Re: [PATCH tip/core/rcu 44/55] rcu: wire up RCU_BOOST_PRIO for rcutree

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, Sep 13, 2011 at 02:02:14PM +0200, Mike Galbraith wrote:
> Hi Paul,
> 
> This patch causes RCU thread priority funnies, with some help from rcun.
> 
> On Tue, 2011-09-06 at 11:00 -0700, Paul E. McKenney wrote:
> > 	return 0;
> > @@ -1466,6 +1474,7 @@ static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
> >  {
> >  	struct sched_param sp;
> >  	struct timer_list yield_timer;
> > +	int prio = current->normal_prio;
> >  
> >  	setup_timer_on_stack(&yield_timer, f, arg);
> >  	mod_timer(&yield_timer, jiffies + 2);
> 
> There's a thinko there, prio either needs to be inverted before feeding
> it to __setscheduler().. or just use ->rt_priority.  I did the latter,
> and twiddled rcun to restore it's priority instead of RCU_KTHREAD_PRIO.
> RCU threads now stay put.

Very good -- some comments below.  I had to hand-apply this due to
conflicts with my current patch stack: https://github.com/paulmckrcu/linux

> rcu: wire up RCU_BOOST_PRIO for rcutree
> 
> RCU boost threads start life at RCU_BOOST_PRIO, while others remain at
> RCU_KTHREAD_PRIO.  Adjust rcu_yield() to preserve priority across the
> yield, and if the node thread restores RT policy for a yielding thread,
> it sets priority to it's own priority.  This sets the stage for user
> controlled runtime changes to priority in the -rt tree.
> 
> While here, change thread names to match other kthreads.
> 
> Signed-off-by: Mike Galbraith <efault@gmx.de>
> 
> ---
>  kernel/rcutree.c        |    2 --
>  kernel/rcutree_plugin.h |   22 ++++++++++++++++------
>  2 files changed, 16 insertions(+), 8 deletions(-)
> 
> Index: linux-3.0-tip/kernel/rcutree.c
> ===================================================================
> --- linux-3.0-tip.orig/kernel/rcutree.c
> +++ linux-3.0-tip/kernel/rcutree.c
> @@ -128,8 +128,6 @@ static void rcu_node_kthread_setaffinity
>  static void invoke_rcu_core(void);
>  static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
> 
> -#define RCU_KTHREAD_PRIO 1	/* RT priority for per-CPU kthreads. */
> -
>  /*
>   * Track the rcutorture test sequence number and the update version
>   * number within a given test.  The rcutorture_testseq is incremented
> Index: linux-3.0-tip/kernel/rcutree_plugin.h
> ===================================================================
> --- linux-3.0-tip.orig/kernel/rcutree_plugin.h
> +++ linux-3.0-tip/kernel/rcutree_plugin.h
> @@ -27,6 +27,14 @@
>  #include <linux/delay.h>
>  #include <linux/stop_machine.h>
> 
> +#define RCU_KTHREAD_PRIO 1
> +
> +#ifdef CONFIG_RCU_BOOST
> +#define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
> +#else
> +#define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
> +#endif
> +
>  /*
>   * Check the RCU kernel configuration parameters and print informative
>   * messages about anything out of the ordinary.  If you like #ifdef, you
> @@ -1345,13 +1353,13 @@ static int __cpuinit rcu_spawn_one_boost
>  	if (rnp->boost_kthread_task != NULL)
>  		return 0;
>  	t = kthread_create(rcu_boost_kthread, (void *)rnp,
> -			   "rcub%d", rnp_index);
> +			   "rcub/%d", rnp_index);
>  	if (IS_ERR(t))
>  		return PTR_ERR(t);
>  	raw_spin_lock_irqsave(&rnp->lock, flags);
>  	rnp->boost_kthread_task = t;
>  	raw_spin_unlock_irqrestore(&rnp->lock, flags);
> -	sp.sched_priority = RCU_KTHREAD_PRIO;
> +	sp.sched_priority = RCU_BOOST_PRIO;
>  	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
>  	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
>  	return 0;
> @@ -1446,6 +1454,7 @@ static void rcu_yield(void (*f)(unsigned
>  {
>  	struct sched_param sp;
>  	struct timer_list yield_timer;
> +	int prio = current->rt_priority;

This makes sense, and I have merged it into your previous patch.

>  	setup_timer_on_stack(&yield_timer, f, arg);
>  	mod_timer(&yield_timer, jiffies + 2);
> @@ -1453,7 +1462,8 @@ static void rcu_yield(void (*f)(unsigned
>  	sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
>  	set_user_nice(current, 19);
>  	schedule();
> -	sp.sched_priority = RCU_KTHREAD_PRIO;
> +	set_user_nice(current, 0);
> +	sp.sched_priority = prio;
>  	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
>  	del_timer(&yield_timer);
>  }
> @@ -1562,7 +1572,7 @@ static int __cpuinit rcu_spawn_one_cpu_k
>  	if (!rcu_scheduler_fully_active ||
>  	    per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
>  		return 0;
> -	t = kthread_create(rcu_cpu_kthread, (void *)(long)cpu, "rcuc%d", cpu);
> +	t = kthread_create(rcu_cpu_kthread, (void *)(long)cpu, "rcuc/%d", cpu);
>  	if (IS_ERR(t))
>  		return PTR_ERR(t);
>  	if (cpu_online(cpu))
> @@ -1608,7 +1618,7 @@ static int rcu_node_kthread(void *arg)
>  				continue;
>  			}
>  			per_cpu(rcu_cpu_has_work, cpu) = 1;
> -			sp.sched_priority = RCU_KTHREAD_PRIO;
> +			sp.sched_priority = current->rt_priority;

This is broken -- the per-node kthread runs at RT prio 99, but we usually
would not want to boost that high.

Seems like we should have a global variable that tracks the current
priority.  This global variable could then be set in a manner similar
to the softirq priorities -- or, perhaps better, simply set whenever
the softirq priority is changed.

Thoughts?

>  			sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
>  			preempt_enable();
>  		}
> @@ -1671,7 +1681,7 @@ static int __cpuinit rcu_spawn_one_node_
>  		return 0;
>  	if (rnp->node_kthread_task == NULL) {
>  		t = kthread_create(rcu_node_kthread, (void *)rnp,
> -				   "rcun%d", rnp_index);
> +				   "rcun/%d", rnp_index);
>  		if (IS_ERR(t))
>  			return PTR_ERR(t);
>  		raw_spin_lock_irqsave(&rnp->lock, flags);
> 
>