Re: [PATCH tip/core/rcu 1/4] rcu: Expedite deboost in case of deferred quiescent state

[Boqun Feng]

Hi Paul,

On Tue, Jan 19, 2021 at 08:32:33PM -0800, paulmck@xxxxxxxxxx wrote:
> From: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
> 
> Historically, a task that has been subjected to RCU priority boosting is
> deboosted at rcu_read_unlock() time.  However, with the advent of deferred
> quiescent states, if the outermost rcu_read_unlock() was invoked with
> either bottom halves, interrupts, or preemption disabled, the deboosting
> will be delayed for some time.  During this time, a low-priority process
> might be incorrectly running at a high real-time priority level.
> 
> Fortunately, rcu_read_unlock_special() already provides mechanisms for
> forcing a minimal deferral of quiescent states, at least for kernels
> built with CONFIG_IRQ_WORK=y.  These mechanisms are currently used
> when expedited grace periods are pending that might be blocked by the
> current task.  This commit therefore causes those mechanisms to also be
> used in cases where the current task has been or might soon be subjected
> to RCU priority boosting.  Note that this applies to all kernels built
> with CONFIG_RCU_BOOST=y, regardless of whether or not they are also
> built with CONFIG_PREEMPT_RT=y.
> 
> This approach assumes that kernels build for use with aggressive real-time
> applications are built with CONFIG_IRQ_WORK=y.  It is likely to be far
> simpler to enable CONFIG_IRQ_WORK=y than to implement a fast-deboosting
> scheme that works correctly in its absence.
> 
> While in the area, alphabetize the rcu_preempt_deferred_qs_handler()
> function's local variables.
> 
> Cc: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> Cc: Scott Wood <swood@xxxxxxxxxx>
> Cc: Lai Jiangshan <jiangshanlai@xxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> ---
>  kernel/rcu/tree_plugin.h | 26 ++++++++++++++------------
>  1 file changed, 14 insertions(+), 12 deletions(-)
> 
> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index 8b0feb2..fca31c6 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -660,9 +660,9 @@ static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
>  static void rcu_read_unlock_special(struct task_struct *t)
>  {
>  	unsigned long flags;
> +	bool irqs_were_disabled;
>  	bool preempt_bh_were_disabled =
>  			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
> -	bool irqs_were_disabled;
>  
>  	/* NMI handlers cannot block and cannot safely manipulate state. */
>  	if (in_nmi())
> @@ -671,30 +671,32 @@ static void rcu_read_unlock_special(struct task_struct *t)
>  	local_irq_save(flags);
>  	irqs_were_disabled = irqs_disabled_flags(flags);
>  	if (preempt_bh_were_disabled || irqs_were_disabled) {
> -		bool exp;
> +		bool expboost; // Expedited GP in flight or possible boosting.
>  		struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
>  		struct rcu_node *rnp = rdp->mynode;
>  
> -		exp = (t->rcu_blocked_node &&
> -		       READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
> -		      (rdp->grpmask & READ_ONCE(rnp->expmask));
> +		expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
> +			   (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
> +			   (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
> +			    t->rcu_blocked_node);

I take it that you check whether possible boosting is in progress via
the last expression of "||", ie:

	(IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
	t->rcu_blocked_node)

if so, I don't see the point of using the new "expboost" in the
raise_softirq_irqoff() branch, because if in_irq() is false, we only
raise softirq if irqs_were_disabled is false (otherwise, we may take the
risk of doing a wakeup with a pi or rq lock held, IIRC), and the
boosting part of the "expboost" above is only true if irqs_were_disabled
is true, so using expboost makes no different here.

>  		// Need to defer quiescent state until everything is enabled.
> -		if (use_softirq && (in_irq() || (exp && !irqs_were_disabled))) {
> +		if (use_softirq && (in_irq() || (expboost && !irqs_were_disabled))) {
>  			// Using softirq, safe to awaken, and either the
> -			// wakeup is free or there is an expedited GP.
> +			// wakeup is free or there is either an expedited
> +			// GP in flight or a potential need to deboost.

and this comment will be incorrect, we won't enter here solely because
there is a potential need to deboost.

That said, why the boosting condition has a "irqs_were_disabled" in it?
What if a task gets boosted because of RCU boosting, and exit the RCU
read-side c.s. with irq enabled and there is no expedited GP in flight,
will the task get deboosted quickly enough?

Maybe I'm missing some subtle?

Regards,
Boqun

>  			raise_softirq_irqoff(RCU_SOFTIRQ);
>  		} else {
>  			// Enabling BH or preempt does reschedule, so...
> -			// Also if no expediting, slow is OK.
> -			// Plus nohz_full CPUs eventually get tick enabled.
> +			// Also if no expediting and no possible deboosting,
> +			// slow is OK.  Plus nohz_full CPUs eventually get
> +			// tick enabled.
>  			set_tsk_need_resched(current);
>  			set_preempt_need_resched();
>  			if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
> -			    !rdp->defer_qs_iw_pending && exp && cpu_online(rdp->cpu)) {
> +			    expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
>  				// Get scheduler to re-evaluate and call hooks.
>  				// If !IRQ_WORK, FQS scan will eventually IPI.
> -				init_irq_work(&rdp->defer_qs_iw,
> -					      rcu_preempt_deferred_qs_handler);
> +				init_irq_work(&rdp->defer_qs_iw, rcu_preempt_deferred_qs_handler);
>  				rdp->defer_qs_iw_pending = true;
>  				irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
>  			}
> -- 
> 2.9.5
>