[PATCH] kernel/irq/proc: /proc/interrupts irq-off latency reduction

[PATCH] kernel/irq/proc: /proc/interrupts irq-off latency reduction

[Nicholas Piggin]

Reading /proc/interrupts shows up as a latency source of several
hundred microseconds on a 2 socket 176 thread system, because it
iterates twice over all CPUs under an irq lock pulling in remote
cachelines, doing lots of seq_printf, etc.

On a 16 socket system with higher latencies to remote cachelines,
latencies around 10ms would be seen here. irqbalanced periodically
reads this file, and the latency spikes are noticed on smaller
systems. Also the file is unprivileged, so it's important to reduce
this latency.

Avoid holding the lock while iterating over CPUs to get their stats.
The desc should be protected with irq_lock_sparse().

Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 kernel/irq/proc.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/kernel/irq/proc.c b/kernel/irq/proc.c
index e8f374971e37..bbc4004b74bc 100644
--- a/kernel/irq/proc.c
+++ b/kernel/irq/proc.c
@@ -544,17 +544,22 @@ int show_interrupts(struct seq_file *p, void *v)
 	if (!desc)
 		goto outsparse;
 
-	raw_spin_lock_irqsave(&desc->lock, flags);
 	for_each_online_cpu(j)
 		any_count |= kstat_irqs_cpu(i, j);
+
 	action = desc->action;
-	if ((!action || irq_desc_is_chained(desc)) && !any_count)
-		goto out;
+	if (!any_count) {
+		raw_spin_lock_irqsave(&desc->lock, flags);
+		if (!action || irq_desc_is_chained(desc))
+			goto out;
+		raw_spin_unlock_irqrestore(&desc->lock, flags);
+	}
 
 	seq_printf(p, "%*d: ", prec, i);
 	for_each_online_cpu(j)
 		seq_printf(p, "%10u ", kstat_irqs_cpu(i, j));
 
+	raw_spin_lock_irqsave(&desc->lock, flags);
 	if (desc->irq_data.chip) {
 		if (desc->irq_data.chip->irq_print_chip)
 			desc->irq_data.chip->irq_print_chip(&desc->irq_data, p);
-- 
2.16.3

Re: [PATCH] kernel/irq/proc: /proc/interrupts irq-off latency reduction

[Thomas Gleixner]

On Thu, 5 Apr 2018, Nicholas Piggin wrote:
> Reading /proc/interrupts shows up as a latency source of several
> hundred microseconds on a 2 socket 176 thread system, because it
> iterates twice over all CPUs under an irq lock pulling in remote
> cachelines, doing lots of seq_printf, etc.
> 
> On a 16 socket system with higher latencies to remote cachelines,
> latencies around 10ms would be seen here. irqbalanced periodically
> reads this file, and the latency spikes are noticed on smaller
> systems. Also the file is unprivileged, so it's important to reduce
> this latency.
> 
> Avoid holding the lock while iterating over CPUs to get their stats.
> The desc should be protected with irq_lock_sparse().

s/should/_IS_/

> 
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> ---
>  kernel/irq/proc.c | 11 ++++++++---
>  1 file changed, 8 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/irq/proc.c b/kernel/irq/proc.c
> index e8f374971e37..bbc4004b74bc 100644
> --- a/kernel/irq/proc.c
> +++ b/kernel/irq/proc.c
> @@ -544,17 +544,22 @@ int show_interrupts(struct seq_file *p, void *v)
>  	if (!desc)
>  		goto outsparse;
>  
> -	raw_spin_lock_irqsave(&desc->lock, flags);
>  	for_each_online_cpu(j)
>  		any_count |= kstat_irqs_cpu(i, j);
> +
>  	action = desc->action;

This is a bad idea. The action can vanish between here and the code further
below which derefences it.

You need to take desc->request_mutex for protection. That serializes
against concurrent request/free_irq().

> -	if ((!action || irq_desc_is_chained(desc)) && !any_count)
> -		goto out;
> +	if (!any_count) {
> +		raw_spin_lock_irqsave(&desc->lock, flags);
> +		if (!action || irq_desc_is_chained(desc))
> +			goto out;
> +		raw_spin_unlock_irqrestore(&desc->lock, flags);
> +	}
>  
>  	seq_printf(p, "%*d: ", prec, i);
>  	for_each_online_cpu(j)
>  		seq_printf(p, "%10u ", kstat_irqs_cpu(i, j));
>  
> +	raw_spin_lock_irqsave(&desc->lock, flags);
>  	if (desc->irq_data.chip) {
>  		if (desc->irq_data.chip->irq_print_chip)
>  			desc->irq_data.chip->irq_print_chip(&desc->irq_data, p);

And if you hold request_mutex you can drop the spinlock right before
printing desc->name.

Thanks,

	tglx