Date:	Wed, 28 Sep 2011 14:31:21 +0200
From:	Frederic Weisbecker <fweisbec@...il.com>
To:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
Cc:	"Kirill A. Shutemov" <kirill@...temov.name>,
	linux-kernel@...r.kernel.org, Dipankar Sarma <dipankar@...ibm.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	Ingo Molnar <mingo@...e.hu>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Lai Jiangshan <laijs@...fujitsu.com>
Subject: Re: linux-next-20110923: warning kernel/rcutree.c:1833

On Tue, Sep 27, 2011 at 11:01:42AM -0700, Paul E. McKenney wrote:
> On Tue, Sep 27, 2011 at 02:16:50PM +0200, Frederic Weisbecker wrote:
> > On Mon, Sep 26, 2011 at 03:50:32PM -0700, Paul E. McKenney wrote:
> > > On Mon, Sep 26, 2011 at 11:20:55AM +0200, Frederic Weisbecker wrote:
> > > > On Sun, Sep 25, 2011 at 06:26:11PM -0700, Paul E. McKenney wrote:
> > > > > On Mon, Sep 26, 2011 at 03:10:33AM +0200, Frederic Weisbecker wrote:
> > > > > > 2011/9/26 Frederic Weisbecker <fweisbec@...il.com>:
> > > > > > > On Sun, Sep 25, 2011 at 09:48:04AM -0700, Paul E. McKenney wrote:
> > > > > > >> This is required for RCU_FAST_NO_HZ, which checks to see whether the
> > > > > > >> current CPU can accelerate the current grace period so as to enter
> > > > > > >> dyntick-idle mode sooner than it would otherwise.  This takes effect
> > > > > > >> in the situation where rcu_needs_cpu() sees that there are callbacks.
> > > > > > >> It then notes a quiescent state (which is illegal in an RCU read-side
> > > > > > >> critical section), calls force_quiescent_state(), and so on.  For this
> > > > > > >> to work, the current CPU must be in an RCU read-side critical section.
> > > > > > >
> > > > > > > You mean it must *not* be in an RCU read-side critical section (ie: in a
> > > > > > > quiescent state)?
> > > > > > >
> > > > > > > That assumption at least fails anytime in idle for the RCU
> > > > > > > sched flavour given that preemption is disabled in the idle loop.
> > > > > > >
> > > > > > >> If this cannot be made to work, another option is to call a new RCU
> > > > > > >> function in the case where rcu_needs_cpu() returned false, but after
> > > > > > >> the RCU read-side critical section has exited.
> > > > > > >
> > > > > > > You mean when rcu_needs_cpu() returns true (when we have callbacks
> > > > > > > enqueued)?
> > > > > > >
> > > > > > >> This new RCU function
> > > > > > >> could then attempt to rearrange RCU so as to allow the CPU to enter
> > > > > > >> dyntick-idle mode more quickly.  It is more important for this to
> > > > > > >> happen when the CPU is going idle than when it is executing a user
> > > > > > >> process.
> > > > > > >>
> > > > > > >> So, is this doable?
> > > > > > >
> > > > > > > At least not when we have RCU sched callbacks enqueued, given preemption
> > > > > > > is disabled. But that sounds plausible in order to accelerate the switch
> > > > > > > to dyntick-idle mode when we only have rcu and/or rcu bh callbacks.
> > > > > > 
> > > > > > But the RCU sched case could be dealt with if we embrace every use of
> > > > > > it with rcu_read_lock_sched() and rcu_read_unlock_sched(), or some light
> > > > > > version that just increases a local counter that rcu_needs_cpu() could check.
> > > > > > 
> > > > > > It's an easy thing to add: we can ensure preempt is disabled when we call it
> > > > > > and we can force rcu_dereference_sched() to depend on it.
> > > > > 
> > > > > Or just check to see if this is the first level of interrupt from the
> > > > > idle task after the scheduler is up.
> > > > 
> > > > I believe it's always the case. tick_nohz_stop_sched_tick() is only called
> > > > from the first level of interrupt in irq_exit().
> > > 
> > > OK, good, let me see if I really understand this...
> > > 
> > > Case 1: The interrupt interrupted non-dyntick-idle code.  In this case,
> > > 	rcu_needs_cpu() can look at the dyntick-idle state and determine
> > > 	that it might not be in a quiescent state.
> > 
> > I guess by dyntick idle code you mean the fact that the RCU in is
> > extended quiescent state? (Not just the tick is stopped)
> > 
> > If so yeah that looks good.
> > 
> > > 
> > > Case 2: The interrupt interrupted dyntick-idle code.  In this case,
> > > 	the interrupted code had better not be in an RCU read-side
> > > 	critical section, and rcu_needs_cpu() should be able to
> > > 	detect this as well.
> > 
> > Yeah.
> > 
> > We already do the appropriate debug checks from the RCU read side
> > APIs so I guess rcu_needs_cpu() doesn't even need to do its own
> > debugging checks here about extended qs.
> > 
> > But indeed it can return right away if we are in extended qs.
> > 
> > > 
> > > Case 3: The interrupt interrupted the process of transitioning to
> > > 	or from dyntick-idle mode.  This should be prohibited by
> > > 	the local_irq_save() calls, right?
> > 
> > Indeed.
> > 
> > > 
> > > > There is always some race window, as it's based on preempt offset: between
> > > > the sub_preempt_count and the softirqs begin and between softirqs end and the end
> > > > of the interrupt. But an "idle_cpu() || in_interrupt()" check in rcu_read_lock_sched_held()
> > > > should catch those offenders.
> > > 
> > > But all of this stuff looks to me to be called from the context
> > > of the idle task, so that idle_cpu() will always return "true"...
> > 
> > I meant "idle_cpu() && !in_interrupt()" that should return false in
> > rcu_read_lock_sched_held().
> 
> The problem is that the idle tasks now seem to make quite a bit of use
> of RCU on entry to and exit from the idle loop itself, for example,
> via tracing.  So it seems like it is time to have the idle loop
> explictly tell RCU when the idle extended quiescent state is in effect.
> 
> An experimental patch along these lines is included below.  Does this
> approach seem reasonable, or am I missing something subtle (or even
> not so subtle) here?
> 
> 							Thanx, Paul
> 
> ------------------------------------------------------------------------
> 
> rcu: Explicitly track idle CPUs.
> 
> In the good old days, RCU simply checked to see if it was running in
> the context of an idle task to determine whether or not it was in the
> idle extended quiescent state.  However, the entry to and exit from
> idle has become more ornate over the years, and some of this processing
> now uses RCU while running in the context of the idle task.  It is
> therefore no longer reasonable to assume that anything running in the
> context of one of the idle tasks is in an extended quiscent state.
> 
> This commit therefore explicitly tracks whether each CPU is in the
> idle loop, allowing the idle task to use RCU anywhere except in those
> portions of the idle loops where RCU has been explicitly informed that
> it is in a quiescent state.
> 
> Signed-off-by: Paul E. McKenney <paulmck@...ux.vnet.ibm.com>

I fear we indeed need that now.

Just some comments:

> 
> diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> index 9d40e42..5b7e62c 100644
> --- a/include/linux/rcupdate.h
> +++ b/include/linux/rcupdate.h
> @@ -177,6 +177,9 @@ extern void rcu_sched_qs(int cpu);
>  extern void rcu_bh_qs(int cpu);
>  extern void rcu_check_callbacks(int cpu, int user);
>  struct notifier_block;
> +extern void rcu_idle_enter(void);
> +extern void rcu_idle_exit(void);
> +extern int rcu_is_cpu_idle(void);
>  
>  #ifdef CONFIG_NO_HZ
>  
> @@ -187,10 +190,12 @@ extern void rcu_exit_nohz(void);
>  
>  static inline void rcu_enter_nohz(void)
>  {
> +	rcu_idle_enter();
>  }
>  
>  static inline void rcu_exit_nohz(void)
>  {
> +	rcu_idle_exit();
>  }
>  
>  #endif /* #else #ifdef CONFIG_NO_HZ */
> diff --git a/include/linux/tick.h b/include/linux/tick.h
> index 375e7d8..cd9e2d1 100644
> --- a/include/linux/tick.h
> +++ b/include/linux/tick.h
> @@ -131,8 +131,16 @@ extern ktime_t tick_nohz_get_sleep_length(void);
>  extern u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time);
>  extern u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time);
>  # else
> -static inline void tick_nohz_idle_enter(bool rcu_ext_qs) { }
> -static inline void tick_nohz_idle_exit(void) { }
> +static inline void tick_nohz_idle_enter(bool rcu_ext_qs)
> +{
> +	if (rcu_ext_qs())
> +		rcu_idle_enter();
> +}

rcu_ext_qs is not a function.

> +static inline void tick_nohz_idle_exit(void)
> +{
> +	if (rcu_ext_qs())
> +		rcu_idle_exit();
> +}

So we probably need to track whether we entered in rcu_ext_qs
so that we can know if we cann rcu_idle_exit(). Or may
be pass the rcu_ext_qs parameter down to tick_nohz_idle_exit() as well.

>  static inline ktime_t tick_nohz_get_sleep_length(void)
>  {
>  	ktime_t len = { .tv64 = NSEC_PER_SEC/HZ };
> diff --git a/kernel/rcu.h b/kernel/rcu.h
> index f600868..220b4fe 100644
> --- a/kernel/rcu.h
> +++ b/kernel/rcu.h
> @@ -23,6 +23,8 @@
>  #ifndef __LINUX_RCU_H
>  #define __LINUX_RCU_H
>  
> +/* Avoid tracing overhead if not configure, mostly for RCU_TINY's benefit. */
> +
>  #ifdef CONFIG_RCU_TRACE
>  #define RCU_TRACE(stmt) stmt
>  #else /* #ifdef CONFIG_RCU_TRACE */
<snip>
> diff --git a/kernel/rcutiny.c b/kernel/rcutiny.c
> index 9e493b9..6d7207d 100644
> --- a/kernel/rcutiny.c
> +++ b/kernel/rcutiny.c
> @@ -65,8 +65,10 @@ static long rcu_dynticks_nesting = 1;
>   */
>  void rcu_enter_nohz(void)
>  {
> -	if (--rcu_dynticks_nesting == 0)
> +	if (--rcu_dynticks_nesting == 0) {
>  		rcu_sched_qs(0); /* implies rcu_bh_qsctr_inc(0) */
> +		rcu_idle_enter();

Although idle and rcu/nohz are still close notions, it sounds
more logical the other way around in the ordering:

tick_nohz_idle_enter() {
	rcu_idle_enter() {
		rcu_enter_nohz();
	}
}

tick_nohz_irq_exit() {
        rcu_idle_enter() {
                rcu_enter_nohz();
        }
}

Because rcu ext qs is something used by idle, not the opposite.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists