sched/deadline: support dl task migration during cpu hotplug

I observe that dl task can't be migrated to other cpus during cpu hotplug, in addition, task may/may not be running again if cpu is added back. The root cause which I found is that dl task will be throtted and removed from dl rq after comsuming all budget, which leads to stop task can't pick it up from dl rq and migrate to other cpus during hotplug.

The method to reproduce:
schedtool -E -t 50000:100000 -e ./test
Actually test is just a simple for loop. Then observe which cpu the test task is on.
echo 0 > /sys/devices/system/cpu/cpuN/online

This patch fix it by push the task to another cpu in dl_task_timer() if rq is offline.

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Peterz, Juri?

Signed-off-by Wanpeng Li <[email protected]>

---
v1 -> v2:
* push the task to another cpu in dl_task_timer() if rq is offline.

 kernel/sched/deadline.c | 39 ++++++++++++++++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 3 deletions(-)

kernel/sched/deadline.c +35 -1

index 04c2cbb..233e482 100644
@@ -487,6 +487,7 @@ static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
 	return hrtimer_active(&dl_se->dl_timer);
 }
 
+static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
 /*
  * This is the bandwidth enforcement timer callback. If here, we know
  * a task is not on its dl_rq, since the fact that the timer was running

@@ -538,6 +539,39 @@ again:
 	update_rq_clock(rq);
 	dl_se->dl_throttled = 0;
 	dl_se->dl_yielded = 0;
+
+	/*
+	 * So if we find that the rq the task was on is no longer
+	 * available, we need to select a new rq.
+	 */
+	if (!rq->online) {
+		struct rq *later_rq = NULL;
+
+		/* We will release rq lock */
+		get_task_struct(p);
+
+		raw_spin_unlock(&rq->lock);
+
+		later_rq = find_lock_later_rq(p, rq);
+
+		if (!later_rq) {
+			put_task_struct(p);
+			goto out;
+		}
+
+		deactivate_task(rq, p, 0);
+		set_task_cpu(p, later_rq->cpu);
+		activate_task(later_rq, p, 0);
+
+		resched_curr(later_rq);
+
+		double_unlock_balance(rq, later_rq);
+
+		put_task_struct(p);
+
+		goto out;
+	}
+
 	if (task_on_rq_queued(p)) {
 		enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
 		if (dl_task(rq->curr))

@@ -555,7 +589,7 @@ again:
 	}
 unlock:
 	raw_spin_unlock(&rq->lock);
-
+out:
 	return HRTIMER_NORESTART;
 }
 

kernel/sched/deadline.c +0 -0

 	cpumask_and(later_mask, later_mask, &task->cpus_allowed);
 	best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
 			task, later_mask);

Hi,

Quoted 90 lines · Wanpeng Li

I observe that dl task can't be migrated to other cpus during cpu hotplug, in addition, task may/may not be running again if cpu is added back. The root cause which I found is that dl task will be throtted and removed from dl rq after comsuming all budget, which leads to stop task can't pick it up from dl rq and migrate to other cpus during hotplug.

The method to reproduce:
schedtool -E -t 50000:100000 -e ./test
Actually test is just a simple for loop. Then observe which cpu the test task is on.
echo 0 > /sys/devices/system/cpu/cpuN/online

This patch fix it by push the task to another cpu in dl_task_timer() if rq is offline.

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Peterz, Juri?

Signed-off-by Wanpeng Li <[email protected]>

---
v1 -> v2:
* push the task to another cpu in dl_task_timer() if rq is offline.

  kernel/sched/deadline.c | 39 ++++++++++++++++++++++++++++++++++++---
  1 file changed, 36 insertions(+), 3 deletions(-)

kernel/sched/deadline.c +35 -1

index 04c2cbb..233e482 100644
@@ -487,6 +487,7 @@ static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
  	return hrtimer_active(&dl_se->dl_timer);
  }
 
+static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
  /*
   * This is the bandwidth enforcement timer callback. If here, we know
   * a task is not on its dl_rq, since the fact that the timer was running

@@ -538,6 +539,39 @@ again:
  	update_rq_clock(rq);
  	dl_se->dl_throttled = 0;
  	dl_se->dl_yielded = 0;
+
+	/*
+	 * So if we find that the rq the task was on is no longer
+	 * available, we need to select a new rq.
+	 */
+	if (!rq->online) {
+		struct rq *later_rq = NULL;
+
+		/* We will release rq lock */
+		get_task_struct(p);
+
+		raw_spin_unlock(&rq->lock);
+
+		later_rq = find_lock_later_rq(p, rq);
+
+		if (!later_rq) {
+			put_task_struct(p);
+			goto out;
+		}
+
+		deactivate_task(rq, p, 0);
+		set_task_cpu(p, later_rq->cpu);
+		activate_task(later_rq, p, 0);
+
+		resched_curr(later_rq);
+
+		double_unlock_balance(rq, later_rq);
+
+		put_task_struct(p);
+
+		goto out;
+	}
+
  	if (task_on_rq_queued(p)) {
  		enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
  		if (dl_task(rq->curr))

@@ -555,7 +589,7 @@ again:
  	}
  unlock:
  	raw_spin_unlock(&rq->lock);
-
+out:
  	return HRTIMER_NORESTART;
  }
 

kernel/sched/deadline.c static int find_later_rq(struct task_struct *task) · line 1219

Quoted 7 lines

kernel/sched/deadline.c +1 -2

@@ -1182,8 +1216,7 @@ static int find_later_rq(struct task_struct *task)
  	 * We have to consider system topology and task affinity
  	 * first, then we can look for a suitable cpu.
  	 */
-	cpumask_copy(later_mask, task_rq(task)->rd->span);
-	cpumask_and(later_mask, later_mask, cpu_active_mask);
+	cpumask_copy(later_mask, cpu_active_mask);

I fear this breaks what I lately fixed in commit 91ec6778ec4f ("sched/deadline: Fix inter- exclusive cpusets migrations"), as

we first have to consider exclusive cpusets topology in looking for a cpu. But, I'd have to test this to see if I'm right, and I'll try to do it soon.

Thanks,

- Juri

kernel/sched/deadline.c static int find_later_rq(struct task_struct *task) · line 1222

Quoted 4 lines

kernel/sched/deadline.c +0 -0

  	cpumask_and(later_mask, later_mask, &task->cpus_allowed);
  	best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
  			task, later_mask);

--
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/

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Peterz, Juri?

Quoted 4 lines · Wanpeng Li

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Quoted 1 line

Peterz, Juri?

So the root domain span is for exclusive cpusets (Documentation/cgroups/cpusets.txt) where we fully separate sets of CPUs and have no load-balancing between the sets.

For 'normal' setups the rd->span is the entire machine, but using cpusets you can create more (smaller) domains.

Is this what you were asking?
--
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/

Hi Peter,

Quoted 10 lines · Peter Zijlstra

Quoted 5 lines · Wanpeng Li

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Peterz, Juri?

So the root domain span is for exclusive cpusets (Documentation/cgroups/cpusets.txt) where we fully separate sets of CPUs and have no load-balancing between the sets.

Quoted 2 lines

For 'normal' setups the rd->span is the entire machine, but using cpusets you can create more (smaller) domains.

I don't setup any cpusets related stuff, however, task_rq(task)->rd->span just include the cpu which the dl task previous running on instead of the entire machine in find_later_rq().

Quoted 12 lines · Wanpeng Li

Hi Peter,

Quoted 10 lines · Peter Zijlstra

Quoted 5 lines · Wanpeng Li

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Peterz, Juri?

So the root domain span is for exclusive cpusets (Documentation/cgroups/cpusets.txt) where we fully separate sets of CPUs and have no load-balancing between the sets.

Quoted 6 lines

Quoted 2 lines

For 'normal' setups the rd->span is the entire machine, but using cpusets you can create more (smaller) domains.

I don't setup any cpusets related stuff, however, task_rq(task)->rd->span just include the cpu which the dl task previous running on instead of the entire machine in find_later_rq().

Ah, it could be that for offline cpus we have a singleton rd. Lemme try

and remember wth we did there.

Quoted 1 line

Ah, it could be that for offline cpus we have a singleton rd. Lemme try

I still cannot find where build the singleton rd in the codes, could you point out?

Quoted 3 lines · Wanpeng Li

Quoted 1 line

Ah, it could be that for offline cpus we have a singleton rd. Lemme try

I still cannot find where build the singleton rd in the codes, could you point out?

So this is all quite horrible code, but what I think happens is that:

	sched_cpu_inactive() -> set_cpu_active(cpu, false);
	cpuset_cpu_inactive() -> cpuset_update_active_cpus(false)
	                      -> partition_sched_domains(1, NULL, NULL)
			      -> build_sched_domains(cpu_active_mask)
			      -> for_each_cpu()
			           cpu_attach_domain() -> rq_attach_root()

Now, that will detach all active cpus from the current root domain and attach them to the new root domain. Which leaves behind the old root domain attached to only the one 'dead' cpu.

I observe that dl task can't be migrated to other cpus during cpu hotplug, in addition, task may/may not be running again if cpu is added back. The root cause which I found is that dl task will be throtted and removed from dl rq after comsuming all budget, which leads to stop task can't pick it up from dl rq and migrate to other cpus during hotplug.

The method to reproduce:
schedtool -E -t 50000:100000 -e ./test
Actually test is just a simple for loop. Then observe which cpu the test task is on.
echo 0 > /sys/devices/system/cpu/cpuN/online

This patch fix it by push the task to another cpu in dl_task_timer() if rq is offline.

Note: dl task can be migrated successfully if rq is offline currently, however, I'm still not sure why task_rq(task)->rd->span just include the cpu which the dl task previous running on, so cpu_active_mask is used in the patch.

Peterz, Juri?

Signed-off-by Wanpeng Li <[email protected]>

---
v1 -> v2:
* push the task to another cpu in dl_task_timer() if rq is offline.

 kernel/sched/deadline.c | 39 ++++++++++++++++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 3 deletions(-)

kernel/sched/deadline.c +35 -1

index 04c2cbb..233e482 100644
@@ -487,6 +487,7 @@ static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
 	return hrtimer_active(&dl_se->dl_timer);
 }
 
+static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
 /*
  * This is the bandwidth enforcement timer callback. If here, we know
  * a task is not on its dl_rq, since the fact that the timer was running

@@ -538,6 +539,39 @@ again:
 	update_rq_clock(rq);
 	dl_se->dl_throttled = 0;
 	dl_se->dl_yielded = 0;
+
+	/*
+	 * So if we find that the rq the task was on is no longer
+	 * available, we need to select a new rq.
+	 */
+	if (!rq->online) {
+		struct rq *later_rq = NULL;
+
+		/* We will release rq lock */
+		get_task_struct(p);
+
+		raw_spin_unlock(&rq->lock);
+
+		later_rq = find_lock_later_rq(p, rq);
+
+		if (!later_rq) {
+			put_task_struct(p);
+			goto out;
+		}
+
+		deactivate_task(rq, p, 0);
+		set_task_cpu(p, later_rq->cpu);
+		activate_task(later_rq, p, 0);
+
+		resched_curr(later_rq);
+
+		double_unlock_balance(rq, later_rq);
+
+		put_task_struct(p);
+
+		goto out;
+	}
+
 	if (task_on_rq_queued(p)) {
 		enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
 		if (dl_task(rq->curr))

@@ -555,7 +589,7 @@ again:
 	}
 unlock:
 	raw_spin_unlock(&rq->lock);
-
+out:
 	return HRTIMER_NORESTART;
 }
 

Quoted 6 lines

(quoted diff) +2 -0

@@ -538,6 +539,39 @@ again:
 	update_rq_clock(rq);
 	dl_se->dl_throttled = 0;
 	dl_se->dl_yielded = 0;
+
+	/*

Review again: · line 551

Quoted 8 lines

(quoted diff) +8 -0

+	 * So if we find that the rq the task was on is no longer
+	 * available, we need to select a new rq.
+	 */
+	if (!rq->online) {
+		struct rq *later_rq = NULL;
+
+		/* We will release rq lock */
+		get_task_struct(p);

Review again: · line 559

Quoted 8 lines

(quoted diff) +8 -0

+		raw_spin_unlock(&rq->lock);
+
+		later_rq = find_lock_later_rq(p, rq);
+
+		if (!later_rq) {
+			put_task_struct(p);
+			goto out;
+		}

isn't this too complicated?

Can't we simply queue throttled tasks in rq_offline_dl() (without clearing of dl_throttled() status)? migrate_tasks() will do the migration right.