_curr_dl(struct rq *rq)
 		struct rt_rq *rt_rq = &rq->rt;
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_time += delta_exec;
 		/*
 		 * We'll let actual RT tasks worry about the overflow here, we
-		 * have our own CBS to keep us inline -- see above.
+		 * have our own CBS to keep us inline; only account when RT
+		 * bandwidth is relevant.
 		 */
+		if (sched_rt_bandwidth_account(rt_rq))
+			rt_rq->rt_time += delta_exec;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 }
kernel/sched/rt.c +8 -0
index a2740b7..1999021 100644
@@ -538,6 +538,14 @@ static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 
 #endif /* CONFIG_RT_GROUP_SCHED */
 
+bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
+{
+	struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
+
+	return (hrtimer_active(&rt_b->rt_period_timer) ||
+		rt_rq->rt_time < rt_b->rt_runtime);
+}
+
 #ifdef CONFIG_SMP
 /*
  * We ran out of runtime, see if we can borrow some from our neighbours.