kernel/irq/proc: /proc/interrupts irq-off latency reduction
PATCH 1/1 kernel/irq/proc: /proc/interrupts irq-off latency reduction
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().
--- kernel/irq/proc.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)
kernel/irq/proc.c +8 -3
@@ -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);
Signature · 1 line
2.16.3
Quoted 5 lines · Nicholas Piggin
Reading
/proc/interruptsshows 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 ofseq_printf, etc.
Quoted 8 lines
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_/
Quoted 12 lines
Cc Thomas Gleixner <[email protected]>Signed-off-by Nicholas Piggin <[email protected]>--- kernel/irq/proc.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)kernel/irq/proc.c +0
-0
kernel/irq/proc.c int show_interrupts(struct seq_file *p, void *v) · line 550
Quoted 8 lines
kernel/irq/proc.c +1
-1@@ -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().
Quoted 7 lines
kernel/irq/proc.c +5
-2-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);
kernel/irq/proc.c int show_interrupts(struct seq_file *p, void *v) · line 565
Quoted 10 lines
kernel/irq/proc.c +2
-0+ }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