[PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[Xunlei Pang]

From: Xunlei Pang <pang.xunlei@linaro.org>

We may suffer from extra rt overload rq due to the affinity,
so when the affinity of any runnable rt task is changed, we
should check to trigger balancing, otherwise it will cause
some unnecessary delayed real-time response. Unfortunately,
current RT global scheduler does nothing about this.

For example: a 2-cpu system with two runnable FIFO tasks(same
rt_priority) bound on CPU0, let's name them rt1(running) and
rt2(runnable) respectively; CPU1 has no RTs. Then, someone sets
the affinity of rt2 to 0x3(i.e. CPU0 and CPU1), but after this,
rt2 still can't be scheduled enters schedule(), this
definitely causes some/big response latency for rt2.

This patch modified set_cpus_allowed_rt(), if the target task
is runnable but not running, it tries to push it away once it
got migratable.

The patch also solves a problem about move_queued_task() called
in set_cpus_allowed_ptr():
When a lower priorioty rt task got migrated due to its curr cpu
isn't in the new affinity mask, after move_queued_task() it will
miss the chance of pushing away, because check_preempt_curr()
called by move_queued_task() doens't set the "need resched flag"
for lower priority tasks.

Signed-off-by: Xunlei Pang <pang.xunlei@linaro.org>
---
 kernel/sched/core.c     | 10 +++++++---
 kernel/sched/deadline.c |  8 +++++---
 kernel/sched/rt.c       | 29 ++++++++++++++++++++++-------
 kernel/sched/sched.h    |  3 ++-
 4 files changed, 36 insertions(+), 14 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index d13fc13..c995a02 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -4768,11 +4768,15 @@ static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
 
 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
 {
+	bool updated = false;
+
 	if (p->sched_class->set_cpus_allowed)
-		p->sched_class->set_cpus_allowed(p, new_mask);
+		updated = p->sched_class->set_cpus_allowed(p, new_mask);
 
-	cpumask_copy(&p->cpus_allowed, new_mask);
-	p->nr_cpus_allowed = cpumask_weight(new_mask);
+	if (!updated) {
+		cpumask_copy(&p->cpus_allowed, new_mask);
+		p->nr_cpus_allowed = cpumask_weight(new_mask);
+	}
 }
 
 /*
diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index 5e95145..3baffb2 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -1574,7 +1574,7 @@ static void task_woken_dl(struct rq *rq, struct task_struct *p)
 	}
 }
 
-static void set_cpus_allowed_dl(struct task_struct *p,
+static bool set_cpus_allowed_dl(struct task_struct *p,
 				const struct cpumask *new_mask)
 {
 	struct rq *rq;
@@ -1610,7 +1610,7 @@ static void set_cpus_allowed_dl(struct task_struct *p,
 	 * it is on the rq AND it is not throttled).
 	 */
 	if (!on_dl_rq(&p->dl))
-		return;
+		return false;
 
 	weight = cpumask_weight(new_mask);
 
@@ -1619,7 +1619,7 @@ static void set_cpus_allowed_dl(struct task_struct *p,
 	 * can migrate or not.
 	 */
 	if ((p->nr_cpus_allowed > 1) == (weight > 1))
-		return;
+		return false;
 
 	/*
 	 * The process used to be able to migrate OR it can now migrate
@@ -1636,6 +1636,8 @@ static void set_cpus_allowed_dl(struct task_struct *p,
 	}
 
 	update_dl_migration(&rq->dl);
+
+	return false;
 }
 
 /* Assumes rq->lock is held */
diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 8885b65..4a49c6a 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -2241,7 +2241,7 @@ static void task_woken_rt(struct rq *rq, struct task_struct *p)
 		push_rt_tasks(rq);
 }
 
-static void set_cpus_allowed_rt(struct task_struct *p,
+static bool set_cpus_allowed_rt(struct task_struct *p,
 				const struct cpumask *new_mask)
 {
 	struct rq *rq;
@@ -2250,18 +2250,19 @@ static void set_cpus_allowed_rt(struct task_struct *p,
 	BUG_ON(!rt_task(p));
 
 	if (!task_on_rq_queued(p))
-		return;
+		return false;
 
 	weight = cpumask_weight(new_mask);
 
+	rq = task_rq(p);
+
 	/*
-	 * Only update if the process changes its state from whether it
-	 * can migrate or not.
+	 * Skip updating the migration stuff if the process doesn't change
+	 * its migrate state, but still need to check if it can be pushed
+	 * away due to its new affinity.
 	 */
 	if ((p->nr_cpus_allowed > 1) == (weight > 1))
-		return;
-
-	rq = task_rq(p);
+		goto check_push;
 
 	/*
 	 * The process used to be able to migrate OR it can now migrate
@@ -2278,6 +2279,20 @@ static void set_cpus_allowed_rt(struct task_struct *p,
 	}
 
 	update_rt_migration(&rq->rt);
+
+check_push:
+	if (weight > 1 &&
+	    !task_running(rq, p) &&
+	    !test_tsk_need_resched(rq->curr) &&
+	    !cpumask_subset(new_mask, &p->cpus_allowed)) {
+		/* Update new affinity and try to push. */
+		cpumask_copy(&p->cpus_allowed, new_mask);
+		p->nr_cpus_allowed = weight;
+		push_rt_tasks(rq);
+		return true;
+	}
+
+	return false;
 }
 
 /* Assumes rq->lock is held */
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index e0e1299..101b359 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1189,7 +1189,8 @@ struct sched_class {
 	void (*task_waking) (struct task_struct *task);
 	void (*task_woken) (struct rq *this_rq, struct task_struct *task);
 
-	void (*set_cpus_allowed)(struct task_struct *p,
+	/* Return true if p's affinity was updated, false otherwise. */
+	bool (*set_cpus_allowed)(struct task_struct *p,
 				 const struct cpumask *newmask);
 
 	void (*rq_online)(struct rq *rq);
-- 
1.9.1

[PATCH v3 2/4] sched/deadline: Check to push the task away after its affinity was changed

[Xunlei Pang]

From: Xunlei Pang <pang.xunlei@linaro.org>

(Sync up the same behaviour as that of RT.)

We may suffer from extra dl overload rq due to the affinity,
so when the affinity of any runnable dl task is changed, we
should check to trigger balancing, otherwise it will cause
some unnecessary delayed real-time response. Unfortunately,
current DL global scheduler does nothing about this.

This patch modified set_cpus_allowed_dl(), if the target task
is runnable but not running and not throttled, it tries to push
it away once it got migratable.

The patch also solves a problem about move_queued_task() called
in set_cpus_allowed_ptr():
When a smaller deadline value dl task got migrated due to its curr
cpu isn't in the new affinity mask, after move_queued_task() it
will miss the chance of pushing away, because check_preempt_curr()
called by move_queued_task() doens't set the "need resched flag"
for smaller deadline value tasks.

Signed-off-by: Xunlei Pang <pang.xunlei@linaro.org>
---
 kernel/sched/deadline.c | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index 3baffb2..968612b 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -1615,11 +1615,12 @@ static bool set_cpus_allowed_dl(struct task_struct *p,
 	weight = cpumask_weight(new_mask);
 
 	/*
-	 * Only update if the process changes its state from whether it
-	 * can migrate or not.
+	 * Skip updating the migration stuff if the process doesn't change
+	 * its migrate state, but still need to check if it can be pushed
+	 * away due to its new affinity.
 	 */
 	if ((p->nr_cpus_allowed > 1) == (weight > 1))
-		return false;
+		goto check_push;
 
 	/*
 	 * The process used to be able to migrate OR it can now migrate
@@ -1637,6 +1638,18 @@ static bool set_cpus_allowed_dl(struct task_struct *p,
 
 	update_dl_migration(&rq->dl);
 
+check_push:
+	if (weight > 1 &&
+	    !task_running(rq, p) &&
+	    !test_tsk_need_resched(rq->curr) &&
+	    !cpumask_subset(new_mask, &p->cpus_allowed)) {
+		/* Update new affinity and try to push. */
+		cpumask_copy(&p->cpus_allowed, new_mask);
+		p->nr_cpus_allowed = weight;
+		push_dl_tasks(rq);
+		return true;
+	}
+
 	return false;
 }
 
-- 
1.9.1

[PATCH v3 3/4] sched/rt: Remove redundant conditions from task_woken_rt()

[Xunlei Pang]

From: Xunlei Pang <pang.xunlei@linaro.org>

- Remove "has_pushable_tasks(rq)".
  Because for queued p, "!task_running(rq, p)" and "p->nr_cpus_allowed > 1"
  already imply that "has_pushable_tasks(rq)" is true.

- Remove "!test_tsk_need_resched(rq->curr)".
  The condtion mainly intends to ensure higher priority rt tasks won't be pushed
  away. I can think of two reasons below for getting rid of it.
  1) With following "rq->curr->prio <= p->prio", we still can guarantee that
     purpose. "rq->curr->prio <= p->prio" implies the "need resched flag" wasn't
     set by check_preempt_curr() except the one set by check_preempt_equal_prio()
     for equal prio cases(In this case, if the condition is removed, it may result
     in an extra push_rt_tasks(), but this doesn't cause the wrong logic, in fact
     this extra push_rt_tasks() will probably return quickly for the case).

     Addtionally, there're also cases the "need resched flag" got set before the
     waking, with current implementation it needn't to push lower priority tasks
     as the cpu will schedule, while it will do an extra pushing if the condition
     is removed. But on the other hand, we can get a timely pushing for the woken
     tasks after the condition is removed(better for the non-preemptible kernel).

  2) With following condtion "rq->curr->nr_cpus_allowed < 2" which was added by
     commit b3bc211cfe7d ("sched: Give CPU bound RT tasks preference"), in the
     scenario descibed in it, "need resched flag" was already set before by
     check_preempt_curr(), thus "!test_tsk_need_resched(rq->curr)" is always false
     which means with current implementation the commit is futile for task_woken_rt().
  So, by removing this condition, we get the right logic.

Signed-off-by: Xunlei Pang <pang.xunlei@linaro.org>
---
 kernel/sched/rt.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 4a49c6a..4bd8551 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -2232,8 +2232,6 @@ out:
 static void task_woken_rt(struct rq *rq, struct task_struct *p)
 {
 	if (!task_running(rq, p) &&
-	    !test_tsk_need_resched(rq->curr) &&
-	    has_pushable_tasks(rq) &&
 	    p->nr_cpus_allowed > 1 &&
 	    (dl_task(rq->curr) || rt_task(rq->curr)) &&
 	    (rq->curr->nr_cpus_allowed < 2 ||
-- 
1.9.1

[PATCH v3 4/4] sched/deadline: Remove redundant conditions from task_woken_dl()

[Xunlei Pang]

From: Xunlei Pang <pang.xunlei@linaro.org>

Sync up the same behaviour as that of RT:
Remove "has_pushable_dl_tasks(rq)" and "!test_tsk_need_resched(rq->curr)".

Signed-off-by: Xunlei Pang <pang.xunlei@linaro.org>
---
 kernel/sched/deadline.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index 968612b..38c08c4 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -1564,8 +1564,6 @@ static void post_schedule_dl(struct rq *rq)
 static void task_woken_dl(struct rq *rq, struct task_struct *p)
 {
 	if (!task_running(rq, p) &&
-	    !test_tsk_need_resched(rq->curr) &&
-	    has_pushable_dl_tasks(rq) &&
 	    p->nr_cpus_allowed > 1 &&
 	    dl_task(rq->curr) &&
 	    (rq->curr->nr_cpus_allowed < 2 ||
-- 
1.9.1

Re: [PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[Peter Zijlstra]

On Tue, May 12, 2015 at 10:46:41PM +0800, Xunlei Pang wrote:
> @@ -2278,6 +2279,20 @@ static void set_cpus_allowed_rt(struct task_struct *p,
>  	}
>  
>  	update_rt_migration(&rq->rt);
> +
> +check_push:
> +	if (weight > 1 &&
> +	    !task_running(rq, p) &&
> +	    !test_tsk_need_resched(rq->curr) &&
> +	    !cpumask_subset(new_mask, &p->cpus_allowed)) {
> +		/* Update new affinity and try to push. */
> +		cpumask_copy(&p->cpus_allowed, new_mask);
> +		p->nr_cpus_allowed = weight;
> +		push_rt_tasks(rq);
> +		return true;
> +	}
> +
> +	return false;
>  }

I think this is broken; push_rt_tasks() will do double_rq_lock() which
will drop rq->lock.

This means load-balancing can come in and move our task p; in fact,
push_rt_task() can do exactly that -- after all that was the point of
this patch.

_However_ this means that after calling ->set_cpus_allowed() we must not
assume @p is on @rt, yet we do. Look at __set_cpus_allowed_ptr(), we'll
call move_queued_task() if (!running || waking) && on_rq, and
move_queued_task() happily calls dequeue_task(rq, p), which will go
*boom*.

I currently do not have a better idea than to repurpose the PUSH_IPI
stuff, that is, send a self IPI to go do the push or somesuch. Lemme
stare at this a little more.

Re: [PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[Peter Zijlstra]

On Fri, May 29, 2015 at 10:04:36PM +0800, pang.xunlei@zte.com.cn wrote:
> Hi Peter,
> 
> Peter Zijlstra <peterz@infradead.org> wrote 2015-05-29 PM 09:16:26:
> > 
> > Re: [PATCH v3 1/4] sched/rt: Check to push the task away after its 
> > affinity was changed
> > 
> > On Tue, May 12, 2015 at 10:46:41PM +0800, Xunlei Pang wrote:
> > > @@ -2278,6 +2279,20 @@ static void set_cpus_allowed_rt(struct 
> > task_struct *p,
> > >     }
> > > 
> > >     update_rt_migration(&rq->rt);
> > > +
> > > +check_push:
> > > +   if (weight > 1 &&
> > > +       !task_running(rq, p) &&
> > > +       !test_tsk_need_resched(rq->curr) &&
> > > +       !cpumask_subset(new_mask, &p->cpus_allowed)) {
> > > +      /* Update new affinity and try to push. */
> > > +      cpumask_copy(&p->cpus_allowed, new_mask);
> > > +      p->nr_cpus_allowed = weight;
> > > +      push_rt_tasks(rq);
> > > +      return true;
> > > +   }
> > > +
> > > +   return false;
> > >  }
> > 
> > I think this is broken; push_rt_tasks() will do double_rq_lock() which
> > will drop rq->lock.
> > 
> > This means load-balancing can come in and move our task p; in fact,
> > push_rt_task() can do exactly that -- after all that was the point of
> > this patch.
> > 
> > _However_ this means that after calling ->set_cpus_allowed() we must not
> > assume @p is on @rt, yet we do. Look at __set_cpus_allowed_ptr(), we'll
> > call move_queued_task() if (!running || waking) && on_rq, and
> > move_queued_task() happily calls dequeue_task(rq, p), which will go
> > *boom*.
> 
> I can't see why this can happen?
> 
> After finishing set_cpus_allowed_rt(), if there happens a successful
> load-balancing (pull or push) action, new task_cpu(@p) will be set, 
> so we will definitely get the following true condition:
> 
>         /* Can the task run on the task's current CPU? If so, we're done 
> */
>         if (cpumask_test_cpu(task_cpu(p), new_mask))
>                 goto out;
> 
> So I think the whole function will simply go out and return normally.

Humm, yes. Missed that. That makes it work by accident; because you
didn't document/Changelog any of this.

Makes me like the thing even less though..

Re: [PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[Peter Zijlstra]

On Sat, 2015-05-30 at 10:20 +0200, Peter Zijlstra wrote:
> Makes me like the thing even less though..

Steven, why do we normally push on schedule()? Would not the natural
location be where we add to pushable_tasks?

Which would be here in set_cpus_allowed() and wakeups. schedule() seems
like a second best location.

Re: [PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[Steven Rostedt]

On Sat, 30 May 2015 12:54:23 +0200
Peter Zijlstra <peterz@infradead.org> wrote:

> On Sat, 2015-05-30 at 10:20 +0200, Peter Zijlstra wrote:
> > Makes me like the thing even less though..
> 
> Steven, why do we normally push on schedule()? Would not the natural
> location be where we add to pushable_tasks?
> 
> Which would be here in set_cpus_allowed() and wakeups. schedule() seems
> like a second best location.

What about when an RT task gets preempted by a higher prio task. We may
need to push the preempted one. You can't do that at wakeup.

-- Steve