Re: [PATCH v2 3/4] irq_work: Handle some irq_work in a per-CPU thread on PREEMPT_RT

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 06, 2021 at 01:18:51PM +0200, Sebastian Andrzej Siewior wrote:
> @@ -104,7 +141,24 @@ bool irq_work_queue_on(struct irq_work *work, int cpu)
>  	if (cpu != smp_processor_id()) {
>  		/* Arch remote IPI send/receive backend aren't NMI safe */
>  		WARN_ON_ONCE(in_nmi());
> -		__smp_call_single_queue(cpu, &work->node.llist);
> +
> +		/*
> +		 * On PREEMPT_RT the items which are not marked as
> +		 * IRQ_WORK_HARD_IRQ are added to the lazy list and a HARD work
> +		 * item is used on the remote CPU to wake the thread.
> +		 */
> +		if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
> +		    !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ) &&
> +		    llist_add(&work->node.llist, &per_cpu(lazy_list, cpu))) {

So if the llist not empty, we'll fail here, and go to the else branch

> +			struct irq_work *wake_work;
> +
> +			wake_work = &per_cpu(irq_work_pending, cpu);
> +			if (irq_work_claim(wake_work))
> +				__smp_call_single_queue(cpu,
> +							&wake_work->node.llist);
> +		} else {

And do this,.. that seems wrong.

> +			__smp_call_single_queue(cpu, &work->node.llist);
> +		}
>  	} else {
>  		__irq_work_queue_local(work);
>  	}

How's this instead?

(work rename due to there already being a percpu variable of that same
name on a number or archs, per 0day)


Index: linux-2.6/kernel/irq_work.c
===================================================================
--- linux-2.6.orig/kernel/irq_work.c
+++ linux-2.6/kernel/irq_work.c
@@ -39,7 +39,7 @@ static void irq_work_wake(struct irq_wor
 	wake_irq_workd();
 }
 
-static DEFINE_PER_CPU(struct irq_work, irq_work_pending) =
+static DEFINE_PER_CPU(struct irq_work, irq_work_wakeup) =
 	IRQ_WORK_INIT_HARD(irq_work_wake);
 
 static int irq_workd_should_run(unsigned int cpu)
@@ -148,20 +148,21 @@ bool irq_work_queue_on(struct irq_work *
 		 * item is used on the remote CPU to wake the thread.
 		 */
 		if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
-		    !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ) &&
-		    llist_add(&work->node.llist, &per_cpu(lazy_list, cpu))) {
-			struct irq_work *wake_work;
-
-			wake_work = &per_cpu(irq_work_pending, cpu);
-			if (irq_work_claim(wake_work))
-				__smp_call_single_queue(cpu,
-							&wake_work->node.llist);
-		} else {
-			__smp_call_single_queue(cpu, &work->node.llist);
+		    !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ)) {
+
+			if (!llist_add(&work->node.llist, &per_cpu(lazy_list, cpu)))
+				goto out;
+
+			work = &per_cpu(irq_work_wakeup, cpu);
+			if (!irq_work_claim(wake_work))
+				goto out;
 		}
+
+		__smp_call_single_queue(cpu, &work->node.llist);
 	} else {
 		__irq_work_queue_local(work);
 	}
+out:
 	preempt_enable();
 
 	return true;