Re: [PATCH] sched/deadline: fix dl bandwidth of root domain overflow after dl task dead

[Juri Lelli]

Hi Wanpeng,

I finally got to review this, sorry about the huge delay.

On 07/04/2015 04:36, Wanpeng Li wrote:
> The total used dl bandwidth of each root domain will be reset to 0 after 
> cpu hotplug when rebuild sched domains, since the call path is:
> 
> _cpu_down
>   cpuset_cpu_inactive() 
>     cpuset_update_active_cpus()
>       partition_sched_domains()
>         build_sched_domains() 
>           init_rootdomain() 
>             init_dl_bw() 
> 
> The bandwidth which dl task occupy will be released when dl task dead,
> it will be minus from total used dl bandwidth of its root domain, 
> however, bandwidth overflow occurs since total used dl bandwidth is 0.
> 

Right, that's a bug.

> This patch fix it by attaching the bandwidth which dl task occupy to 
> the new root domain when the task is migrating since cpu hotplug, and
> attach all the used dl bandwidth of dl tasks to the new root domain 
> when sched domains are rebuild.
> 

But, I think this fix has still a couple of problems:

 - what happens if a DL task is simply sleeping when domains are
   reconfigured?

 - def_root_domain has now multiple accounting problems, as you do
   this thing even when a cpu is moved there in the cpuoff path

Also, runqueue (and throttling) information are dynamic, while we
are trying to fix a static problem. It's probably not a good idea
mixing them.

I'm not sure how (I need more time to think it through), but can
we maybe fix this using cpuset information?

Thanks,

- Juri

> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxxxxxx>
> ---
>  kernel/sched/core.c     |  1 +
>  kernel/sched/deadline.c | 25 +++++++++++++++++++++++++
>  kernel/sched/sched.h    |  1 +
>  3 files changed, 27 insertions(+)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 28b0d75..c940999 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -5586,6 +5586,7 @@ static void rq_attach_root(struct rq *rq, struct root_domain *rd)
>  	rq->rd = rd;
>  
>  	cpumask_set_cpu(rq->cpu, rd->span);
> +	attach_dl_bw(rq);
>  	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
>  		set_rq_online(rq);
>  
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index 5e95145..62680d7 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -224,6 +224,7 @@ static void dl_task_offline_migration(struct rq *rq, struct task_struct *p)
>  {
>  	struct rq *later_rq = NULL;
>  	bool fallback = false;
> +	struct dl_bw *dl_b;
>  
>  	later_rq = find_lock_later_rq(p, rq);
>  
> @@ -258,6 +259,11 @@ static void dl_task_offline_migration(struct rq *rq, struct task_struct *p)
>  	set_task_cpu(p, later_rq->cpu);
>  	activate_task(later_rq, p, ENQUEUE_REPLENISH);
>  
> +	dl_b = dl_bw_of(later_rq->cpu);
> +	raw_spin_lock(&dl_b->lock);
> +	__dl_add(dl_b, p->dl.dl_bw);
> +	raw_spin_unlock(&dl_b->lock);
> +
>  	if (!fallback)
>  		resched_curr(later_rq);
>  
> @@ -1776,6 +1782,25 @@ static void prio_changed_dl(struct rq *rq, struct task_struct *p,
>  		switched_to_dl(rq, p);
>  }
>  
> +void attach_dl_bw(struct rq *rq)
> +{
> +	struct rb_node *next_node = rq->dl.rb_leftmost;
> +	struct sched_dl_entity *dl_se;
> +	struct dl_bw *dl_b;
> +
> +	dl_b = dl_bw_of(rq->cpu);
> +	raw_spin_lock(&dl_b->lock);
> +next_node:
> +	if (next_node) {
> +		dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
> +		__dl_add(dl_b, dl_se->dl_bw);
> +		next_node = rb_next(next_node);
> +
> +		goto next_node;
> +	}
> +	raw_spin_unlock(&dl_b->lock);
> +}
> +
>  const struct sched_class dl_sched_class = {
>  	.next			= &rt_sched_class,
>  	.enqueue_task		= enqueue_task_dl,
> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> index e0e1299..a7b1a59 100644
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -1676,6 +1676,7 @@ extern void init_dl_rq(struct dl_rq *dl_rq);
>  
>  extern void cfs_bandwidth_usage_inc(void);
>  extern void cfs_bandwidth_usage_dec(void);
> +void attach_dl_bw(struct rq *rq);
>  
>  #ifdef CONFIG_NO_HZ_COMMON
>  enum rq_nohz_flag_bits {
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/