lkml.org 
[lkml]   [2011]   [Sep]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH tip/core/rcu 44/55] rcu: wire up RCU_BOOST_PRIO for rcutree
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);
>
>



\
 
 \ /
  Last update: 2011-09-13 17:43    [W:0.109 / U:39.348 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site